more KernelCountException

kill nvidia pids at benchmarks start (#17406 )
remove invalid special case in cast [PR] (#17405 )
2026-08-04 21:37:31 -07:00 · 2026-08-04 23:45:49 -04:00 · 2026-08-04 23:08:07 -04:00 · 2026-08-04 18:46:38 -07:00 · 2026-08-04 21:29:24 -04:00 · 2026-08-04 18:17:29 -07:00
165 changed files with 7430 additions and 2251 deletions
@@ -104,6 +104,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -155,6 +158,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -204,6 +210,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p extra/datasets
@@ -250,6 +259,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -292,6 +304,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -402,6 +417,35 @@ jobs:
      run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py

  testcommalatest:
+    name: comma Benchmark (0.11.2)
+    runs-on: [self-hosted, Linux, comma]
+    timeout-minutes: 12
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: openpilot compile3 0.11.2 supercombo
+      run: BENCHMARK_LOG=openpilot_0_11_2_supercombo PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
+    - name: openpilot compile3 0.11.2 supercombo (from pickle)
+      run: BENCHMARK_LOG=openpilot_0_11_2_supercombo_run_pickle RUN_PICKLE=1 PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 DEV=QCOM taskset -c 4-7 python3 examples/openpilot/compile3.py
+    - name: IR3 openpilot compile3 0.11.2 supercombo
+      run: BENCHMARK_LOG=ir3_openpilot_0_11_2_supercombo PYTHONPATH="." ASSERT_MIN_STEP_TIME=41 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
+    - name: openpilot compile3 0.11.2 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_11_2_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/3e7b31dfbc0a5234f1baf196513b77fc6af12204b8a8ffe8ee0417e48352f316
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testcommaold:
    name: comma Benchmark (0.11.0)
    runs-on: [self-hosted, Linux, comma]
    timeout-minutes: 12
@@ -432,35 +476,6 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

-  testcommaold:
-    name: comma Benchmark (0.10.1)
-    runs-on: [self-hosted, Linux, comma]
-    timeout-minutes: 12
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: reset process replay
-      run: test/external/process_replay/reset.py
-    - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
-      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3.2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
  testqualcommdsp:
    name: DSP Benchmark
    runs-on: [self-hosted, Linux, comma4]
@@ -515,12 +530,12 @@ jobs:
      run: |
        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." GMMU=0 DEV=USB+AMD:LLVM ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot load_pickle 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_load_pickle PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
-    - name: openpilot run_pickle 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py
+    - name: openpilot compile3 big_driving_supercombo
+      run: BENCHMARK_LOG=usbgpu_openpilot_big_driving_supercombo PICKLE_OOB=1 PYTHONPATH="." TC_OPT=2 GMMU=0 DEV=USB+AMD:LLVM ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/10926f2c0911821ca0e72439c1c3bf3ec11f0a08789aa14b7ee8f25379b2afa4 openpilot.pkl
+    - name: openpilot load_pickle big_driving_supercombo
+      run: BENCHMARK_LOG=usbgpu_openpilot_big_driving_supercombo_load_pickle PICKLE_OOB=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_LOAD_TIME=25 python3 examples/openpilot/load_pickle.py openpilot.pkl
+    - name: openpilot run_pickle big_driving_supercombo
+      run: BENCHMARK_LOG=usbgpu_openpilot_big_driving_supercombo_run_pickle RUN_PICKLE=1 PICKLE_OOB=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py - openpilot.pkl
    - name: Test copy speeds
      run: SIZE=64e6 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3 test/external/external_test_usb_asm24.py TestDevCopySpeeds

@@ -219,8 +219,8 @@ jobs:
      run: python3 test/external/external_benchmark_schedule.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
-    - name: Repo line count < 25000 lines
-      run: MAX_LINE_COUNT=25000 python sz.py
+    - name: Repo line count <= 26000 lines
+      run: MAX_LINE_COUNT=26000 python sz.py

  spec:
    strategy:
@@ -353,6 +353,8 @@ jobs:
        run: DEV=NULL NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
        run: DEV=NULL NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+      - name: Test gpt-oss training
+        run: DEV=NULL NULL_ALLOW_COPYOUT=1 SAMPLES=32 BS=2 SEQLEN=128 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 MXFP8=1 VOCAB_SIZE=32000 LAYERS=2 EXPERTS=4 MODEL=gptoss PYTHONPATH=. python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -641,12 +643,8 @@ jobs:
      with:
        key: unittest-macos
        deps: testing_unit
-        amd: 'true'
-        ocelot: 'true'
    - name: Run unit tests
      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
-    - name: Run NULL backend tests
-      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
    - name: Test tensor core ops (fake)
      run: DEV=METAL DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
    - name: Test tensor core ops (real)
@@ -657,6 +655,25 @@ jobs:
      run: DEV=METAL python3 -m pytest test/device/test_metal.py
    #- name: Fuzz Test linearizer
    #  run: DEV=METAL DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  unittestmacosmock:
+    name: MacOS (unit, mock)
+    runs-on: macos-26
+    timeout-minutes: 20
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: unittest-macos-mock
+        deps: testing_unit
+        amd: 'true'
+        ocelot: 'true'
+    - name: Run NULL backend tests
+      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
    - name: Run pytest (amd)
      env:
        DEV: MOCKKFD+AMD
@@ -22,10 +22,6 @@ class Attention:
    self.head_dim = dim // n_heads

  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]) -> Tensor:
-    if mask is not None or start_pos.val == 0:
-      # no symbolic shape qkv when consuming prompts
-      start_pos = start_pos.val
-
    if HALF: x = x.half()
    xqkv = self.c_attn(x).reshape(None, None, 3, self.n_heads, self.head_dim)
    xq, xk, xv = [xqkv[:, :, i, :, :] for i in range(3)]
@@ -38,12 +34,8 @@ class Attention:
    # update the cache
    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()

-    if start_pos > 0:
-      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
-      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
-    else:
-      keys = xk
-      values = xv
+    keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
+    values = self.cache_kv[1][:, :start_pos+seqlen, :, :]

    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
    return self.c_proj(xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2).reshape(bsz, seqlen, self.dim))
@@ -86,15 +78,14 @@ class Transformer:
      seqlen = tokens.shape[1]
      tok_emb = self.wte(tokens)

-    # not symbolic when consuming the prompt
-    selected_pos = (0, seqlen) if start_pos.val == 0 else (start_pos, start_pos+1)
-    pos_emb = self.wpe(self.allpos.shrink((None, selected_pos)))
+    # start_pos is a bound Variable, so everything below it stays symbolic
+    pos_emb = self.wpe(self.allpos.shrink((None, (start_pos, start_pos+seqlen))))

    h = tok_emb + pos_emb

    if HALF: h = h.half()

-    mask = Tensor.full((1, 1, seqlen, start_pos.val+seqlen), float("-inf"), dtype=h.dtype).triu(start_pos.val+1) if seqlen > 1 else None
+    mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype).triu(start_pos+1) if seqlen > 1 else None

    for hi in self.h: h = hi(h, start_pos, mask)

@@ -1282,7 +1282,7 @@ def train_bert():
        previous_step = i

 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8
+  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8, MXFP4
  from examples.llama3 import MODEL_PARAMS
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
  from examples.mlperf.optim import GradAccClipAdamW, clip_grads
@@ -1434,9 +1434,9 @@ def train_llama3():
    load_state_dict(scheduler, safe_load(fn), realize=False)

  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
-  fp8_next_amax = [t for ts in model._fp8_next_amax.values() for t in ts] if hasattr(model, "_fp8_next_amax") else []
-  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts] if hasattr(model, "_fp8_grad_amax") else []
-  fp8_next_grad_amax = [t for ts in model._fp8_next_grad_amax.values() for t in ts] if hasattr(model, "_fp8_next_grad_amax") else []
+  fp8_next_amax = [t for ts in model._fp8_next_amax.values() for t in ts]
+  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts]
+  fp8_next_grad_amax = [t for ts in model._fp8_next_grad_amax.values() for t in ts]
  fp8_inv_scales = list(model._fp8_inv_scale.values()) + list(model._fp8_next_inv_scale.values())

  from tinygrad.nn.state import get_state_dict
@@ -1462,8 +1462,7 @@ def train_llama3():

  @TinyJit
  def minibatch(tokens:Tensor):
-    for nxt in fp8_next_amax: nxt.assign(0)
-    for nxt in fp8_next_grad_amax: nxt.assign(0)
+    model.reset_amax()
    if is_dp: tokens = tokens.to(None).shard(device, 0)
    if is_mp: tokens = tokens.shard(device)
    if not is_sharding: tokens = tokens.to(None)
@@ -1487,8 +1486,7 @@ def train_llama3():
    scheduler.step()

    for g in grads: g.assign(0)
-    for cur, nxt in zip(fp8_amax, fp8_next_amax): cur.assign(nxt)
-    for cur, nxt in zip(fp8_grad_amax, fp8_next_grad_amax): cur.assign(nxt)
+    model.update_amax()

    lr_cpu = optim.lr.float().to("CPU")
    grad_norm_cpu = grad_norm.float().to("CPU")
@@ -1579,7 +1577,7 @@ def train_llama3():

      mem_gb = GlobalCounters.mem_used / 1e9
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 100
+      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * (9.2e15 if MXFP4 else 4.6e15))) * 100
      tqdm.write(
          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
@@ -1711,9 +1709,10 @@ def train_gptoss():
    wandb.init(config=config, **wandb_args, project="MLPerf-gpt-oss")

  model_params = GPT_OSS_20B
-  model_params['vocab_size'] = 128256
+  model_params['vocab_size'] = getenv("VOCAB_SIZE", 128256)
  real_vocab_size = model_params['vocab_size']
  if (layers:=getenv("LAYERS")) != 0: model_params['n_layers'] = layers
+  if (experts:=getenv("EXPERTS")) != 0: model_params['n_experts'] = experts
  print(f"model parameters: {model_params}")

  model = GPTOSS(**model_params, max_context=SEQLEN)
@@ -1748,7 +1747,10 @@ def train_gptoss():

  from extra.gemm.cdna_asm_gemm import _mx_block_scale
  model_state = get_state_dict(model)
-  fp8_scale_names = {n: f"{n}_scale" for n, t in model_state.items() if t.dtype == FP8_DTYPE}
+  def _scale_key(n):
+    if "." in n and (c:=f"{(b:=n.rsplit('.',1))[0]}_scale.{b[1]}") in model_state: return c
+    return f"{n}_scale"
+  fp8_scale_names = {n: _scale_key(n) for n, t in model_state.items() if t.dtype == FP8_DTYPE}
  fp8_inv_scales = [model_state[sname] for sname in fp8_scale_names.values()]
  for wname, sname in fp8_scale_names.items():
    w, scale = model_state[wname], model_state[sname]
@@ -25,6 +25,7 @@ FUSED_SILU_W13 = getenv("FUSED_SILU_W13", 0)
 SPLIT_W13 = getenv("SPLIT_W13", 0)
 COLUMNWISE_WEIGHT_SCALE = getenv("COLUMNWISE_WEIGHT_SCALE", 0)
 MXFP8 = getenv("MXFP8", 0)
+MXFP4 = getenv("MXFP4", 0)

 FP8_DTYPE = dtypes.fp8e4m3
 FP8_GRAD_DTYPE = dtypes.fp8e5m2
@@ -44,6 +45,11 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
      from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
      if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T),)
    return (x @ w.T,)
+  if MXFP4:
+    assert x is not None, "MXFP4 matmul requires an unquantized input"
+    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm
+    if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T, mxfp4=True),)
+    return (x @ w.T,)
  assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
  if MXFP8:
    from extra.gemm.cdna_asm_gemm import asm_gemm, quantize_mxfp8, mx_pack, can_use_asm_gemm, _mx_block_scale
@@ -77,9 +83,9 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
      return out, x_fp8
  return (x_fp8.dot(w.T, dtype=dtypes.float) * ((amax_x.float() + 1e-8) / FP8_MAX) * w_inv_scale).cast(dtypes.bfloat16), x_fp8

-def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor,
-                         next_amax_x:Tensor, grad_amax_state:Tensor, next_grad_amax_state:Tensor):
-  if FUSED_ADD_NORM_MUL_QUANTIZE:
+def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
+                         next_amax_x:Tensor|None, grad_amax_state:Tensor|None, next_grad_amax_state:Tensor|None):
+  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_rmsnorm_mul_quantize_fp8
    x_fp8, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x,
@@ -90,9 +96,9 @@ def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, ep
                     next_grad_amax_state=next_grad_amax_state, next_amax_x=next_amax_x)
  return out, x_normed, rrms, ret

-def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor,
-                             next_amax_x:Tensor, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
-  if FUSED_ADD_NORM_MUL_QUANTIZE:
+def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
+                             next_amax_x:Tensor|None, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
+  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
    x_fp8, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x,
@@ -105,10 +111,10 @@ def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w
  return out, h, x_normed, rrms, ret

 def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
-                             amax_x2:Tensor, next_amax_x2:Tensor,
-                             grad_amax_xw13:Tensor, next_grad_amax_xw13:Tensor,
-                             grad_amax_xout:Tensor, next_grad_amax_xout:Tensor):
-  if FUSED_SILU_W13:
+                             amax_x2:Tensor|None, next_amax_x2:Tensor|None,
+                             grad_amax_xw13:Tensor|None, next_grad_amax_xw13:Tensor|None,
+                             grad_amax_xout:Tensor|None, next_grad_amax_xout:Tensor|None):
+  if FUSED_SILU_W13 and not MXFP4:
    from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
    x2_fp8 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13,
                                                 next_grad_amax_state=next_grad_amax_xw13, amax_out=next_amax_x2)
@@ -158,14 +164,15 @@ class FlatTransformer:
    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).clone().is_param_(False)

    def _amax(): return Tensor.full((), FP8_MAX, dtype=dtypes.float32).contiguous().is_param_(False)
+    n_amax = 0 if MXFP4 else n_layers
    names = ["xqkv", "xo", "x2"]
    names += ["x1", "x3"] if SPLIT_W13 else ["x13"]
-    self._fp8_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
-    self._fp8_next_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
+    self._fp8_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
+    self._fp8_next_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
    grad_names = ["xqkv", "xo", "xout"]
    grad_names += ["xw1", "xw3"] if SPLIT_W13 else ["xw13"]
-    self._fp8_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
-    self._fp8_next_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
+    self._fp8_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
+    self._fp8_next_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
    w_scales = [("wqkv", s_qkv), ("wo", s_o), ("w2", s_2)]
    w_scales += [("w1", s_1), ("w3", s_3)] if SPLIT_W13 else [("w13", s_13)]
    self._fp8_inv_scale = {name: (s if MXFP8 else s.float()).contiguous().is_param_(False) for name, s in w_scales}
@@ -179,6 +186,9 @@ class FlatTransformer:
      from extra.gemm.cdna_asm_gemm import quantize_mxfp8
      w_q, w_e8, _ = quantize_mxfp8(w.reshape(self.n_layers * out_features, in_features))
      return w_q.reshape(self.n_layers, out_features, in_features), w_e8.reshape(self.n_layers, out_features, in_features // 32)
+    if MXFP4:
+      # FP4 is produced dynamically so optimizer updates always start from the current BF16 weight.
+      return w.cast(dtypes.bfloat16), Tensor.ones(self.n_layers)
    amax = (w.abs().max(axis=2) if COLUMNWISE_WEIGHT_SCALE else w.abs().flatten(1).max(1)).detach()
    scale = FP8_MAX / (amax + 1e-8)
    inv_scale = (amax + 1e-8) / FP8_MAX
@@ -186,9 +196,10 @@ class FlatTransformer:
    return (w * scale_b).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE), inv_scale

  def attention(self, x:Tensor, freqs_cis:Tensor, *, attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                amax_xqkv:Tensor, amax_xo:Tensor, s_qkv:Tensor, s_o:Tensor,
-                next_amax_xqkv:Tensor, next_amax_xo:Tensor,
-                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor, next_grad_amax_xqkv:Tensor, next_grad_amax_xo:Tensor):
+                amax_xqkv:Tensor|None, amax_xo:Tensor|None, s_qkv:Tensor, s_o:Tensor,
+                next_amax_xqkv:Tensor|None, next_amax_xo:Tensor|None,
+                grad_amax_xqkv:Tensor|None, grad_amax_xo:Tensor|None,
+                next_grad_amax_xqkv:Tensor|None, next_grad_amax_xo:Tensor|None):
    bsz, seqlen, _ = x.shape
    saves = []

@@ -310,28 +321,33 @@ class FlatTransformer:
          for i in range(len(amax_dict[name])):
            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().is_param_(False)

+  def reset_amax(self):
+    for st in (self._fp8_next_amax, self._fp8_next_grad_amax):
+      for ts in st.values():
+        for t in ts: t.assign(0)
+
+  def update_amax(self):
+    for cur, nxt in ((self._fp8_amax, self._fp8_next_amax), (self._fp8_grad_amax, self._fp8_next_grad_amax)):
+      for name in cur:
+        for c, n in zip(cur[name], nxt[name]): c.assign(n)
+
  def __call__(self, tokens:Tensor, save:bool=True):
    h = self.tok_embeddings(tokens)
    freqs_cis = self.freqs_cis.cast(h.dtype)
    if not getenv("HK_FLASH_ATTENTION"): freqs_cis = freqs_cis[:, :tokens.shape[1], :, :, :]
    a, na, ga, nga, s = self._fp8_amax, self._fp8_next_amax, self._fp8_grad_amax, self._fp8_next_grad_amax, self._fp8_inv_scale
+    def amax_kwargs(i:int, act_names:tuple[str, ...], grad_names:tuple[str, ...]) -> dict[str, Tensor|None]:
+      specs = (("amax_", a, act_names), ("next_amax_", na, act_names), ("grad_amax_", ga, grad_names), ("next_grad_amax_", nga, grad_names))
+      if MXFP4: return dict.fromkeys(f"{prefix}{name}" for prefix, _, names in specs for name in names)
+      return {f"{prefix}{name}":val[name][i] for prefix, val, names in specs for name in names}
    for i in range(self.n_layers):
-      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wo=self.wo[i],
-                         amax_xqkv=a["xqkv"][i], amax_xo=a["xo"][i], s_qkv=s["wqkv"][i], s_o=s["wo"][i],
-                         next_amax_xqkv=na["xqkv"][i], next_amax_xo=na["xo"][i],
-                         grad_amax_xqkv=ga["xqkv"][i], grad_amax_xo=ga["xo"][i],
-                         next_grad_amax_xqkv=nga["xqkv"][i], next_grad_amax_xo=nga["xo"][i])
-      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], w2=self.w2[i],
-                        amax_x2=a["x2"][i], s_2=s["w2"][i], grad_amax_xout=ga["xout"][i], next_grad_amax_xout=nga["xout"][i],
-                        next_amax_x2=na["x2"][i])
+      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wo=self.wo[i], s_qkv=s["wqkv"][i], s_o=s["wo"][i],
+                         **amax_kwargs(i, ("xqkv", "xo"), ("xqkv", "xo")))
+      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], w2=self.w2[i], s_2=s["w2"][i], **amax_kwargs(i, ("x2",), ("xout",)))
      if SPLIT_W13:
-        ffn_kwargs.update(w1=self.w1[i], w3=self.w3[i], amax_x1=a["x1"][i], amax_x3=a["x3"][i],
-                          next_amax_x1=na["x1"][i], next_amax_x3=na["x3"][i],
-                          s_1=s["w1"][i], s_3=s["w3"][i], grad_amax_xw1=ga["xw1"][i], grad_amax_xw3=ga["xw3"][i],
-                          next_grad_amax_xw1=nga["xw1"][i], next_grad_amax_xw3=nga["xw3"][i])
+        ffn_kwargs.update(w1=self.w1[i], w3=self.w3[i], s_1=s["w1"][i], s_3=s["w3"][i], **amax_kwargs(i, ("x1", "x3"), ("xw1", "xw3")))
      else:
-        ffn_kwargs.update(w13=self.w13[i], amax_x13=a["x13"][i], s_13=s["w13"][i], grad_amax_xw13=ga["xw13"][i],
-                          next_grad_amax_xw13=nga["xw13"][i], next_amax_x13=na["x13"][i])
+        ffn_kwargs.update(w13=self.w13[i], s_13=s["w13"][i], **amax_kwargs(i, ("x13",), ("xw13",)))
      h, *_ = self.run_layer(h, freqs_cis, attn_kwargs, ffn_kwargs, save=save)

    logits = matmul(self.norm(h), self.output[0], fp8=False)[0]
@@ -415,9 +431,7 @@ if __name__ == "__main__":
  @TinyJit
  def fwd_bwd(tokens:Tensor):
    with Timing("python forward: "):
-      for amax_dict in (model._fp8_next_amax, model._fp8_next_grad_amax):
-        for ts in amax_dict.values():
-          for nxt in ts: nxt.assign(0)
+      model.reset_amax()
      logits = model(tokens[:, :-1], save=llama_size=="8B")
      loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
    with Timing("python backward: "):
@@ -13,10 +13,14 @@ from tinygrad.uop.ops import Ops, UOp
 from extra.models.llama import apply_rotary_emb
 from extra.llama_kernels.rmsnorm import rmsnorm
 from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8
+from extra.gemm.moe_gemm import grouped_mx_gemm
+from extra.gemm.moe_routing import route, dispatch, combine

 FP8_DTYPE = dtypes.fp8e4m3
 FP8_MAX = 448.0
-INIT_STD = 0.008
+INIT_STD = 0.02
+ASM_GEMM = getenv("ASM_GEMM", 0)
+

 def _quant_dequant_fwd(x:Tensor) -> Tensor:
  # x (2d bf16) -> bf16 value after an mxfp8 round-trip (1x32 block scaling on the last axis)
@@ -50,8 +54,7 @@ def _dequant_fwd_fxn(wq_p, ws_p, device):
  return _dequant_fwd(Tensor(wq_p, device=device), Tensor(ws_p, device=device))

 def _dequant_bwd(grad:UOp, call:UOp) -> tuple:
-  w_scale = Tensor(call.src[2])
-  return ((Tensor(grad).cast(dtypes.bfloat16) * _mx_scale(w_scale).cast(dtypes.bfloat16)).uop, None)
+  return (Tensor(grad).cast(dtypes.bfloat16).uop, None)

 def dequant_weight(w_q:Tensor, w_scale:Tensor) -> Tensor:
  fxn = _dequant_fwd_fxn(w_q.as_param(0).uop, w_scale.as_param(1).uop, w_q.device)
@@ -60,10 +63,34 @@ def dequant_weight(w_q:Tensor, w_scale:Tensor) -> Tensor:

 def matmul_mx(x:Tensor, w_q:Tensor, w_scale:Tensor) -> Tensor:
  l_shape = x.shape[:-1]
+  if ASM_GEMM:
+    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm, mx_pack
+    x2, K, N = x.reshape(-1, x.shape[-1]), x.shape[-1], w_q.shape[0]
+    wq, ws = w_q, w_scale
+    if (pad := (-K) % 256):
+      x2 = x2.pad(((0, 0), (0, pad)))
+      wq = wq.pad(((0, 0), (0, pad)))
+      ws = ws.pad(((0, 0), (0, pad // 32)), value=127).cast(dtypes.uint8)
+    if (npad := (-N) % 256):
+      wq = wq.pad(((0, npad), (0, 0)))
+      ws = ws.pad(((0, npad), (0, 0)), value=127).cast(dtypes.uint8)
+    x_q, x_e8, x_si = quantize_mxfp8(x2)
+    if x_si is not None and can_use_asm_gemm(x_q, wq.T):
+      out = asm_gemm(x_q, wq.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(ws), ws), mx_w_stored=True)
+      return (out[:, :N] if npad else out).reshape(*l_shape, N).cast(dtypes.bfloat16)
  x_phys = quant_dequant_mx(x.reshape(-1, x.shape[-1])).reshape(*l_shape, x.shape[-1])
  w_phys = dequant_weight(w_q, w_scale)
  return (x_phys @ w_phys.T).cast(dtypes.bfloat16)

+def _pad_to_mult(t:Tensor, axis:int, mult:int=256) -> Tensor:
+  if (r := (-t.shape[axis]) % mult) == 0: return t
+  pads = [(0, 0)] * t.ndim
+  pads[axis] = (0, r)
+  return t.pad(tuple(pads))
+
+def _pad_cols(t:Tensor) -> Tensor: return _pad_to_mult(t, -1)
+def _pad_rows(t:Tensor) -> Tensor: return _pad_to_mult(t, -2)
+
 def swiglu(x:Tensor, limit:float=7.0, alpha:float=1.702) -> Tensor:
  x_glu, x_linear = x[..., ::2], x[..., 1::2]
  x_glu = x_glu.clamp(max_=limit)
@@ -100,9 +127,9 @@ class GPTOSS:
    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
    self.gate = Tensor.normal(n_layers, n_experts, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
    self.gate_bias = Tensor.zeros(n_layers, n_experts, dtype=dtypes.bfloat16).contiguous()
-    self.w_gate_up, self.w_gate_up_scale = self._quant_weight(n_layers, n_experts, intermediate_size * 2, dim)
+    self.w_gate_up, self.w_gate_up_scale = self._quant_weight(n_layers, n_experts, intermediate_size * 2, dim, moe=True)
    self.w_gate_up_bias = Tensor.zeros(n_layers, n_experts, intermediate_size * 2, dtype=dtypes.bfloat16).contiguous()
-    self.w_down, self.w_down_scale = self._quant_weight(n_layers, n_experts, dim, intermediate_size, std=scaled_std)
+    self.w_down, self.w_down_scale = self._quant_weight(n_layers, n_experts, dim, intermediate_size, std=scaled_std, moe=True)
    self.w_down_bias = Tensor.zeros(n_layers, n_experts, dim, dtype=dtypes.bfloat16).contiguous()

    # output
@@ -112,10 +139,15 @@ class GPTOSS:
    self.output = Tensor.normal(vocab_size, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
    self.freqs_cis = precompute_freqs_cis(head_dim, max_context * 2, rope_theta).contiguous().is_param_(False)

-  def _quant_weight(self, *shape:int, std:float=INIT_STD):
-    w = Tensor.zeros(*shape) if getenv("ZEROS") else Tensor.normal(*shape, mean=0.0, std=std)
-    w_q, w_e8, _ = quantize_mxfp8(w)
-    return w_q, w_e8.is_param_(False)
+  def _quant_weight(self, *shape:int, std:float=INIT_STD, moe:bool=False):
+    def _one(*s:int):
+      w = Tensor.zeros(*s) if getenv("ZEROS") else Tensor.normal(*s, mean=0.0, std=std)
+      w_q, w_e8, _ = quantize_mxfp8(_pad_cols(_pad_rows(w)) if moe else w)
+      return w_q, w_e8.is_param_(False)
+    if moe:
+      qs = [_one(*shape[1:]) for _ in range(shape[0])]
+      return [q[0] for q in qs], [q[1] for q in qs]
+    return _one(*shape)

  def _attn_mask(self, seqlen:int, dtype) -> Tensor:
    i, j = Tensor.arange(seqlen).reshape(seqlen, 1), Tensor.arange(seqlen).reshape(1, seqlen)
@@ -174,17 +206,32 @@ class GPTOSS:
                   w_down:Tensor, w_down_scale:Tensor, w_down_bias:Tensor):
    x_normed, rrms = rmsnorm(x, self.norm_eps)
    inp = x_normed * ffn_norm
-
    logits = inp.float() @ gate.float().T + gate_bias.float()
-    thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
-    weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
+    dim, inter = self.dim, self.intermediate_size

-    out = None
-    for e in range(self.n_experts):
-      gate_up = matmul_mx(inp, w_gate_up[e], w_gate_up_scale[e]) + w_gate_up_bias[e]
-      y = (matmul_mx(swiglu(gate_up, self.swiglu_limit), w_down[e], w_down_scale[e]) + w_down_bias[e]).contiguous()
-      contrib = weights[..., e:e+1].cast(y.dtype) * y
-      out = contrib if out is None else out + contrib
+    if getenv("GROUPED_MOE", 0):
+      bsz, seqlen = x.shape[:2]
+      inp, logits = inp.reshape(-1, dim), logits.reshape(-1, self.n_experts)
+      r = route(logits, self.experts_per_tok, self.n_experts)
+      onehot = r.rows_e.one_hot(self.n_experts).float()
+      xg = dispatch(_pad_cols(inp.cast(dtypes.bfloat16)), r)
+      h = grouped_mx_gemm(xg, (w_gate_up, w_gate_up_scale), r.off)[:, :2*inter] + (onehot @ w_gate_up_bias.float()).cast(dtypes.bfloat16)
+      y = swiglu(h, self.swiglu_limit)
+      z = grouped_mx_gemm(_pad_cols(y.cast(dtypes.bfloat16)), (w_down, w_down_scale), r.off)[:, :dim] \
+          + (onehot @ w_down_bias.float()).cast(dtypes.bfloat16)
+      out = combine(z, r, inp.shape[0], self.experts_per_tok).reshape(bsz, seqlen, dim)
+    else:
+      thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
+      weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
+
+      out = None
+      for e in range(self.n_experts):
+        gu_q, gu_s = w_gate_up[e][:2*inter, :dim].contiguous(), w_gate_up_scale[e][:2*inter, :dim//32].contiguous()
+        dn_q, dn_s = w_down[e][:dim, :inter].contiguous(), w_down_scale[e][:dim, :inter//32].contiguous()
+        gate_up = matmul_mx(inp, gu_q, gu_s) + w_gate_up_bias[e]
+        y = (matmul_mx(swiglu(gate_up, self.swiglu_limit), dn_q, dn_s) + w_down_bias[e]).contiguous()
+        contrib = weights[..., e:e+1].cast(y.dtype) * y
+        out = contrib if out is None else out + contrib
    return out, [x_normed, rrms]

  @function(precompile=True, precompile_backward=True)
@@ -2,7 +2,7 @@ from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes
 from tinygrad.nn.optim import Optimizer, OptimizerGroup
 from tinygrad.helpers import FUSE_OPTIM, getenv
-from tinygrad.uop.ops import UOp, Ops
+from tinygrad.uop.ops import UOp, Ops, AxisType

 STOCHASTIC_ROUND = getenv("STOCHASTIC_ROUND", 0)
 MASTER_WEIGHTS = getenv("MASTER_WEIGHTS", 0)
@@ -42,8 +42,8 @@ class GradAccClipAdamW(Optimizer):
      self.master_params = None

  def _zero_shard(self, t:Tensor) -> Tensor:
-    if not self.zero or (t.shape[0] % len(self.device)) != 0: return t
-    return Tensor(t.uop._shard(0, len(self.device)).unshard(0)).clone()
+    if not self.zero or t.ndim < 2 or (t.shape[0] % len(self.device)) != 0: return t
+    return Tensor(t.uop._shard(0, UOp.range(len(self.device), -1, AxisType.DEVICE)).unshard(0)).clone()

  def _zero_gather(self, t:Tensor) -> Tensor:
    if not isinstance(t.device, tuple) or t.uop.axis != 0: return t
@@ -123,6 +123,9 @@ class GradAccClipAdamW(Optimizer):
    return out.shard_like(t) if offloaded else out

 class GradAccClipAdamWGroup(OptimizerGroup):
+  def __init__(self, *optimizers:GradAccClipAdamW):
+    super().__init__(*optimizers)
+    for o in self.optimizers[1:]: o.lr = self.optimizers[0].lr
  def fstep(self, grads:list[Tensor], grad_norm:Tensor|None=None):
    offset = 0
    to_realize = []
@@ -10,6 +10,7 @@ export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -10,6 +10,7 @@ export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -1,4 +1,4 @@
-import os, sys, pickle, time, re
+import os, sys, pickle, time, re, tempfile, struct, shutil, io
 import numpy as np
 if "JIT_BATCH_SIZE" not in os.environ: os.environ["JIT_BATCH_SIZE"] = "0"

@@ -9,6 +9,39 @@ from tinygrad.nn.onnx import OnnxRunner

 OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
 OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
+PICKLE_OOB = getenv("PICKLE_OOB")
+
+def dump_pickle(obj, f):
+  if PICKLE_OOB:
+    # allows pickling when buffers don't fit in (CPU) RAM
+    # from openpilot/selfdrive/modeld/helpers.py
+    with tempfile.TemporaryFile(dir=".") as tmp:
+      def buffer_callback(pb: pickle.PickleBuffer):
+        m = pb.raw()
+        tmp.write(struct.pack('<q', m.nbytes))
+        tmp.write(m)
+        pb.release() # keep peak ram at ~1 buffer
+      stream = io.BytesIO()
+      pickle.Pickler(stream, protocol=5, buffer_callback=buffer_callback).dump(obj)
+      opcodes = stream.getvalue()
+      f.write(struct.pack('<q', len(opcodes)))
+      f.write(opcodes)
+      tmp.seek(0)
+      shutil.copyfileobj(tmp, f)
+  else: pickle.dump(obj, f)
+
+def load_pickle(f):
+  if PICKLE_OOB:
+    # allows unpickling when buffers don't fit in (CPU) RAM
+    # from openpilot/selfdrive/modeld/helpers.py
+    opcodes = f.read(struct.unpack('<q', f.read(8))[0])
+    def buffers():
+      while (h := f.read(8)):
+        pb = pickle.PickleBuffer(bytearray(struct.unpack('<q', h)[0]))
+        f.readinto(pb)
+        yield pb
+    return pickle.load(io.BytesIO(opcodes), buffers=buffers())
+  else: return pickle.load(f)

 def compile(onnx_file):
  run_onnx = OnnxRunner(onnx_file)
@@ -28,8 +61,8 @@ def compile(onnx_file):
    inputs = {k:Tensor(v.numpy(), device=Device.DEFAULT).realize() if 'img' in k else v for k,v in inputs.items()}
  print("created tensors")

-  run_onnx_jit = TinyJit(lambda **kwargs:
-                         next(iter(run_onnx({k:v.to(Device.DEFAULT) for k,v in kwargs.items()}).values())).cast('float32'), prune=True)
+  @TinyJit(prune=True)
+  def run_onnx_jit(**kwargs): return next(iter(run_onnx({k:v.to(Device.DEFAULT) for k,v in kwargs.items()}).values())).cast('float32')
  for i in range(3):
    GlobalCounters.reset()
    print(f"run {i}")
@@ -65,8 +98,7 @@ def compile(onnx_file):
  if (allowed_gated_read_image:=getenv("ALLOWED_GATED_READ_IMAGE", -1)) != -1:
    assert gated_read_image_count == allowed_gated_read_image, f"different gated read_image! {gated_read_image_count=}, {allowed_gated_read_image=}"

-  with open(OUTPUT, "wb") as f:
-    pickle.dump(run_onnx_jit, f)
+  with open(OUTPUT, "wb") as f: dump_pickle(run_onnx_jit, f)
  mdl_sz = os.path.getsize(onnx_file)
  pkl_sz = os.path.getsize(OUTPUT)
  print(f"mdl size is {mdl_sz/1e6:.2f}M")
@@ -136,7 +168,7 @@ def bench(run, inputs):

 if __name__ == "__main__":
  if getenv("RUN_PICKLE"):
-    with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)
+    with open(OUTPUT, "rb") as f: pickle_loaded = load_pickle(f)
    inputs = {name: Tensor(Tensor.randn(*view.shape, dtype=dtype).numpy(), device=device)
              for name, (view, _vars, dtype, device) in zip(pickle_loaded.captured.expected_names, pickle_loaded.captured.expected_input_info)}
    test_vs_compile(pickle_loaded, inputs)
@@ -144,7 +176,7 @@ if __name__ == "__main__":
    onnx_file = fetch(OPENPILOT_MODEL)
    inputs, outputs = compile(onnx_file)

-    with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)
+    with open(OUTPUT, "rb") as f: pickle_loaded = load_pickle(f)

    test_vs_compile(pickle_loaded, inputs, outputs)
    if getenv("SELFTEST"):
@@ -1,5 +1,6 @@
-import sys, pickle
+import sys
 from extra.bench_log import WallTimeEvent, BenchEvent
+from examples.openpilot.compile3 import load_pickle
 from tinygrad.helpers import getenv

 PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
@@ -7,7 +8,7 @@ PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
 load_times = []

 for _ in range(10):
-  with WallTimeEvent(BenchEvent.STEP) as wte: pickle.load(open(PKL, 'rb'))
+  with WallTimeEvent(BenchEvent.STEP) as wte: load_pickle(open(PKL, 'rb'))
  load_times.append(wte.time)
  print(f"pickle load: {wte.time:6.2f} s")

@@ -264,7 +264,7 @@ def export_model(model, target:str, *inputs, model_name: Optional[str] = "model"
        if getattr(dim, "op", None) is Ops.ADD and len(dim.src) == 2 and \
           any(s.op is Ops.PARAM and s.addrspace is AddrSpace.ALU for s in dim.src) and any(s.op is Ops.CONST for s in dim.src):
          name, val = dim.src if dim.src[1].op is Ops.CONST else reversed(dim.src)
-          global_size[j] = f"_{name.expr}[0] + {val.arg}"
+          global_size[j] = f"_{name.expr}[0] + {val.val}"

  prg = ""
  if target == "clang":
@@ -18,9 +18,9 @@ def custom_matmul(output: UOp, inp: UOp, weight: UOp) -> UOp:
  SEQ = inp.shape[1]
  OUT = weight.shape[0]
  IN = weight.shape[-1]
-  seq_idx = UOp.range(SEQ, 2, AxisType.LOOP)
-  out_idx = UOp.range(OUT, 3, AxisType.LOOP)
-  batch_idx = UOp.range(output.size//SEQ//OUT, 1, AxisType.LOOP)
+  seq_idx = UOp.range(SEQ, 2)
+  out_idx = UOp.range(OUT, 3)
+  batch_idx = UOp.range(output.size//SEQ//OUT, 1)
  reduce_idx = UOp.range(IN, 0, AxisType.REDUCE)
  product = (inp.index((seq_idx*IN+reduce_idx+batch_idx*IN*SEQ)) * weight.index((out_idx*IN+reduce_idx))).cast(dtypes.float)
  reduced = product.reduce(reduce_idx, arg=Ops.ADD)
@@ -70,8 +70,8 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:

  if use_wmma:
    k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)
-    tile_m = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-    tile_n = UOp.range(TN, 201, AxisType.LOOP)
+    tile_m = UOp.range(TM // WMMA_ACC, 200)
+    tile_n = UOp.range(TN, 201)

    acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0,2,1)[tile_m, tile_n]
    a_frag = A_local.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_K // WMMA_K, WMMA_K)[wave_m, tile_m, lane_n, k]
@@ -1,203 +0,0 @@
-from tinygrad import Tensor, UOp, getenv
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-from tinygrad.dtype import AddrSpace, dtypes
-from tinygrad.helpers import DEBUG, GlobalCounters, Context
-import math
-
-BLOCK_M, BLOCK_N = 64, 64
-WARP_SIZE = 32
-WMMA_M, WMMA_N, WMMA_K = 16, 16, 16
-WAVES_M, WAVES_N = 4, 1
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 2, 16
-WMMA_ACC = WMMA_M // LANES_PER_WAVE_M
-THREADS_PER_BLOCK = WARP_SIZE * WAVES_M * WAVES_N
-LDS_PAD = 4  # pad LDS rows to reduce bank conflicts
-
-WMMA_ARG = (WMMA_M, WMMA_N, WMMA_K), 'AMD', 32
-LOG2E = math.log2(math.e)
-
-def warp_shfl_xor(val, offset, lane):
-  """Read val from lane ^ offset using ds_bpermute."""
-  idx = ((lane ^ offset) * 4).cast(dtypes.int)
-  if val.op is Ops.INDEX and val.addrspace == AddrSpace.REG: val = val.load()
-  return UOp(Ops.CUSTOM, dtypes.float, (idx, val),
-             arg="__builtin_bit_cast(float, __builtin_amdgcn_ds_bpermute({0}, __builtin_bit_cast(int, {1})))")
-
-def warp_reduce_max(val, lane):
-  """Tree reduce MAX across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = UOp(Ops.MAX, dtypes.float, (val, warp_shfl_xor(val, offset, lane)))
-  return val
-
-def warp_reduce_sum(val, lane):
-  """Tree reduce SUM across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = val + warp_shfl_xor(val, offset, lane)
-  return val
-
-def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
-  # inputs are (B*H, N, D)
-  BH, N, D = q.shape
-  assert N % BLOCK_M == 0 and N % BLOCK_N == 0, f"N={N} must be divisible by BLOCK_M={BLOCK_M} and BLOCK_N={BLOCK_N}"
-  assert D % WMMA_K == 0 and D % LANES_PER_WAVE_N == 0, f"D={D} must be divisible by WMMA_K={WMMA_K} and LANES_PER_WAVE_N={LANES_PER_WAVE_N}"
-  assert BLOCK_M % (WAVES_M * WMMA_M) == 0 and BLOCK_N % LANES_PER_WAVE_N == 0
-  TM = BLOCK_M // (WAVES_M * LANES_PER_WAVE_M)
-  TN = BLOCK_N // (WAVES_N * LANES_PER_WAVE_N)
-  TD = D // (WAVES_N * LANES_PER_WAVE_N)
-  SCALE = 1.0 / math.sqrt(D)
-
-  block_bh = UOp.range(BH, 0, AxisType.GLOBAL)
-  block_m = UOp.range(N // BLOCK_M, 1, AxisType.GLOBAL)
-
-  q = q.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-  k = k.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  v = v.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  o = o.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-
-  wave_m = UOp.range(WAVES_M, 2, AxisType.LOCAL)
-  wave_n = UOp.range(WAVES_N, 3, AxisType.LOCAL)
-  lane = UOp.range(WARP_SIZE, -1, AxisType.WARP)
-  tid = (wave_m * WAVES_N + wave_n) * WARP_SIZE + lane
-  lane_m = lane // LANES_PER_WAVE_N
-  lane_n = lane % LANES_PER_WAVE_N
-
-  # LDS allocation: slot 0 = Q then P (shared), slot 1 = K then V
-  # TODO: the memory planner should be able to find this reuse
-  ELEMS_PER_THREAD = BLOCK_M * D // THREADS_PER_BLOCK
-  QP_lds = UOp.placeholder((BLOCK_M, D + LDS_PAD), dtypes.half, slot=0, addrspace=AddrSpace.LOCAL)
-  KV_lds = UOp.placeholder((BLOCK_N, D + LDS_PAD), dtypes.half, slot=1, addrspace=AddrSpace.LOCAL)[:, :D]
-
-  # register state
-  acc = UOp.placeholder((TM, TD), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  m_i = UOp.placeholder((TM,), dtypes.float, slot=3, addrspace=AddrSpace.REG)
-  l_i = UOp.placeholder((TM,), dtypes.float, slot=4, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(acc.const_like(0)))
-  m_i = m_i.after(m_i.store(m_i.const_like(-math.inf)))
-  l_i = l_i.after(l_i.store(l_i.const_like(0)))
-
-  # ====== KV tile loop ======
-  n_tile = UOp.range(N // BLOCK_N, 100, AxisType.REDUCE)
-
-  # load Q + K into LDS (Q reloaded each iteration since P overwrites slot 0)
-  Q_lds = QP_lds[:, :D]
-  Q_store = Q_lds.after(n_tile).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    q.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  K_store = KV_lds.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    k[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  # NOTE: no explicit barrier needed, the AFTER on the LOCAL buffers implies it in late codegen
-  Q_lds = Q_lds.after(UOp.group(Q_store, K_store))
-  KV_lds_k = KV_lds.after(UOp.group(Q_store, K_store))
-
-  # -- S = Q @ K^T via WMMA (re-init each n_tile) --
-  S_reg = UOp.placeholder((TM, TN), dtypes.float, slot=6, addrspace=AddrSpace.REG)
-  S_reg = S_reg.after(S_reg.after(n_tile).store(S_reg.const_like(0)))
-  k_qk = UOp.range(D // WMMA_K, 101, AxisType.REDUCE)
-  tm1 = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-  tn1 = UOp.range(TN, 201, AxisType.LOOP)
-  S_frag = S_reg.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0, 2, 1)[tm1, tn1]
-  q_frag = Q_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, D // WMMA_K, WMMA_K)[wave_m, tm1, lane_n, k_qk]
-  k_frag = KV_lds_k.reshape(WAVES_N, TN, WMMA_N, D // WMMA_K, WMMA_K)[wave_n, tn1, lane_n, k_qk]
-  qk = UOp.wmma(q_frag, k_frag, S_frag.after(k_qk), *WMMA_ARG)
-  qk_done = S_frag.store(qk).end(tm1, tn1).end(k_qk)
-  S_reg = S_reg.after(qk_done)
-
-  # -- softmax in registers with warp shuffles --
-  S_reg = S_reg.after(S_reg.store(S_reg * SCALE))
-
-  # per-thread local row max over TN=4 elements, then warp reduce across 16 lanes
-  m_ij = UOp.placeholder((TM,), dtypes.float, slot=7, addrspace=AddrSpace.REG)
-  m_ij = m_ij.after(m_ij.after(n_tile).store(m_ij.const_like(-math.inf)))
-  rm2 = UOp.range(TN, 261, AxisType.REDUCE)
-  m_ij = m_ij.after(m_ij.store(m_ij.after(rm2).maximum(S_reg[:, rm2])).end(rm2))
-  # warp reduce max (in-place)
-  ri_w = UOp.range(TM, 270, AxisType.LOOP)
-  m_ij = m_ij.after(m_ij[ri_w].store(warp_reduce_max(m_ij[ri_w], lane)).end(ri_w))
-
-  # compute P = exp(S - m_ij) in S_reg
-  S_reg = S_reg.after(S_reg.store(((S_reg - m_ij.reshape(TM, 1).expand(TM, TN)) * LOG2E).exp2()))
-
-  p_local = UOp.placeholder((TM,), dtypes.float, slot=8, addrspace=AddrSpace.REG)
-  p_local = p_local.after(p_local.after(n_tile).store(p_local.const_like(0)))
-  rp2 = UOp.range(TN, 291, AxisType.REDUCE)
-  p_local = p_local.after(p_local.store(p_local.after(rp2) + S_reg[:, rp2]).end(rp2))
-  ri_ws = UOp.range(TM, 295, AxisType.LOOP)
-  p_sum = p_local.after(p_local[ri_ws].store(warp_reduce_sum(p_local[ri_ws], lane)).end(ri_ws))
-
-  # write P = exp(S - m_ij) to P_lds (reuses slot 0, Q no longer needed)
-  P_lds = QP_lds[:, :BLOCK_N]
-  P_write = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TN, LANES_PER_WAVE_N)
-  P_write = P_write.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TN)
-  P_store = P_write[tid].store(S_reg.cast(dtypes.half))
-
-  # -- online softmax correction --
-  ri4 = UOp.range(TM, 330, AxisType.LOOP)
-  m_new_val = m_i[ri4].maximum(m_ij[ri4])
-  alpha_val = ((m_i[ri4] - m_new_val) * LOG2E).exp2()
-  beta_val = ((m_ij[ri4] - m_new_val) * LOG2E).exp2()
-  rj4 = UOp.range(TD, 331, AxisType.LOOP)
-  correction = UOp.group(
-    acc[ri4, rj4].store(alpha_val * acc[ri4, rj4]).end(rj4),
-    l_i[ri4].store(alpha_val * l_i[ri4] + beta_val * p_sum[ri4]),
-    m_i[ri4].store(m_new_val),
-  ).end(ri4)
-  acc = acc.after(correction)
-  l_i = l_i.after(correction)
-  m_i = m_i.after(correction)
-
-  # load V into KV_lds (must wait for QK WMMA to finish reading K from KV_lds)
-  V_store = KV_lds.after(qk_done).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    v[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  # NOTE: no explicit barrier needed, the AFTER on the LOCAL buffers implies it in late codegen
-  P_lds = P_lds.after(UOp.group(P_store, V_store))
-  KV_lds_v = KV_lds.after(UOp.group(P_store, V_store))
-
-  # -- acc += P @ V via WMMA --
-  k_pv = UOp.range(BLOCK_N // WMMA_K, 400, AxisType.REDUCE)
-  tm2 = UOp.range(TM // WMMA_ACC, 401, AxisType.LOOP)
-  tn2 = UOp.range(TD, 402, AxisType.LOOP)
-  acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TD).permute(0, 2, 1)[tm2, tn2]
-  p_frag = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_N // WMMA_K, WMMA_K)[wave_m, tm2, lane_n, k_pv]
-  v_frag = KV_lds_v.reshape(WAVES_N, TD, WMMA_N, BLOCK_N // WMMA_K, WMMA_K)[wave_n, tn2, lane_n, k_pv]
-  pv = UOp.wmma(p_frag, v_frag, acc_frag.after(k_pv), *WMMA_ARG)
-
-  # end KV tile loop
-  n_tile_end = acc_frag.store(pv).end(tm2, tn2).end(k_pv).end(n_tile)
-  acc = acc.after(n_tile_end)
-  l_i = l_i.after(n_tile_end)
-  m_i = m_i.after(n_tile_end)
-
-  # normalize: acc /= l_i
-  acc = acc.after(acc.store(acc * (1 / l_i).reshape(TM, 1).expand(TM, TD)))
-
-  # store output
-  o = o.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TD, LANES_PER_WAVE_N)
-  o = o.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TD)
-  return o[tid].store(acc).end(wave_m, wave_n, lane).end(block_m, block_bh).sink(arg=KernelInfo(opts_to_apply=()))
-
-if __name__ == "__main__":
-  B, H, N, D = getenv("B", 1), getenv("H", 32), getenv("N", 1024), getenv("D", 64)
-  q = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  k = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  v = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  o = Tensor.empty(B, H, N, D, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(q, k, v)
-
-  q_flat, k_flat, v_flat, o_flat = q.reshape(B*H, N, D), k.reshape(B*H, N, D), v.reshape(B*H, N, D), o.reshape(B*H, N, D)
-  NUM_RUNS = getenv("CNT", 5)
-  ets = []
-  with Context(DEBUG=2):
-    for _ in range(NUM_RUNS):
-      GlobalCounters.reset()
-      tst = Tensor.custom_kernel(o_flat, q_flat, k_flat, v_flat, fxn=amd_flash_attention)[0].realize()
-      ets.append(GlobalCounters.time_sum_s)
-  print(f"best time: {min(ets)*1e3:.2f}ms")
-
-  if getenv("VERIFY", 1):
-    with Context(DEBUG=0):
-      ref = q.float().scaled_dot_product_attention(k.float(), v.float()).reshape(B*H, N, D).realize()
-      err = (ref - tst).square().mean().item()
-    print(f"mean squared error {err}")
-    if err > 1e-2:
-      raise RuntimeError("flash attention is wrong!")
-    else:
-      print("flash attention is correct!")
@@ -28,10 +28,10 @@ REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
 assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
 assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"

-def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
+def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.WEAK): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
 def copy(dest:UOp, src:UOp, rng:int, upcast=False):
  assert dest.shape == src.shape
-  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
+  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.WEAK)
  return dest[*rngs].store(src[*rngs]).end(*rngs)

 def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
@@ -1,8 +1,9 @@
-import atexit, functools, pathlib
+import atexit, functools, math, pathlib
 from tinygrad import Tensor, Device, dtypes
+from tinygrad.dtype import AddrSpace
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
 from tinygrad.renderer import Estimates
-from tinygrad.helpers import getenv, all_same, DEBUG
+from tinygrad.helpers import getenv, all_same, DEBUG, ceildiv
 from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8

@@ -107,6 +108,23 @@ def custom_hk_mxfp8_gemm(C:UOp, A:UOp, B:UOp, scale_A:UOp, scale_B:UOp, *extra:U
  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
                               UOp(Ops.BINARY, arg=lib)))

+# ** MXFP4 GEMM custom kernel
+
+@functools.cache
+def custom_mxfp4_gemm(C:UOp, A:UOp, B:UOp, scale_a:UOp, scale_b:UOp, *extra:UOp, tile_m:int, tile_n:int) -> UOp:
+  from extra.gemm.gemm_mxfp4 import build_kernel
+  M, half_k = math.prod(A.shape[:-1]), A.shape[-1]
+  N, half_k_b = math.prod(B.shape[:-1]), B.shape[-1]
+  K = half_k * 2
+  assert half_k == half_k_b and math.prod(C.shape[:-1]) == M and C.shape[-1] == N
+  threads = UOp.special(256, "lidx0")
+  groups_x, groups_y = UOp.special(ceildiv(N, tile_n), "gidx0"), UOp.special(ceildiv(M, tile_m), "gidx1")
+  lds = UOp.placeholder((163840,), dtypes.uint8, 0, AddrSpace.LOCAL)
+  sink = UOp.sink(C.base, A.base, B.base, scale_a.base, scale_b.base, *(x.base for x in extra), lds, threads, groups_x, groups_y,
+                  arg=KernelInfo(f"custom_mxfp4_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K)))
+  insts = build_kernel(M, N, K, tile_m, tile_n)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=x) for x in insts))))
+
 def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
  # 1x32 block scaling along the last axis
  *batch, K = x.shape
@@ -119,6 +137,50 @@ def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
  packed = mx_pack(e8) if len(batch) == 1 and scale_K % 4 == 0 else None
  return x_clamped.cast(FP8_DTYPE), e8, packed

+def _mxfp4_shuffle_weight(x:Tensor) -> Tensor:
+  # shuffle_weight(x, layout=(16, 16)) on the packed uint8 buffer.
+  if x.ndim == 3:
+    ndev, rows, half_k = x.shape
+    return x.reshape(ndev, rows//16, 16, half_k//32, 2, 16).permute(0, 1, 3, 4, 2, 5).reshape(ndev, rows, half_k).contiguous()
+  rows, half_k = x.shape
+  return x.reshape(rows//16, 16, half_k//32, 2, 16).permute(0, 2, 3, 1, 4).reshape(rows, half_k).contiguous()
+
+def _mxfp4_shuffle_scales(x:Tensor) -> Tensor:
+  # e8m0_shuffle: each 256x8 scale tile is arranged for the raw MFMA scale loads.
+  if x.ndim == 3:
+    ndev, rows, scale_k = x.shape
+    return x.reshape(ndev, rows//32, 2, 16, scale_k//8, 2, 4).permute(0, 1, 4, 6, 3, 5, 2).reshape(ndev, rows, scale_k).contiguous()
+  rows, scale_k = x.shape
+  return x.reshape(rows//32, 2, 16, scale_k//8, 2, 4).permute(0, 3, 5, 2, 4, 1).reshape(rows, scale_k).contiguous()
+
+def quantize_mxfp4(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  # OCP MXFP4: 1x32 blocks, E2M1 values packed low-nibble first, and E8M0 scales.
+  *batch, K = x.shape
+  rows = math.prod(batch)
+  assert x.ndim >= 2 and K % 256 == 0 and rows % 32 == 0, \
+    f"mxfp4 quantization needs rows%32 and K%256, got {x.shape}"
+  xb = x.float().reshape(*batch, K//32, 32)
+  amax = xb.abs().max(axis=-1)
+
+  # even scale rounding: round the fp32 significand before choosing 2^(floor(log2)-2).
+  amax_rounded = ((amax.bitcast(dtypes.uint32) + 0x200000) & 0xFF800000).bitcast(dtypes.float32)
+  scale_exp = (amax_rounded.maximum(2**-126).log2().floor() - 2).clamp(-127, 127)
+  e8 = (scale_exp + 127).cast(dtypes.uint8)
+  scaled = xb * (-scale_exp).exp2().reshape(*batch, K//32, 1)
+
+  mag = scaled.abs()
+  code = sum(x.cast(dtypes.uint8) for x in
+             (mag > .25, mag >= .75, mag > 1.25, mag >= 1.75, mag > 2.5, mag >= 3.5, mag > 5.0))
+  code = code | ((scaled < 0).cast(dtypes.uint8) << 3)
+  code = code.reshape(*batch, K)
+  packed = code[..., 0::2] | (code[..., 1::2] << 4)
+  if isinstance(x.device, tuple) and x.uop.axis == x.ndim-2 and x.shape[x.uop.axis] == len(x.device):
+    axis = x.uop.axis
+    order = (axis, *range(axis), *range(axis+1, e8.ndim))
+    e8_local = e8.permute(order)
+    return packed, e8, _mxfp4_shuffle_scales(e8_local.reshape(e8_local.shape[0], -1, K//32))
+  return packed, e8, _mxfp4_shuffle_scales(e8.reshape(rows, K//32))
+
 def mx_pack(e8:Tensor) -> Tensor:
  rows, scale_K = e8.shape
  return e8.reshape(rows, scale_K // 4, 4).bitcast(dtypes.uint32).reshape(rows, scale_K // 4).permute(1, 0).contiguous()
@@ -171,13 +233,13 @@ def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
  M, K = A.shape[0]*A.shape[1], A.shape[2]
  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
  assert K == K2
-  m = UOp.range(M, 1, AxisType.LOOP)
-  n = UOp.range(N, 2, AxisType.LOOP)
+  m = UOp.range(M, 1)
+  n = UOp.range(N, 2)
  k = UOp.range(K, 0, AxisType.REDUCE)
-  mul = (A.flatten().index((m*UOp.const(dtypes.weakint, K)+k))*
-         B.flatten().index((k*UOp.const(dtypes.weakint, N)+n))).cast(dtypes.float32)
+  mul = (A.flatten().index((m*UOp.const(K)+k))*
+         B.flatten().index((k*UOp.const(N)+n))).cast(dtypes.float32)
  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype)
-  store = C.flatten().index((m*UOp.const(dtypes.weakint, N)+n)).store(red).end(m, n)
+  store = C.flatten().index((m*UOp.const(N)+n)).store(red).end(m, n)
  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))

 # ** bf16 A @ B.T kernel in C
@@ -244,7 +306,6 @@ def hk_bf16_atb_gemm(a:Tensor, b:Tensor) -> Tensor:
  if reduce_out: out = out.sum(0)
  return out.squeeze(0) if out.ndim == 3 else out

-
 # ** backward gemm, might use the asm gemm

 def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=False, has_w_post:bool=False):
@@ -341,13 +402,29 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F
  if wp is not None: grad_b = grad_b / wp.reshape(-1, 1)
  return (None, grad_a.uop, grad_b.uop) + tuple(None for _ in inputs[3:])

+# ** mxfp4 gemm backward
+
+def custom_mxfp4_gemm_bw(gradient:UOp, kernel:UOp):
+  # The raw kernel consumes quantized buffers, while the final two inputs retain the BF16 operands for STE gradients.
+  inputs = kernel.src[1:]  # (out, a_q, b_q, scale_a, scale_b, a, w)
+  assert len(inputs) == 7
+  a, w = Tensor(inputs[5], device=inputs[5].device), Tensor(inputs[6], device=inputs[6].device)
+  g = Tensor(gradient, device=a.device)[:a.shape[0]].cast(dtypes.bfloat16)
+  grad_a = asm_gemm(g, w, mxfp4=True)
+  a_flat, g_flat = a.reshape(-1, a.shape[-1]), g.reshape(-1, g.shape[-1])
+  grad_w = asm_gemm(g_flat.T, a_flat, mxfp4=True)
+  return (None, None, None, None, None, grad_a.uop, grad_w.uop)
+
 # ** main gemm function

 def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=None, grad_amax_state:Tensor|None=None,
             next_grad_amax_state:Tensor|None=None,
             w_post_scale:Tensor|None=None, mx:bool=False, mx_scales:tuple|None=None, mx_w_stored:bool=False, g_amax:Tensor|None=None,
-             a_pretranspose:Tensor|None=None) -> Tensor:
+             a_pretranspose:Tensor|None=None, mxfp4:bool=False) -> Tensor:
  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
+  if mxfp4:
+    assert not mx and mx_scales is None, "mxfp4 owns quantization; mx/mx_scales are for mxfp8"
+    assert a.dtype == dtypes.bfloat16, f"cannot quantize {a.dtype} to mxfp4"
  counters["used"] += 1
  unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
  if unfold_batch:
@@ -355,7 +432,7 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
    a = a.reshape(a.shape[0]*a.shape[1], a.shape[2])
  squeeze = a.ndim == 2
  if squeeze: a = a.unsqueeze(0)
-  out_dtype = dtypes.bfloat16 if a.dtype == FP8_DTYPE else a.dtype
+  out_dtype = dtypes.bfloat16 if a.dtype == FP8_DTYPE or mxfp4 else a.dtype

  batch, M, K = a.shape
  N = b.shape[1]
@@ -378,7 +455,21 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
  renderer = Device[dname:=(a.device[0] if is_multi else a.device)].renderer
  dname, arch = dname.split(":")[0], renderer.target.arch
  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
-    if mx:
+    if mxfp4:
+      tile_m, tile_n = next((tm, tn) for tm, tn in ((256, 256), (192, 256), (128, 512)) if (batch*M) % tm == N % tn == 0)
+      fxn = functools.partial(custom_mxfp4_gemm, tile_m=tile_m, tile_n=tile_n)
+      w = b.T
+      if k_sharded:
+        ndev = len(a.device)
+        a_q, _, scale_a = quantize_mxfp4(a.reshape(batch, M, ndev, K))
+        b_q, _, scale_b = quantize_mxfp4(w.reshape(w.shape[0], ndev, K))
+        b_q = _mxfp4_shuffle_weight(b_q.permute(1, 0, 2))
+      else:
+        a_q, _, scale_a = quantize_mxfp4(a.reshape(batch*M, K))
+        b_q, _, scale_b = quantize_mxfp4(w)
+        a_q, b_q = a_q.reshape(batch, M, K//2).contiguous(), _mxfp4_shuffle_weight(b_q)
+      out = Tensor.custom_kernel(out, a_q, b_q, scale_a, scale_b, a, w, fxn=fxn, grad_fxn=custom_mxfp4_gemm_bw)[0]
+    elif mx:
      # mxfp8 1x32 block scaling
      if mx_scales is not None:
        a_si, a_e8, b_si, b_e8 = mx_scales
@@ -79,7 +79,7 @@ def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
  # this is the big accumulator
  acc = UOp.placeholder((BLOCK_N//TC_N, BLOCK_M//TC_M//WARPGROUP_SIZE), dtypes.float, 0, AddrSpace.REG)
  assert acc.size*WARP_SIZE*WARPGROUP_SIZE*4 == BLOCK_M*BLOCK_N
-  acc = acc[init_l:=UOp.range(acc.size, 500)].set(UOp.const(dtypes.float, (0.0,)*4), end=init_l)
+  acc = acc[init_l:=UOp.range(acc.size, 500)].set(UOp.const((0.0,)*4, dtypes.float), end=init_l)

  # create locals (note A is permuted, and the stride is changed to avoid bank conflicts)
  def make_locals(slot) -> tuple[UOp, UOp]:
@@ -29,7 +29,7 @@ TID_SIZE = WARPGROUP_SIZE*WARP_SIZE

 def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=()):
  assert dest.shape == src.shape
-  rngs = [UOp.range(s, rng+i, AxisType.UPCAST if i in upcast else AxisType.LOOP) for i,s in enumerate(src.shape)]
+  rngs = [UOp.range(s, rng+i, AxisType.UPCAST if i in upcast else AxisType.WEAK) for i,s in enumerate(src.shape)]
  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
  return dest.after(copy) if set else copy

@@ -0,0 +1,130 @@
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+
+BLOCK_ROW = 256
+
+def _sharded_invalids(shape:tuple[int, ...], dtype, device) -> Tensor:
+  if isinstance(device, tuple):
+    return Tensor.invalids(*shape, dtype=dtype, device=device[0]).shard(device, axis=0)
+  return Tensor.invalids(*shape, dtype=dtype, device=device)
+
+def _atomic_add(device:str) -> str:
+  return "__hip_atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);" if device == "AMD" \
+    else "__atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED);"
+
+def _blk_for(D:int) -> int:
+  blk = 64
+  while D % blk: blk //= 2
+  return blk
+
+def _kv_ranges(G, N, D, BLK):
+  g = UOp.range(G, 0)
+  m = UOp.range(N, 1)
+  jo = UOp.range(D // BLK, 2)
+  ji = UOp.range(BLK, 3, AxisType.LOCAL)
+  return g, m, jo * BLK + ji, jo, ji
+
+def _ggather_fwd_kernel(out:UOp, table:UOp, idx:UOp) -> UOp:
+  G, M, D = out.shape
+  g, m, j, jo, ji = _kv_ranges(G, M, D, _blk_for(D))
+  row = idx.index(g, m).cast(dtypes.weakint)
+  val = table.index(g, row, j).load()
+  return out.index(g, m, j).store(val).end(g, m, jo, ji).sink(
+    arg=KernelInfo(name=f"ggather_fwd_{M}_{D}", opts_to_apply=()))
+
+def _ggather_zero_kernel(out:UOp) -> UOp:
+  i = UOp.range(out.numel(), 0)
+  return out.flatten().index(i).store(UOp.const(0.0, out.dtype)).end(i).sink(arg=KernelInfo(name="ggather_zero"))
+
+def _sharded_zeros(shape:tuple[int, ...], dtype, device) -> Tensor:
+  return Tensor.custom_kernel(_sharded_invalids(shape, dtype, device), fxn=_ggather_zero_kernel)[0]
+
+def _ggather_bwd(gradient:UOp, kernel:UOp) -> tuple:
+  _, table_u, idx_u = kernel.src[1:4]
+  dev = table_u.device
+  device = (dev[0] if isinstance(dev, tuple) else dev).split(":")[0]
+  G, R, D = table_u.shape
+  gt = _sharded_zeros((G, R, D), dtypes.float32, dev)
+  go = Tensor(gradient, device=dev)
+  atomic_str = _atomic_add(device)
+  def _bwd_kernel(gtab:UOp, gout:UOp, idx:UOp) -> UOp:
+    Gk, M, Dk = gout.shape
+    g, m, j, jo, ji = _kv_ranges(Gk, M, Dk, _blk_for(Dk))
+    row = idx.index(g, m).cast(dtypes.weakint)
+    val = gout.index(g, m, j).load().cast(dtypes.float32)
+    atomic = UOp(Ops.CUSTOM, dtypes.void, (gtab.index(g, row, j), val), arg=atomic_str)
+    return atomic.end(g, m, jo, ji).sink(arg=KernelInfo(name=f"ggather_bwd_{M}_{Dk}", opts_to_apply=()))
+  grad_table = Tensor.custom_kernel(gt, go, Tensor(idx_u, device=dev), fxn=_bwd_kernel)[0]
+  return (None, grad_table.cast(table_u.dtype).uop, None)
+
+def grouped_gather_rows(table:Tensor, idx:Tensor, n_groups:int) -> Tensor:
+  G, R, D = table.shape
+  M = idx.shape[1]
+  out = _sharded_invalids((G, M, D), table.dtype, table.device)
+  return Tensor.custom_kernel(out, table, idx, fxn=_ggather_fwd_kernel, grad_fxn=_ggather_bwd)[0]
+
+def _gscatter_fwd_kernel(out:UOp, src:UOp, idx:UOp) -> UOp:
+  G, M, D = out.shape
+  k = idx.shape[1] // src.shape[1]
+  g, m, j, jo, ji = _kv_ranges(G, idx.shape[1], D, _blk_for(D))
+  row = idx.index(g, m).cast(dtypes.weakint)
+  val = src.index(g, (m // k).cast(dtypes.weakint), j).load()
+  return out.index(g, row, j).store(val).end(g, m, jo, ji).sink(
+    arg=KernelInfo(name=f"gscatter_fwd_{idx.shape[1]}_{D}", opts_to_apply=()))
+
+def _gscatter_bwd(gradient:UOp, kernel:UOp) -> tuple:
+  _, src_u, idx_u = kernel.src[1:4]
+  dev = src_u.device
+  G, T_l, D = src_u.shape
+  k = idx_u.shape[1] // T_l
+  sel = grouped_gather_rows(Tensor(gradient, device=dev), Tensor(idx_u, device=dev), G)
+  return (None, sel.reshape(G, T_l, k, D).sum(2).cast(src_u.dtype).uop, None)
+
+def grouped_scatter_rows(src:Tensor, idx:Tensor, m_l:int) -> Tensor:
+  G, T_l, D = src.shape
+  zero = _sharded_zeros((G, m_l, D), src.dtype, src.device)
+  return Tensor.custom_kernel(zero, src, idx, fxn=_gscatter_fwd_kernel, grad_fxn=_gscatter_bwd)[0]
+
+def m_max_for(t_local:int, experts_per_tok:int, n_experts:int) -> int:
+  return (-(-t_local * experts_per_tok // BLOCK_ROW) + n_experts) * BLOCK_ROW
+
+class Routing:
+  def __init__(self, weights:Tensor, dest_row:Tensor, off:Tensor, m_l:int, n_groups:int, t_local:int):
+    self.weights, self.dest_row = weights, dest_row
+    self.off = off
+    self.m_l, self.n_groups, self.t_local = m_l, n_groups, t_local
+
+  @property
+  def rows_e(self) -> Tensor:
+    G, E = self.off.shape[0], self.off.shape[1] - 1
+    tr = Tensor.arange(self.m_l // BLOCK_ROW, dtype=dtypes.int32).reshape(1, -1, 1) * BLOCK_ROW
+    tr = tr.shard(self.off.device) if isinstance(self.off.device, tuple) else tr.to(self.off.device)
+    tile_e = ((tr >= self.off[:, :E].reshape(G, 1, E)).sum(-1) - 1).cast(dtypes.int32)
+    return tile_e.reshape(-1, 1).expand(-1, BLOCK_ROW).reshape(-1)
+
+def n_groups_of(t:Tensor) -> int:
+  return len(t.device) if isinstance(t.device, tuple) else 1
+
+def route(logits:Tensor, experts_per_tok:int, n_experts:int) -> Routing:
+  T, E = logits.shape
+  k, G = experts_per_tok, n_groups_of(logits)
+  assert T % G == 0, f"tokens {T} must split across {G} devices"
+  T_l, m_l = T // G, m_max_for(T // G, k, n_experts)
+
+  topv, topi = logits.reshape(G, T_l, E).topk(k)
+  weights = topv.softmax(-1)
+  m = topi.reshape(G, T_l * k).cast(dtypes.int32).one_hot(E).cast(dtypes.int32)
+
+  pad = ((m.sum(1) + (BLOCK_ROW - 1)) // BLOCK_ROW) * BLOCK_ROW
+  off = pad.cumsum(1).pad(((0, 0), (1, 0)))
+  dest_row = ((m.cumsum(1) + off[:, :E].reshape(G, 1, E)) * m).sum(-1).sub(1).cast(dtypes.int32)
+  return Routing(weights, dest_row, off, m_l, G, T_l)
+
+def dispatch(x:Tensor, r:Routing) -> Tensor:
+  G, D = r.n_groups, x.shape[-1]
+  return grouped_scatter_rows(x.reshape(G, r.t_local, D), r.dest_row, r.m_l).reshape(G * r.m_l, D)
+
+def combine(y:Tensor, r:Routing, n_tokens:int, experts_per_tok:int) -> Tensor:
+  G, D, k = r.n_groups, y.shape[-1], experts_per_tok
+  sel = grouped_gather_rows(y.reshape(G, r.m_l, D), r.dest_row, G).reshape(G, r.t_local, k, D)
+  return (sel * r.weights.reshape(G, r.t_local, k, 1).cast(sel.dtype)).sum(2).reshape(n_tokens, D).cast(y.dtype)
@@ -0,0 +1,139 @@
+"""
+tilelang-style matmul_relu written with tinygrad UOp APIs.
+
+Reference tilelang kernel:
+  @tilelang.jit
+  def matmul_relu(A, B, block_M=64, block_N=64, block_K=64,
+                  dtype=T.float16, accum_dtype=T.float32):
+    M, N, K = T.const('M, N, K')
+    C = T.empty([M, N], dtype)
+    with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=128) as (bx, by):
+      A_shared = T.alloc_shared((block_M, block_K), dtype)
+      B_shared = T.alloc_shared((block_K, block_N), dtype)
+      C_local  = T.alloc_fragment((block_M, block_N), accum_dtype)
+      T.clear(C_local)
+      for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=3):
+        T.copy(A[by * block_M, ko * block_K], A_shared)
+        T.copy(B[ko * block_K, bx * block_N], B_shared)
+        T.gemm(A_shared, B_shared, C_local)
+      for i, j in T.Parallel(block_M, block_N):
+        C_local[i, j] = T.max(C_local[i, j], 0)
+      T.copy(C_local, C[by * block_M, bx * block_N])
+    return C
+"""
+
+from tinygrad.dtype import dtypes, AddrSpace, DType
+from tinygrad.uop.ops import UOp, Ops, AxisType, KernelInfo
+from tinygrad.helpers import cdiv, getenv
+from tinygrad.tensor import Tensor
+
+# ---------------------------------------------------------------------------
+# tilelang builtins, expressed with tinygrad UOp APIs
+# ---------------------------------------------------------------------------
+
+def alloc_shared(shape:tuple[int, ...], dtype:DType, slot:int) -> UOp:
+  """T.alloc_shared: one LOCAL buffer shared by all threads in the block."""
+  return UOp.placeholder(tuple(shape), dtype, slot, AddrSpace.LOCAL)
+
+def alloc_fragment(shape:tuple[int, ...], dtype:DType, slot:int, axes:tuple[int, ...], rngs:tuple[UOp, ...]) -> UOp:
+  """T.alloc_fragment: per-thread REG fragment + UNSHARD over the LOCAL thread grid."""
+  assert len(axes) == len(rngs)
+  assert all(tnum.op is Ops.RANGE and tnum.arg[-1] is AxisType.LOCAL for tnum in rngs), "fragments shard over LOCAL ranges"
+  by_axis = dict(zip(axes, rngs))
+  shard_shape = tuple(s // (int(by_axis[i].vmax)+1) if i in by_axis else s for i, s in enumerate(shape))
+  fragment = UOp.placeholder(shard_shape, dtype, slot, AddrSpace.REG)
+  return fragment.unshard(axes, rngs)
+
+# ---------------------------------------------------------------------------
+# GEMM kernel: C = relu(A @ B), float inputs (fp16 or fp32), fp32 fragment accumulator, no WMMA
+# ---------------------------------------------------------------------------
+
+# 64x64 output tile per block, 128 threads as an 8x16 grid; each thread owns an 8x4 fragment sub-tile
+# (the 2-D per-thread layout tilelang infers for this GEMM). The 4 contiguous columns (TN=4) are what
+# let codegen vectorize loads/stores to float4, matching tilelang's lowering exactly.
+BLOCK_M = BLOCK_N = BLOCK_K = 64
+TY = 8
+TX = 16
+THREADS = TY * TX
+TM = BLOCK_M // TY   # fragment rows per thread  (8)
+TN = BLOCK_N // TX   # fragment columns per thread (4)
+
+def matmul_relu_kernel(c:UOp, a:UOp, b:UOp) -> UOp:
+  """C[M, N] = relu(A[M, K] @ B[K, N]) -- one 64x64 tile per block, locals + a 2-D fragment."""
+  M, K = a.shape
+  K2, N = b.shape
+  assert K == K2 and a.dtype == b.dtype == c.dtype and not dtypes.is_int(a.dtype)
+  assert not (K % BLOCK_K or M % BLOCK_M or N % BLOCK_N), "test sizes must be multiples of the block sizes"
+
+  # with T.Kernel(T.ceildiv(N, BLOCK_N), T.ceildiv(M, BLOCK_M), threads=128) as (bx, by):
+  bx = UOp.range(cdiv(N, BLOCK_N), 0, AxisType.GLOBAL)
+  by = UOp.range(cdiv(M, BLOCK_M), 1, AxisType.GLOBAL)
+
+  # 16*8 threads = 128 threads
+  tx = UOp.range(TX, 2, AxisType.LOCAL)
+  ty = UOp.range(TY, 3, AxisType.LOCAL)
+
+  # shared + fragment (regs)
+  A_shared = alloc_shared((BLOCK_M, BLOCK_K), a.dtype, 0)
+  B_shared = alloc_shared((BLOCK_K, BLOCK_N), b.dtype, 1)
+  C_local = alloc_fragment((TM, TY, TX, TN), dtypes.float32, 0, (1, 2), (ty, tx))
+
+  # zero out the regs to start. this is expanded by the devectorizer
+  C_local = C_local.after(C_local.store(0.0))
+
+  # for ko in T.Pipelined(T.ceildiv(K, BLOCK_K), num_stages=3):
+  ko = UOp.range(cdiv(K, BLOCK_K), 6, AxisType.LOOP)
+
+  # index the outer matrices
+  a = a.rearrange("(m bm) (k bk) -> m k bm bk", bm=BLOCK_M, bk=BLOCK_K)[by, ko]
+  b = b.rearrange("(k bk) (n bn) -> k n bk bn", bk=BLOCK_K, bn=BLOCK_N)[ko, bx]
+  c = c.rearrange("(m bm) (n bn) -> m n bm bn", bm=BLOCK_M, bn=BLOCK_N)[by, bx]
+
+  # T.copy: A_shared <- a, B_shared <- b
+  def with_threads(x:UOp): return x.rearrange("(tm ty) (tx tn) -> ty tx tm tn", tm=TM, tn=TN)[ty, tx]
+  A_shared = A_shared.after(with_threads(A_shared).store(with_threads(a)))
+  B_shared = B_shared.after(with_threads(B_shared).store(with_threads(b)))
+
+  # T.gemm(A_shared, B_shared, C_local), no WMMA
+  kk = UOp.range(BLOCK_K, 11, AxisType.LOOP)
+  ir = UOp.range(TM, 12, AxisType.LOOP)
+  jj = UOp.range(TN, 13, AxisType.UPCAST)
+  acc = C_local.after(kk)[ir, ty, tx, jj] + A_shared[ir*TM + ty, kk].cast(dtypes.float32) * B_shared[kk, tx*TN + jj].cast(dtypes.float32)
+  # closing the ko loop here too; codegen adds the barrier so no thread overwrites the tiles while others still read them
+  C_local = C_local[ir, ty, tx, jj].set(acc, end=(kk, ir, jj, ko))
+
+  # c <- C_local (with relu and cast): every thread stores its shard's sub-view of the output tile
+  c_st = c.reshape(C_local.shape).store(C_local.relu().cast(c.dtype))
+
+  # close the locals and globals
+  return c_st.end(tx, ty, bx, by).sink(arg=KernelInfo(name="matmul_relu", opts_to_apply=()))
+
+# ---------------------------------------------------------------------------
+# python wrapper: same signature as the tilelang function
+# ---------------------------------------------------------------------------
+
+def matmul_relu(a:Tensor, b:Tensor) -> Tensor:
+  """C = relu(A @ B), fp16 in/out with an fp32 fragment accumulator."""
+  c = Tensor.empty(a.shape[0], b.shape[1], dtype=a.dtype, device=a.device)
+  return c.custom_kernel(a, b, fxn=matmul_relu_kernel)[0]
+
+# ---------------------------------------------------------------------------
+# test
+# ---------------------------------------------------------------------------
+
+if __name__ == "__main__":
+  from tinygrad import Device
+  assert Device[Device.DEFAULT].renderer.has_local, "this GPU-style kernel needs a backend with local memory (LOCAL ranges + barriers)"
+  M = K = N = getenv("N", 256)  # 4x4 grid of 64x64 tiles, 4 K chunks
+  dtype_in = dtypes.half if getenv("HALF") else dtypes.float
+
+  a = Tensor.randn(M, K, dtype=dtype_in).contiguous()
+  b = Tensor.randn(K, N, dtype=dtype_in).contiguous()
+  ref = (a @ b).relu().realize()
+
+  for _ in range(10):
+    out = matmul_relu(a, b).realize()
+
+  import numpy as np
+  np.testing.assert_allclose(out.numpy(), ref.numpy(), atol=1e-1, rtol=1e-2)
+  print("matmul_relu passed!")
@@ -4,7 +4,7 @@ import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, co
 assert sys.platform != 'win32'
 from dataclasses import dataclass
 from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_cmdbuf
-from tinygrad.runtime.support.hcq2 import make_binary_patch, make_patch
+from tinygrad.runtime.support.hcq2 import make_binary_patch
 from tinygrad.uop.ops import sint, UOp
 from tinygrad.device import Compiled, BufferSpec, Buffer, Device
 from tinygrad.dtype import dtypes
@@ -37,7 +37,7 @@ class PM4Ops(FastEnum):
  RELEASE_MEM = auto(); DISPATCH_DIRECT = auto(); EVENT_WRITE = auto()  # noqa: E702

 def pkt3(ctx, op:PM4Ops, *vals):
-  return UOp(Ops.INS, arg=op, src=tuple(UOp.const(dtypes.uint32, x)
+  return UOp(Ops.INS, arg=op, src=tuple(UOp.const(x, dtypes.uint32)
    for x in (ctx.pm4.PACKET3(getattr(ctx.pm4, f"PACKET3_{op.name}"), len(vals) - 1), *vals)))

 def wreg(ctx, reg:AMDReg, *args:sint, **kwargs:int):
@@ -152,28 +152,22 @@ def pm4_submit(ctx, lin):
  ring, wptr, doorbell, put_ptr = (UOp.placeholder((b.size,), b.dtype, 0, device=devs).rtag(f"COMPUTE:0_{name}")
    for name, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))

-  # two tail dwords coordinate safe IB reuse: GPU completions and host submits
-  size_dw = sum(len(ins.src) for ins in lin.src) + len(release_mem(ctx, 0, 0).src)
+  # the host fence at the start of the batch guarantees the ib is free to reuse
+  size_dw = sum(len(ins.src) for ins in lin.src)
  assert size_dw < (1 << 20), f"indirect buffer of {size_dw} dwords doesn't fit one packet"

-  ib = UOp.placeholder((size_dw + 2,), dtypes.uint32, next(UOp.unique_num), device=devs, volatile=True).rtag("cmdbuf")
-  done_idx, submit_idx = UOp.const(dtypes.int, size_dw + 0), UOp.const(dtypes.int, size_dw + 1)
-  submitted = (counter:=ib.after(*[make_patch(ib, (size_dw + i) * 4, UOp.const(dtypes.uint32, 0)) for i in range(2)]).index(submit_idx)).load()
-  completed = ib.after(loop:=UOp.loop(0)).index(done_idx).load()
-  ib_free = completed.end(loop, completed != submitted)
-
-  bump_fence = pm4_store(ctx, UOp(Ops.SLICE, dtypes.uint32, (ib, UOp.const(dtypes.weakint, size_dw)), 2), (submitted + 1).cast(dtypes.uint64))
-  cmdbuf = make_cmdbuf(lin.replace(src=lin.src + (bump_fence,)), devs, buf=ib, dep=ib_free)
+  ib = UOp.placeholder((size_dw,), dtypes.uint32, next(UOp.unique_num), device=devs, volatile=True).rtag("cmdbuf")
+  cmdbuf = make_cmdbuf(lin, devs, buf=ib)

  # the ring itself only carries a packet pointing at the ib, wrapping the ring
-  put = put_ptr.index(zero:=UOp.const(dtypes.int, 0))
+  put = put_ptr.index(zero:=UOp.const(0, dtypes.int))
  pkt = (ctx.pm4.PACKET3(ctx.pm4.PACKET3_INDIRECT_BUFFER, 2), *data64_le(cmdbuf.getaddr(devs)), size_dw | ctx.pm4.INDIRECT_BUFFER_VALID)
-  write_pkt = UOp.barrier(*[ring.index(((put + off) % q.ring.size).cast(dtypes.int)).store(UOp.const(dtypes.uint32, x)) for off,x in enumerate(pkt)])
+  write_pkt = UOp.barrier(*[ring.index(((put + off) % q.ring.size).cast(dtypes.int)).store(UOp.const(x, dtypes.uint32)) for off,x in enumerate(pkt)])

  # advance the put/write pointers past the packet
  bump_put_ptr = put_ptr.index(zero).store(put + len(pkt))
  bump_wptr = wptr.index(zero).store(put + len(pkt))
-  flush = UOp.barrier(write_pkt, bump_put_ptr, bump_wptr, counter.store(submitted + 1))
+  flush = UOp.barrier(write_pkt, bump_put_ptr, bump_wptr)
  return doorbell.after(flush).index(zero).store(put + len(pkt))

 pm_pm4_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"), pm4_submit)])
@@ -186,26 +180,26 @@ class SDMAOps(FastEnum): COPY = auto(); POLL_REGMEM = auto(); FENCE = auto(); TR
 def sdma_copy(ctx, call):
  sz = call.src[2].max_numel() * call.src[2].dtype.itemsize
  src_addr, dst_addr = call.src[2].getaddr(ctx.devs), call.src[1].getaddr(ctx.devs)
-  return call.ins(SDMAOps.COPY, src=tuple(UOp.const(dtypes.uint32, x) for off in range(0, sz, ctx.max_copy_size) for x in (
+  return call.ins(SDMAOps.COPY, src=tuple(UOp.const(x, dtypes.uint32) for off in range(0, sz, ctx.max_copy_size) for x in (
    ctx.sdma.SDMA_OP_COPY | ctx.sdma.SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_COPY_LINEAR),
    ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz-off, ctx.max_copy_size)-1), 0, *data64_le(src_addr+off), *data64_le(dst_addr+off))))

 def sdma_wait(ctx, ins, dst, val):
  op = ctx.sdma.SDMA_OP_POLL_REGMEM | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_FUNC(WAIT_REG_MEM_FUNCTION_GEQ) \
     | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)
-  return ins.ins(SDMAOps.POLL_REGMEM, src=tuple(UOp.const(dtypes.uint32, x) for x in (
+  return ins.ins(SDMAOps.POLL_REGMEM, src=tuple(UOp.const(x, dtypes.uint32) for x in (
    op, *data64_le(dst.getaddr(ctx.devs)), val, 0xffffffff,
    ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(0x04) | ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff))))

 def sdma_store(ctx, ins, dst, val):
  op = ctx.sdma.SDMA_OP_FENCE | (ctx.sdma.SDMA_PKT_FENCE_HEADER_MTYPE(3) if ctx.target[0] != 9 else 0)
  return UOp(Ops.LINEAR, src=(
-    ins.ins(SDMAOps.FENCE, src=tuple(UOp.const(dtypes.uint32, x) for x in (op, *data64_le(dst.getaddr(ctx.devs)), val))),
-    ins.ins(SDMAOps.TRAP, src=tuple(UOp.const(dtypes.uint32, x) for x in (ctx.sdma.SDMA_OP_TRAP, 0)))))
+    ins.ins(SDMAOps.FENCE, src=tuple(UOp.const(x, dtypes.uint32) for x in (op, *data64_le(dst.getaddr(ctx.devs)), val))),
+    ins.ins(SDMAOps.TRAP, src=tuple(UOp.const(x, dtypes.uint32) for x in (ctx.sdma.SDMA_OP_TRAP, 0)))))

 def sdma_timestamp(ctx, ins, dst):
  op = ctx.sdma.SDMA_OP_TIMESTAMP | ctx.sdma.SDMA_PKT_TIMESTAMP_GET_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_TIMESTAMP_GET_GLOBAL)
-  return ins.ins(SDMAOps.TIMESTAMP, src=tuple(UOp.const(dtypes.uint32, x) for x in (op, *data64_le(dst.getaddr(ctx.devs)))))
+  return ins.ins(SDMAOps.TIMESTAMP, src=tuple(UOp.const(x, dtypes.uint32) for x in (op, *data64_le(dst.getaddr(ctx.devs)))))

 pm_sdma_opsel = PatternMatcher([
  (UPat(Ops.CALL, src=(UPat(Ops.COPY),), name="call", allow_any_len=True), sdma_copy),
@@ -218,7 +212,7 @@ pm_sdma_opsel = PatternMatcher([

 def sdma_submit(cmdbuf, devs):
  # the cmdbuf to submit + the patch writes that fill it
-  size_dw, zero = cmdbuf.nbytes() // dtypes.uint32.itemsize, UOp.const(dtypes.int, 0)
+  size_dw, zero = cmdbuf.nbytes() // dtypes.uint32.itemsize, UOp.const(0, dtypes.int)

  # the sdma queue's ring and its host-side ring/write/put pointers
  for d in devs: q = Device[d].sdma_queue(0)
@@ -234,8 +228,8 @@ def sdma_submit(cmdbuf, devs):

  # zero the wrapped tail, then copy the cmdbuf into the ring
  zi = UOp.range(zero_amt_dw, 0, dtype=dtypes.int, src=(cmdbuf,))
-  zero_tail = ring.index(tail_off_dw + zi).store(UOp.const(dtypes.uint32, 0)).end(zi)
-  i = UOp.range(UOp.const(dtypes.int, size_dw), 0, dtype=dtypes.int, src=(cmdbuf,))
+  zero_tail = ring.index(tail_off_dw + zi).store(UOp.const(0, dtypes.uint32)).end(zi)
+  i = UOp.range(UOp.const(size_dw, dtypes.int), 0, dtype=dtypes.int, src=(cmdbuf,))
  copy_to_ring = ring.index(start_dw + i).store(cmdbuf.index(i).load()).end(i)

  # advance the put/write pointers past the zeroed tail and the cmdbuf
@@ -517,8 +511,7 @@ class PCIIface(PCIIfaceBase):
      cq = d.compute_queue
      for b in (cq.put_value, cq.read_ptr, cq.write_ptr): b._buf.view.view(fmt='Q')[0] = 0
      d.iface.dev_impl.gfx.setup_ring(*cq.params)
-      d.timeline_signal('COMPUTE:0')._buf.cpu_view().mv.cast('Q')[0] = \
-        d.timeline_value('COMPUTE:0').as_memoryview(force_zero_copy=True).cast('Q')[0] - 1
+      d.signal('timeline')._buf.cpu_view().mv.cast('Q')[0] = d.signal('value', 1).as_memoryview(force_zero_copy=True).cast('Q')[0] - 1

  def sleep(self, timeout):
    if hasattr(self.pci_dev, 'irq_poller') and self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
@@ -638,9 +631,6 @@ class AMDDevice(HCQ2Compiled):
    qname = f"{'COPY' if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA else 'COMPUTE'}:{idx}"
    self.pm_bufferize = PatternMatcher([
      (UPat(Ops.PARAM, tag=f"{qname}_{name}"), lambda ctx, b=getattr(queue, name): b) for name in ["ring", "write_ptr", "doorbell", "put_value"]
-    ] + [
-      (UPat(Ops.PARAM, tag=f"{qname}_timeline_signal"), lambda ctx, q=qname: ctx[0].timeline_signal(q)),
-      (UPat(Ops.PARAM, tag=f"{qname}_timeline_value"), lambda ctx, q=qname: ctx[0].timeline_value(q)),
    ]) + self.pm_bufferize

    return queue
@@ -16,7 +16,7 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_out:UOp, x:UOp, amax_state:

  wg = UOp.range(NUM_WG, 0, AxisType.GLOBAL)
  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
-  it = UOp.range((n_elems // VEC) // (NUM_WG * THREADS_PER_WG), 2, AxisType.LOOP)
+  it = UOp.range((n_elems // VEC) // (NUM_WG * THREADS_PER_WG), 2, AxisType.WEAK)
  lane = UOp.range(VEC, 3, AxisType.UNROLL)

  idx = (((it * NUM_WG + wg) * THREADS_PER_WG + tid) * VEC) + lane
@@ -48,7 +48,7 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_out:UOp, x:UOp, amax_state:
  device = device[0].split(":")[0] if isinstance(device, tuple) else device.split(":")[0]
  if device in {"AMD", "NULL"}: atomic_arg = "if ({2} > {3}) __hip_atomic_fetch_max((int*){0}, {1}, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);"
  else: raise NotImplementedError(f"no atomic max for device {device}")
-  amax_idx = amax_out.reshape((1,)).index(UOp.const(dtypes.weakint, 0))
+  amax_idx = amax_out.reshape((1,)).index(UOp.const(0))
  max_val = lds[0].load()
  atomic = UOp(Ops.CUSTOM, dtypes.void, (amax_idx, max_val.bitcast(dtypes.int32), max_val, amax_idx.load()), arg=atomic_arg)
  return atomic.end(tid, wg).sink(arg=KernelInfo(f"quantize_fp8_with_amax_{n_elems}", opts_to_apply=()))
@@ -4,7 +4,7 @@ from hexdump import hexdump
 from copy import deepcopy
 import pathlib, sys
 from tinygrad.helpers import to_mv, getenv
-from tinygrad.runtime.autogen import adreno
+from tinygrad.runtime.autogen import mesa
 sys.path.append(pathlib.Path(__file__).parent.parent.parent.as_posix())

 IOCTL = getenv("IOCTL", 0)
@@ -23,7 +23,7 @@ for child in xml.getroot():
 CAPTURED_STATE = {}

 REGS = {}
-for k, v in adreno.__dict__.items():
+for k, v in mesa.__dict__.items():
  if k.startswith("REG_") and isinstance(v, int) and v > 1024: REGS[v] = k

 from extra.qcom_gpu_driver import msm_kgsl
@@ -42,7 +42,7 @@ def get_struct(argp, stype):

 def format_struct(s):
  sdats = []
-  for field_name, *_ in s._real_fields_:
+  for field_name, *_ in s._fields_:
    if field_name in {"__pad", "PADDING_0"}: continue
    dat = getattr(s, field_name)
    if isinstance(dat, int): sdats.append(f"{field_name}:0x{dat:X}")
@@ -96,9 +96,9 @@ def parse_cmd_buf(dat):
        CAPTURED_STATE['LOAD_FRAGS'].append((state_block, state_type, num_unit, dst_off))

        if state_block == SB6_CS_SHADER:
-          from extra.disassemblers.adreno import disasm_raw
+          from tinygrad.runtime.support.compiler_mesa import disas_adreno
          if state_type == ST6_SHADER and IOCTL > 3:
-            disasm_raw(get_mem(((vals[2] << 32) | vals[1]), num_unit * 128))
+            disas_adreno(get_mem(((vals[2] << 32) | vals[1]), num_unit * 128))
          if state_type == ST6_CONSTANTS:
            x = get_mem(((vals[2] << 32) | vals[1]), num_unit*4)
            CAPTURED_STATE['constants'] = x[:]
@@ -142,7 +142,7 @@ def parse_cmd_buf(dat):
      vals = struct.unpack("I"*size, dat[ptr+4:ptr+4+4*size])
      if IOCTL > 0: print(f"{ptr:3X} -- typ 4: {size=:3d}, {reg_name}", hprint(vals))
      for vi,v in enumerate(vals): CAPTURED_STATE[offset+vi] = v
-      if offset == adreno.REG_A6XX_SP_CS_CONFIG:
+      if offset == mesa.REG_A6XX_SP_CS_CONFIG:
        val = vals[0]
        if IOCTL > 0:
          print(f"\tBINDLESS_TEX={(val >> 0) & 0b1}")
@@ -215,79 +215,3 @@ def install_hook(c_function, python_function):

 libc = ctypes.CDLL(ctypes.util.find_library("libc"))
 install_hook(libc.ioctl, ioctl)
-
-def before_launch():
-  global CAPTURED_STATE
-  CAPTURED_STATE.clear()
-def collect_last_launch_state():
-  global CAPTURED_STATE
-  return deepcopy(CAPTURED_STATE)
-def compare_launch_state(state, good_state):
-  cmp = [
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_NTEX__MASK),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_NSAMP__MASK),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_NIBO__MASK),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_ENABLED),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_TEX),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_SAMP),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_IBO),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_UBO),
-
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_HALFREGFOOTPRINT__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_BRANCHSTACK__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_THREADMODE__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_EARLYPREAMBLE),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_MERGEDREGS),
-
-    (adreno.REG_A6XX_SP_CS_PVT_MEM_PARAM, adreno.A6XX_SP_CS_PVT_MEM_PARAM_MEMSIZEPERITEM__MASK),
-    (adreno.REG_A6XX_SP_CS_PVT_MEM_PARAM, adreno.A6XX_SP_CS_PVT_MEM_PARAM_HWSTACKSIZEPERTHREAD__MASK),
-
-    (adreno.REG_A6XX_SP_CS_UNKNOWN_A9B1, adreno.A6XX_SP_CS_UNKNOWN_A9B1_UNK5),
-    (adreno.REG_A6XX_SP_CS_UNKNOWN_A9B1, adreno.A6XX_SP_CS_UNKNOWN_A9B1_UNK6),
-
-    (adreno.REG_A6XX_SP_CS_BRANCH_COND, 0xffffffff),
-
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_KERNELDIM__MASK),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEX__MASK),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEY__MASK),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEZ__MASK),
-
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_1, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_2, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_3, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_4, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_5, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_6, 0xffffffff),
-
-    (adreno.REG_A6XX_HLSQ_CS_CNTL_0, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_CNTL_1, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_KERNEL_GROUP_X, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_KERNEL_GROUP_Y, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_KERNEL_GROUP_Z, 0xffffffff),
-  ]
-
-  for x,m in cmp:
-    print(f"Field {REGS[x]}, mask: 0x{m:X} cmp: {state.get(x, 0) & m} vs {good_state.get(x, 0) & m}")
-    if state.get(x, 0) & m != good_state.get(x, 0) & m:
-      return False, f"Field {REGS[x]}, mask: 0x{m:X} mismatch: {state.get(x, 0) & m} vs {good_state.get(x, 0) & m}"
-
-  for n in ['descriptors', 'ibos']:
-    if n not in good_state: continue
-    mv1, mv2 = state.get(n), good_state.get(n)
-
-    if len(mv1) != len(mv2): return False, f"{n}: len mismatch {len(mv1)} != {len(mv2)}"
-    mv1 = memoryview(bytearray(mv1)).cast('I')
-    mv2 = memoryview(bytearray(mv2)).cast('I')
-    for i in range(len(mv2)):
-      if i % 8 == 5 or i % 8 == 4: continue # addresses
-      if mv1[i]!=mv2[i]: return False, f"{n}: content mismatch {i} {mv1[i]} {mv2[i]}"
-
-  for n in ['samplers']:
-    if n not in good_state: continue
-    mv1, mv2 = state.get(n), good_state.get(n)
-    if len(mv1) != len(mv2): return False, f"{n}: len mismatch {len(mv1)} != {len(mv2)}"
-    if any(mv1[i]!=mv2[i] for i in range(len(mv1))): return False, f"{n}: content mismatch"
-
-  return True, "PASS"
@@ -533,6 +533,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -791,6 +793,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1048,6 +1052,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1303,6 +1309,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1582,6 +1590,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1842,6 +1852,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -2099,6 +2111,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -2354,6 +2368,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -48,14 +48,14 @@ class Kernel(AbstractContextManager):
  @property
  def warpgroup(self): return self.group(4)

-  def range(self, start:int, end:int=0, step:int=1, axis_type:AxisType=AxisType.LOOP, track:bool=True):
+  def range(self, start:int, end:int=0, step:int=1, axis_type:AxisType=AxisType.WEAK, track:bool=True):
    if end == 0: start, end = 0, start
    rng = _tk_range(start, end, step, axis_type, self.range_id)
    self.range_id += 1
    if track: self.range_stack.append(rng)
    return rng

-  def raw_range(self, end:int=0, axis_type:AxisType=AxisType.LOOP):
+  def raw_range(self, end:int=0, axis_type:AxisType=AxisType.WEAK):
    rng = UOp.range(end, self.range_id, axis_type=axis_type)
    self.range_id += 1
    return rng
@@ -80,7 +80,7 @@ All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op
 \op{Index} & $(T, i_0, i_1, \ldots)$ & --- & Index from left. $()$-shaped $i$ removes dim; $(k,)$-shaped makes it $k$. \\
 \op{Stack} & $(T_0, T_1, \ldots)$ & --- & Join along a newly created leading axis. All shapes must match. \\
 \op{Bitcast} & $(T,)$ & dtype & Reinterpret storage as target dtype; preserve total bytes. \\
-\op{Unshard} & $(T, R)$ & axis $a$ & Concatenate the shards indexed by \op{Range} $R$ along $a$; $R$ is outer. \\
+\op{Unshard} & $(T, R_0, R_1, \ldots)$ & axes $(a_0, a_1, \ldots)$ & Concatenate shards of \op{Range} $R_k$ along axis $a_k$; $R_k$ is outer. \\
 \bottomrule
 \end{tabular}

@@ -260,7 +260,7 @@ Every UOp has a \textbf{dtype}, \textbf{shape}, \textbf{device}, \textbf{addrspa
 \op{Const} & from arg & $()$ & \textsc{null} & $[v, v]$ \\
 \op{Param} & from arg & from $\mathrm{src}[0]$ & from arg & from src or dtype range \\[3pt]
 Movement ops & $\mathrm{src}[0].\mathrm{dtype}$ & (see op) & $\mathrm{src}[0].\mathrm{device}$ & $\mathrm{src}[0]$ \\
-\op{Unshard} & $\mathrm{src}[0].\mathrm{dtype}$ & $\mathrm{src}[0]$, axis $\times n$ & $\mathrm{src}[0].\mathrm{device}$ & $\mathrm{src}[0]$ \\
+\op{Unshard} & $\mathrm{src}[0].\mathrm{dtype}$ & $\mathrm{src}[0]$, each $a_k \times n_k$ & $\mathrm{src}[0].\mathrm{device}$ & $\mathrm{src}[0]$ \\
 \op{Reduce} & $\mathrm{src}[0].\mathrm{dtype}$ & remove first $n$ axes & $\mathrm{src}[0].\mathrm{device}$ & dtype range \\[3pt]
 \op{Cast} & from arg & $\mathrm{src}[0].\mathrm{shape}$ & $\mathrm{src}[0].\mathrm{device}$ & clamped to dtype \\
 \op{Bitcast} & from arg & $\mathrm{src}[0].\mathrm{shape}$ & $\mathrm{src}[0].\mathrm{device}$ & dtype range \\
@@ -286,9 +286,9 @@ $[a,A]$, $[b,B]$, $[c,C]$ denote min\_max of $\mathrm{src}[0]$, $\mathrm{src}[1]
 Default \emph{dtype range}: $[\mathrm{dtype\_min},\, \mathrm{dtype\_max}]$.

 \medskip
-\textbf{axis} tracks the multi-device sharding dimension. \op{Unshard} defines it (axis $=$ arg). \op{Buffer} with $n$-tuple device: axis $= 0$ (device dim).
-\op{Reshape} remaps axis to preserve the shard boundary. \op{Permute} follows the permutation. \op{Expand} shifts axis right by $|\mathbf{n}|$.
-\op{Reduce} on the shard axis $\to$ \textsc{null} (shard axis is among the first $n$ axes). \op{Replicated} on the shard axis $\to$ \textsc{null}. \op{Copy} $\to$ \textsc{null}. ALU ops inherit from sources. Default: \textsc{null}.
+\textbf{sharding} tracks multi-device sharding as a set of (axis, \op{Range}) pairs. \op{Unshard} defines it: arg is the tuple of sharded axes, one \op{Range} in src per axis (positional: the $k$-th \op{Range} shards the $k$-th axis). \op{Buffer} with $n$-tuple device: sharded on axis $0$ (device dim). The single-axis convenience \textbf{axis} is \textsc{null} unless exactly one axis is sharded.
+\op{Reshape} remaps each sharded axis to preserve its shard boundary. \op{Permute} follows the permutation. \op{Expand} shifts all sharded axes right by $|\mathbf{n}|$.
+\op{Reduce} on a sharded axis drops it. \op{Replicated} on the shard axis $\to$ \textsc{null}. \op{Copy} $\to$ \textsc{null}. ALU ops inherit from sources. Default: \textsc{null}.

 %% ============================================================
 \subsection*{Kernel Optimizations (OptOps) \normalfont\small--- schedule-level transforms on kernel ranges}
@@ -382,7 +382,7 @@ def scatter_add(T, idx, val):
 Let $D = (d_0, \ldots, d_{n-1})$ be an $n$-tuple device.
 \op{Copy} to an $n$-tuple device reshards with axis $= 0$. \op{Copy} never changes shape.

-\textbf{Sharding} splits a tensor along an axis across $n$ devices. It opens a \op{Range} of type \texttt{DEVICE} (a symbolic per-device index $d$), shrinks each device's view to its piece, then closes the range with \op{Unshard}$(T, R, a)$. The result is a logical tensor whose shape along axis $a$ is the full size; each device holds $1/n$ of it. \op{Unshard} is the inverse of sharding --- it marks the boundary between per-device computation and the logical multi-device tensor. The range need not be \texttt{DEVICE}; e.g.\ a \texttt{WARP} range closes the same way, concatenating per-lane shards along $a$ with the range as the outer factor.
+\textbf{Sharding} splits a tensor along an axis across $n$ devices. It opens a \op{Range} of type \texttt{DEVICE} (a symbolic per-device index $d$), shrinks each device's view to its piece, then closes the range with \op{Unshard}$(T, R, a)$. The result is a logical tensor whose shape along axis $a$ is the full size; each device holds $1/n$ of it. \op{Unshard} is the inverse of sharding --- it marks the boundary between per-device computation and the logical multi-device tensor. The range need not be \texttt{DEVICE}; e.g.\ a \texttt{WARP} range closes the same way, concatenating per-lane shards along $a$ with the range as the outer factor. A tensor may be sharded along several axes at once: \op{Unshard}$(T, R_0, R_1, \ldots;\; a_0, a_1, \ldots)$ carries one \op{Range} per sharded axis, and every movement op maps each sharded axis independently.

 \begin{lstlisting}
 # T has shape (s,) on a single device.
@@ -11,8 +11,8 @@ from tinygrad.runtime.autogen.amd.rdna3.enum import VOP1Op, VOP2Op, SOP2Op, DSOp

 def _srcs():
  """Create minimal source variables for pcode parsing."""
-  def u32(v=0): return UOp.const(dtypes.uint32, v)
-  return {'S0': u32(), 'S1': u32(), 'S2': u32(), 'SCC': u32(), 'VCC': UOp.const(dtypes.uint64, 0), 'laneId': u32()}
+  def u32(v=0): return UOp.const(v, dtypes.uint32)
+  return {'S0': u32(), 'S1': u32(), 'S2': u32(), 'SCC': u32(), 'VCC': UOp.const(0, dtypes.uint64), 'laneId': u32()}

 class TestBasicParsing(unittest.TestCase):
  """Test basic pcode parsing for common instruction patterns."""
@@ -44,8 +44,8 @@ class TestWithSources(unittest.TestCase):

  def test_v_add_f32_with_sources(self):
    """Test V_ADD_F32 with actual float constants."""
-    s0 = UOp.const(dtypes.uint32, 0x3f800000)  # 1.0f
-    s1 = UOp.const(dtypes.uint32, 0x40000000)  # 2.0f
+    s0 = UOp.const(0x3f800000, dtypes.uint32)  # 1.0f
+    s1 = UOp.const(0x40000000, dtypes.uint32)  # 2.0f
    _, assigns = parse_pcode(PCODE[VOP2Op.V_ADD_F32_E32], {'S0': s0, 'S1': s1})
    self.assertEqual(len(assigns), 1)
    dest, val = assigns[0]
@@ -55,8 +55,8 @@ class TestWithSources(unittest.TestCase):

  def test_v_mul_f32_with_sources(self):
    """Test V_MUL_F32 with actual float constants."""
-    s0 = UOp.const(dtypes.uint32, 0x40000000)  # 2.0f
-    s1 = UOp.const(dtypes.uint32, 0x40400000)  # 3.0f
+    s0 = UOp.const(0x40000000, dtypes.uint32)  # 2.0f
+    s1 = UOp.const(0x40400000, dtypes.uint32)  # 3.0f
    _, assigns = parse_pcode(PCODE[VOP2Op.V_MUL_F32_E32], {'S0': s0, 'S1': s1})
    self.assertEqual(len(assigns), 1)
    dest, val = assigns[0]
@@ -67,36 +67,36 @@ class TestParseExpr(unittest.TestCase):

  def test_integer_literals(self):
    """Test parsing integer literals."""
-    self.assertEqual(parse_expr('0', {}).arg, 0)
-    self.assertEqual(parse_expr('42', {}).arg, 42)
-    self.assertEqual(parse_expr('42U', {}).arg, 42)
+    self.assertEqual(parse_expr('0', {}).val, 0)
+    self.assertEqual(parse_expr('42', {}).val, 42)
+    self.assertEqual(parse_expr('42U', {}).val, 42)

  def test_negative_integers(self):
    """Test parsing negative integer literals."""
    result = parse_expr('-1', {})
-    self.assertEqual(result.arg, -1)
+    self.assertEqual(result.val, -1)
    self.assertEqual(result.dtype, dtypes.int)

  def test_float_literals(self):
    """Test parsing float literals."""
    result = parse_expr('1.0F', {})
-    self.assertEqual(result.arg, 1.0)
+    self.assertEqual(result.val, 1.0)
    self.assertEqual(result.dtype, dtypes.float32)

  def test_hex_literals(self):
    """Test parsing hex literals."""
    result = parse_expr('0xFF', {})
-    self.assertEqual(result.arg, 255)
+    self.assertEqual(result.val, 255)

  def test_variable_lookup(self):
    """Test variable lookup in parse_expr."""
-    vrs = {'x': UOp.const(dtypes.uint32, 42)}
+    vrs = {'x': UOp.const(42, dtypes.uint32)}
    result = parse_expr('x', vrs)
-    self.assertEqual(result.arg, 42)
+    self.assertEqual(result.val, 42)

  def test_binary_ops(self):
    """Test parsing binary operations."""
-    vrs = {'a': UOp.const(dtypes.uint32, 10), 'b': UOp.const(dtypes.uint32, 5)}
+    vrs = {'a': UOp.const(10, dtypes.uint32), 'b': UOp.const(5, dtypes.uint32)}

    # Addition
    result = parse_expr('a + b', vrs)
@@ -105,11 +105,11 @@ class TestParseExpr(unittest.TestCase):
    # Subtraction with constant folding
    result = parse_expr('10 - 5', {})
    self.assertEqual(result.op, Ops.CONST)
-    self.assertEqual(result.arg, 5)
+    self.assertEqual(result.val, 5)

  def test_ternary(self):
    """Test parsing ternary expressions."""
-    vrs = {'cond': UOp.const(dtypes.bool, True), 'a': UOp.const(dtypes.uint32, 1), 'b': UOp.const(dtypes.uint32, 0)}
+    vrs = {'cond': UOp.const(True), 'a': UOp.const(1, dtypes.uint32), 'b': UOp.const(0, dtypes.uint32)}
    result = parse_expr('cond ? a : b', vrs)
    self.assertEqual(result.op, Ops.WHERE)

@@ -127,7 +127,7 @@ class TestForLoopParsing(unittest.TestCase):
  def test_clz_parsing(self):
    """Test CLZ pcode parsing produces correct structure."""
    pcode = PCODE[VOP1Op.V_CLZ_I32_U32_E32]
-    S0 = UOp.const(dtypes.uint32, 0xFFFFFFFF)  # All ones - CLZ should be 0
+    S0 = UOp.const(0xFFFFFFFF, dtypes.uint32)  # All ones - CLZ should be 0
    _vrs, assigns = parse_pcode(pcode, {'S0': S0})

    self.assertEqual(len(assigns), 1)
@@ -139,7 +139,7 @@ class TestForLoopParsing(unittest.TestCase):
  def test_clz_with_zero(self):
    """Test CLZ with input 0 - should return -1."""
    pcode = PCODE[VOP1Op.V_CLZ_I32_U32_E32]
-    S0 = UOp.const(dtypes.uint32, 0)
+    S0 = UOp.const(0, dtypes.uint32)
    _vrs, assigns = parse_pcode(pcode, {'S0': S0})

    # Check that the innermost value (default) is -1 (may be wrapped in CAST)
@@ -150,7 +150,7 @@ class TestForLoopParsing(unittest.TestCase):
    # Unwrap CAST if present
    while val.op == Ops.CAST:
      val = val.src[0]
-    self.assertEqual(val.arg, -1)
+    self.assertEqual(val.val, -1)

  def test_ctz_parsing(self):
    """Test CTZ pcode parsing."""
@@ -158,7 +158,7 @@ class TestForLoopParsing(unittest.TestCase):
    if pcode is None:
      self.skipTest("V_CTZ_I32_B32_E32 pcode not available")

-    S0 = UOp.const(dtypes.uint32, 1)  # LSB set - CTZ should be 0
+    S0 = UOp.const(1, dtypes.uint32)  # LSB set - CTZ should be 0
    _vrs, assigns = parse_pcode(pcode, {'S0': S0})
    self.assertEqual(len(assigns), 1)

@@ -169,8 +169,8 @@ class TestDSPcodePatterns(unittest.TestCase):
    """Test GLOBAL_ATOMIC_ADD_F32 keeps memory values in float dtype."""
    vmem = UOp.param(2, dtypes.uint32, (1024,))
    srcs = {
-      'ADDR': UOp.const(dtypes.uint64, 0),
-      'DATA': UOp.const(dtypes.uint32, 0x3f800000),
+      'ADDR': UOp.const(0, dtypes.uint64),
+      'DATA': UOp.const(0x3f800000, dtypes.uint32),
      '_vmem': vmem,
    }

@@ -199,8 +199,8 @@ class TestDSPcodePatterns(unittest.TestCase):
    """Test MEM[addr].type read expression parsing."""
    # Create a mock LDS buffer
    lds = UOp.param(3, dtypes.uint32, (16384,))
-    addr = UOp.const(dtypes.uint32, 0)
-    vrs = {'_lds': lds, 'ADDR': addr, 'OFFSET': UOp.const(dtypes.uint32, 0)}
+    addr = UOp.const(0, dtypes.uint32)
+    vrs = {'_lds': lds, 'ADDR': addr, 'OFFSET': UOp.const(0, dtypes.uint32)}

    result = parse_expr('MEM[ADDR + OFFSET].b32', vrs)
    # Should be an INDEX operation into LDS
@@ -212,13 +212,13 @@ class TestDSPcodePatterns(unittest.TestCase):
    self.assertIsNotNone(pcode)
    assert pcode is not None
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 0),
-      'OFFSET0': UOp.const(dtypes.uint32, 0),
-      'OFFSET1': UOp.const(dtypes.uint32, 1),
-      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
-      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+      'ADDR': UOp.const(0, dtypes.uint32),
+      'OFFSET0': UOp.const(0, dtypes.uint32),
+      'OFFSET1': UOp.const(1, dtypes.uint32),
+      'DATA': UOp.const(0xAAAAAAAA, dtypes.uint32),
+      'DATA2': UOp.const(0xBBBBBBBB, dtypes.uint32),
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # Should have 2 MEM write assignments
    self.assertEqual(len(assigns), 2)
@@ -235,12 +235,12 @@ class TestDSPcodePatterns(unittest.TestCase):
    assert pcode is not None
    lds = UOp.param(3, dtypes.uint32, (16384,))
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 0),
-      'OFFSET0': UOp.const(dtypes.uint32, 0),
-      'OFFSET1': UOp.const(dtypes.uint32, 1),
+      'ADDR': UOp.const(0, dtypes.uint32),
+      'OFFSET0': UOp.const(0, dtypes.uint32),
+      'OFFSET1': UOp.const(1, dtypes.uint32),
      '_lds': lds,
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # Should have 2 RETURN_DATA assignments
    self.assertEqual(len(assigns), 2)
@@ -252,36 +252,36 @@ class TestDSPcodePatterns(unittest.TestCase):
    pcode = PCODE.get(DSOp.DS_STORE_2ADDR_B32)
    assert pcode is not None
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 100),
-      'OFFSET0': UOp.const(dtypes.uint32, 2),
-      'OFFSET1': UOp.const(dtypes.uint32, 5),
-      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
-      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+      'ADDR': UOp.const(100, dtypes.uint32),
+      'OFFSET0': UOp.const(2, dtypes.uint32),
+      'OFFSET1': UOp.const(5, dtypes.uint32),
+      'DATA': UOp.const(0xAAAAAAAA, dtypes.uint32),
+      'DATA2': UOp.const(0xBBBBBBBB, dtypes.uint32),
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # Check addresses: 100 + 2*4 = 108, 100 + 5*4 = 120
    # assigns[i][1] is (addr, val) tuple for MEM writes; mypy sees UOp
-    self.assertEqual(assigns[0][1][0].simplify().arg, 108)  # type: ignore[index]
-    self.assertEqual(assigns[1][1][0].simplify().arg, 120)  # type: ignore[index]
+    self.assertEqual(assigns[0][1][0].simplify().val, 108)  # type: ignore[index]
+    self.assertEqual(assigns[1][1][0].simplify().val, 120)  # type: ignore[index]

  def test_ds_store_data_values(self):
    """Test DS_STORE_2ADDR_B32 uses correct data values."""
    pcode = PCODE.get(DSOp.DS_STORE_2ADDR_B32)
    assert pcode is not None
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 0),
-      'OFFSET0': UOp.const(dtypes.uint32, 0),
-      'OFFSET1': UOp.const(dtypes.uint32, 1),
-      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
-      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+      'ADDR': UOp.const(0, dtypes.uint32),
+      'OFFSET0': UOp.const(0, dtypes.uint32),
+      'OFFSET1': UOp.const(1, dtypes.uint32),
+      'DATA': UOp.const(0xAAAAAAAA, dtypes.uint32),
+      'DATA2': UOp.const(0xBBBBBBBB, dtypes.uint32),
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # assigns[i][1] is (addr, val) tuple for MEM writes; mypy sees UOp
    # DATA[31:0] should preserve the value
-    self.assertEqual(assigns[0][1][1].simplify().arg, 0xAAAAAAAA)  # type: ignore[index]
-    self.assertEqual(assigns[1][1][1].simplify().arg, 0xBBBBBBBB)  # type: ignore[index]
+    self.assertEqual(assigns[0][1][1].simplify().val, 0xAAAAAAAA)  # type: ignore[index]
+    self.assertEqual(assigns[1][1][1].simplify().val, 0xBBBBBBBB)  # type: ignore[index]

 class TestConditionalParsing(unittest.TestCase):
  """Test conditional (if/elsif/else) pcode parsing."""
@@ -290,9 +290,9 @@ class TestConditionalParsing(unittest.TestCase):
    """Test parsing ternary expression (which becomes WHERE)."""
    # S_CSELECT_B32: D0.u32 = SCC ? S0.u32 : S1.u32
    pcode = PCODE[SOP2Op.S_CSELECT_B32]
-    s0 = UOp.const(dtypes.uint32, 10)
-    s1 = UOp.const(dtypes.uint32, 20)
-    scc = UOp.const(dtypes.uint32, 1)
+    s0 = UOp.const(10, dtypes.uint32)
+    s1 = UOp.const(20, dtypes.uint32)
+    scc = UOp.const(1, dtypes.uint32)
    _vrs, assigns = parse_pcode(pcode, {'S0': s0, 'S1': s1, 'SCC': scc})
    self.assertEqual(len(assigns), 1)
    dest, val = assigns[0]
@@ -305,26 +305,26 @@ class TestConcatWidthParsing(unittest.TestCase):

  def test_permlanex16_altrow_concat(self):
    for row, expected in [(0, 1), (1, 0), (2, 3), (3, 2)]:
-      parsed = parse_expr('{ row[1], ~row[0] }', {'row': UOp.const(dtypes.uint32, row)})
-      self.assertEqual(parsed.simplify().arg, expected)
+      parsed = parse_expr('{ row[1], ~row[0] }', {'row': UOp.const(row, dtypes.uint32)})
+      self.assertEqual(parsed.simplify().val, expected)

  def test_permlane64_altlane_concat(self):
    for lane, expected in [(0, 32), (1, 33), (31, 63), (32, 0), (63, 31)]:
-      parsed = parse_expr('{ ~lane[5], lane[4:0] }', {'lane': UOp.const(dtypes.uint32, lane)})
-      self.assertEqual(parsed.simplify().arg, expected)
+      parsed = parse_expr('{ ~lane[5], lane[4:0] }', {'lane': UOp.const(lane, dtypes.uint32)})
+      self.assertEqual(parsed.simplify().val, expected)

  def test_permlane64_wave64_pcode_indices(self):
    vgpr = UOp.param(0, dtypes.uint32, (256,))
    srcs = {
-      'SRC0': UOp.const(dtypes.uint32, 0),
-      'VDST': UOp.const(dtypes.uint32, 1),
-      'EXEC_LO': UOp.const(dtypes.uint32, 0xFFFFFFFF),
-      'EXEC': UOp.const(dtypes.uint64, 0xFFFFFFFFFFFFFFFF),
+      'SRC0': UOp.const(0, dtypes.uint32),
+      'VDST': UOp.const(1, dtypes.uint32),
+      'EXEC_LO': UOp.const(0xFFFFFFFF, dtypes.uint32),
+      'EXEC': UOp.const(0xFFFFFFFFFFFFFFFF, dtypes.uint64),
      '_vgpr': vgpr,
      '_wave_size': 64,
-      'S0': UOp.const(dtypes.uint32, 0),
-      'S1': UOp.const(dtypes.uint32, 0),
-      'S2': UOp.const(dtypes.uint32, 0),
+      'S0': UOp.const(0, dtypes.uint32),
+      'S1': UOp.const(0, dtypes.uint32),
+      'S2': UOp.const(0, dtypes.uint32),
    }

    def load_idx(v: UOp) -> int:
@@ -333,12 +333,12 @@ class TestConcatWidthParsing(unittest.TestCase):
      self.assertEqual(simp.src[0].op, Ops.INDEX)
      idx = simp.src[0].src[1].simplify()
      self.assertEqual(idx.op, Ops.CONST)
-      return idx.arg
+      return idx.val

    _, assigns = parse_pcode(PCODE[VOP1Op.V_PERMLANE64_B32_E32], srcs)
    self.assertEqual(len(assigns), 64)
    for lane, (dst_idx, src_idx) in {0: (64, 32), 31: (95, 63), 32: (96, 0), 63: (127, 31)}.items():
-      self.assertEqual(assigns[lane][1][0].simplify().arg, dst_idx)  # type: ignore[index]
+      self.assertEqual(assigns[lane][1][0].simplify().val, dst_idx)  # type: ignore[index]
      self.assertEqual(load_idx(assigns[lane][1][1]), src_idx)  # type: ignore[index]

 class TestAllPcode(unittest.TestCase):
@@ -346,7 +346,7 @@ class TestAllPcode(unittest.TestCase):

  def _make_srcs(self):
    """Create dummy source variables for pcode parsing."""
-    u32, u64 = lambda v=0: UOp.const(dtypes.uint32, v), lambda v=0: UOp.const(dtypes.uint64, v)
+    u32, u64 = lambda v=0: UOp.const(v, dtypes.uint32), lambda v=0: UOp.const(v, dtypes.uint64)
    lds = UOp.param(3, dtypes.uint32, (16384,))
    return {'laneId': u32(), 'laneID': u32(), 'S0': u32(), 'S1': u32(), 'S2': u32(), 'S3': u32(), 'SRC0': u32(),
            'D0': u32(), 'D1': u32(), 'DST': u32(), 'VDST': u32(), 'SDST': u32(),
@@ -358,7 +358,7 @@ class TestAllPcode(unittest.TestCase):
            'M0': u32(), 'PC': u64(), 'DENORM': u32(1), 'ROUND_MODE': u32(), 'ROUND_TOWARD_ZERO': u32(),
            'ROUND_NEAREST_EVEN': u32(), 'WAVE_STATUS': u32(),
            'MAX_FLOAT_F32': u32(0x7f7fffff), 'Unsigned': u32(1), 'clampedLOD': u32(),
-            '_lds': lds, '_vmem': lds, '_active': UOp.const(dtypes.bool, True)}
+            '_lds': lds, '_vmem': lds, '_active': UOp.const(True)}

  def _parse_all_pcode(self, pcode_dict, arch: str, min_pct: float):
    """Parse all pcode. RuntimeError = parser limitation (ok), other exceptions = real bugs."""
@@ -4,13 +4,13 @@ from tinygrad import Tensor, GlobalCounters, dtypes, nn, Device, Variable
 from tinygrad.helpers import Context, getenv, DEV
 from tinygrad.engine.realize import run_linear, estimate_uop, compile_linear
 from tinygrad.renderer.ptx import PTXRenderer
-from test.helpers import needs_second_gpu
+from test.helpers import needs_second_gpu, check_schedule, assert_kernel_count, KernelCountException

 class TestArange(unittest.TestCase):
  def _get_flops(self, tensor, desired):
    GlobalCounters.reset()
    linear = compile_linear(tensor.schedule_linear())
-    self.assertEqual(len(linear.src), 1)
+    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
    run_linear(linear)
    np.testing.assert_equal(tensor.numpy(), desired)
    return estimate_uop(linear.src[-1]).ops
@@ -55,8 +55,7 @@ class TestIndexing(unittest.TestCase):
    with Context(NOOPT=1):
      GlobalCounters.reset()
      out = ((Tensor.arange(1,16385)-1)*needle).sum()
-      linear, var_vals = out.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(out, 1)
      run_linear(linear, var_vals)
    self.assertEqual(out.item(), 1337)

@@ -72,8 +71,7 @@ class TestIndexing(unittest.TestCase):
      reshape_dataset = dataset.T.reshape(1, DDIM, DSET, 1).expand(4, DDIM, DSET, 1)
      full = (rng==idxs).where(reshape_dataset, Tensor.zeros(4, DDIM, DSET, 1, buffer=False))
      X = full.sum(axis=(2,3))
-      linear, var_vals = X.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(X, 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -98,8 +96,7 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      X = dataset[idxs]
      assert X.shape == (4,DDIM)
-      linear, var_vals = X.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(X, 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -113,8 +110,7 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      X = dataset[idxs]
      assert X.shape == (4,DDIM)
-      linear, var_vals = X.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(X, 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops} != {4*DSET}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -157,7 +153,7 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      z = emb(x).realize()
      self.assertLessEqual(GlobalCounters.global_ops, op_limit)
-      self.assertEqual(GlobalCounters.kernel_count, 2)
+      assert_kernel_count(2)
    if getenv("CHECK", 1):
      import torch
      with torch.no_grad():
@@ -257,7 +253,7 @@ class TestIndexing(unittest.TestCase):
    xq_rope, _ = apply_rotary_emb(xq, xq, freqs_cis)
    xq_rope.sum().backward()
    linear = compile_linear(wq.grad.schedule_linear())
-    assert len(linear.src) == 1, f"expected one kernel for backward, got: {len(linear.src)}"
+    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
    bwd_ops = estimate_uop(linear.src[0]).ops
    expected_ops = bs*seqlen*dim*dim*ops_scale
    print(f"rope matmul bwd ({dtype}): {GlobalCounters.kernel_count} kernels, {bwd_ops:,} ops")
@@ -1,7 +1,7 @@
 import unittest
 from tinygrad import Tensor, Device, dtypes, Context
 from tinygrad.helpers import getenv, system, DEV
-from extra.gemm.cdna_asm_gemm import asm_gemm, hk_bf16_atb_gemm
+from extra.gemm.cdna_asm_gemm import asm_gemm, hk_bf16_atb_gemm, quantize_mxfp4
 from test.helpers import needs_second_gpu
 from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8, FP8_MAX

@@ -150,6 +150,45 @@ class TestAsmGEMM(unittest.TestCase):
    with self.assertRaisesRegex(AssertionError, "not a multiple"):
      verify_asm_gemm(1, 256, 1000, 256)

+class TestMXFP4(unittest.TestCase):
+  def setUp(self):
+    if not is_cdna4() or DEV.interface.startswith("MOCK"):
+      self.skipTest("requires real amd machine")
+
+  def test_quantize(self):
+    import numpy as np
+    block = np.array([0, .26, .74, .75, 1.26, 1.75, 2.51, 3.5, 5.1, 6, -6] + [0] * 21, dtype=np.float32)
+    x = Tensor(np.tile(block, (32, 8)), dtype=dtypes.bfloat16)
+    packed, scale, _ = quantize_mxfp4(x)
+    p = packed.numpy()
+    codes = np.stack((p & 0xF, p >> 4), axis=-1).reshape(32, 256)
+    np.testing.assert_array_equal(codes[0, :11], [0, 1, 1, 2, 3, 4, 5, 6, 7, 7, 15])
+    np.testing.assert_array_equal(scale.numpy(), np.full((32, 8), 127, dtype=np.uint8))
+
+  def test_correctness(self):
+    import numpy as np
+    M = N = K = 256
+    rng = np.random.default_rng(1)
+    a = Tensor(rng.standard_normal((M, K), dtype=np.float32), dtype=dtypes.bfloat16)
+    b = Tensor(rng.standard_normal((N, K), dtype=np.float32), dtype=dtypes.bfloat16)
+    out = asm_gemm(a, b.T, mxfp4=True).realize()
+    # reference gemm
+    a_packed, scale_a, _ = quantize_mxfp4(a)
+    b_packed, scale_b, _ = quantize_mxfp4(b)
+    def unpack(x): return np.stack((x & 0xF, x >> 4), axis=-1).reshape(x.shape[0], -1)
+    code_a, code_b = unpack(a_packed.numpy()), unpack(b_packed.numpy())
+    lut = np.array([0, .5, 1, 1.5, 2, 3, 4, 6, -0., -.5, -1, -1.5, -2, -3, -4, -6], dtype=np.float32)
+    a_dequant = lut[code_a] * np.repeat(np.exp2(scale_a.numpy().astype(np.int16)-127), 32, axis=1)
+    b_dequant = lut[code_b] * np.repeat(np.exp2(scale_b.numpy().astype(np.int16)-127), 32, axis=1)
+    ref = Tensor(a_dequant @ b_dequant.T, dtype=dtypes.bfloat16).realize().numpy()
+    np.testing.assert_array_equal(out.numpy(), ref)
+
+  def test_empty(self):
+    M, N, K = getenv("M", 16384), getenv("N", 4096), getenv("K", 14336)
+    a = Tensor.empty(M, K, dtype=dtypes.bfloat16)
+    b = Tensor.empty(N, K, dtype=dtypes.bfloat16)
+    asm_gemm(a, b.T, mxfp4=True).realize()
+
 # test the Asm GEMM with Llama shapes, only run on the real machine for speed

@unittest.skipUnless(has_hipcc(), "requires hipcc to compile")
@@ -6,11 +6,11 @@ from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.uop.ops import KernelInfo

 def call_out_kernel(F:UOp, C:UOp) -> UOp:
-  call = F[0].load().call(UOp.const(dtypes.int, 3), C[0], ret_dtype=dtypes.void)
+  call = F[0].load().call(UOp.const(3).cast(dtypes.int), C[0], ret_dtype=dtypes.void)
  return C.after(call)[1].store(C.after(call)[0].load() + 1).sink(arg=KernelInfo(name="call_out"))

 def call_ret_kernel(F:UOp, C:UOp) -> UOp:
-  val = F[0].load().call(UOp.const(dtypes.int, 21), ret_dtype=dtypes.int)
+  val = F[0].load().call(UOp.const(21).cast(dtypes.int), ret_dtype=dtypes.int)
  return C[0].store(val * 2).sink(arg=KernelInfo(name="call_ret"))

@unittest.skipUnless(isinstance(Device["CPU"].renderer, CStyleLanguage), "TODO: CALL is rendered in C style only")
@@ -16,7 +16,7 @@ def _check_ast_count(desired_count:int, t:Tensor):

 class TestMovedConstFolding(unittest.TestCase):
  def test_contiguous_deviceless_const(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0)).contiguous()
+    t = Tensor(UOp.const(2.0, dtypes.float)).contiguous()
    self.assertIs(t.uop.op, Ops.CONST)
    self.assertIsNone(t.uop.device)

@@ -169,7 +169,7 @@ class TestMultiConstFolding(unittest.TestCase):
 class TestThreefryConstFolding(unittest.TestCase):
  def test_threefry(self):
    # THREEFRY(const,const) folds to a const once decomposed
-    x = threefry2x32(UOp.const(dtypes.uint64, 5), UOp.const(dtypes.uint64, 10))
+    x = threefry2x32(UOp.const(5, dtypes.uint64), UOp.const(10, dtypes.uint64))
    self.assertIs(x.simplify().op, Ops.CONST)

 class TestTautologicalCompare(unittest.TestCase):
@@ -1,7 +1,10 @@
 import unittest
 from tinygrad import Tensor, UOp, GlobalCounters, Context, Device
+import numpy as np
 from tinygrad.dtype import AddrSpace, dtypes, Invalid
 from tinygrad.uop.ops import KernelInfo, AxisType, Ops
+from tinygrad.renderer.ptx import PTXRenderer
+from test.helpers import assert_kernel_count

 # **** kernels ****

@@ -274,7 +277,7 @@ class TestCustomKernel(unittest.TestCase):

    GlobalCounters.reset()
    out.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 5)
+    assert_kernel_count(5)

  def test_simple_reshape(self):
    a = Tensor.ones(2,3,4).realize()
@@ -284,7 +287,7 @@ class TestCustomKernel(unittest.TestCase):
    GlobalCounters.reset()
    c.realize()
    assert all(i == 3. for i in c.flatten().tolist()), f"all 3 {c.tolist()}"
-    self.assertEqual(GlobalCounters.kernel_count, 3)
+    assert_kernel_count(3)

  def test_multi_after_schedule_order(self):
    """Test correct scheduling order when custom_kernel has multiple outputs.
@@ -334,12 +337,12 @@ class TestCustomKernel(unittest.TestCase):
    c = Tensor.custom_kernel(c, a, fxn=custom_add_one_kernel)[0]
    GlobalCounters.reset()
    c.realize()
-    self.assertEqual(GlobalCounters.kernel_count, len(devs))
+    assert_kernel_count(len(devs))
    self.assertTrue((c == 2).all().item())

  def test_partial_invalid_store_keeps_uncovered_reads(self):
    x = Tensor([10., 20., 30., 40.])
-    after = x.uop.after(x.uop.shrink(((0, 2),)).store(UOp.const(dtypes.float, Invalid, shape=(2,))))
+    after = x.uop.after(x.uop.shrink(((0, 2),)).store(Invalid))
    self.assertEqual(Tensor(after).contiguous().tolist(), [10., 20., 30., 40.])

  def test_multi_after_invalid_store_dep_removed(self):
@@ -399,7 +402,7 @@ class TestCustomKernel(unittest.TestCase):
    else: z = y.T.T+1
    GlobalCounters.reset()
    z.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)
    self.assertEqual(z.tolist(), x.add(2).tolist())

  @unittest.expectedFailure
@@ -416,24 +419,196 @@ class TestCustomKernel(unittest.TestCase):
    GlobalCounters.reset()
    y = run(x[0]).realize()
    # it's copying the input and the output
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(y.tolist(), [1, 2, 3, 4])

  @Context(DEV="CPU")
  def test_simple_from_source(self):
-    a = Tensor([0., 1., 2.]).realize()
-
-    src = "void test_src(float* restrict a) { a[0] = 1.0; }"
+    a = Tensor.arange(4).clone().realize()
+    src = "void test_src(int* restrict a) { a[0] = 1; }"
    # TODO: it currently requires a compiler for Ops.BINARY
    from tinygrad.device import Device
    binary = Device[a.device].renderer.compiler.compile(src)
    def custom_src_kernel(A:UOp) -> UOp:
      sink = UOp.sink(A, arg=KernelInfo(name="test_src"))
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(sink.toposort())),
-                                   UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=binary)))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(sink.toposort())), UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=binary)))
+    a = Tensor.custom_kernel(a.reshape(2, 2).T, fxn=custom_src_kernel)[0]
+    self.assertEqual(a.tolist(), [[1, 2], [1, 3]])

-    a = Tensor.custom_kernel(a, fxn=custom_src_kernel)[0]
-    self.assertEqual(a.tolist(), [1., 1., 2.])
+  def test_inplace_transpose(self):
+    def custom_assign_row_max_kernel(A:UOp) -> UOp:
+      row = UOp.range(A.shape[0], 0)
+      col = UOp.range(A.shape[1], 1)
+      return A[row, col].store(A[row].max(axis=0)).end(col).end(row).sink(arg=KernelInfo(name=f"assign_row_max_{A.numel()}"))
+    a = Tensor.arange(4).clone().realize()
+    a = Tensor.custom_kernel(a.reshape(2, 2).T, fxn=custom_assign_row_max_kernel)[0]
+    self.assertEqual(a.flatten().tolist(), [2, 2, 3, 3])
+    self.assertEqual(a.shape, (2, 2))
+
+class TestCustomKernelInput(unittest.TestCase):
+  def _test_mop(self, mop_fxn, max_kernels):
+    # default: input is BUFFER
+    x = mop_fxn(Tensor.arange(32).clone("CPU").realize())
+    y = Tensor.custom_kernel(Tensor.empty_like(x), x, fxn=custom_add_one_kernel)[0]
+    GlobalCounters.reset()
+    y.realize()
+    kernel_count = GlobalCounters.kernel_count
+    self.assertEqual(y.tolist(), x.add(1).tolist())
+    self.assertLessEqual(kernel_count, max_kernels)
+    # same test with @function, input is PARAM
+    from tinygrad import function
+    x0 = Tensor.arange(32).clone("CPU").realize()
+    @function(precompile=True)
+    def run(a:Tensor) -> Tensor:
+      xv = mop_fxn(a)
+      y = Tensor.invalids(*xv.shape, dtype=xv.dtype, device=a.device)
+      return Tensor.custom_kernel(y, xv, fxn=custom_add_one_kernel)[0]
+    GlobalCounters.reset()
+    y = run(x0).realize()
+    kernel_count = GlobalCounters.kernel_count
+    self.assertEqual(y.tolist(), mop_fxn(x0).add(1).tolist())
+    self.assertLessEqual(kernel_count, max_kernels)
+
+  def test_reshape(self): self._test_mop(lambda x: x.reshape(16, 2), max_kernels=2)
+  def test_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T, max_kernels=3)
+  def test_double_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T.T, max_kernels=3)
+  def test_shrink(self): self._test_mop(lambda x: x[:4], max_kernels=2)
+  def test_pad(self): self._test_mop(lambda x: x[:4].pad(((0, 4),)), max_kernels=2)
+  def test_flip(self): self._test_mop(lambda x: x.flip(0), max_kernels=2)
+  def test_offset_shrink(self): self._test_mop(lambda x: x[4:8], max_kernels=2)
+  def test_2d_shrink(self): self._test_mop(lambda x: x.reshape(4, 8)[:, 2:6], max_kernels=3)
+  def test_expand(self): self._test_mop(lambda x: x.reshape(16, 2)[:, :1].expand(16, 2), max_kernels=3)
+
+class TestUnshardIndex(unittest.TestCase):
+  """Regression tests for INDEX on UNSHARD (fragment) resolution in schedule/multi.py.
+
+  A fragment is a per-thread REG buffer wrapped in UNSHARD over LOCAL thread ranges.
+  index_multi must resolve an INDEX on the UNSHARD view into an INDEX on the per-thread
+  shard. Two ownership patterns must work:
+    contiguous: idx = rng*shard_sz + local   (thread rng owns [rng*shard_sz, ...))
+    strided:    idx = rng + ir*shard_sz      (thread rng owns {rng, rng+shard_sz, ...})
+  """
+  def _run(self, kernel, shape=(8, 8)):
+    c = Tensor.empty(*shape)
+    out = Tensor.custom_kernel(c, fxn=kernel)[0]
+    try: return out.numpy()
+    except RuntimeError as e:
+      if isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and "dynamic register indexing" in str(e):
+        self.skipTest("PTX does not support dynamic register indexing")
+      raise
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_contiguous_fragment_index(self):
+    # thread ty owns rows [ty*8, ty*8+8) of a 64-row fragment -- contiguous ownership.
+    # This is the pre-existing case that index_multi always handled.
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      ir = UOp.range(8, 1, AxisType.LOOP)
+      j = UOp.range(8, 2, AxisType.LOOP)
+      # 8x8 fragment, 8 threads -> 64x8 full tile. thread ty owns rows [ty*8, ty*8+8).
+      frag = UOp.placeholder((8, 8), dtypes.float32, 0, AddrSpace.REG).unshard((0,), (ty,))
+      return C[ty*8 + ir, j].store(frag[ty*8 + ir, j]).end(j, ir, ty).sink(arg=KernelInfo(name="contig_frag"))
+    out = self._run(kernel, (64, 8))
+    assert out.shape == (64, 8)
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_strided_fragment_index(self):
+    # thread ty owns rows {ty, ty+8, ty+16, ty+24, ..., ty+56} of a 64-row fragment --
+    # strided ownership. idx = ty + ir*8 where shard_sz=8 (8 threads, shard rows=8).
+    # The contiguous check (idx - rng*shard_sz) fails; the strided check
+    # (idx-rng) % shard_sz == 0 must succeed. This is the pattern the index_multi fix adds.
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      ir = UOp.range(8, 1, AxisType.LOOP)
+      j = UOp.range(8, 2, AxisType.LOOP)
+      # 8x8 fragment, 8 threads -> 64x8 full tile. thread ty owns rows {ty, ty+8, ..., ty+56}.
+      frag = UOp.placeholder((8, 8), dtypes.float32, 0, AddrSpace.REG).unshard((0,), (ty,))
+      return C[ty + ir*8, j].store(frag[ty + ir*8, j]).end(j, ir, ty).sink(arg=KernelInfo(name="strided_frag"))
+    out = self._run(kernel, (64, 8))
+    assert out.shape == (64, 8)
+
+  def test_fragment_index_cannot_shard(self):
+    # thread ty indexing rows [ty, ty+8) overlaps with other threads' rows -- this matches neither
+    # the contiguous nor the strided ownership pattern, so index_multi must raise.
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      ir = UOp.range(8, 1, AxisType.LOOP)
+      j = UOp.range(8, 2, AxisType.LOOP)
+      frag = UOp.placeholder((8, 8), dtypes.float32, 0, AddrSpace.REG).unshard((0,), (ty,))
+      return C[ty + ir, j].store(frag[ty + ir, j]).end(j, ir, ty).sink(arg=KernelInfo(name="bad_frag"))
+    with self.assertRaisesRegex(RuntimeError, "cannot shard index"):
+      self._run(kernel, (64, 8))
+
+def _run_fragment_kernel(testcase, kernel, out_shape, inputs=()):
+  c = Tensor.empty(*out_shape)
+  out = Tensor.custom_kernel(c, *inputs, fxn=kernel)[0]
+  try: return out.numpy()
+  except RuntimeError as e:
+    if isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and "dynamic register indexing" in str(e):
+      testcase.skipTest("PTX does not support dynamic register indexing")
+    raise
+
+class TestUnshardAlu(unittest.TestCase):
+  """Tests for ALU on (fragment) UNSHARD values in schedule/multi.py's alu_multi.
+
+  An ALU with UNSHARD srcs lowers to per-shard ops when every src is one of:
+    same sharding:  peel the UNSHARD, keep the layout
+    scalar:         broadcast to every shard
+    whole unsharded same-shape value: takes its per-shard sub-view (shard_subview)
+  """
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_alu_scalar_broadcast(self):
+    # scalar srcs broadcast to every shard: frag*2.0 where frag is 1.5 per thread -> 3.0 everywhere
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      # 8 values per thread, 8 threads -> 64-value full view
+      frag = UOp.placeholder((8,), dtypes.float32, 0, AddrSpace.LOCAL).unshard((0,), (ty,))
+      v = frag.after(frag.store(1.5)) * 2.0
+      return C.store(v).end(ty).sink(arg=KernelInfo(name="alu_scalar", opts_to_apply=()))
+    out = _run_fragment_kernel(self, kernel, (64,))
+    np.testing.assert_allclose(out, 3.0)
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_alu_whole_value_subview(self):
+    # UNSHARD + whole unsharded same-shape value: each shard adds its own sub-view of A.
+    def kernel(C:UOp, A:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      frag = UOp.placeholder((8,), dtypes.float32, 0, AddrSpace.LOCAL).unshard((0,), (ty,))
+      v = frag.after(frag.store(0.0)) + A
+      return C.store(v).end(ty).sink(arg=KernelInfo(name="alu_subview", opts_to_apply=()))
+    a = Tensor(np.arange(64, dtype=np.float32))
+    out = _run_fragment_kernel(self, kernel, (64,), inputs=(a,))
+    np.testing.assert_allclose(out, a.numpy(), atol=1e-4)
+
+class TestUnshardStore(unittest.TestCase):
+  """Tests for STORE of a sharded value into an unsharded dest (store_value_multi in schedule/multi.py).
+
+  Every shard stores its value into its own contiguous sub-view of the dest, one SHRINK per sharded axis.
+  """
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_store_unshard_value(self):
+    # single-axis: 8 threads each own 8 values of the 64-value output tile
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      frag = UOp.placeholder((8,), dtypes.float32, 0, AddrSpace.LOCAL).unshard((0,), (ty,))
+      v = frag.after(frag.store(0.0)) + 2.5
+      return C.store(v).end(ty).sink(arg=KernelInfo(name="store_unshard", opts_to_apply=()))
+    out = _run_fragment_kernel(self, kernel, (64,))
+    np.testing.assert_allclose(out, 2.5)
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_store_unshard_value_2axis(self):
+    # two sharded axes (the gemm fragment layout): thread (ty, tx) owns the (2, 1, 1, 2) sub-view of the
+    # (2, 4, 2, 2) output tile; the store must SHRINK dest on both sharded axes
+    def kernel(C:UOp, A:UOp) -> UOp:
+      ty = UOp.range(4, 0, AxisType.LOCAL)
+      tx = UOp.range(2, 1, AxisType.LOCAL)
+      frag = UOp.placeholder((2, 1, 1, 2), dtypes.float32, 0, AddrSpace.REG).unshard((1, 2), (ty, tx))
+      v = frag.after(frag.store(0.0)) + A
+      return C.store(v).end(tx, ty).sink(arg=KernelInfo(name="store_unshard_2axis", opts_to_apply=()))
+    a = Tensor(np.arange(32, dtype=np.float32).reshape(2, 4, 2, 2))
+    out = _run_fragment_kernel(self, kernel, (2, 4, 2, 2), inputs=(a,))
+    np.testing.assert_allclose(out, a.numpy(), atol=1e-4)

 class TestUOpReduce(unittest.TestCase):
  def test_uop_sum(self):
@@ -7,7 +7,7 @@ from tinygrad.renderer.isa.x86 import X86Renderer, X86Ops
 from tinygrad.renderer.isa import IselContext

 # INDEX on a register value with a constant index extracts a single element (the old GEP)
-def lane(y:UOp, i:int) -> UOp: return y.index(UOp.const(dtypes.int, i), dtype=y.dtype.scalar())
+def lane(y:UOp, i:int) -> UOp: return y.index(UOp.const(i, dtypes.int), dtype=y.dtype.scalar())

@unittest.skipUnless(isinstance(Device[Device.DEFAULT].renderer, X86Renderer), "only x86")
 class TestIselX86(unittest.TestCase):
@@ -49,7 +49,7 @@ class TestIselX86(unittest.TestCase):
    load = UOp.param(0, dtypes.int32, (16,)).index(a + 1).load()
    n = self.isel_rewrite(load)
    # displacement is the constant in "a" scaled to the buffer element size, dtype is int8 when the value fits otherwise int32
-    self.assertTrue(n.src[2].op is Ops.CONST and n.src[2].dtype is dtypes.int8 and n.src[2].arg == 4)
+    self.assertTrue(n.src[2].op is Ops.CONST and n.src[2].dtype is dtypes.int8 and n.src[2].val == 4)

 if __name__ == "__main__":
  unittest.main()
@@ -2,7 +2,7 @@
 import unittest
 import numpy as np

-from test.helpers import assert_jit_cache_len, call_is_graph, not_support_multi_device, needs_second_gpu
+from test.helpers import assert_jit_cache_len, call_is_graph, not_support_multi_device, needs_second_gpu, KernelCountException
 from test.unit.test_jit import _simple_test
 from tinygrad import Tensor, Variable, TinyJit, Device, dtypes
 from tinygrad.engine.jit import graph_class
@@ -97,7 +97,7 @@ class TestJit(unittest.TestCase):
      prev = o

    # Checking that 2 graphs are inited.
-    assert len(jf.captured.linear.src) == 2
+    if len(jf.captured.linear.src) != 2: raise KernelCountException(2, len(jf.captured.linear.src))
    for si in jf.captured.linear.src:
      assert call_is_graph(si)

@@ -12,7 +12,8 @@ from tinygrad.dtype import DType, dtypes, AddrSpace
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.cstyle import CUDARenderer
 from tinygrad.renderer.isa import ISARenderer
-from test.helpers import replace_opts
+from test.helpers import replace_opts, check_schedule
+from test.backend.test_softmax_fusion import single_kernel_softmax
 MOCKGPU = DEV.interface.startswith("MOCK")

 from tinygrad.uop.render import print_uops # noqa: F401 # pylint: disable=unused-import
@@ -267,9 +268,9 @@ class TestLinearizer(unittest.TestCase):
    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
    idxs = dedup([uop for uop in uops if uop.op is Ops.SPECIAL])
    idxs = sorted(idxs, key=lambda uop: uop.arg)
-    assert (idxs[0].arg, idxs[0].src[0].arg) == ('gidx0', 6), idxs[0]
-    assert (idxs[1].arg, idxs[1].src[0].arg) == ('gidx1', 5), idxs[1].arg
-    assert (idxs[2].arg, idxs[2].src[0].arg) == ('gidx2', 4), idxs[2].arg
+    assert (idxs[0].arg, idxs[0].src[0].val) == ('gidx0', 6), idxs[0]
+    assert (idxs[1].arg, idxs[1].src[0].val) == ('gidx1', 5), idxs[1].arg
+    assert (idxs[2].arg, idxs[2].src[0].val) == ('gidx2', 4), idxs[2].arg

  def test_sum_collapse(self):
    t = Tensor([2]).reshape(1, 1).expand(256, 256).sum()
@@ -292,8 +293,7 @@ class TestLinearizer(unittest.TestCase):
    a = Tensor.ones(4, 4).contiguous().realize()
    b = a.shrink(((1, 2), None)).pad(((1, 2), None)).bool()
    a.assign(b.where(2, a))
-    linear, var_vals = a.linear_with_vars()
-    assert len(linear.src) == 1
+    linear, var_vals = check_schedule(a, 1)
    run_linear(linear, var_vals)
    np.testing.assert_equal(a.flatten().numpy(), [1.,1.,1.,1.,2.,2.,2.,2.,1.,1.,1.,1.,1.,1.,1.,1.])
    program = to_program(replace_opts(linear.src[-1].src[0], []), renderer=Device[Device.DEFAULT].renderer)
@@ -392,6 +392,16 @@ class TestLinearizer(unittest.TestCase):
    # the global store doesn't change
    assert stores[1].src[1].dtype == dtypes.float

+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  def test_two_grouped_stores_local(self):
+    # GROUP on both reduces puts two LOCAL buffers in one kernel, and the store to each needs its own barrier
+    a = Tensor.rand(32, 32).realize()
+    opts = [Opt(OptOps.GROUP, 1, 4), Opt(OptOps.GROUP, 2, 4)]
+    ast = helper_linearizer_opt(single_kernel_softmax(a), [opts])
+    uops = to_program(replace_opts(ast, opts), renderer=Device[Device.DEFAULT].renderer).src[1].src
+    self.assertEqual(len([u for u in uops if u.op is Ops.BARRIER]), 2)
+
 # *** helpers ***

 def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
@@ -12,18 +12,18 @@ class TestLinearizerFailure(unittest.TestCase):
  @unittest.skipUnless(Device.DEFAULT == "METAL", "only tested on METAL")
  def test_failure_beam_mnist(self):
    c0 = UOp.param(0, dtypes.uchar, (4014080,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 0, AxisType.GLOBAL)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 784), 1, AxisType.GLOBAL)
-    c3 = UOp.range(UOp.const(dtypes.weakint, 10), 3, AxisType.GLOBAL)
+    c1 = UOp.range(UOp.const(512), 0, AxisType.GLOBAL)
+    c2 = UOp.range(UOp.const(784), 1, AxisType.GLOBAL)
+    c3 = UOp.range(UOp.const(10), 3, AxisType.GLOBAL)
    c4 = UOp.param(1, dtypes.int, (512,))
-    c5 = c4.index(c1.valid(UOp.const(dtypes.bool, True)))
-    c6 = UOp.range(UOp.const(dtypes.weakint, 6000), 1004, AxisType.REDUCE)
-    c7 = UOp.range(UOp.const(dtypes.weakint, 3750), 2006, AxisType.REDUCE)
-    c8 = UOp.range(UOp.const(dtypes.weakint, 16), 2007, AxisType.GROUP_REDUCE)
+    c5 = c4.index(c1.valid(UOp.const(True)))
+    c6 = UOp.range(UOp.const(6000), 1004, AxisType.REDUCE)
+    c7 = UOp.range(UOp.const(3750), 2006, AxisType.REDUCE)
+    c8 = UOp.range(UOp.const(16), 2007, AxisType.GROUP_REDUCE)
    c9 = UOp.param(2, dtypes.uchar, (47040000,))
-    c10 = c9.index((((c3*UOp.const(dtypes.weakint, 4704000))+c2)+(c6*UOp.const(dtypes.weakint, 784))).valid(UOp.const(dtypes.bool, True)))
-    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.weakint, 6000))+c6)+((c7*UOp.const(dtypes.weakint, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.weakint, 59999)).where(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(dtypes.int, -1))).where(UOp.const(dtypes.uchar, 0), c10).reduce(c6, arg=Ops.ADD)
-    c12 = c0.index((((c1*UOp.const(dtypes.weakint, 7840))+(c2*UOp.const(dtypes.weakint, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
+    c10 = c9.index((((c3*UOp.const(4704000))+c2)+(c6*UOp.const(784))).valid(UOp.const(True)))
+    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(6000))+c6)+((c7*UOp.const(16))+c8)).alu(Ops.CMPLT, UOp.const(59999)).where(UOp.const(0).cast(dtypes.int), UOp.const(1).cast(dtypes.int)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(-1).cast(dtypes.int))).where(UOp.const(0).cast(dtypes.uchar), c10).reduce(c6, arg=Ops.ADD)
+    c12 = c0.index((((c1*UOp.const(7840))+(c2*UOp.const(10)))+c3).valid(UOp.const(True))).store(c11).end(c1, c2, c3)
    ast = c12.sink(arg=KernelInfo(name='test', axis_types=(), dont_use_locals=False, applied_opts=(Opt(op=OptOps.GROUP, axis=1, arg=16),), opts_to_apply=None))
    _ = to_program(ast, Device["METAL"].renderer)

@@ -7,7 +7,7 @@ from extra.llama_kernels import local_abs_max
 from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed, quantize_fp8_scalar
 from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
 from extra.thunder.amd.fa import custom_fused_qkv_rope_backward, fused_qkv_rope
-from test.helpers import needs_second_gpu
+from test.helpers import needs_second_gpu, assert_kernel_count
 from test.backend.test_asm_gemm import has_hipcc

 def run_fused_ce(bs:int, seqlen:int, vocab:int, label_smoothing:float=0.0) -> None:
@@ -95,7 +95,7 @@ class TestLocalAmax(unittest.TestCase):
    x = Tensor.arange(16).reshape(4, 4).cast(dtypes.float).clone(devices[0]).realize().shard(devices, axis=0).realize()
    GlobalCounters.reset()
    out = (x * local_abs_max(x)).clone().realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)
    self.assertEqual(out.tolist(), [[0., 7., 14., 21.], [28., 35., 42., 49.], [120., 135., 150., 165.], [180., 195., 210., 225.]])

@unittest.skipUnless(has_hipcc() and Device.DEFAULT == "AMD", "requires hipcc to compile and amd device to run")
@@ -6,7 +6,7 @@ from tinygrad.nn.state import get_parameters
 from tinygrad.engine.realize import run_linear, compile_linear
 import numpy as np
 from hypothesis import given, strategies as strat, settings
-from test.helpers import not_support_multi_device, needs_second_gpu, slow, call_is_graph
+from test.helpers import not_support_multi_device, needs_second_gpu, slow, call_is_graph, check_schedule, assert_kernel_count

 settings.register_profile("my_profile", max_examples=200, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
 settings.load_profile("my_profile")
@@ -62,7 +62,7 @@ class TestMultiTensor(unittest.TestCase):
  def test_shard_empty(self):
    GlobalCounters.reset()
    X = Tensor.empty(256).shard(devices_2, 0).realize()
-    assert GlobalCounters.kernel_count == 0
+    assert_kernel_count(0)
    (X + X).realize()

  # TODO: fix this to not copy on the src device
@@ -76,6 +76,13 @@ class TestMultiTensor(unittest.TestCase):
    run_linear(linear)
    self.assertEqual(len(set(names)), 1, "function was relinearized")

+  def test_shard_beam(self):
+    cpu_2 = ("CPU:1", "CPU:2")
+    src = Tensor.ones(16).shard(cpu_2, 0).realize()
+    pad = src.to(cpu_2[::-1]).schedule_linear().src[0]
+    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): prg = compile_linear(UOp(Ops.LINEAR, src=(pad,))).src[0].src[0]
+    self.assertNotEqual(prg.src[0].arg.applied_opts, ())
+
  def test_shard_same_device(self):
    X = Tensor.ones(256).contiguous().realize()
    X.shard_((d1, X.device), 0)
@@ -348,8 +355,7 @@ class TestMultiTensor(unittest.TestCase):
  def test_const_like_shrink_on_shard_axis(self):
    t = Tensor.ones(16, 16, dtype=dtypes.int).shard(devices_2, axis=0)
    out = t.const_like(2)[:, :8]
-    linear, var_vals = out.linear_with_vars()
-    self.assertEqual(len(linear.src), 0)
+    linear, var_vals = check_schedule(out, 0)
    run_linear(linear, var_vals)
    self.assertEqual(out.tolist(), [[2]*8]*16)

@@ -425,6 +431,49 @@ class TestMultiBufferView(unittest.TestCase):
    run_linear(linear, var_vals)
    np.testing.assert_equal(out.numpy(), ref[5].numpy())

+@unittest.skipIf(not_support_multi_device(), "need multi")
+class Test2DShard(unittest.TestCase):
+  def setUp(self):
+    self.devices_4 = tuple(f"{Device.DEFAULT}:{i}" for i in range(4))
+    self.rng = UOp.range(4, -1, AxisType.DEVICE)
+    self.rng0, self.rng1 = self.rng // 2, self.rng % 2
+
+  def _shard_2d(self, t:Tensor) -> Tensor:
+    u = t.uop.copy_to_device(self.devices_4)._shard(0, self.rng0)._shard(1, self.rng1).unshard((0, 1), (self.rng0, self.rng1))
+    return Tensor(u)
+
+  def test_2d_shard_basic(self):
+    ref = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = t.contiguous().realize()
+    np.testing.assert_equal(out.numpy(), ref.numpy())
+
+  def test_2d_shard_elementwise(self):
+    ref = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = (t + 1).contiguous().realize()
+    np.testing.assert_equal(out.numpy(), ref.numpy() + 1)
+
+  def test_2d_shard_sum_all(self):
+    ref = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = t.sum().contiguous().realize()
+    np.testing.assert_equal(out.numpy(), np.array(ref.numpy().sum()))
+
+  def test_2d_shard_sum_non_sharded_axis(self):
+    ref = Tensor.arange(4*4*2).reshape(4, 4, 2).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = t.sum(axis=2).contiguous().realize()
+    np.testing.assert_equal(out.numpy(), ref.numpy().sum(axis=2))
+
+  def test_2d_shard_matmul(self):
+    a = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    b = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    a_s = self._shard_2d(a)
+    b_s = self._shard_2d(b)
+    out = (a_s @ b_s).contiguous().realize()
+    np.testing.assert_equal(out.numpy(), a.numpy() @ b.numpy())
+
@unittest.skipIf(not_support_multi_device(), "need multi")
 class TestMultiTransformer(unittest.TestCase):
  @needs_second_gpu
@@ -3,11 +3,11 @@ import unittest
 import numpy as np
 import torch
 from tinygrad import Tensor, Device, TinyJit, dtypes
-from tinygrad.uop.ops import Ops
 from tinygrad.helpers import GlobalCounters, Context
 from tinygrad.nn import Conv1d, ConvTranspose1d, Conv2d, ConvTranspose2d, Linear, Embedding
 from tinygrad.nn import BatchNorm, LayerNorm, LayerNorm2d, GroupNorm, InstanceNorm, RMSNorm, LSTMCell
 from tinygrad.nn.state import load_state_dict
+from test.helpers import check_schedule
 from tinygrad.engine.realize import run_linear
 from test.helpers import not_support_multi_device, needs_second_gpu, slow

@@ -428,18 +428,14 @@ class TestNN(unittest.TestCase):
    a = Tensor([[1, 5, 9, 11],
                [12, 19, 8, 1]])
    result = layer(a)
-    linear, var_vals = result.linear_with_vars()
-    self.assertEqual(len([call for call in linear.src if call.src[0].op is Ops.SINK]), kcount,
-                     "first run realizes weight and embedding")
+    linear, var_vals = check_schedule(result, kcount)
    run_linear(linear, var_vals)

    b = Tensor([[1, 2, 3],
                [4, 5, 6],
                [7, 8, 9]])
    result = layer(b)
-    linear, var_vals = result.linear_with_vars()
-    self.assertEqual(1, len([call for call in linear.src if call.src[0].op is Ops.SINK]),
-                     "second run realizes embedding only")
+    linear, var_vals = check_schedule(result, 1)
    run_linear(linear, var_vals)
    print(f"Embedding used {GlobalCounters.global_ops} ops")
    self.assertLessEqual(GlobalCounters.global_ops, ops)
@@ -728,6 +728,17 @@ class TestOps(unittest.TestCase):
        else:
          self.assertAlmostEqual(tiny_out, torch_out, msg=f"{x}, {c}")

+  def test_pow_neg_inf_frac_exponent(self):
+    # pow(-inf, 0.3) is +inf, so the gradient 0.3*pow(-inf, -0.7) is 0, never nan
+    helper_test_op(None, lambda x: x**0.3, vals=[[-math.inf]])
+    # is_odd truncates, so it calls 3.3 odd: only the non_int guard keeps pow(-inf, 3.3) from negating to -inf
+    helper_test_op(None, lambda x: x**3.3, vals=[[-math.inf]])
+
+  def test_pow_zero_exponent(self):
+    # x ** 0 is the constant 1 for every x, so the gradient with respect to the base is 0, never nan
+    # TODO: nan ** 0, failed on WEBGPU
+    helper_test_op(None, lambda x,y: x**y, vals=[[-math.inf, math.inf, 0.0], [0.0, 0.0, 0.0]])
+
  def test_pow_zero_tensor(self):
    helper_test_op(None, lambda x,y: x**y, vals=[[0.0], [0.0]])
    # TODO: fix WEBGPU
@@ -1728,6 +1739,8 @@ class TestOps(unittest.TestCase):
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(0), atol=1e-7, grad_atol=1e-7)
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(1), atol=1e-7, grad_atol=1e-7)
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(2), atol=1e-7, grad_atol=1e-7)
+  def test_softmin(self):
+    helper_test_op([(45,65)], torch.nn.Softmin(dim=1), Tensor.softmin, atol=1e-7, grad_atol=1e-7)

  def test_normalize(self):
    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x), lambda x: x.normalize(), atol=1e-7, grad_atol=1e-7)
@@ -13,7 +13,7 @@ class TestPickle(unittest.TestCase):

  def test_pickle_pattern_matcher(self):
    pm = PatternMatcher([(UPat.cvar('x'), lambda x: x*2)])
-    sink = UOp.const(dtypes.int, 2)
+    sink = UOp.const(2)
    tt = pm.rewrite(sink)
    pm_str = pickle.dumps(pm)
    pm2 = pickle.loads(pm_str)
@@ -8,6 +8,7 @@ from tinygrad.helpers import prod
 from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.wgsl import WGSLRenderer
+from test.helpers import check_schedule
 from tinygrad.runtime.ops_python import PythonRenderer
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, python_alu
 from tinygrad.tensor import Tensor
@@ -24,7 +25,7 @@ def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
  dtype = alu_src_uops[0].dtype
  a = UOp.param(0, dtype, (1,))
  b = UOp.param(1, dtype, (1,))
-  idx = UOp.const(dtypes.int, 0)
+  idx = UOp.const(0)
  ld = b.index(idx).load()
  alu = ld.alu(alu_op, *alu_src_uops)
  store = UOp.store(a.index(idx), alu)
@@ -35,7 +36,7 @@ class TestRendererFailures(unittest.TestCase):
  def test_gated_store_with_alu(self):
    a = UOp.param(0, dtypes.int, (4,))
    gate_alu = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0.valid(gate_alu)), UOp.const(dtypes.int, 1)))
+    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0.valid(gate_alu)), UOp.const(1).cast(dtypes.int)))
    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
    ret = _test_uop_result([], sink, local_size=[4, 1, 1])[0]
    np.testing.assert_equal(ret, [0, 1, 1, 1])
@@ -45,7 +46,7 @@ class TestRendererFailures(unittest.TestCase):
    a = UOp.param(0, dtypes.int, (8,))
    gate_alu_0 = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
    gate_alu_1 = (lidx1:=UOp.special(2, 'lidx1')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, src=(a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(dtypes.int, 1)))
+    gated_alu_store = UOp(Ops.STORE, src=(a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(1).cast(dtypes.int)))
    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
    ret = _test_uop_result([], sink, local_size=[4, 2, 1])[0]
    np.testing.assert_equal(ret, [0, 0, 0, 0, 0, 1, 1, 1])
@@ -54,15 +55,14 @@ class TestRendererFailures(unittest.TestCase):
 class TestCStyleFailures(unittest.TestCase):
  def test_inline_const_alu(self):
    # CPU doesn't use the max function
-    ret = _setup_and_test_alu(Ops.MAX, 1, UOp.const(dtypes.int, dtypes.int.min+1))
+    ret = _setup_and_test_alu(Ops.MAX, 1, UOp.const(dtypes.int.min+1).cast(dtypes.int))
    self.assertEqual(ret[0], 1)

  def _test_src_strip_paren(self, op: Ops, should_strip_paren:bool=True):
    dtype = "bool" if op in (Ops.OR, Ops.XOR, Ops.AND) else None
    ret = Tensor.empty(1, dtype=dtype)
    for _ in range(5): ret = python_alu[op](ret, Tensor.empty(1, dtype=dtype))
-    linear = ret.schedule_linear()
-    assert len(linear.src) == 1
+    linear, _ = check_schedule(ret, 1)
    src = to_program(linear.src[0].src[0], Device[Device.DEFAULT].renderer).src[2].arg
    self.assertEqual("("*5 not in src, should_strip_paren)

@@ -80,7 +80,7 @@ class TestWGSLFailures(unittest.TestCase):
  def test_multiply_infinity(self):
    # multiplying a positive constant by infinity should return infinity
    # WGSL pipelines do not handle this reliably, some of which return zero, unless infinity always comes from a read on a dynamic buffer
-    ret = _setup_and_test_alu(Ops.MUL, 5.0, UOp.const(dtypes.float32, float("inf")))
+    ret = _setup_and_test_alu(Ops.MUL, 5.0, UOp.const(float("inf")).cast(dtypes.float32))
    self.assertEqual(ret[0], float("inf"))

  # WGSL has a specific select(alt, val, gate) ternary operator instead of gate?val:alt
@@ -104,7 +104,7 @@ class TestPTXFailures(unittest.TestCase):
  def test_gated_store_with_if(self):
    a = UOp.param(0, dtypes.int, (4,))
    gate_alu = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
-    val = UOp.const(dtypes.int, 1)
+    val = UOp.const(1).cast(dtypes.int)
    if_uop = UOp(Ops.IF, src=(gate_alu,))
    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0, if_uop), val))
    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
@@ -6,34 +6,13 @@ import unittest, time
 import numpy as np

 from tinygrad import nn, dtypes, Device, Tensor, Variable
-from tinygrad.uop.ops import UOp, Ops, UPat
-from tinygrad.helpers import DEBUG, DEV, GlobalCounters, Context, all_same, temp
-from tinygrad.engine.realize import compile_linear, run_linear
+from tinygrad.uop.ops import Ops, UPat
+from tinygrad.helpers import DEV, GlobalCounters, Context, all_same, temp
+from tinygrad.engine.realize import run_linear
+from test.helpers import check_schedule, assert_kernel_count

 supported_dtypes = Device[Device.DEFAULT].renderer.supported_dtypes()

-class KernelCountException(Exception): pass
-def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Tensor]|None=None, filter_sink=True):
-  if to_prerealize:
-    with Context(DEBUG=0, TRACK_MATCH_STATS=0): Tensor.realize(*to_prerealize)
-  if isinstance(t, Tensor): linear, var_vals = t.linear_with_vars()
-  elif isinstance(t, list) and isinstance(t[0], Tensor): linear, var_vals = Tensor.linear_with_vars(*t)
-  else:
-    assert isinstance(t, UOp), f"can't schedule {t}"
-    linear, var_vals = Tensor(t).linear_with_vars()
-  kernel_cnt = sum((len(call.device) if isinstance(call.device, tuple) else 1)
-                   for call in linear.src if call.src[0].op is Ops.SINK or not filter_sink)
-  if kernel_cnt != allowed:
-    print(f"SCHEDULE ISSUE, expecting {allowed} got {kernel_cnt}")
-    if DEBUG >= 3:
-      for i,call in enumerate(linear.src):
-        print("kernel", i+1)
-        print(call.src[0])
-    raise KernelCountException(f"{kernel_cnt} != {allowed}")
-  # test compiling the linear
-  compile_linear(linear)
-  return linear, var_vals
-
 def _realize_weights(m):
  for p in nn.state.get_parameters(m): p.realize()

@@ -113,11 +92,9 @@ class TestSchedule(unittest.TestCase):
    a2 = mop(a)
    expected = (a+a2).tolist()
    a.assign(a+a2)
-    linear, var_vals = a.linear_with_vars()
-    kcount = len(linear.src)
+    linear, var_vals = check_schedule(a, expected_kcount)
    run_linear(linear, var_vals)
    self.assertListEqual(a.tolist(), expected)
-    self.assertEqual(kcount, expected_kcount)
  def test_setitem_permuted_sched(self): self.test_setitem_sched(lambda x: x.T, 2)
  def test_setitem_paddded_sched(self): self.test_setitem_sched(lambda x: x.shrink_to(4, 1).pad_to(4, 4), 1)

@@ -126,9 +103,9 @@ class TestSchedule(unittest.TestCase):
    a = Tensor.arange(16).clone().realize()
    GlobalCounters.reset()
    a[4] = 3
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    a.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertListEqual(a.tolist(), [0, 1, 2, 3, 3, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15])

  def test_no_extra_contiguous_on_setitem_assign_back(self):
@@ -4,6 +4,7 @@ from tinygrad import Tensor, GlobalCounters, Context, Device
 from tinygrad.dtype import DTypeLike, dtypes
 from tinygrad.engine.realize import run_linear
 from tinygrad.helpers import DEBUG, get_single_element
+from test.helpers import check_schedule

 def single_kernel_softmax(x_in:Tensor, axis=-1, dtype:DTypeLike|None=None) -> Tensor:
  # only support axis =-1
@@ -103,8 +104,7 @@ class TestFuse(unittest.TestCase):
    k = (x @ wk).contiguous()
    v = (x @ wv).contiguous()
    attn = q.scaled_dot_product_attention(k, v)
-    s = attn.schedule_linear()
-    self.assertEqual(len(s.src), 4) # 3 matmul and 1 attention
+    check_schedule(attn, 4) # 3 matmul and 1 attention

  @unittest.skip("needs RANGEIFY>1")
  def test_flash_attention(self):
@@ -24,13 +24,13 @@ class TestTinygrad(unittest.TestCase):
    self.assertEqual(Tensor(3.14).shape, ())

  def test_deviceless_const_construct_device_repr(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0))
+    t = Tensor(UOp.const(2.0).cast(dtypes.float))
    self.assertIsNone(t.uop.device)
    self.assertIsNone(t.device)
    self.assertIn("<UOp None", repr(t))

  def test_deviceless_const_realize_noop(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0))
+    t = Tensor(UOp.const(2.0).cast(dtypes.float))
    uop = t.uop
    t.realize()
    self.assertIs(t.uop, uop)
@@ -728,12 +728,12 @@ class TestZeroShapeTensor(unittest.TestCase):
    self.assertIsNot(a.uop.base.buffer, b.uop.base.buffer)

  def test_clone_deviceless_const(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0)).clone()
+    t = Tensor(UOp.const(2.0).cast(dtypes.float)).clone()
    np.testing.assert_equal(t.numpy(), 2.0)
    self.assertTrue(t.uop.has_buffer_identity())

  def test_numpy_deviceless_const(self):
-    np.testing.assert_equal(Tensor(UOp.const(dtypes.float, 2.0)).numpy(), 2.0)
+    np.testing.assert_equal(Tensor(UOp.const(2.0).cast(dtypes.float)).numpy(), 2.0)

  def test_clone_with_shrink(self):
    a = Tensor.rand(16, 16)
@@ -756,7 +756,7 @@ class TestZeroShapeTensor(unittest.TestCase):
    np.testing.assert_allclose(a.grad.numpy(), b.grad.numpy())

  def test_clone_deviceless_const_to_cpu(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0)).clone(device="CPU")
+    t = Tensor(UOp.const(2.0).cast(dtypes.float)).clone(device="CPU")
    self.assertEqual(t.device, "CPU")
    np.testing.assert_equal(t.numpy(), 2.0)

@@ -1,6 +1,6 @@
 import unittest
 import numpy as np
-from tinygrad import Tensor, Variable, dtypes
+from tinygrad import Device, Tensor, Variable, TinyJit, dtypes
 from tinygrad.helpers import CHECK_OOB

 class TestTensorVariable(unittest.TestCase):
@@ -18,10 +18,18 @@ class TestTensorVariable(unittest.TestCase):
      self.assertListEqual((vv * t).tolist(), [2, 2, 2])
    except RuntimeError: pass

-  # TODO: a Variable PARAM lowers to int32, so a bound value that doesn't fit int32 truncates or fails to bind
-  @unittest.expectedFailure
+  @unittest.skipUnless(dtypes.long in Device[Device.DEFAULT].renderer.supported_dtypes(), "requires long support")
  def test_large_range_variable(self):
-    self.assertEqual(Tensor(Variable("b", 0, 2**40).bind(2**35)).item(), 2**35)
+    self.assertEqual(Tensor(Variable("b", 0, 2**40, dtype=dtypes.long).bind(2**35)).clone(Device.DEFAULT).item(), 2**35)
+
+  @unittest.skipUnless(dtypes.long in Device[Device.DEFAULT].renderer.supported_dtypes(), "requires long support")
+  def test_large_range_variable_jit(self):
+    @TinyJit
+    def f(a,b): return (Tensor(a+b).clone(Device.DEFAULT) * 2).realize()
+    for i in range(3):
+      a = Variable("a", 0, 2**10, dtype=dtypes.int).bind(i)
+      b = Variable("b", 0, 2**40, dtype=dtypes.long).bind(2**35)
+      self.assertEqual(f(a,b).item(), (2**35 + i) * 2)

  def test_variable_defers_like_a_literal(self):
    vv = Variable("a", 1, 10).bind(2)
@@ -2,7 +2,7 @@ import unittest
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.helpers import Context, getenv, DEV, OSX
-from test.backend.test_schedule import check_schedule
+from test.helpers import check_schedule
 from test.backend.test_dtype_alu import ht, dtypes_float
 import numpy as np
 import math
@@ -19,7 +19,7 @@ def run_uops(uops_list:list[UOp], bufs:list[Buffer]):
  run_linear(UOp(Ops.LINEAR, src=(UOp.sink(*uops_list, arg=KernelInfo()).call(*buf_uops),)))

 def uop(uops:list[UOp], op:Ops, dtype:Optional[DType], src:tuple[UOp, ...], arg:Any=None) -> UOp:
-  if op is Ops.CONST: uops.append(UOp.const(dtype, arg))
+  if op is Ops.CONST: uops.append(UOp.const(arg).cast(dtype))
  elif op is Ops.PARAM: uops.append(UOp.param(arg, dtype, shape=(1,)))
  else: uops.append(UOp(op, dtype, tuple(src), arg))
  return uops[-1]
@@ -43,7 +43,7 @@ def _test_single_value_const(vals, op, dts):
  buf_store = uop(uops, Ops.PARAM, output_dtype, (), 0)
  loads = (uop(uops, Ops.CONST, dtype, [], a) for a,dtype in zip(vals, dts))
  alu = uop(uops, op, output_dtype, loads)
-  out = buf_store[UOp.const(dtypes.int32, 0)].store(alu)
+  out = buf_store[UOp.const(0).cast(dtypes.int32)].store(alu)
  buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
  run_uops([out], [buf])
  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
@@ -221,12 +221,12 @@ class TestAssembly(unittest.TestCase):
  def test_bitshift_left(self):
    g1 = UOp.param(0, dtypes.int32, shape=(3,))
    out = UOp.param(1, dtypes.int32, shape=(2,))
-    c1 = UOp.const(dtypes.int, 2)
-    c2 = UOp.const(dtypes.int, 3)
+    c1 = UOp.const(2)
+    c2 = UOp.const(3)
    l1 = g1.index(c1)
    a1 = UOp(Ops.MUL, src=(l1, c1))
    a2 = UOp(Ops.MUL, src=(l1, c2))
-    uops = to_uops_list([out.index(UOp.const(dtypes.int, 0)).store(a1), out.index(UOp.const(dtypes.int, 1)).store(a2)],
+    uops = to_uops_list([out.index(UOp.const(0)).store(a1), out.index(UOp.const(1)).store(a2)],
                        ren=Device[Device.DEFAULT].renderer)
    Device[Device.DEFAULT].renderer.render(uops)
    ops = [x.op for x in uops]
@@ -249,16 +249,16 @@ class TestAssembly(unittest.TestCase):

  def test_mulacc_shl(self):
    g1 = UOp.param(0, dtypes.int32, shape=(2,))
-    c1 = UOp.const(dtypes.int, 0)
-    c2 = UOp.const(dtypes.int, 1)
-    expr = g1.index(c1) * UOp.const(dtypes.int, 4096) + g1.index(c2)
+    c1 = UOp.const(0)
+    c2 = UOp.const(1)
+    expr = g1.index(c1) * UOp.const(4096) + g1.index(c2)
    uops = to_uops_list([expr], ren=Device[Device.DEFAULT].renderer)
    Device[Device.DEFAULT].renderer.render(uops)
    self.assertIn(Ops.MULACC, [x.op for x in uops])

  def test_use_cmpeq(self):
    g = UOp.param(0, dtypes.uint32, shape=(8,))
-    c = UOp.const(dtypes.uint, 7)
+    c = UOp.const(7)
    comp = g.index(c).ne(c).ne(True)
    uops = to_uops_list([comp], ren=Device[Device.DEFAULT].renderer)
    Device[Device.DEFAULT].renderer.render(uops)
@@ -330,6 +330,7 @@ class TestHCQ(unittest.TestCase):
  # Test profile api
  def test_speed_exec_time(self):
    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
+    st = time.perf_counter()
    TestHCQ.d0.hw_compute_queue_t().timestamp(sig_st) \
                                   .exec(TestHCQ.runtime, TestHCQ.kernargs_ba_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size) \
                                   .timestamp(sig_en) \
@@ -337,11 +338,13 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
+    host_us = (time.perf_counter() - st) * 1e6

    et = float(sig_en.timestamp - sig_st.timestamp)

    print(f"exec kernel time: {et:.2f} us")
-    assert 0.1 <= et <= (3000000 if MOCKGPU or Device.DEFAULT in {"CPU"} else 100)
+    # emulated devices are only bounded by the host window around submit+wait
+    assert 0.1 <= et <= (host_us if MOCKGPU or Device.DEFAULT in {"CPU"} else 100)

  def test_speed_copy_bandwidth(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -352,6 +355,7 @@ class TestHCQ(unittest.TestCase):
    b = Buffer(Device.DEFAULT, SZ, dtypes.uint8, options=BufferSpec(nolru=True)).allocate()

    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
+    st = time.perf_counter()
    TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
                                .copy(a._buf, b._buf, SZ) \
                                .timestamp(sig_en) \
@@ -359,13 +363,14 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
+    host_ms = (time.perf_counter() - st) * 1e3

-    et = float(sig_en.timestamp - sig_st.timestamp)
-    et_ms = et / 1e3
+    et_ms = float(sig_en.timestamp - sig_st.timestamp) / 1e3
+    assert 0 < et_ms <= host_ms # timestamps are in us and cover only the copy

    gb_s = ((SZ / 1e9) / et_ms) * 1e3
    print(f"same device copy:  {et_ms:.2f} ms, {gb_s:.2f} GB/s")
-    assert (0.2 if MOCKGPU else 10) <= gb_s <= 1000
+    assert (0 if MOCKGPU else 10) <= gb_s <= 1000

  def test_speed_cross_device_copy_bandwidth(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -379,6 +384,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.allocator._map(b._buf)

    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
+    st = time.perf_counter()
    TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
                                .copy(a._buf, b._buf, SZ) \
                                .timestamp(sig_en) \
@@ -386,13 +392,14 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
+    host_ms = (time.perf_counter() - st) * 1e3

-    et = float(sig_en.timestamp - sig_st.timestamp)
-    et_ms = et / 1e3
+    et_ms = float(sig_en.timestamp - sig_st.timestamp) / 1e3
+    assert 0 < et_ms <= host_ms # timestamps are in us and cover only the copy

    gb_s = ((SZ / 1e9) / et_ms) * 1e3
    print(f"cross device copy: {et_ms:.2f} ms, {gb_s:.2f} GB/s")
-    assert (0.2 if MOCKGPU else 2) <= gb_s <= 100
+    assert (0 if MOCKGPU else 2) <= gb_s <= 100

  def test_timeline_signal_rollover(self):
    for queue_type in [TestHCQ.d0.hw_compute_queue_t, TestHCQ.d0.hw_copy_queue_t]:
@@ -13,9 +13,9 @@ from tinygrad.dtype import Invalid

 def vision_conv_143():
  c0 = UOp.param(0, dtypes.half, shape=(16, 1024, 4))
-  c2 = UOp.range(32, 3, AxisType.LOOP)
-  c5 = UOp.range(128, 4, AxisType.LOOP)
-  c8 = UOp.range(16, 2, AxisType.LOOP)
+  c2 = UOp.range(32, 3)
+  c5 = UOp.range(128, 4)
+  c8 = UOp.range(16, 2)
  c16 = UOp.range(7, 0, AxisType.REDUCE)
  c17 = c8*2+c16
  c24 = ((c17<3)!=True)&(c17<35)
@@ -24,8 +24,8 @@ def vision_conv_143():
  c32 = ((c27<3)!=True)&(c27<67)
  c34 = UOp.param(1, dtypes.half, shape=(32, 1024, 4))
  c38 = c5//2
-  c45 = (c32&c24).where((c27*64+c38+c17*4096+-12480), UOp.const(dtypes.weakint, Invalid))
-  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
+  c45 = (c32&c24).where((c27*64+c38+c17*4096+-12480), UOp.const(Invalid, dtypes.weakint))
+  c48 = (c24&c32).where(c34.index(c45), UOp.const(0.0, dtypes.float))
  c49 = UOp.param(2, dtypes.half, shape=(64, 49, 4))
  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
  c63 = UOp.param(3, dtypes.float, (128,))
@@ -39,9 +39,9 @@ def vision_conv_143():

 def vision_conv_153():
  c0 = UOp.param(0, dtypes.half, shape=(8, 1024, 4))
-  c2 = UOp.range(16, 3, AxisType.LOOP)
-  c5 = UOp.range(256, 4, AxisType.LOOP)
-  c8 = UOp.range(8, 2, AxisType.LOOP)
+  c2 = UOp.range(16, 3)
+  c5 = UOp.range(256, 4)
+  c8 = UOp.range(8, 2)
  c16 = UOp.range(7, 0, AxisType.REDUCE)
  c17 = c8*2+c16
  c24 = ((c17<3)!=True)&(c17<19)
@@ -50,8 +50,8 @@ def vision_conv_153():
  c32 = ((c27<3)!=True)&(c27<35)
  c34 = UOp.param(1, dtypes.half, shape=(16, 1024, 4))
  c38 = c5//2
-  c45 = (c32&c24).where((c27*128+c38+c17*4096+-12672), UOp.const(dtypes.weakint, Invalid))
-  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
+  c45 = (c32&c24).where((c27*128+c38+c17*4096+-12672), UOp.const(Invalid, dtypes.weakint))
+  c48 = (c24&c32).where(c34.index(c45), UOp.const(0.0, dtypes.float))
  c49 = UOp.param(2, dtypes.half, shape=(128, 49, 4))
  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
  c63 = UOp.param(3, dtypes.float, (256,))
@@ -65,7 +65,7 @@ def vision_conv_153():

 def dm_conv_172():
  c0 = UOp.param(0, dtypes.half, shape=(1, 240, 4))
-  c2 = UOp.range(960, 4, AxisType.LOOP)
+  c2 = UOp.range(960, 4)
  c5 = UOp.param(1, dtypes.half, shape=(8, 384, 4))
  c7 = UOp.range(32, 0, AxisType.REDUCE)
  c10 = UOp.range(4, 1, AxisType.REDUCE)
@@ -40,7 +40,7 @@ def random_int_expr(depth=10):
 def random_bool_expr(depth=10, expr1=None):
  if depth == 0: return True
  if expr1 is None: expr1 = random_int_expr(depth-1)
-  expr2 = random.choice([random_or_sub_expression_int(depth-1, expr1), UOp.const(dtypes.weakint, random.randint(-10, 10))])
+  expr2 = random.choice([random_or_sub_expression_int(depth-1, expr1), UOp.const(random.randint(-10, 10))])
  return random.choice(comp_ops)(expr1, expr2)


@@ -82,8 +82,8 @@ if __name__ == "__main__":
            f"v2=Variable(\"{u2.arg[0]}\", {u2.arg[1]}, {u2.arg[2]})\n" +\
            f"v3=Variable(\"{u3.arg[0]}\", {u3.arg[1]}, {u3.arg[2]})\n" +\
            f"expr = {expr}\n" +\
-            f"v1_val, v2_val, v3_val = UOp.const(dtypes.weakint, {n1.as_long()}), UOp.const(dtypes.weakint, {n2.as_long()})," +\
-                f"UOp.const(dtypes.weakint, {n3.as_long()})\n" +\
+            f"v1_val, v2_val, v3_val = UOp.const({n1.as_long()}), UOp.const({n2.as_long()})," +\
+                f"UOp.const({n3.as_long()})\n" +\
            "num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()\n" +\
            "rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()\n" +\
            "assert num==rn, f\"{num} != {rn}\"\n"
@@ -8,10 +8,11 @@ from tinygrad.tensor import _to_np_dtype
 from tinygrad.codegen import to_program
 from tinygrad.dtype import DType, truncate
 from tinygrad.nn.state import get_parameters
-from tinygrad.helpers import T, Target, DEV
+from tinygrad.helpers import T, Target, DEV, DEBUG, Context, GlobalCounters
 from tinygrad.renderer import Renderer
 from tinygrad.codegen import full_rewrite_to_sink, line_rewrite, pm_linearize_cleanups
 from tinygrad.codegen.late.linearizer import linearize
+from tinygrad.engine.realize import compile_linear

 # decorator to skip slow tests by default, run with RUN_SLOW=1 to include them
 slow = unittest.skipUnless(os.getenv("RUN_SLOW"), "slow test, set RUN_SLOW=1 to run")
@@ -34,6 +35,36 @@ def derandomize_model(model):
    p.replace(Tensor.empty(p.shape, device=p.device, dtype=p.dtype))
    p.realize()

+class KernelCountException(Exception):
+  def __init__(self, expected:int, got:int):
+    self.expected, self.got = expected, got
+    super().__init__(f"expected {expected}, got {got}")
+
+def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Tensor]|None=None, filter_sink=True):
+  if to_prerealize:
+    with Context(DEBUG=0, TRACK_MATCH_STATS=0): Tensor.realize(*to_prerealize)
+  if isinstance(t, Tensor): linear, var_vals = t.linear_with_vars()
+  elif isinstance(t, list) and isinstance(t[0], Tensor): linear, var_vals = Tensor.linear_with_vars(*t)
+  else:
+    assert isinstance(t, UOp), f"can't schedule {t}"
+    linear, var_vals = Tensor(t).linear_with_vars()
+  kernel_cnt = sum((len(call.device) if isinstance(call.device, tuple) else 1)
+                   for call in linear.src if call.src[0].op is Ops.SINK or not filter_sink)
+  if kernel_cnt != allowed:
+    print(f"SCHEDULE ISSUE, expecting {allowed} got {kernel_cnt}")
+    if DEBUG >= 3:
+      for i,call in enumerate(linear.src):
+        print("kernel", i+1)
+        print(call.src[0])
+    raise KernelCountException(allowed, kernel_cnt)
+  # test compiling the linear
+  compile_linear(linear)
+  return linear, var_vals
+
+def assert_kernel_count(expected:int):
+  got = GlobalCounters.kernel_count
+  if got != expected: raise KernelCountException(expected, got)
+
 def call_is_graph(call:UOp) -> bool:
  ast = call.src[0]
  return ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph"
@@ -53,15 +84,15 @@ def jit_cache_count(linear:UOp) -> int:
 def assert_jit_cache_len(fxn, expected_len):
  linear = fxn.captured.linear if fxn.captured is not None else None
  if linear is None or not linear.src:
-    assert expected_len == 0, expected_len
+    if expected_len != 0: raise KernelCountException(expected_len, 0)
    return
  if expected_len and all(call_is_hcq(call) for call in linear.src): expected_len = 3 # HCQ2: merged same-queue calls + finalizer + bumps
  if call_is_graph(linear.src[0]):
-    assert len(linear.src) == 1, len(linear.src)
+    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
    inner = linear.src[0].src[0].src[0]  # LINEAR UOp inside CUSTOM_FUNCTION
-    assert len(inner.src) == expected_len, f"expected {expected_len}, got {len(inner.src)}"
+    if len(inner.src) != expected_len: raise KernelCountException(expected_len, len(inner.src))
  else:
-    assert len(linear.src) == expected_len, f"expected {expected_len}, got {len(linear.src)}"
+    if len(linear.src) != expected_len: raise KernelCountException(expected_len, len(linear.src))

 def min_normal(dt:DType) -> float: return 2.0 ** (2 - (1 << (dtypes.finfo(dt)[0] - 1)))

@@ -90,7 +121,7 @@ def eval_uop(uop:UOp, inputs:list[tuple[DType, list[Any]]]|None=None, vals:tuple
    bufs.append(buf:=allocator.alloc(len(data) * buf_dt.itemsize))
    allocator._copyin(buf, memoryview(struct.pack(str(len(data)) + (buf_dt.fmt or ""), *data)))
  g = UOp.param(0, uop.dtype, (1,))
-  prg = to_program(UOp.store(g.index(UOp.const(dtypes.int, 0)), uop).sink(arg=KernelInfo()), PythonRenderer(Target("PYTHON")))
+  prg = to_program(UOp.store(g.index(UOp.const(0)), uop).sink(arg=KernelInfo()), PythonRenderer(Target("PYTHON")))
  prog = dev.runtime(prg.to_elf())
  prog(out_buf:=allocator.alloc(uop.dtype.itemsize), *bufs, vals=vals)
  return out_buf.cast(uop.dtype.fmt or "").tolist()[0]
@@ -51,7 +51,7 @@ class _MXCSRContext:
    if lib is None or not hasattr(self, '_saved'): return
    lib.set_fpcr(self._saved)

-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.device import Buffer, BufferSpec, Device
 from tinygrad.runtime.autogen import hsa
@@ -192,16 +192,16 @@ def _init_sqtt_encoder():

  return emit, finish, finalize

-def _c(val, dtype=dtypes.uint32): return UOp.const(dtype, val)
+def _c(val, dtype=dtypes.uint32): return UOp.const(val, dtype)

 def _u64(lo: UOp, hi: UOp) -> UOp:
  """Combine two 32-bit UOps into a 64-bit UOp."""
-  return lo.cast(dtypes.uint64) | (hi.cast(dtypes.uint64) << UOp.const(dtypes.uint64, 32))
+  return lo.cast(dtypes.uint64) | (hi.cast(dtypes.uint64) << UOp.const(32, dtypes.uint64))

 def _split64(val: UOp) -> tuple[UOp, UOp]:
  """Split a 64-bit value into (lo, hi) 32-bit values."""
  v64 = val.bitcast(dtypes.uint64) if val.dtype == dtypes.float64 else val.cast(dtypes.uint64) if val.dtype != dtypes.uint64 else val
-  return v64.cast(dtypes.uint32), (v64 >> UOp.const(dtypes.uint64, 32)).cast(dtypes.uint32)
+  return v64.cast(dtypes.uint32), (v64 >> UOp.const(32, dtypes.uint64)).cast(dtypes.uint32)

 _SRC_MOD_TYPES = {16: (dtypes.uint16, dtypes.half, 0x7FFF), 32: (dtypes.uint32, dtypes.float32, 0x7FFFFFFF),
                  64: (dtypes.uint64, dtypes.float64, 0x7FFFFFFFFFFFFFFF)}
@@ -210,7 +210,7 @@ def _apply_src_mods(val: UOp, mod_bit: int, abs_bits: int, neg_bits: int, bits:
  if not (abs_bits & (1 << mod_bit)) and not (neg_bits & (1 << mod_bit)): return val
  ut, ft, mask = _SRC_MOD_TYPES[bits]
  fv = val.cast(ut).bitcast(ft) if bits == 16 else val.bitcast(ft) if val.dtype == ut else val
-  if abs_bits & (1 << mod_bit): fv = (fv.bitcast(ut) & UOp.const(ut, mask)).bitcast(ft)
+  if abs_bits & (1 << mod_bit): fv = (fv.bitcast(ut) & UOp.const(mask, ut)).bitcast(ft)
  if neg_bits & (1 << mod_bit): fv = fv.neg()
  return fv.bitcast(ut).cast(dtypes.uint32) if bits == 16 else fv.bitcast(ut)

@@ -251,7 +251,7 @@ def _to_u32(val: UOp) -> UOp:
  if val.dtype.itemsize == 4: return val.bitcast(dtypes.uint32)  # same size: bitcast (float32->uint32)
  return val.cast(dtypes.uint32)  # different size: cast (bool, int16, etc)
 def _lane_active(exec_mask: UOp, lane: UOp) -> UOp:
-  if exec_mask.dtype == dtypes.uint64: return ((exec_mask >> lane.cast(dtypes.uint64)) & UOp.const(dtypes.uint64, 1)).ne(UOp.const(dtypes.uint64, 0))
+  if exec_mask.dtype == dtypes.uint64: return ((exec_mask >> lane.cast(dtypes.uint64)) & UOp.const(1, dtypes.uint64)).ne(UOp.const(0, dtypes.uint64))
  return ((exec_mask >> lane.cast(dtypes.uint32)) & _c(1)).ne(_c(0))
 def _hi16(v: UOp) -> UOp: return (v >> _c(16)) & _c(0xFFFF)
 def _cond(cond, if_true, if_false):
@@ -264,9 +264,9 @@ def _set_lane_bit(old: UOp, lane: UOp, val: UOp, exec_mask: UOp) -> UOp:
  """Set/clear a single bit in a mask based on lane index, respecting exec mask."""
  if old.dtype in (dtypes.uint64, dtypes.int64):
    dt = dtypes.uint64
-    mask = UOp.const(dt, 1) << lane.cast(dt)
+    mask = UOp.const(1, dt) << lane.cast(dt)
    new_bit = _to_u32(val).cast(dt) << lane.cast(dt)
-    cleared = old.cast(dt) & (mask ^ UOp.const(dt, 0xFFFFFFFFFFFFFFFF))
+    cleared = old.cast(dt) & (mask ^ UOp.const(0xFFFFFFFFFFFFFFFF, dt))
    return _lane_active(exec_mask, lane).where(cleared | new_bit, old.cast(dt))
  mask = _c(1) << lane.cast(dtypes.uint32)
  new_bit = _to_u32(val) << lane.cast(dtypes.uint32)
@@ -365,7 +365,7 @@ def _write_64bit(val: UOp, wfn, reg_or_addr, is_mem: bool, *args) -> list[UOp]:
  """Write a 64-bit value as two 32-bit writes. args passed to wfn after reg/addr and lo/hi value."""
  lo, hi = _split64(val)
  incr = 4 if is_mem else 1  # 4 bytes for memory addresses, 1 for register indices
-  return [wfn(reg_or_addr, lo, *args), wfn(reg_or_addr + (UOp.const(reg_or_addr.dtype, incr) if isinstance(reg_or_addr, UOp) else incr), hi, *args)]
+  return [wfn(reg_or_addr, lo, *args), wfn(reg_or_addr + (UOp.const(incr, reg_or_addr.dtype) if isinstance(reg_or_addr, UOp) else incr), hi, *args)]

 def _write_val(bits: int, val: UOp, wfn, reg_or_addr, *args, is_mem: bool = False) -> list[UOp]:
  """Write value, splitting 64-bit if needed. bits=64 for 64-bit writes, otherwise 32-bit."""
@@ -374,7 +374,7 @@ def _write_val(bits: int, val: UOp, wfn, reg_or_addr, *args, is_mem: bool = Fals
 def _mem_store(mem: UOp, addr: UOp, val: UOp, active: UOp, addr_bits: int = 32, data_bits: int = 32) -> list[UOp]:
  """Conditional memory store with sub-word support. Returns list of store UOps."""
  adt = dtypes.uint64 if addr_bits == 64 else dtypes.uint32
-  word_addr = addr >> UOp.const(adt, 2)
+  word_addr = addr >> UOp.const(2, adt)
  idx = mem.index(word_addr.valid(active))
  if data_bits == 32: return [idx.store(active.where(_to_u32(val), idx))]
  # Sub-word store: read-modify-write with mask
@@ -388,7 +388,7 @@ def _mem_store(mem: UOp, addr: UOp, val: UOp, active: UOp, addr_bits: int = 32,
  is_cross = byte_pos.eq(_c(3))
  cross_word0 = (idx & _c(0x00FFFFFF)) | ((val_u32 & _c(0xFF)) << _c(24))
  store0 = idx.store(active.where(is_cross.where(cross_word0, new_word), idx))
-  next_idx = mem.index((word_addr + UOp.const(adt, 1)).valid(active & is_cross))
+  next_idx = mem.index((word_addr + UOp.const(1, adt)).valid(active & is_cross))
  cross_word1 = (next_idx & _c(0xFFFFFF00)) | ((val_u32 >> _c(8)) & _c(0xFF))
  return [store0, next_idx.store((active & is_cross).where(cross_word1, next_idx))]

@@ -397,8 +397,8 @@ def _mem_store_bytes(mem: UOp, addr: UOp, val: UOp, active: UOp, data_bits: int
  stores = []
  val_u32 = val.cast(dtypes.uint32) if val.dtype != dtypes.uint32 else val
  for i in range(data_bits // 8):
-    byte_val = (val_u32 >> UOp.const(dtypes.uint32, i * 8)) & UOp.const(dtypes.uint32, 0xFF)
-    stores.append(mem.index((addr + UOp.const(dtypes.uint64, i)).valid(active)).store(byte_val.cast(dtypes.uint8)))
+    byte_val = (val_u32 >> UOp.const(i * 8, dtypes.uint32)) & UOp.const(0xFF, dtypes.uint32)
+    stores.append(mem.index((addr + UOp.const(i, dtypes.uint64)).valid(active)).store(byte_val.cast(dtypes.uint8)))
  return stores

 def _collect_data_slices(assigns: list[tuple[str, UOp]], data_prefix: str, pcode_vars: dict | None = None, op_name: str = "") -> dict[int, UOp]:
@@ -446,7 +446,7 @@ class _Ctx:
    """Create a lane range UOp with unique axis ID."""
    if n is None: n = self.wave_size
    self._axis_id += 1
-    return UOp.range(n, self._axis_id, AxisType.LOOP, dtype=dtypes.int)
+    return UOp.range(n, self._axis_id, dtype=dtypes.int)

  def unroll_lanes(self, get_lane_bit, exec_mask: UOp, apply_exec: bool = True) -> UOp:
    """Combine lane bits into a mask using RANGE+REDUCE (32-bit for RDNA, 64-bit for CDNA)."""
@@ -462,8 +462,8 @@ class _Ctx:
  def inst_word(self, dword_idx: int) -> UOp:
    """Read instruction dword from vmem at PC + dword_idx*4."""
    pc = self.rpc()
-    addr = pc if dword_idx == 0 else pc + UOp.const(dtypes.uint64, dword_idx * 4)
-    return self.vmem.index(addr >> UOp.const(dtypes.uint64, 2)).load()
+    addr = pc if dword_idx == 0 else pc + UOp.const(dword_idx * 4, dtypes.uint64)
+    return self.vmem.index(addr >> UOp.const(2, dtypes.uint64)).load()

  def inst_field(self, field) -> UOp:
    """Extract field bits from instruction encoding. Tracks field for canonical key computation."""
@@ -475,15 +475,15 @@ class _Ctx:
    word = self.inst_word(dword_idx)
    if lo // 32 == hi // 32:  # Same dword
      mask = (1 << (hi - lo + 1)) - 1
-      shifted = word if lo_in_dword == 0 else word >> UOp.const(dtypes.uint32, lo_in_dword)
-      return shifted & UOp.const(dtypes.uint32, mask)
+      shifted = word if lo_in_dword == 0 else word >> UOp.const(lo_in_dword, dtypes.uint32)
+      return shifted & UOp.const(mask, dtypes.uint32)
    else:  # Spans two dwords
      lo_bits = 32 - lo_in_dword
      lo_mask = (1 << lo_bits) - 1
      hi_mask = (1 << (hi_in_dword + 1)) - 1
-      lo_part = (word >> UOp.const(dtypes.uint32, lo_in_dword)) & UOp.const(dtypes.uint32, lo_mask)
-      hi_part = self.inst_word(dword_idx + 1) & UOp.const(dtypes.uint32, hi_mask)
-      return lo_part | (hi_part << UOp.const(dtypes.uint32, lo_bits))
+      lo_part = (word >> UOp.const(lo_in_dword, dtypes.uint32)) & UOp.const(lo_mask, dtypes.uint32)
+      hi_part = self.inst_word(dword_idx + 1) & UOp.const(hi_mask, dtypes.uint32)
+      return lo_part | (hi_part << UOp.const(lo_bits, dtypes.uint32))

  def inst_field_signed(self, field) -> UOp:
    """Extract field and sign-extend based on field width."""
@@ -584,7 +584,7 @@ class _Ctx:
      inline = is_float_const.where(float_inline.bitcast(dtypes.uint64), int_inline.bitcast(dtypes.uint64))
      # Literal handling: F64 VOP puts literal in high 32 bits; B64/I64/U64 VOP and SOP zero-extend
      if literal is not None:
-        lit_val = literal.cast(dtypes.uint64) << UOp.const(dtypes.uint64, 32) if is_f64 else literal.cast(dtypes.uint64)
+        lit_val = literal.cast(dtypes.uint64) << UOp.const(32, dtypes.uint64) if is_f64 else literal.cast(dtypes.uint64)
        inline = off.eq(_c(255)).where(lit_val, inline)
      scalar_val = (off < _c(128)).where(sgpr_val, inline)
    else:
@@ -602,7 +602,7 @@ class _Ctx:

  def inc_pc(self) -> list[UOp]:
    """Increment PC by instruction size in bytes. Returns [store]."""
-    new_pc = self.rpc() + UOp.const(dtypes.uint64, self.inst_size)
+    new_pc = self.rpc() + UOp.const(self.inst_size, dtypes.uint64)
    lo, hi = _split64(new_pc)
    return [self.wsgpr_dyn(_c(PC_LO_IDX), lo), self.wsgpr_dyn(_c(PC_HI_IDX), hi)]

@@ -668,7 +668,7 @@ class _Ctx:
    if 'VCC' not in srcs: srcs['VCC'] = self.rmask(_c(vcc_reg))
    srcs.update({'EXEC': exec_mask, 'SCC': self.rsgpr_dyn(_c(SCC.offset)), 'laneId': lane, 'VDST': vdst_reg,
                  'ROUND_MODE': _c(0), 'ROUND_TOWARD_ZERO': _c(0), 'ROUND_NEAREST_EVEN': _c(0), '_vgpr': self.vgpr, '_wave_size': self.wave_size,
-                  'MAX_FLOAT_F32': UOp.const(dtypes.float32, 3.4028234663852886e38),
+                  'MAX_FLOAT_F32': UOp.const(3.4028234663852886e38, dtypes.float32),
                  # CDNA SDWA byte/word select constants (E32 always uses BYTE0/WORD0 defaults)
                  'SDWA_SRC0_SEL': _c(0), 'BYTE0': _c(0), 'BYTE1': _c(1), 'BYTE2': _c(2), 'BYTE3': _c(3),
                  'WORD0': _c(0), 'WORD1': _c(1)})  # rounding mode and SDWA constants
@@ -709,7 +709,7 @@ class _Ctx:
      # VGPR bit-slice assignment: VGPR[lane][reg][hi:lo] = (vgpr_idx, rhs_val, hi, lo[, cond]) -> read-modify-write
      if dest.startswith('VGPR[') and re.search(r'\[\d+:\d+\]', dest):
        # VGPR bit-slice: (vgpr_idx, rhs_val, hi_bit, lo_bit) - hi/lo are UOp constants
-        hi_bit, lo_bit = int(val[2].arg), int(val[3].arg)
+        hi_bit, lo_bit = int(val[2].val), int(val[3].val)
        width = hi_bit - lo_bit + 1
        old = self.vgpr.index(val[0]).load()
        new_val = _set_bits(old, _val_to_bits(val[1]), width, lo_bit).cast(dtypes.uint32)
@@ -732,24 +732,24 @@ class _Ctx:
            d0_width, slice_mask = d0_hi_bit - d0_lo_bit + 1, (1 << (d0_hi_bit - d0_lo_bit + 1)) - 1
            val_bits = val.bitcast(dtypes.uint16).cast(dtypes.uint32) if val.dtype == dtypes.half else \
                       val.cast(dtypes.uint32) if val.dtype in (dtypes.uint16, dtypes.int16) else \
-                       val.cast(dtypes.uint32) & UOp.const(dtypes.uint32, slice_mask)
+                       val.cast(dtypes.uint32) & UOp.const(slice_mask, dtypes.uint32)
            raw_stores.append(('vgpr_slice', (d0_lo_bit, d0_width, val_bits)))
            continue
        # For integer ops with clamp, use pre-computed saturated value; for floats, clamp to [0,1]
        if int_saturate is not None: val = int_saturate
        elif clmp and val.dtype in (dtypes.float32, dtypes.half, dtypes.float64):
-          clamped = val.maximum(UOp.const(val.dtype, 0.0)).minimum(UOp.const(val.dtype, 1.0))
-          val = _FUNCS['isNAN'](val).where(UOp.const(val.dtype, 0.0), clamped)
+          clamped = val.maximum(UOp.const(0.0, val.dtype)).minimum(UOp.const(1.0, val.dtype))
+          val = _FUNCS['isNAN'](val).where(UOp.const(0.0, val.dtype), clamped)
        if val.dtype in (dtypes.uint64, dtypes.int64, dtypes.float64):
          lo, hi = _split64(val)
          raw_stores.extend([('vgpr', self.wvgpr_dyn(vdst_reg, lane, lo, exec_mask)),
                             ('vgpr', self.wvgpr_dyn(vdst_reg + _c(1), lane, hi, exec_mask))])
        elif val.dtype in (dtypes.half, dtypes.uint16, dtypes.int16):
          result, old_val = _val_to_u32(val), self.rvgpr_dyn(vdst_reg, lane)
-          hi_result = (old_val & UOp.const(dtypes.uint32, 0xFFFF)) | (result << UOp.const(dtypes.uint32, 16))
+          hi_result = (old_val & UOp.const(0xFFFF, dtypes.uint32)) | (result << UOp.const(16, dtypes.uint32))
          # GFX9/CDNA zeroes upper 16 bits on lo-half write; RDNA preserves them
-          lo_result = (result & UOp.const(dtypes.uint32, 0xFFFF)) if self.wave_size == 64 else \
-                      (old_val & UOp.const(dtypes.uint32, 0xFFFF0000)) | (result & UOp.const(dtypes.uint32, 0xFFFF))
+          lo_result = (result & UOp.const(0xFFFF, dtypes.uint32)) if self.wave_size == 64 else \
+                      (old_val & UOp.const(0xFFFF0000, dtypes.uint32)) | (result & UOp.const(0xFFFF, dtypes.uint32))
          result = opsel_dst_hi.where(hi_result, lo_result) if isinstance(opsel_dst_hi, UOp) else hi_result if opsel_dst_hi else lo_result
          raw_stores.append(('vgpr', self.wvgpr_dyn(vdst_reg, lane, result, exec_mask)))
        else: raw_stores.append(('vgpr', self.wvgpr_dyn(vdst_reg, lane, _val_to_u32(val), exec_mask)))
@@ -767,8 +767,8 @@ class _Ctx:
    if slice_stores:
      result = self.rvgpr_dyn(vdst_reg, lane)
      for lo_bit, width, val_bits in slice_stores:
-        mask = UOp.const(dtypes.uint32, ((1 << width) - 1) << lo_bit)
-        result = (result & (mask ^ UOp.const(dtypes.uint32, 0xFFFFFFFF))) | (val_bits << UOp.const(dtypes.uint32, lo_bit))
+        mask = UOp.const(((1 << width) - 1) << lo_bit, dtypes.uint32)
+        result = (result & (mask ^ UOp.const(0xFFFFFFFF, dtypes.uint32))) | (val_bits << UOp.const(lo_bit, dtypes.uint32))
      lane_stores.append(self.wvgpr_dyn(vdst_reg, lane, result, exec_mask))
    # VCC/EXEC mask writes must be computed BEFORE VGPR stores to avoid reading modified VGPRs.
    # When vdst overlaps with src operands (e.g. v_add_co_u32 v[0], vcc, s[8], v[0]), the carry
@@ -790,8 +790,8 @@ class _Ctx:
 def _compile_sopp(inst: ir3.SOPP | ir4.SOPP, ctx: _Ctx) -> UOp:
  simm16 = ctx.inst_field_signed(type(inst).simm16).cast(dtypes.int16)
  if inst.op in (ir3.SOPPOp.S_ENDPGM, ir4.SOPPOp.S_ENDPGM, irc.SOPPOp.S_ENDPGM):
-    return UOp.sink(ctx.wsgpr_dyn(_c(PC_LO_IDX), UOp.const(dtypes.uint32, 0xFFFFFFFF)),
-                          ctx.wsgpr_dyn(_c(PC_HI_IDX), UOp.const(dtypes.uint32, 0xFFFFFFFF)))
+    return UOp.sink(ctx.wsgpr_dyn(_c(PC_LO_IDX), UOp.const(0xFFFFFFFF, dtypes.uint32)),
+                          ctx.wsgpr_dyn(_c(PC_HI_IDX), UOp.const(0xFFFFFFFF, dtypes.uint32)))
  # S_BARRIER: advance PC past the barrier instruction. The execution loop detects barriers before executing and handles synchronization.
  barrier_ops = {ir3.SOPPOp.S_BARRIER, irc.SOPPOp.S_BARRIER}
  if hasattr(ir4.SOPPOp, 'S_BARRIER_WAIT'): barrier_ops.add(ir4.SOPPOp.S_BARRIER_WAIT)
@@ -804,8 +804,8 @@ def _compile_sopp(inst: ir3.SOPP | ir4.SOPP, ctx: _Ctx) -> UOp:
    pc_bytes = ctx.rpc()  # PC is already 64-bit byte address
    vcc, exec_val = ctx.rmask(_c(VCC_LO.offset)), ctx.rexec()
    srcs: dict[str, UOp|int] = {'PC': pc_bytes.cast(dtypes.int64), 'SIMM16': simm16, 'SCC': ctx.rsgpr_dyn(_c(SCC.offset)), 'VCC': vcc,
-            'VCCZ': vcc.eq(UOp.const(vcc.dtype, 0)).cast(dtypes.uint32),
-            'EXECZ': exec_val.eq(UOp.const(exec_val.dtype, 0)).cast(dtypes.uint32)}
+            'VCCZ': vcc.eq(UOp.const(0, vcc.dtype)).cast(dtypes.uint32),
+            'EXECZ': exec_val.eq(UOp.const(0, exec_val.dtype)).cast(dtypes.uint32)}
    for dest, val in parse_pcode(pcode, srcs)[1]:
      if dest == 'PC' or dest.startswith('PC.'):
        lo, hi = _split64(val.cast(dtypes.uint64))
@@ -833,12 +833,12 @@ def _compile_smem(inst: ir3.SMEM | ir4.SMEM, ctx: _Ctx) -> UOp:
  part = op_name.rsplit('_', 1)[1]  # B32, DWORD, DWORDX2, U8, I8, etc.
  nval = int(part.removeprefix('DWORD').removeprefix('X') or '1') if 'DWORD' in part else int(part[1:]) / 32 * (-1 if part[0] == 'I' else 1)
  ndwords = max(1, int(abs(nval)))
-  dword_base = addr >> UOp.const(dtypes.uint64, 2)
-  vals = [ctx.vmem.index(dword_base + UOp.const(dtypes.uint64, i)) for i in range(ndwords)]
+  dword_base = addr >> UOp.const(2, dtypes.uint64)
+  vals = [ctx.vmem.index(dword_base + UOp.const(i, dtypes.uint64)) for i in range(ndwords)]
  if abs(nval) < 1:
    nbits = int(abs(nval) * 32)
-    byte_off = (addr & UOp.const(dtypes.uint64, 3)).cast(dtypes.uint32) * UOp.const(dtypes.uint32, 8)
-    extracted = (vals[0] >> byte_off) & UOp.const(dtypes.uint32, (1 << nbits) - 1)
+    byte_off = (addr & UOp.const(3, dtypes.uint64)).cast(dtypes.uint32) * UOp.const(8, dtypes.uint32)
+    extracted = (vals[0] >> byte_off) & UOp.const((1 << nbits) - 1, dtypes.uint32)
    vals[0] = extracted.cast({8: dtypes.int8, 16: dtypes.int16}[nbits]).cast(dtypes.int32).bitcast(dtypes.uint32) if nval < 0 else extracted
  stores = [ctx.wsgpr_dyn(sdata_reg + _c(i), vals[i]) for i in range(ndwords)]
  return UOp.sink(*stores, *ctx.inc_pc())
@@ -941,7 +941,7 @@ def _dpp16_ctrl(lane: UOp, dpp: int, row_mask: int, bank_mask: int, wave_size: i
  enabled = (((_c(row_mask) >> row.cast(dtypes.uint32)) & _c(1)).ne(_c(0)) &
             (((_c(bank_mask) >> bank.cast(dtypes.uint32)) & _c(1)).ne(_c(0))))
  op, arg = decode_dpp16(dpp)
-  src_lane, valid = lane_i, UOp.const(dtypes.bool, True)
+  src_lane, valid = lane_i, UOp.const(True)

  if op == 'quad_perm':
    assert isinstance(arg, tuple)
@@ -967,7 +967,7 @@ def _load_dpp16_src0(ctx: _Ctx, inst, lane: UOp, fallback: UOp) -> UOp:
                                         getattr(inst, 'bank_mask', 0xf) or 0xf, ctx.wave_size)
  safe_src_lane = (enabled & valid).where(src_lane, _c(0, dtypes.int))
  swizzled = ctx.rvgpr_dyn(ctx.inst_field(type(inst).vsrc0), safe_src_lane)
-  invalid = UOp.const(fallback.dtype, 0) if getattr(inst, 'bc', 0) else fallback
+  invalid = UOp.const(0, fallback.dtype) if getattr(inst, 'bc', 0) else fallback
  return enabled.where(valid.where(swizzled, invalid), fallback)

 def _compile_sdwa(inst: irc.VOP1_SDWA | irc.VOP2_SDWA | irc.VOP2_SDWA_SDST | irc.VOPC_SDWA_SDST, ctx: _Ctx) -> UOp:
@@ -1159,7 +1159,7 @@ def _compile_vopc(inst: ir3.VOPC|ir3.VOPC_DPP16|ir3.VOP3|ir4.VOPC|ir4.VOPC_DPP16
        s1 = _apply_src_mods(s1, 0, 1 if getattr(inst, 'src1_abs', 0) else 0, 1 if getattr(inst, 'src1_neg', 0) else 0, bits['s1'])
      s0 = _apply_src_mods(s0, 0, abs_bits, neg_bits, bits['s0'])
      s1 = _apply_src_mods(s1, 1, abs_bits, neg_bits, bits['s1'])
-    for dest, val in parse_pcode(pcode, {'S0': s0, 'S1': s1, 'laneId': lc, 'D0': UOp.const(dtypes.uint64, 0)})[1]:
+    for dest, val in parse_pcode(pcode, {'S0': s0, 'S1': s1, 'laneId': lc, 'D0': UOp.const(0, dtypes.uint64)})[1]:
      if '[laneId]' in dest and ('D0' in dest or 'EXEC' in dest): return val.cast(dtypes.uint32)
    return _c(0)

@@ -1189,8 +1189,8 @@ def _compile_bitop3(inst, ctx: _Ctx, exec_mask: UOp, bits: dict, op_name: str) -
  is_16 = 'B16' in op_name
  dt, mask = (dtypes.uint16, 0xFFFF) if is_16 else (dtypes.uint32, 0xFFFFFFFF)
  s0, s1, s2 = src0.cast(dt), src1.cast(dt), src2.cast(dt)
-  def bnot(v): return v ^ UOp.const(dt, mask)
-  result = UOp.const(dt, 0)
+  def bnot(v): return v ^ UOp.const(mask, dt)
+  result = UOp.const(0, dt)
  for i in range(8):
    if not (ttbl & (1 << i)): continue
    result = result | ((s0 if i & 4 else bnot(s0)) & (s1 if i & 2 else bnot(s1)) & (s2 if i & 1 else bnot(s2)))
@@ -1244,7 +1244,7 @@ def _compile_vop3(inst: ir3.VOP3 | ir4.VOP3 | irc.VOP3, ctx: _Ctx) -> UOp:
  src0 = _apply_src_mods(src0, 0, abs_bits, neg_bits, bits['s0'])
  src1 = _apply_src_mods(src1, 1, abs_bits, neg_bits, bits['s1'])
  src2 = _apply_src_mods(src2, 2, abs_bits, neg_bits, bits['s2'])
-  srcs = {'S0': src0, 'S1': src1, 'S2': src2, 'OPSEL': UOp.const(dtypes.uint32, opsel)}
+  srcs = {'S0': src0, 'S1': src1, 'S2': src2, 'OPSEL': UOp.const(opsel, dtypes.uint32)}
  if 'CNDMASK' in op_name and src2 is not None: srcs['VCC'] = src2
  # FMAC instructions need D0 (accumulator) from destination register
  if 'FMAC' in op_name: srcs['D0'] = ctx.rvgpr_dyn(vdst_reg, lane)
@@ -1398,7 +1398,7 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
  # Layout: tmp[0..n_a_elems-1] = A[m][k], tmp[n_a_elems..n_a_elems+n_b_elems-1] = B[n][k]
  # Within each group of lanes, lane%grp_sub gives M/N index, lane//grp_sub gives sub-block
  grp_sub = min(M, 16)  # lanes within group mapped to M/N dimension
-  b_off = UOp.const(dtypes.int, n_a_elems)
+  b_off = UOp.const(n_a_elems, dtypes.int)
  acc_dt = dtypes.int32 if is_int_out else dtypes.float32
  # Use uint32 temp array to prevent optimizer from eliminating f16→f32 bitcast chains.
  # The optimizer folds bitcast(uint32→float32) stores to float32 arrays, losing the conversion.
@@ -1407,44 +1407,44 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
  def cvt_elem(raw: UOp, sub_idx: int) -> UOp:
    if is_i8:
      # Extract i8, sign-extend to i32
-      byte_val = (raw >> UOp.const(dtypes.uint32, sub_idx * 8)) & UOp.const(dtypes.uint32, 0xFF)
-      return (byte_val.cast(dtypes.int32) ^ UOp.const(dtypes.int32, 0x80)) - UOp.const(dtypes.int32, 0x80)
+      byte_val = (raw >> UOp.const(sub_idx * 8, dtypes.uint32)) & UOp.const(0xFF, dtypes.uint32)
+      return (byte_val.cast(dtypes.int32) ^ UOp.const(0x80, dtypes.int32)) - UOp.const(0x80, dtypes.int32)
    elif is_f32_src:
      return raw  # already uint32 (f32 bit pattern)
    elif is_fp8:
-      return ((raw >> UOp.const(dtypes.uint32, sub_idx * 8)) & UOp.const(dtypes.uint32, 0xFF)).cast(dtypes.uint32)
+      return ((raw >> UOp.const(sub_idx * 8, dtypes.uint32)) & UOp.const(0xFF, dtypes.uint32)).cast(dtypes.uint32)
    elif is_bf16:
      # bf16→f32 bits: just shift left by 16 (bf16 is upper 16 bits of f32)
-      return ((raw >> UOp.const(dtypes.uint32, sub_idx * 16)) & UOp.const(dtypes.uint32, 0xFFFF)) << UOp.const(dtypes.uint32, 16)
+      return ((raw >> UOp.const(sub_idx * 16, dtypes.uint32)) & UOp.const(0xFFFF, dtypes.uint32)) << UOp.const(16, dtypes.uint32)
    else:
      # f16→f32 conversion using float arithmetic to avoid UOp optimizer eliminating the conversion.
      # The optimizer folds bitcast(uint32→float32) chains, so we compute the float value directly.
-      h = (raw >> UOp.const(dtypes.uint32, sub_idx * 16)) & UOp.const(dtypes.uint32, 0xFFFF)
-      sign = (h >> UOp.const(dtypes.uint32, 15)) & UOp.const(dtypes.uint32, 1)
-      exp = (h >> UOp.const(dtypes.uint32, 10)) & UOp.const(dtypes.uint32, 0x1F)
-      mant = h & UOp.const(dtypes.uint32, 0x3FF)
+      h = (raw >> UOp.const(sub_idx * 16, dtypes.uint32)) & UOp.const(0xFFFF, dtypes.uint32)
+      sign = (h >> UOp.const(15, dtypes.uint32)) & UOp.const(1, dtypes.uint32)
+      exp = (h >> UOp.const(10, dtypes.uint32)) & UOp.const(0x1F, dtypes.uint32)
+      mant = h & UOp.const(0x3FF, dtypes.uint32)
      # Use bf16 path: shift left by 16 to create bf16 bits, then shift mantissa and adjust exponent in float domain
      # bf16 bits = (sign << 15) | (exp_bf16 << 7) | mant_bf16 -- but f16 and bf16 have different formats
      # Instead: construct f32 bits properly, use a local uint32 array to force materialization
-      f32_bits = (sign << UOp.const(dtypes.uint32, 31)) | \
-                 ((exp + UOp.const(dtypes.uint32, 112)) << UOp.const(dtypes.uint32, 23)) | \
-                 (mant << UOp.const(dtypes.uint32, 13))
-      is_zero = exp.eq(UOp.const(dtypes.uint32, 0))
+      f32_bits = (sign << UOp.const(31, dtypes.uint32)) | \
+                 ((exp + UOp.const(112, dtypes.uint32)) << UOp.const(23, dtypes.uint32)) | \
+                 (mant << UOp.const(13, dtypes.uint32))
+      is_zero = exp.eq(UOp.const(0, dtypes.uint32))
      # Return uint32 (f32 bit pattern) — stored directly to uint32 temp array, bitcast to float on read
-      return is_zero.where(UOp.const(dtypes.uint32, 0), f32_bits)
+      return is_zero.where(UOp.const(0, dtypes.uint32), f32_bits)

  read_lane = ctx.range()
  # For 32x32: lane%16 = M/N index within 16-wide block, lane//16 = which of 4 quarter-waves
  # Groups: lanes 0-31 = group 0, lanes 32-63 = group 1
  # Within group: (lane%32)%16 = M/N[0-15], (lane%32)//16 selects M/N[0-15] or [16-31]
-  lane_in_grp = read_lane % UOp.const(dtypes.int, grp_size)
-  grp_idx = read_lane // UOp.const(dtypes.int, grp_size)
+  lane_in_grp = read_lane % UOp.const(grp_size, dtypes.int)
+  grp_idx = read_lane // UOp.const(grp_size, dtypes.int)

  if M == 32:
    # 32x32: lane_in_grp%16 = sub-row/col (0-15), lane_in_grp//16 = block (0=rows 0-15, 1=rows 16-31)
-    sub_mn = lane_in_grp % UOp.const(dtypes.int, 16)
-    block_mn = lane_in_grp // UOp.const(dtypes.int, 16)
-    mn_idx = block_mn * UOp.const(dtypes.int, 16) + sub_mn  # actual M/N index (0-31)
+    sub_mn = lane_in_grp % UOp.const(16, dtypes.int)
+    block_mn = lane_in_grp // UOp.const(16, dtypes.int)
+    mn_idx = block_mn * UOp.const(16, dtypes.int) + sub_mn  # actual M/N index (0-31)
  else:
    mn_idx = lane_in_grp  # for 16x16 and 4x4

@@ -1456,18 +1456,18 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
                                ctx.rsrc_dyn(src0_off, _c(0, dtypes.int), 32))
    a_val = cvt_elem(a_raw, sub_idx)
    if M == 4:
-      a_idx = grp_idx * UOp.const(dtypes.int, M * K) + mn_idx * UOp.const(dtypes.int, K) + UOp.const(dtypes.int, kl)
+      a_idx = grp_idx * UOp.const(M * K, dtypes.int) + mn_idx * UOp.const(K, dtypes.int) + UOp.const(kl, dtypes.int)
    else:
-      a_idx = mn_idx * UOp.const(dtypes.int, K) + grp_idx * UOp.const(dtypes.int, k_per_grp) + UOp.const(dtypes.int, kl)
+      a_idx = mn_idx * UOp.const(K, dtypes.int) + grp_idx * UOp.const(k_per_grp, dtypes.int) + UOp.const(kl, dtypes.int)
    read_stores.append(tmp.index(a_idx).store(a_val))

    b_raw = src1_is_vgpr.where(ctx.rvgpr_dyn(src1_r + _c(reg_idx), read_lane),
                                ctx.rsrc_dyn(src1_off, _c(0, dtypes.int), 32))
    b_val = cvt_elem(b_raw, sub_idx)
    if M == 4:
-      b_idx = b_off + grp_idx * UOp.const(dtypes.int, N * K) + mn_idx * UOp.const(dtypes.int, K) + UOp.const(dtypes.int, kl)
+      b_idx = b_off + grp_idx * UOp.const(N * K, dtypes.int) + mn_idx * UOp.const(K, dtypes.int) + UOp.const(kl, dtypes.int)
    else:
-      b_idx = b_off + mn_idx * UOp.const(dtypes.int, K) + grp_idx * UOp.const(dtypes.int, k_per_grp) + UOp.const(dtypes.int, kl)
+      b_idx = b_off + mn_idx * UOp.const(K, dtypes.int) + grp_idx * UOp.const(k_per_grp, dtypes.int) + UOp.const(kl, dtypes.int)
    read_stores.append(tmp.index(b_idx).store(b_val))

  read_phase = UOp.group(*read_stores).end(read_lane)
@@ -1488,11 +1488,11 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
    # Lane mapping: n = (lane%32)%16 + ((lane%32)//16)*16, gives column 0-31
    # Row groups: 4 groups of 4, covering rows 0-31. Group g (0-3): rows g*4 .. g*4+3
    # group assignment: lane//16 gives quarter (0-3), each quarter maps to 4 rows
-    c_lane_in_32 = compute_lane % UOp.const(dtypes.int, 32)
-    c_sub = c_lane_in_32 % UOp.const(dtypes.int, 16)
-    c_block = c_lane_in_32 // UOp.const(dtypes.int, 16)
-    n_idx = c_block * UOp.const(dtypes.int, 16) + c_sub
-    c_half = compute_lane // UOp.const(dtypes.int, 32)  # 0 or 1
+    c_lane_in_32 = compute_lane % UOp.const(32, dtypes.int)
+    c_sub = c_lane_in_32 % UOp.const(16, dtypes.int)
+    c_block = c_lane_in_32 // UOp.const(16, dtypes.int)
+    n_idx = c_block * UOp.const(16, dtypes.int) + c_sub
+    c_half = compute_lane // UOp.const(32, dtypes.int)  # 0 or 1

    for out_reg in range(16):
      # Each half covers 8 rows. out_reg 0-3: rows 0-3 (half0) or 16-19 (half1)
@@ -1503,7 +1503,7 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
      # acc[r] at lane l -> D[m][n] where n = (l%32)%16 + ((l%32)//16)*16
      # m = (l//32)*16 + (r//4)*4 + (r%4)  ... giving rows in blocks of 4
      # So: m_base = half * 16 + (out_reg // 4) * 4 + (out_reg % 4)
-      m_base = c_half * UOp.const(dtypes.int, 16) + UOp.const(dtypes.int, (out_reg // 4) * 4 + (out_reg % 4))
+      m_base = c_half * UOp.const(16, dtypes.int) + UOp.const((out_reg // 4) * 4 + (out_reg % 4), dtypes.int)

      acc_v = (ctx.raccvgpr_dyn if use_acc else ctx.rvgpr_dyn)(src2_r + _c(out_reg), compute_lane, src2_is_vgpr)
      if is_int_out: acc_v = acc_v.cast(dtypes.int32)
@@ -1511,8 +1511,8 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
      acc = src2_is_vgpr.where(acc_v, acc_scalar)

      for k in range(K):
-        a_val = tmp2.index(m_base * UOp.const(dtypes.int, K) + UOp.const(dtypes.int, k)).bitcast(acc_dt)
-        b_val = tmp2.index(b_off + n_idx * UOp.const(dtypes.int, K) + UOp.const(dtypes.int, k)).bitcast(acc_dt)
+        a_val = tmp2.index(m_base * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+        b_val = tmp2.index(b_off + n_idx * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
        acc = acc + a_val * b_val

      if is_int_out:
@@ -1523,8 +1523,8 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
          vdst_reg + _c(out_reg), compute_lane, acc.bitcast(dtypes.uint32), exec_mask))
  else:
    # 16x16 and 4x4: each lane computes out_per_lane outputs
-    n_idx = compute_lane % UOp.const(dtypes.int, grp_sub)
-    c_grp = compute_lane // UOp.const(dtypes.int, grp_sub)
+    n_idx = compute_lane % UOp.const(grp_sub, dtypes.int)
+    c_grp = compute_lane // UOp.const(grp_sub, dtypes.int)

    for out_reg in range(out_per_lane):
      acc_v = (ctx.raccvgpr_dyn if use_acc else ctx.rvgpr_dyn)(src2_r + _c(out_reg), compute_lane, src2_is_vgpr)
@@ -1534,17 +1534,17 @@ def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:

      if M == 4:
        # 4x4: each group is independent. A/B indexed per-group.
-        m_base = c_grp * UOp.const(dtypes.int, M * K) + UOp.const(dtypes.int, out_reg * K)
+        m_base = c_grp * UOp.const(M * K, dtypes.int) + UOp.const(out_reg * K, dtypes.int)
        for k in range(K):
-          a_val = tmp2.index(m_base + UOp.const(dtypes.int, k)).bitcast(acc_dt)
-          b_val = tmp2.index(b_off + c_grp * UOp.const(dtypes.int, N*K) + n_idx * UOp.const(dtypes.int, K)+UOp.const(dtypes.int, k)).bitcast(acc_dt)
+          a_val = tmp2.index(m_base + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+          b_val = tmp2.index(b_off + c_grp * UOp.const(N*K, dtypes.int) + n_idx * UOp.const(K, dtypes.int)+UOp.const(k, dtypes.int)).bitcast(acc_dt)
          acc = acc + a_val * b_val
      else:
        # 16x16: K is split across groups. Shared MxK/NxK arrays.
-        m_base = c_grp * UOp.const(dtypes.int, out_per_lane) + UOp.const(dtypes.int, out_reg)
+        m_base = c_grp * UOp.const(out_per_lane, dtypes.int) + UOp.const(out_reg, dtypes.int)
        for k in range(K):
-          a_val = tmp2.index(m_base * UOp.const(dtypes.int, K) + UOp.const(dtypes.int, k)).bitcast(acc_dt)
-          b_val = tmp2.index(b_off + n_idx * UOp.const(dtypes.int, K) + UOp.const(dtypes.int, k)).bitcast(acc_dt)
+          a_val = tmp2.index(m_base * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+          b_val = tmp2.index(b_off + n_idx * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
          acc = acc + a_val * b_val

      if is_int_out:
@@ -1570,8 +1570,8 @@ def _compile_wmma(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P, ctx: _Ctx) -> UOp:
  is_rdna4 = isinstance(inst, ir4.VOP3P)
  # read 16x16 F16/BF16 matrix from VGPRs → flat f32 array[row*16+k]
  def read_f16_val(src, lane, vgpr, half):
-    v = ctx.rvgpr_dyn(src + _c(vgpr), UOp.const(dtypes.int, lane))
-    return cvt((v >> UOp.const(dtypes.uint32, 16)) if half else (v & UOp.const(dtypes.uint32, 0xFFFF)))
+    v = ctx.rvgpr_dyn(src + _c(vgpr), UOp.const(lane, dtypes.int))
+    return cvt((v >> UOp.const(16, dtypes.uint32)) if half else (v & UOp.const(0xFFFF, dtypes.uint32)))

  # RDNA3: 16 lanes × 8 VGPRs × 2 halves, k maps linearly
  # RDNA4: 32 lanes × 4 VGPRs × 2 halves, k bits are scrambled (k[2] goes to lane bit 4)
@@ -1593,20 +1593,20 @@ def _compile_wmma(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P, ctx: _Ctx) -> UOp:
             for m in range(16) for n in range(16) for lane, vgpr in [d_map(m, n)]]
    mat_d = [sum(mat_a[r*16+k] * mat_b[c*16+k] for k in range(16)) + mat_c[r*16+c] for r in range(16) for c in range(16)]
    def f32_to_f16_bits(v: UOp) -> UOp: return v.cast(dtypes.half).bitcast(dtypes.uint16).cast(dtypes.uint32)
-    def f32_to_bf16_bits(v: UOp) -> UOp: return (v.bitcast(dtypes.uint32) >> UOp.const(dtypes.uint32, 16)) & UOp.const(dtypes.uint32, 0xFFFF)
+    def f32_to_bf16_bits(v: UOp) -> UOp: return (v.bitcast(dtypes.uint32) >> UOp.const(16, dtypes.uint32)) & UOp.const(0xFFFF, dtypes.uint32)
    out_cvt = f32_to_bf16_bits if is_bf16 else f32_to_f16_bits
    if is_rdna4:  # pack 2 f16 per VGPR: adjacent m values share (lane, vgpr) since vgpr=m&7, half=m&1
-      stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1] // 2), UOp.const(dtypes.int, d_map(m, n)[0]),
-                out_cvt(mat_d[m*16+n]) | (out_cvt(mat_d[(m+1)*16+n]) << UOp.const(dtypes.uint32, 16)), exec_mask)
+      stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1] // 2), UOp.const(d_map(m, n)[0], dtypes.int),
+                out_cvt(mat_d[m*16+n]) | (out_cvt(mat_d[(m+1)*16+n]) << UOp.const(16, dtypes.uint32)), exec_mask)
                for n in range(16) for m in range(0, 16, 2)]
    else:  # (rdna3) 1 f16 per VGPR (lo half only)
-      stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1]), UOp.const(dtypes.int, d_map(m, n)[0]), out_cvt(mat_d[m*16+n]), exec_mask)
+      stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1]), UOp.const(d_map(m, n)[0], dtypes.int), out_cvt(mat_d[m*16+n]), exec_mask)
                for m in range(16) for n in range(16)]
  else: # f32
-    mat_c = [ctx.rvgpr_dyn(src2_r + _c(d_map(m, n)[1]), UOp.const(dtypes.int, d_map(m, n)[0])).bitcast(dtypes.float32)
+    mat_c = [ctx.rvgpr_dyn(src2_r + _c(d_map(m, n)[1]), UOp.const(d_map(m, n)[0], dtypes.int)).bitcast(dtypes.float32)
             for m in range(16) for n in range(16)]
    mat_d = [sum(mat_a[r*16+k] * mat_b[c*16+k] for k in range(16)) + mat_c[r*16+c] for r in range(16) for c in range(16)]
-    stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1]), UOp.const(dtypes.int, d_map(m, n)[0]), mat_d[m*16+n].bitcast(dtypes.uint32), exec_mask)
+    stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1]), UOp.const(d_map(m, n)[0], dtypes.int), mat_d[m*16+n].bitcast(dtypes.uint32), exec_mask)
              for m in range(16) for n in range(16)]
  return UOp.sink(*stores, *ctx.inc_pc())

@@ -1688,8 +1688,8 @@ def _compile_vop3p(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P, ctx: _Ctx) -> UOp:
      scalar_hi_sel = src_lo if not opsel_hi_bit else is_sgpr_pair.where(sgpr_hi, src_lo)
      lo = is_vgpr.where(vgpr_hi if opsel_lo else vgpr_lo, scalar_lo_sel)
      hi = is_vgpr.where(vgpr_hi if opsel_hi_bit else vgpr_lo, scalar_hi_sel)
-      if neg_lo: lo = lo ^ UOp.const(dtypes.uint32, 0x80000000)
-      if neg_hi_bit: hi = hi ^ UOp.const(dtypes.uint32, 0x80000000)
+      if neg_lo: lo = lo ^ UOp.const(0x80000000, dtypes.uint32)
+      if neg_hi_bit: hi = hi ^ UOp.const(0x80000000, dtypes.uint32)
      return _u64(lo, hi)
    srcs = {'S0': build_pk_f32(src0, src_offs[0], opsel & 1, opsel_hi & 1, neg & 1, neg_hi & 1),
            'S1': build_pk_f32(src1, src_offs[1], opsel & 2, opsel_hi & 2, neg & 2, neg_hi & 2),
@@ -1700,35 +1700,35 @@ def _compile_vop3p(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P, ctx: _Ctx) -> UOp:
    def apply_abs(v, bit, opsel_hi_bit, opsel_bit):
      if not (neg_hi & bit): return v
      # Apply abs based on whether source is f32 or f16
-      if not (combined_opsel_hi & opsel_hi_bit): return v & UOp.const(dtypes.uint32, 0x7FFFFFFF)  # f32 abs
-      if opsel & opsel_bit: return v & UOp.const(dtypes.uint32, 0x7FFF0000)  # f16 hi abs (preserve lo)
-      return v & UOp.const(dtypes.uint32, 0xFFFF7FFF)  # f16 lo abs (preserve hi)
+      if not (combined_opsel_hi & opsel_hi_bit): return v & UOp.const(0x7FFFFFFF, dtypes.uint32)  # f32 abs
+      if opsel & opsel_bit: return v & UOp.const(0x7FFF0000, dtypes.uint32)  # f16 hi abs (preserve lo)
+      return v & UOp.const(0xFFFF7FFF, dtypes.uint32)  # f16 lo abs (preserve hi)
    def apply_neg_mix(v, bit, opsel_hi_bit, opsel_bit):
      if not (neg & bit): return v
-      if not (combined_opsel_hi & opsel_hi_bit): return v ^ UOp.const(dtypes.uint32, 0x80000000)  # f32 neg
-      if opsel & opsel_bit: return v ^ UOp.const(dtypes.uint32, 0x80000000)  # f16 hi neg
-      return v ^ UOp.const(dtypes.uint32, 0x00008000)  # f16 lo neg
+      if not (combined_opsel_hi & opsel_hi_bit): return v ^ UOp.const(0x80000000, dtypes.uint32)  # f32 neg
+      if opsel & opsel_bit: return v ^ UOp.const(0x80000000, dtypes.uint32)  # f16 hi neg
+      return v ^ UOp.const(0x00008000, dtypes.uint32)  # f16 lo neg
    s0_mod = apply_neg_mix(apply_abs(src0, 1, 1, 1), 1, 1, 1)
    s1_mod = apply_neg_mix(apply_abs(src1, 2, 2, 2), 2, 2, 2)
    s2_mod = apply_neg_mix(apply_abs(src2, 4, 4, 4), 4, 4, 4)
    srcs = {'S@0': s0_mod, 'S@1': s1_mod, 'S@2': s2_mod,
-            'OPSEL_HI': UOp.const(dtypes.uint32, combined_opsel_hi), 'OPSEL': UOp.const(dtypes.uint32, opsel)}
+            'OPSEL_HI': UOp.const(combined_opsel_hi, dtypes.uint32), 'OPSEL': UOp.const(opsel, dtypes.uint32)}
  else:
    def get_half_bits(val: UOp, use_hi: bool, apply_neg: bool = False) -> UOp:
-      bits = ((val >> UOp.const(dtypes.uint32, 16)) if use_hi else val) & UOp.const(dtypes.uint32, 0xFFFF)
+      bits = ((val >> UOp.const(16, dtypes.uint32)) if use_hi else val) & UOp.const(0xFFFF, dtypes.uint32)
      if apply_neg: bits = bits.cast(dtypes.uint16).bitcast(dtypes.half).neg().bitcast(dtypes.uint16).cast(dtypes.uint32)
      return bits
    def build_remapped_src(src: UOp, opsel_lo_bit: int, opsel_hi_bit: int, neg_lo_bit: int, neg_hi_bit: int) -> UOp:
      lo = get_half_bits(src, bool(opsel_lo_bit), bool(neg_lo_bit))
      hi = get_half_bits(src, bool(opsel_hi_bit), bool(neg_hi_bit))
-      return lo | (hi << UOp.const(dtypes.uint32, 16))
+      return lo | (hi << UOp.const(16, dtypes.uint32))
    # DOT IU instructions use NEG bits for signed/unsigned selection, not fp16 negation
    is_dot_iu = 'DOT' in op_name and 'IU' in op_name
    n0, n1, n2, nh0, nh1, nh2 = (0, 0, 0, 0, 0, 0) if is_dot_iu else (neg & 1, neg & 2, neg & 4, neg_hi & 1, neg_hi & 2, neg_hi & 4)
    srcs = {'S0': build_remapped_src(src0, opsel & 1, opsel_hi & 1, n0, nh0),
            'S1': build_remapped_src(src1, opsel & 2, opsel_hi & 2, n1, nh1),
            'S2': build_remapped_src(src2, opsel & 4, 1 if opsel_hi2 else 0, n2, nh2)}
-    if is_dot_iu: srcs['NEG'] = UOp.const(dtypes.uint32, neg)
+    if is_dot_iu: srcs['NEG'] = UOp.const(neg, dtypes.uint32)
  return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask)

 def _compile_vopd(inst: ir3.VOPD | ir4.VOPD, ctx: _Ctx) -> UOp:
@@ -1781,7 +1781,7 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
  # CDNA acc bit: when set, VGPR operands (vdst/vdata) target ACCVGPR file instead of VGPR
  use_acc = bool(getattr(inst, 'acc', 0))
  mem = ctx.lds if is_lds else ctx.scratch if is_scratch else ctx.vmem
-  addr_shift = UOp.const(dtypes.uint32 if is_lds else dtypes.uint64, 2)
+  addr_shift = UOp.const(2, dtypes.uint32 if is_lds else dtypes.uint64)

  # Extract register info - all dynamic for deduplication
  if is_lds:
@@ -1831,20 +1831,20 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
      return addr
    offset64 = offset.cast(dtypes.uint64)
    # Dynamic saddr check: saddr < 124 means valid SGPR, otherwise use VGPR pair for address
-    use_saddr = (saddr_reg < _c(124)) if saddr_reg is not None else UOp.const(dtypes.bool, False)
+    use_saddr = (saddr_reg < _c(124)) if saddr_reg is not None else UOp.const(False)
    if is_scratch:
      scratch_stride = ctx.rsgpr_dyn(_c(SCRATCH_STRIDE_IDX)).cast(dtypes.uint64)
      base = lane.cast(dtypes.uint64) * scratch_stride
      # SVE (Scratch VGPR Enable): when SVE=1, VADDR is used as offset; when SVE=0, VADDR is ignored
      sve = getattr(inst, 'sve', 0)
      vaddr = ctx.rvgpr_dyn(addr_reg, lane).cast(dtypes.uint64)
-      addr_offset = vaddr if sve == 1 else UOp.const(dtypes.uint64, 0)
+      addr_offset = vaddr if sve == 1 else UOp.const(0, dtypes.uint64)
      # Add saddr value only if use_saddr is true (saddr < 124)
-      saddr_contrib = use_saddr.where(ctx.rsgpr_dyn(saddr_reg).cast(dtypes.uint64), UOp.const(dtypes.uint64, 0)) \
-        if saddr_reg is not None else UOp.const(dtypes.uint64, 0)
+      saddr_contrib = use_saddr.where(ctx.rsgpr_dyn(saddr_reg).cast(dtypes.uint64), UOp.const(0, dtypes.uint64)) \
+        if saddr_reg is not None else UOp.const(0, dtypes.uint64)
      return base + addr_offset + saddr_contrib + offset64
    # FLAT/GLOBAL: choose between SGPR base (saddr) or VGPR pair (addr) based on saddr validity
-    saddr_base = _u64(ctx.rsgpr_dyn(saddr_reg), ctx.rsgpr_dyn(saddr_reg + _c(1))) if saddr_reg is not None else UOp.const(dtypes.uint64, 0)
+    saddr_base = _u64(ctx.rsgpr_dyn(saddr_reg), ctx.rsgpr_dyn(saddr_reg + _c(1))) if saddr_reg is not None else UOp.const(0, dtypes.uint64)
    vaddr_base = _u64(ctx.rvgpr_dyn(addr_reg, lane), ctx.rvgpr_dyn(addr_reg + _c(1), lane))
    # When saddr is valid: base = saddr pair, vaddr is 32-bit offset; otherwise: base = 0, vaddr is 64-bit address
    base_addr = use_saddr.where(saddr_base + ctx.rvgpr_dyn(addr_reg, lane).cast(dtypes.uint64), vaddr_base)
@@ -1874,18 +1874,18 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
        data = {'DATA': ctx.rvgpr_dyn(vdata_reg, lane), 'DATA1': ctx.rvgpr_dyn(vdata_reg + _c(1), lane),
                'DATA2': ctx.rvgpr_dyn(vdata_reg + _c(2), lane)}
      elif data_bits_mem <= 32:
-        data = {'DATA': ctx.rvgpr_dyn(vdata_reg, lane), 'DATA2': ctx.rvgpr_dyn(data1_reg, lane) if has_data1 else UOp.const(dtypes.uint32, 0)}
+        data = {'DATA': ctx.rvgpr_dyn(vdata_reg, lane), 'DATA2': ctx.rvgpr_dyn(data1_reg, lane) if has_data1 else UOp.const(0, dtypes.uint32)}
      else:
        data = {'DATA': _u64(ctx.rvgpr_dyn(vdata_reg, lane), ctx.rvgpr_dyn(vdata_reg + _c(1), lane)),
-                'DATA2': _u64(ctx.rvgpr_dyn(data1_reg, lane), ctx.rvgpr_dyn(data1_reg + _c(1), lane)) if has_data1 else UOp.const(dtypes.uint64, 0)}
+                'DATA2': _u64(ctx.rvgpr_dyn(data1_reg, lane), ctx.rvgpr_dyn(data1_reg + _c(1), lane)) if has_data1 else UOp.const(0, dtypes.uint64)}
      # RDNA3 uses ADDR/OFFSET, RDNA4 uses vgpr_a/offset (lowercase) + CalcDsAddr function
      return {'ADDR': addr, 'ADDR_BASE': addr, 'OFFSET': offset, 'OFFSET0': offset0, 'OFFSET1': offset1, '_lds': mem, 'laneId': lane,
              'vgpr_a': ctx.rvgpr_dyn(addr_reg, lane), 'offset': offset, 'offset0': offset0, 'offset1': offset1, **data}
    active = _lane_active(exec_mask, lane)
    # saddr < 124 means valid SGPR pair, otherwise use 0 (NULL means no saddr contribution)
-    use_saddr = (saddr_reg < _c(124)) if saddr_reg is not None else UOp.const(dtypes.bool, False)
-    saddr_raw = _u64(ctx.rsgpr_dyn(saddr_reg), ctx.rsgpr_dyn(saddr_reg + _c(1))) if saddr_reg is not None else UOp.const(dtypes.uint64, 0)
-    saddr_base = use_saddr.where(saddr_raw, UOp.const(dtypes.uint64, 0))
+    use_saddr = (saddr_reg < _c(124)) if saddr_reg is not None else UOp.const(False)
+    saddr_raw = _u64(ctx.rsgpr_dyn(saddr_reg), ctx.rsgpr_dyn(saddr_reg + _c(1))) if saddr_reg is not None else UOp.const(0, dtypes.uint64)
+    saddr_base = use_saddr.where(saddr_raw, UOp.const(0, dtypes.uint64))
    # Sign-extend offset to 64-bit for the final address calculation
    ioffset64 = offset.cast(dtypes.int64).cast(dtypes.uint64)
    # v_addr for CalcGlobalAddr: when saddr valid, use low 32 bits as offset; otherwise full 64-bit address. Include ioffset.
@@ -1900,13 +1900,13 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
    # acc bit: read/write ACCVGPR instead of VGPR for data operands
    _rvdata = (lambda r, l, *a: ctx.raccvgpr_dyn(r, l)) if use_acc else ctx.rvgpr_dyn
    vdata = _rvdata(vdata_reg, lane).cast(dtypes.uint64) if 'STORE' in op_name \
-      else _rvdata(vdst_reg, lane) if 'D16' in op_name else UOp.const(dtypes.uint32, 0)
+      else _rvdata(vdst_reg, lane) if 'D16' in op_name else UOp.const(0, dtypes.uint32)
    if 'STORE' in op_name and data_bits_mem >= 64:
-      vdata = vdata | (_rvdata(vdata_reg + _c(1), lane).cast(dtypes.uint64) << UOp.const(dtypes.uint64, 32))
+      vdata = vdata | (_rvdata(vdata_reg + _c(1), lane).cast(dtypes.uint64) << UOp.const(32, dtypes.uint64))
    srcs = {'ADDR': addr, 'VDATA': vdata, '_vmem': mem, '_active': active,
            'laneId': lane, 'v_addr': vaddr_base, 's_saddr': saddr_base, 'SADDR': saddr_base, 'OFFSET': offset}
    for i in range(data_bits_mem // 32):
-      srcs[f'VDATA{i}'] = _rvdata(vdata_reg + _c(i), lane) if 'STORE' in op_name else UOp.const(dtypes.uint32, 0)
+      srcs[f'VDATA{i}'] = _rvdata(vdata_reg + _c(i), lane) if 'STORE' in op_name else UOp.const(0, dtypes.uint32)
    return srcs

  def make_stores(dest: str, val: UOp, lane: UOp, active: UOp, writes_return_data: bool) -> list[UOp]:
@@ -1983,7 +1983,7 @@ def _compile_mubuf(inst: irc.MUBUF, ctx: _Ctx) -> UOp:
  offset, offen, idxen = ctx.inst_field(type(inst).offset), ctx.inst_field(type(inst).offen), ctx.inst_field(type(inst).idxen)

  # V# descriptor: base[0:1], num_records[2], stride=word3[13:0]
-  base = _u64(ctx.rsgpr_dyn(srsrc), ctx.rsgpr_dyn(srsrc + _c(1))) & UOp.const(dtypes.uint64, 0xFFFFFFFFFFFF)
+  base = _u64(ctx.rsgpr_dyn(srsrc), ctx.rsgpr_dyn(srsrc + _c(1))) & UOp.const(0xFFFFFFFFFFFF, dtypes.uint64)
  num_records = ctx.rsgpr_dyn(srsrc + _c(2))
  stride = (ctx.rsgpr_dyn(srsrc + _c(3)) & _c(0x3FFF)).cast(dtypes.uint64)

@@ -2000,7 +2000,7 @@ def _compile_mubuf(inst: irc.MUBUF, ctx: _Ctx) -> UOp:
  buffer_offset = (stride * index + voff + offset.cast(dtypes.uint64)).cast(dtypes.uint32)
  in_bounds = active & buffer_offset.__lt__(num_records)
  addr = base + soff + buffer_offset.cast(dtypes.uint64)
-  addr = in_bounds.where(addr, UOp.const(dtypes.uint64, 0))  # safe address when OOB
+  addr = in_bounds.where(addr, UOp.const(0, dtypes.uint64))  # safe address when OOB
  mem = ctx.vmem

  stores: list[UOp] = []
@@ -2009,20 +2009,20 @@ def _compile_mubuf(inst: irc.MUBUF, ctx: _Ctx) -> UOp:
    lds_base = ctx.rsgpr_dyn(_c(124)) & _c(0x3FFFF)
    lds_addr = lds_base + lane.cast(dtypes.uint32) * _c(n_dwords * 4)
    for i in range(n_dwords):
-      word_addr = (addr + UOp.const(dtypes.uint64, i * 4)) >> UOp.const(dtypes.uint64, 2)
+      word_addr = (addr + UOp.const(i * 4, dtypes.uint64)) >> UOp.const(2, dtypes.uint64)
      val = in_bounds.where(mem.index(word_addr.cast(dtypes.int64)).load(), _c(0))
      lds_idx = (lds_addr + _c(i * 4)) >> _c(2)
      lds_slot = ctx.lds.index(lds_idx.valid(active))
      stores.append(lds_slot.store(active.where(val, lds_slot)))
  elif is_store:
    for i in range(n_dwords):
-      word_addr = (addr + UOp.const(dtypes.uint64, i * 4)) >> UOp.const(dtypes.uint64, 2)
+      word_addr = (addr + UOp.const(i * 4, dtypes.uint64)) >> UOp.const(2, dtypes.uint64)
      idx = mem.index(word_addr.cast(dtypes.int64).valid(in_bounds))
      val = (ctx.raccvgpr_dyn if use_acc else ctx.rvgpr_dyn)(vdata + _c(i), lane)
      stores.append(idx.store(in_bounds.where(_to_u32(val), idx)))
  else:
    for i in range(n_dwords):
-      word_addr = (addr + UOp.const(dtypes.uint64, i * 4)) >> UOp.const(dtypes.uint64, 2)
+      word_addr = (addr + UOp.const(i * 4, dtypes.uint64)) >> UOp.const(2, dtypes.uint64)
      val = in_bounds.where(mem.index(word_addr.cast(dtypes.int64).valid(in_bounds)).load(), _c(0))
      stores.append((ctx.waccvgpr_dyn if use_acc else ctx.wvgpr_dyn)(vdata + _c(i), lane, val, exec_mask))
  return UOp.sink(UOp.group(*stores).end(lane), *ctx.inc_pc())
@@ -7,7 +7,7 @@ from tinygrad.codegen.decomp.dtype import f2f
 # Type alias for vars dict: stores UOps and tuples for lambda definitions
 VarVal = UOp | tuple[str, list[str], str]

-def _const(dt, v): return UOp.const(dt, v)
+def _const(dt, v): return UOp.const(v, dt)
 def _u32(v): return _const(dtypes.uint32, v)
 def _u64(v): return _const(dtypes.uint64, v)
 def _to_u32(v): return v if v.dtype == dtypes.uint32 else v.bitcast(dtypes.uint32) if v.dtype.itemsize == 4 else v.cast(dtypes.uint32)
@@ -55,8 +55,8 @@ def _expr_bits(v: UOp) -> int:
  if v.op in (Ops.AND, Ops.XOR):
    widths: list[int] = []
    for src in v.src:
-      if src.op == Ops.CONST and isinstance(src.arg, int) and src.arg > 0 and (src.arg & (src.arg + 1)) == 0:
-        widths.append(src.arg.bit_length())
+      if src.op == Ops.CONST and isinstance(src.val, int) and src.val > 0 and (src.val & (src.val + 1)) == 0:
+        widths.append(src.val.bit_length())
    if widths: return max(widths)
  return v.dtype.bitsize

@@ -144,9 +144,9 @@ def _minmax_reduce(is_max: bool, dt, *args: UOp) -> UOp:
 def _find_two_pi_mul(x):
  if x.op != Ops.MUL or len(x.src) != 2: return None
  for i, s in enumerate(x.src):
-    if s.op == Ops.CONST and abs(s.arg - 6.283185307179586) < 1e-5: return (x.src[1-i], 6.283185307179586)
+    if s.op == Ops.CONST and abs(s.val - 6.283185307179586) < 1e-5: return (x.src[1-i], 6.283185307179586)
    if s.op == Ops.MUL and len(s.src) == 2:
-      vals = [ss.arg for ss in s.src if ss.op == Ops.CONST] + [ss.src[0].arg for ss in s.src if ss.op == Ops.CAST and ss.src[0].op == Ops.CONST]
+      vals = [ss.val for ss in s.src if ss.op == Ops.CONST] + [ss.src[0].val for ss in s.src if ss.op == Ops.CAST and ss.src[0].op == Ops.CONST]
      if len(vals) == 2 and abs(vals[0] * vals[1] - 6.283185307179586) < 1e-5: return (x.src[1-i], vals[0] * vals[1])
  return None

@@ -163,7 +163,7 @@ def _trig_reduce(x, phase=0.0):

 def _signext(val: UOp) -> UOp:
  for bits, mask, ext in [(4, 0xF, 0xFFFFFFF0), (8, 0xFF, 0xFFFFFF00), (16, 0xFFFF, 0xFFFF0000)]:
-    if (val.op == Ops.AND and len(val.src) == 2 and val.src[1].op == Ops.CONST and val.src[1].arg == mask) or val.dtype.itemsize == bits // 8:
+    if (val.op == Ops.AND and len(val.src) == 2 and val.src[1].op == Ops.CONST and val.src[1].val == mask) or val.dtype.itemsize == bits // 8:
      v32 = val.cast(dtypes.uint32) if val.dtype != dtypes.uint32 else val
      sb = (v32 >> _u32(bits - 1)) & _u32(1)
      return sb.ne(_u32(0)).where(v32 | _u32(ext), v32).cast(dtypes.int)
@@ -497,7 +497,7 @@ class Parser:
        if not dtypes.is_int(right.dtype): right = right.cast(dtypes.uint32)
        return (left >> right) if op == '>>' else (left << right)
      case '+' | '-':
-        if op == '-' and left.op == Ops.CONST and right.op == Ops.CONST: return _const(left.dtype, left.arg - right.arg)
+        if op == '-' and left.op == Ops.CONST and right.op == Ops.CONST: return _const(left.dtype, left.val - right.val)
        return (left + right) if op == '+' else (left - right)
      case '*' | '/':
        # Integer promotion: promote 16-bit integers to 32-bit before multiply to avoid overflow
@@ -507,7 +507,7 @@ class Parser:
          left, right = left.cast(pdt), right.cast(pdt)
        if op == '*': return left * right
        return (left // right) if dtypes.is_int(left.dtype) else (left / right)
-      case '**': return UOp(Ops.EXP2, src=(right.cast(left.dtype),)) if left.op == Ops.CONST and left.arg == 2.0 else left
+      case '**': return UOp(Ops.EXP2, src=(right.cast(left.dtype),)) if left.op == Ops.CONST and left.val == 2.0 else left

  _PREC = [('||',), ('&&',), ('|',), ('^',), ('&',), ('==', '!=', '<>'), ('>=', '<=', '>', '<'), ('>>', '<<'), ('+', '-'), ('*', '/'), ('**',)]

@@ -530,7 +530,7 @@ class Parser:
    if self.try_eat_val('-', 'OP'):
      inner = self.unary()
      if inner.op == Ops.CONST:
-        return _const(dtypes.int if inner.dtype == dtypes.uint32 else inner.dtype, -inner.arg)
+        return _const(dtypes.int if inner.dtype == dtypes.uint32 else inner.dtype, -inner.val)
      return inner.neg()
    if self.try_eat_val('+', 'OP'): return self.unary()
    return self.postfix()
@@ -670,14 +670,14 @@ class Parser:
      width = self.parse()
      self.eat('RBRACKET')
      if width.op == Ops.CONST:
-        w = int(width.arg)
+        w = int(width.val)
        return (base >> _to_u32(first)) & _const(base.dtype, (1 << w) - 1)
      return base
    if self.try_eat('COLON'):
      second = self.parse()
      self.eat('RBRACKET')
      if first.op == Ops.CONST and second.op == Ops.CONST:
-        a, b = int(first.arg), int(second.arg)
+        a, b = int(first.val), int(second.val)
        if a < b: return _bitreverse(base, b - a + 1)
        hi, lo = a, b
        if lo >= base.dtype.itemsize * 8:
@@ -699,7 +699,7 @@ class Parser:
    if var_name is None:
      var_name = self._find_var_name(base)
    if first.op == Ops.CONST:
-      idx = int(first.arg)
+      idx = int(first.val)
      # Check for array element (var@idx)
      if var_name and f'{var_name}@{idx}' in self.vars:
        v = self.vars[f'{var_name}@{idx}']
@@ -866,13 +866,13 @@ class Parser:
        idx_hi_native = ((addr + _const(adt, 4)) >> _const(adt, 2)).cast(dtypes.int64)
        safe_idx_hi = is_unaligned.where(idx_hi_native, idx_native)
        hi = mindex(safe_idx_hi)
-        combined = val.cast(dtypes.uint64) | (hi.cast(dtypes.uint64) << UOp.const(dtypes.uint64, 32))
-        val = is_unaligned.where((combined >> (byte_off.cast(dtypes.uint64) * UOp.const(dtypes.uint64, 8))).cast(dtypes.uint32), val)
+        combined = val.cast(dtypes.uint64) | (hi.cast(dtypes.uint64) << UOp.const(32, dtypes.uint64))
+        val = is_unaligned.where((combined >> (byte_off.cast(dtypes.uint64) * UOp.const(8, dtypes.uint64))).cast(dtypes.uint32), val)
    return _cast_to(val, dt)

  def _coerce_cmp(self, l: UOp, r: UOp) -> tuple[UOp, UOp]:
    if l.dtype != r.dtype:
-      if r.dtype == dtypes.int and r.op == Ops.CONST and r.arg < 0: l = l.cast(dtypes.int)
+      if r.dtype == dtypes.int and r.op == Ops.CONST and r.val < 0: l = l.cast(dtypes.int)
      else: r = r.cast(l.dtype)
    return l, r

@@ -970,7 +970,7 @@ def parse_block(lines: list[str], start: int, env: dict[str, VarVal], funcs: dic
        if p.at('NUM'): return int(p.eat('NUM').val.rstrip('UuLl'))
        expr = p.parse().simplify()
        assert expr.op == Ops.CONST, f"loop bound must be constant, got {expr}"
-        return int(expr.arg)
+        return int(expr.val)
      start_val = parse_bound()
      p.eat('COLON')
      end_val = parse_bound()
@@ -1258,7 +1258,7 @@ def parse_block(lines: list[str], start: int, env: dict[str, VarVal], funcs: dic
      def parse_cond(s, kw):
        ll = s.lower()
        return _to_bool(parse_expr(s[ll.find(kw) + len(kw):ll.rfind('then')].strip(), env, funcs))
-      def is_const(c, v): return c.op == Ops.CONST and c.arg is v
+      def is_const(c, v): return c.op == Ops.CONST and c.val is v
      cond = parse_cond(line, 'if')
      conditions: list[tuple[UOp, UOp | dict[str, VarVal] | None]] = [(cond, None)] if not is_const(cond, False) else []
      branch_assigns: list[tuple[UOp, list]] = []  # (cond, assigns_list) for side-effect merging
@@ -1339,7 +1339,7 @@ def parse_block(lines: list[str], start: int, env: dict[str, VarVal], funcs: dic
              return (dest, (val[0], cnd.where(val[1], val[1])))
            return (dest, val)
          # Build combined condition: each branch fires when its cond is true AND no earlier cond was true
-          remaining = UOp.const(dtypes.bool, True)
+          remaining = UOp.const(True)
          for bc, bse in branch_assigns:
            effective = remaining & bc if remaining.op != Ops.CONST else bc
            for dest, val in bse: assigns.append(_cond_side_effect(effective, dest, val))
@@ -91,7 +91,7 @@ class GPFIFO:
    args_cnt, vals_cnt = const0[80], const0[81]
    args_addr = qmd.constant_buffer_addr_lower_0 + (qmd.constant_buffer_addr_upper_0 << 32) + 0x160
    args = to_mv(args_addr, args_cnt*8).cast('Q')
-    vals = to_mv(args_addr + args_cnt*8, vals_cnt*4).cast('I')
+    vals = to_mv(args_addr + args_cnt*8, vals_cnt*8).cast('Q')
    cargs = [ctypes.cast(args[i], ctypes.c_void_p) for i in range(args_cnt)] + [ctypes.cast(vals[i], ctypes.c_void_p) for i in range(vals_cnt)]
    gx, gy, gz = qmd.cta_raster_width, qmd.cta_raster_height, qmd.cta_raster_depth
    lx, ly, lz = qmd.cta_thread_dimension0, qmd.cta_thread_dimension1, qmd.cta_thread_dimension2
@@ -1,7 +1,7 @@
 import unittest
 from tinygrad import Tensor, dtypes, TinyJit, UOp
 from tinygrad.llm.model import apply_rope as apply_rope_new, precompute_freqs_cis
-from test.helpers import assert_jit_cache_len
+from test.helpers import assert_jit_cache_len, check_schedule

 def apply_rope(x:Tensor, start_pos:int):
  B, H, T, Hd = x.shape
@@ -16,9 +16,8 @@ class TestAttention(unittest.TestCase):
    k = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
    v = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
    attn = q.scaled_dot_product_attention(k, v)
-    sched = attn.schedule_linear()
    # attention has 4 kernels now
-    self.assertEqual(len(sched.src), 4)
+    check_schedule(attn, 4)

  def test_apply_rope_jit_prune(self):
    def rope_fn(x_in, pos): return apply_rope(x_in, pos)
@@ -1,6 +1,6 @@
 import unittest, itertools, math
 from tinygrad import Tensor, dtypes, Context
-from tinygrad.dtype import DType, ConstType, Invalid
+from tinygrad.dtype import DType, ConstType
 from tinygrad.uop.ops import Ops, UOp
 from test.helpers import full_rewrite
 import numpy as np
@@ -36,21 +36,20 @@ class TestUnaryOpsConstFolding(unittest.TestCase):

 class TestWeakConstFolding(unittest.TestCase):
  def test_weakint_math(self):
-    out = (UOp.const(dtypes.weakint, 2**40) + UOp.const(dtypes.weakint, 2**40)).simplify()
-    self.assertEqual((out.op, out.dtype, out.arg), (Ops.CONST, dtypes.weakint, 2**41))
+    out = (UOp.const(2**40) + UOp.const(2**40)).simplify()
+    self.assertEqual((out.op, out.dtype, out.val), (Ops.CONST, dtypes.weakint, 2**41))

  def test_float_unaries(self):
-    for dtype in (dtypes.weakfloat,):
-      for op in (Ops.SIN, Ops.LOG2, Ops.EXP2, Ops.SQRT, Ops.RECIPROCAL):
-        out = UOp.const(dtype, 4).alu(op).simplify()
-        self.assertEqual((out.op, out.dtype), (Ops.CONST, dtypes.weakfloat))
+    for op in (Ops.SIN, Ops.LOG2, Ops.EXP2, Ops.SQRT, Ops.RECIPROCAL):
+      out = UOp.const(4.0).alu(op).simplify()
+      self.assertEqual((out.op, out.dtype), (Ops.CONST, dtypes.weakfloat))

  def test_weakfloat_math(self):
-    out = (UOp.const(dtypes.weakfloat, 1.25) + UOp.const(dtypes.weakfloat, 2.5)).simplify()
-    self.assertEqual((out.op, out.dtype, out.arg), (Ops.CONST, dtypes.weakfloat, 3.75))
+    out = (UOp.const(1.25) + UOp.const(2.5)).simplify()
+    self.assertEqual((out.op, out.dtype, out.val), (Ops.CONST, dtypes.weakfloat, 3.75))

  def test_invalid_poison(self):
-    self.assertIs(UOp.invalid().alu(Ops.CDIV, UOp.const(dtypes.weakint, 0)).simplify().arg, Invalid)
+    self.assertTrue(UOp.invalid().alu(Ops.CDIV, UOp.const(0)).simplify().is_invalid)

 class TestBinaryOpsConstFolding(unittest.TestCase):
  def test_add_literal_zero(self):
@@ -121,10 +120,10 @@ class TestBitcastConstFolding(unittest.TestCase):
    def t(cases: dict[DType, ConstType]):
      for (from_dt, from_v), (to_dt, to_v) in itertools.product(cases.items(), cases.items()):
        if not math.isnan(from_v):
-          r = full_rewrite(UOp.const(from_dt, from_v).bitcast(to_dt).sink()).src[0]
+          r = full_rewrite(UOp.const(from_v, from_dt).bitcast(to_dt).sink()).src[0]
          self.assertEqual(r.op, Ops.CONST, msg:=f"{from_dt} -> {to_dt} ({from_v} -> {to_v})")
          self.assertEqual(r.dtype, to_dt, msg)
-          np.testing.assert_equal(r.arg, to_v, msg)
+          np.testing.assert_equal(r.val, to_v, msg)

    t({dtypes.int8: 0, dtypes.uint8: 0, dtypes.bool: False})
    t({dtypes.int8: 1, dtypes.uint8: 1, dtypes.bool: True})
@@ -145,9 +144,9 @@ class TestBitcastConstFolding(unittest.TestCase):

  def test_vec_bitcast(self):
    with Context(SPEC=0):
-      srcs = full_rewrite(UOp.const(dtypes.int32, (-1, -2**31, 75)).bitcast(dtypes.uint32).sink()).src
+      srcs = full_rewrite(UOp.const((-1, -2**31, 75), dtypes.int32).bitcast(dtypes.uint32).sink()).src
    self.assertTrue(all(r.op is Ops.CONST and r.dtype == dtypes.uint32 for r in srcs))
-    self.assertEqual(tuple(x.arg for x in srcs), (2**32-1, 2**31, 75))
+    self.assertEqual(tuple(x.val for x in srcs), (2**32-1, 2**31, 75))

 # folds advance indexing into basic indexing
 class TestIndexingConstFolding(unittest.TestCase):
@@ -12,7 +12,7 @@ class TestGroupedDims(unittest.TestCase):
    idxs = get_grouped_dims(prefix, dims, max_sizes, reverse)
    loop_idxs = dedup(flatten([[y for y in x.toposort() if y.op is Ops.SPECIAL] for x in idxs]))
    loop_idxs = sorted(loop_idxs, key=lambda uop: uop.arg)
-    sizes = [x.src[0].arg for x in loop_idxs]
+    sizes = [x.src[0].val for x in loop_idxs]
    assert len(idxs) == len(dims), f"expected idxs to have same length as dims {len(dims)}, got {len(idxs)}"
    if assert_same_length:
      assert len(loop_idxs) == min(len(sizes), len(dims)), f"expected idxs to have length {min(len(sizes), len(dims))}, got {len(loop_idxs)}"
@@ -24,7 +24,7 @@ class TestGroupedDims(unittest.TestCase):
    total = math.prod(dims)
    specials = sorted(dedup(flatten([[y for y in x.toposort() if y.op is Ops.SPECIAL] for x in idxs])), key=lambda u: u.arg)
    # build flat index and primed flat (same expression with renamed SPECIALs)
-    flat = UOp.const(dtypes.weakint, 0)
+    flat = UOp.const(0)
    for i, idx in enumerate(idxs):
      flat = flat + idx * int(math.prod(dims[i+1:]))
    flat_p = flat.substitute({s: UOp(Ops.SPECIAL, src=s.src, arg=s.arg+"_p") for s in specials})
@@ -107,7 +107,7 @@ class TestGroupedDims(unittest.TestCase):

  def test_global_prod_max(self):
    g, l = UOp.range(256, 0, AxisType.GLOBAL), UOp.range(256, 1, AxisType.LOCAL)
-    sink = UOp.param(0, dtypes.float, (512,)).index(g + l).store(UOp.const(dtypes.float, 1.0)).end(g, l).sink(arg=KernelInfo())
+    sink = UOp.param(0, dtypes.float, (512,)).index(g + l).store(UOp.const(1.0)).end(g, l).sink(arg=KernelInfo())
    class R(Renderer): global_max, local_max, global_prod_max = (256, 256, 256), (128, 128, 128), (128, 128, 128)
    specials = [u for u in add_gpudims(R(Target()), sink).toposort() if u.op is Ops.SPECIAL]
    self.assertGreater(len([s for s in specials if "lidx" in s.arg]), 1)
@@ -14,10 +14,10 @@ class TestGradient(unittest.TestCase):
  def _test_one_input_function(self, f:Callable, jf:Callable|None=None):
    if jf is None: jf = f
    x = UOp.variable('x', -math.inf, math.inf, dtype=dtypes.float)
-    gx = compute_gradient(f(x), UOp.const(dtypes.float, 1.0), set([x]))[x]
+    gx = compute_gradient(f(x), UOp.const(1.0), set([x]))[x]

    for val in [-5., -2.0, 0.0, 2.0, 5.]:
-      tg_out = gx.substitute({x: x.const_like(val)}).ssimplify()
+      tg_out = gx.substitute({x: UOp.const(val)}).ssimplify()
      tx = torch.tensor([val], dtype=torch.float, requires_grad=True)
      torch_out = torch.autograd.grad(jf(tx), tx)[0].item()
      self._cmp_nan_okay(tg_out, torch_out)
@@ -26,13 +26,13 @@ class TestGradient(unittest.TestCase):
    if jf is None: jf = f
    x = UOp.variable('x', -math.inf, math.inf, dtype=dtypes.float)
    y = UOp.variable('y', -math.inf, math.inf, dtype=dtypes.float)
-    grads = compute_gradient(f(x, y), UOp.const(dtypes.float, 1.0), set([x, y]))
+    grads = compute_gradient(f(x, y), UOp.const(1.0), set([x, y]))
    gx, gy = grads[x], grads[y]

    for valx in [-5., -2.0, 0.0, 2.0, 5.]:
      for valy in [-5., -2.0, 0.0, 2.0, 5.]:
        # Substitute the values into the gradient expressions
-        substitutions = {x: x.const_like(valx), y: y.const_like(valy)}
+        substitutions = {x: UOp.const(valx), y: UOp.const(valy)}
        tg_out_x = gx.substitute(substitutions).ssimplify()
        tg_out_y = gy.substitute(substitutions).ssimplify()

@@ -15,13 +15,13 @@ def apply_rewrite(expr):
 def apply_rewrite_values(expr):
  srcs = full_rewrite(expr.sink()).src
  if len(srcs) == 1:
-    if srcs[0].op is Ops.CONST: return (srcs[0].arg,) if not isinstance(srcs[0].arg, tuple) else srcs[0].arg
-    if srcs[0].op is Ops.STACK: return tuple(s.arg for s in srcs[0].src)
-  return tuple(s.arg for s in srcs)
+    if srcs[0].op is Ops.CONST: return (srcs[0].val,)
+    if srcs[0].op is Ops.STACK: return tuple(s.val for s in srcs[0].src)
+  return tuple(s.val for s in srcs)

 def evaluate_uop(uop, variables):
  if uop.op == Ops.CONST:
-    return uop.arg
+    return uop.val
  elif uop.op == Ops.PARAM and uop.arg.addrspace is AddrSpace.ALU:
    return variables[uop.expr]
  elif uop.op in GroupOp.ALU:
@@ -32,67 +32,67 @@ def evaluate_uop(uop, variables):

 class TestArithmeticSimplifications(unittest.TestCase):
  def test_full_graph_rewrite_division_by_zero(self):
-    optimized_div_uop = apply_rewrite(UOp.const(dtypes.float32, 10.0) / UOp.const(dtypes.float32, 0.0))
+    optimized_div_uop = apply_rewrite(UOp.const(10.0) / UOp.const(0.0))
    self.assertEqual(optimized_div_uop.op, Ops.CONST)
-    self.assertTrue(math.isinf(optimized_div_uop.arg) or math.isnan(optimized_div_uop.arg))
+    self.assertTrue(math.isinf(optimized_div_uop.val) or math.isnan(optimized_div_uop.val))

  def test_full_graph_rewrite_redundant_operations(self):
-    optimized_uop = apply_rewrite((UOp.const(dtypes.float32, 10.0) + UOp.const(dtypes.float32, 0.0)) * UOp.const(dtypes.float32, 1.0))
+    optimized_uop = apply_rewrite((UOp.const(10.0) + UOp.const(0.0)) * UOp.const(1.0))
    self.assertEqual(optimized_uop.op, Ops.CONST)
-    self.assertEqual(optimized_uop.arg, 10.0)
+    self.assertEqual(optimized_uop.val, 10.0)

  def test_full_graph_rewrite_large_graph(self):
-    prev_uop = UOp.const(dtypes.int32, 0)
+    prev_uop = UOp.const(0)
    for i in range(1, 101):
-      prev_uop += UOp.const(dtypes.int32, i)
+      prev_uop += UOp.const(i)
    optimized_uop = apply_rewrite(prev_uop)
    self.assertEqual(optimized_uop.op, Ops.CONST)
-    self.assertEqual(optimized_uop.arg, sum(range(1, 101)))
+    self.assertEqual(optimized_uop.val, sum(range(1, 101)))

  def test_full_graph_rewrite_division_by_one(self):
-    optimized_uop = apply_rewrite(UOp.const(dtypes.float32, 42.0) / UOp.const(dtypes.float32, 1.0))
+    optimized_uop = apply_rewrite(UOp.const(42.0) / UOp.const(1.0))
    self.assertEqual(optimized_uop.op, Ops.CONST)
-    self.assertEqual(optimized_uop.arg, 42.0)
+    self.assertEqual(optimized_uop.val, 42.0)

  def test_full_graph_rewrite_modulo_by_one(self):
-    optimized_uop = apply_rewrite(UOp.const(dtypes.int32, 42) % UOp.const(dtypes.int32, 1))
+    optimized_uop = apply_rewrite(UOp.const(42) % UOp.const(1))
    self.assertEqual(optimized_uop.op, Ops.CONST)
-    self.assertEqual(optimized_uop.arg, 0)
+    self.assertEqual(optimized_uop.val, 0)


 class TestFoldingAndReduction(unittest.TestCase):
  @unittest.skip("reduce is removed now")
  def test_full_graph_rewrite_constant_reduction_folding(self):
-    const1 = UOp.const(dtypes.int32, 5)
-    const2 = UOp.const(dtypes.int32, 10)
-    const3 = UOp.const(dtypes.int32, 20)
+    const1 = UOp.const(5)
+    const2 = UOp.const(10)
+    const3 = UOp.const(20)
    optimized_sink = apply_rewrite((const1 + const2 + const3).reduce(Ops.ADD))
    expected_sum = 5 + 10 + 20
-    self.assertEqual(optimized_sink.arg, expected_sum)
+    self.assertEqual(optimized_sink.val, expected_sum)

  @unittest.skip("reduce is removed now")
  def test_full_graph_rewrite_reduction_with_unused_range(self):
-    const1 = UOp.const(dtypes.int32, 15)
-    const2 = UOp.const(dtypes.int32, 25)
+    const1 = UOp.const(15)
+    const2 = UOp.const(25)
    rng = UOp.range(10, idx=0)
    optimized_sink = apply_rewrite((const1 + const2).reduce(Ops.ADD, rng))
    expected_sum = 10 * (15 + 25)
-    self.assertEqual(optimized_sink.arg, expected_sum)
+    self.assertEqual(optimized_sink.val, expected_sum)

  @unittest.skip("currently failing")
  def test_full_graph_rewrite_range_reduction(self):
    simple_range = UOp.range(5, idx=0)
    optimized_sink = apply_rewrite(simple_range.reduce(Ops.ADD, simple_range))
    expected_sum = sum(range(5))
-    self.assertEqual(optimized_sink.arg, expected_sum)
+    self.assertEqual(optimized_sink.val, expected_sum)

  @unittest.skip("currently failing")
  def test_full_graph_rewrite_simple_reduction_folding(self):
    simple_range = UOp.range(4, idx=0)
-    add_uop = simple_range + UOp.const(dtypes.int32, 1)
+    add_uop = simple_range + UOp.const(1)
    optimized_sink = apply_rewrite(add_uop.reduce(Ops.ADD, simple_range))
    expected_sum = sum(i + 1 for i in range(4))
-    self.assertEqual(optimized_sink.arg, expected_sum)
+    self.assertEqual(optimized_sink.val, expected_sum)

  @unittest.skip("currently failing")
  def test_full_graph_rewrite_nested_loop_collapse(self):
@@ -101,7 +101,7 @@ class TestFoldingAndReduction(unittest.TestCase):
    expr = (outer_range * 10) + inner_range
    optimized_reduce_uop = apply_rewrite(expr.reduce(Ops.ADD, outer_range, inner_range))
    self.assertEqual(optimized_reduce_uop.op, Ops.CONST)
-    self.assertEqual(optimized_reduce_uop.arg, sum((i * 10) + j for i in range(8) for j in range(4)))
+    self.assertEqual(optimized_reduce_uop.val, sum((i * 10) + j for i in range(8) for j in range(4)))


 class TestModuloAndDivisionFolding(unittest.TestCase):
@@ -110,27 +110,27 @@ class TestModuloAndDivisionFolding(unittest.TestCase):
    x_var_uop = UOp.variable('x', 0, 100).cast(dtypes.weakint)
    optimized_mod_uop = apply_rewrite(((x_var_uop * 4) + 2) % 4)
    self.assertEqual(optimized_mod_uop.op, Ops.CONST)
-    self.assertEqual(optimized_mod_uop.arg, 2)
+    self.assertEqual(optimized_mod_uop.val, 2)

  def test_full_graph_rewrite_division_folding_with_define_var(self):
    # index dtype because div-mod rules only work on index
    n_var_uop = UOp.variable('n', 1, 1000).cast(dtypes.weakint)
    optimized_div_uop = apply_rewrite((n_var_uop * 6) // 3)
    self.assertEqual(optimized_div_uop.op, Ops.MUL)
-    self.assertEqual(optimized_div_uop.src[1].arg, 2)
+    self.assertEqual(optimized_div_uop.src[1].val, 2)

  def test_full_graph_rewrite_complex_mod_div_folding(self):
    # index dtype because div-mod rules only work on index
    k_var_uop = UOp.variable('k', 0, 50).cast(dtypes.weakint)
    optimized_div_uop = apply_rewrite(((k_var_uop * 12 + 8) % 6) // 2)
    self.assertEqual(optimized_div_uop.op, Ops.CONST)
-    self.assertEqual(optimized_div_uop.arg, 1)
+    self.assertEqual(optimized_div_uop.val, 1)

  def test_graph_rewrite_div_folding_bug(self):
    lhs = UOp(Ops.ADD, src=(
-      UOp(Ops.STACK, arg=None, src=(UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 32),), arg='lidx0'),)*4),
-      UOp.const(dtypes.int, (0, 256, 512, 768))))
-    rhs = UOp.const(dtypes.int, (2,)*4)
+      UOp(Ops.STACK, arg=None, src=(UOp(Ops.SPECIAL, src=(UOp.const(32),), arg='lidx0'),)*4),
+      UOp.const((0, 256, 512, 768))))
+    rhs = UOp.const((2,)*4)
    unopt = lhs<rhs
    opt = apply_rewrite(unopt)
    print(unopt)
@@ -159,11 +159,11 @@ class TestModuloAndDivisionFolding(unittest.TestCase):

 class TestEdgeCasesAndSpecialOperations(unittest.TestCase):
  def test_full_graph_rewrite_transcendental_edge_cases(self):
-    optimized_sink = full_rewrite(UOp.const(dtypes.float32, -1.0).log2().sink(UOp.const(dtypes.float32, 0.0).reciprocal()))
+    optimized_sink = full_rewrite(UOp.const(-1.0).log2().sink(UOp.const(0.0).reciprocal()))
    optimized_log2_neg, optimized_recip_zero = optimized_sink.src
-    self.assertTrue(math.isnan(optimized_log2_neg.arg), f"Expected NaN for log2(-1.0), got {optimized_log2_neg.arg}")
-    self.assertTrue(math.isinf(optimized_recip_zero.arg) and optimized_recip_zero.arg > 0,
-                    f"Expected +inf for reciprocal(0.0), got {optimized_recip_zero.arg}")
+    self.assertTrue(math.isnan(optimized_log2_neg.val), f"Expected NaN for log2(-1.0), got {optimized_log2_neg.val}")
+    self.assertTrue(math.isinf(optimized_recip_zero.val) and optimized_recip_zero.val > 0,
+                    f"Expected +inf for reciprocal(0.0), got {optimized_recip_zero.val}")

  @unittest.skip("broken")
  def test_full_graph_rewrite_modulo_negative_dividend(self):
@@ -182,27 +182,27 @@ class TestEdgeCasesAndSpecialOperations(unittest.TestCase):
 class TestGEPAndVectorizeRewrite(unittest.TestCase):
  def test_gep_single_element_extraction(self):
    # GEP on a vector dtype to extract a single element
-    base_vector = UOp.const(dtypes.float32, (1.0, 2.0, 3.0, 4.0))
-    self.assertEqual(apply_rewrite(base_vector.index(2)).arg, 3.0)
+    base_vector = UOp.const((1.0, 2.0, 3.0, 4.0))
+    self.assertEqual(apply_rewrite(base_vector.index(2)).val, 3.0)

  def test_gep_tuple_extraction(self):
    # GEP on a vector dtype to extract multiple elements as a vector
-    base_vector = UOp.const(dtypes.float32, (1.0, 2.0, 3.0, 4.0))
+    base_vector = UOp.const((1.0, 2.0, 3.0, 4.0))
    self.assertEqual(list(apply_rewrite_values(UOp.stack(*[base_vector.index(i) for i in (2, 3)]))), [3.0, 4.0])

  def test_gep_on_const_stack(self):
    # GEP on a const STACK to extract a single element
-    const_stack = UOp.const(dtypes.float32, (1.0, 2.0, 3.0, 4.0))
-    self.assertEqual(apply_rewrite(const_stack.index(2)).arg, 3.0)
+    const_stack = UOp.const((1.0, 2.0, 3.0, 4.0))
+    self.assertEqual(apply_rewrite(const_stack.index(2)).val, 3.0)

  def test_gep_tuple_on_const_stack(self):
    # GEP on a const STACK using a tuple to extract multiple elements
-    const_stack = UOp.const(dtypes.float32, (7.0, 8.0, 9.0, 10.0))
+    const_stack = UOp.const((7.0, 8.0, 9.0, 10.0))
    self.assertEqual(list(apply_rewrite_values(UOp.stack(*[const_stack.index(i) for i in (1, 3)]))), [8.0, 10.0])

  def test_vectorize_multiple_elements(self):
    # Vectorizing multiple elements using GEP
-    base_vector = UOp.const(dtypes.float32, (5.0, 10.0, 15.0, 20.0))
+    base_vector = UOp.const((5.0, 10.0, 15.0, 20.0))
    vectorized_uop = UOp(Ops.STACK, src=tuple(base_vector.index(i) for i in range(4)))
    self.assertEqual(list(apply_rewrite_values(vectorized_uop)), [5.0, 10.0, 15.0, 20.0])

@@ -213,7 +213,7 @@ from tinygrad.uop.symbolic import symbolic_simple

 class TestBottomUpRewrite(unittest.TestCase):
  def test_const_folding(self):
-    a = UOp.const(dtypes.int, 5)
+    a = UOp.const(5)
    ret = (a*3) + (a*7)
    gt = graph_rewrite(ret, symbolic_simple)
    ret = graph_rewrite(ret, symbolic_simple, bottom_up=True)
@@ -305,28 +305,28 @@ class TestRecurse(unittest.TestCase):
    graph_rewrite(a, pm, bottom_up=True)

  def test_inf_loop(self):
-    a = UOp.const(dtypes.int, 3)
+    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
    ])
    with self.assertRaises(RuntimeError):
      graph_rewrite(a, pm)

  def test_inf_loop_bottom_up(self):
-    a = UOp.const(dtypes.int, 3)
+    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
    ])
    with self.assertRaises(RuntimeError):
      graph_rewrite(a, pm, bottom_up=True)

-def bidir_append(ctx, x, b): ctx.append((x.arg if x.op is Ops.CONST else "+", b))
+def bidir_append(ctx, x, b): ctx.append((x.val if x.op is Ops.CONST else "+", b))
 class TestBidirectional(unittest.TestCase):
  def test_simple(self):
-    a = UOp.const(dtypes.int, 1)
-    b = UOp.const(dtypes.int, 2)
+    a = UOp.const(1)
+    b = UOp.const(2)
    c = a + b
    pm = PatternMatcher([ (UPat(GroupOp.All, name="x"), lambda ctx,x: bidir_append(ctx, x, False)) ])
    bpm = PatternMatcher([ (UPat(GroupOp.All, name="x"), lambda ctx,x: bidir_append(ctx, x, True)) ])
@@ -336,14 +336,14 @@ class TestBidirectional(unittest.TestCase):

 class TestStopEarly(unittest.TestCase):
  def test_stop_early(self):
-    a = UOp.const(dtypes.int, 3)
-    b = UOp.const(dtypes.int, 4)
+    a = UOp.const(3)
+    b = UOp.const(4)
    c = a+b
-    cn = UOp.const(dtypes.int, 7)
-    d = UOp.const(dtypes.int, 2)
+    cn = UOp.const(7)
+    d = UOp.const(2)
    def visit_const(c:UOp):
-      print(f"visit {c.arg}")
-      assert c.arg not in (3,4)
+      print(f"visit {c.val}")
+      assert c.val not in (3,4)
    pm_cvisit = PatternMatcher([(UPat(Ops.CONST, name="c"), visit_const),])
    ret = (c+d).substitute({c:cn}, extra_pm=pm_cvisit)
    assert ret == cn+d
@@ -376,15 +376,15 @@ class TestWalkRewrite(unittest.TestCase):

  def test_walk_topdown_no_fixed_point(self):
    """A bouncing pattern applies once and stops instead of looping."""
-    a = UOp.const(dtypes.int, 3)
+    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
    ])
    with self.assertRaises(RuntimeError):
      graph_rewrite(a, pm, bottom_up=True)
    ret = graph_rewrite(a, pm, walk=True)
-    self.assertIs(ret, UOp.const(dtypes.int, 4))
+    self.assertIs(ret, UOp.const(4))

  def test_walk_topdown_rewrites_children(self):
    a = UOp.variable('a', 0, 10)
@@ -418,11 +418,11 @@ class TestWalkRewrite(unittest.TestCase):
    """Top-down walk fires pm after children are processed (post-order)."""
    visited = []
    def track_visit(ctx, x):
-      ctx.append(x.arg if x.op is Ops.CONST else x.op)
+      ctx.append(x.val if x.op is Ops.CONST else x.op)
      return None
    pm = PatternMatcher([(UPat(GroupOp.All, name="x"), track_visit)])
-    a = UOp.const(dtypes.int, 1)
-    b = UOp.const(dtypes.int, 2)
+    a = UOp.const(1)
+    b = UOp.const(2)
    graph_rewrite(a + b, pm, ctx=visited, walk=True)
    self.assertEqual(visited, [1, 2, Ops.ADD])

@@ -454,23 +454,23 @@ class TestWalkRewrite(unittest.TestCase):

  def test_walk_bottomup_no_fixed_point(self):
    """Bottom-up walk also applies once per node, no fixed-point iteration."""
-    a = UOp.const(dtypes.int, 3)
+    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
    ])
    ret = graph_rewrite(a, pm, bottom_up=True, walk=True)
-    self.assertIs(ret, UOp.const(dtypes.int, 4))
+    self.assertIs(ret, UOp.const(4))

  def test_walk_bottomup_visit_order(self):
    """Bottom-up walk fires bpm before descending (pre-order)."""
    visited = []
    def track_visit(ctx, x):
-      ctx.append(x.arg if x.op is Ops.CONST else x.op)
+      ctx.append(x.val if x.op is Ops.CONST else x.op)
      return None
    pm = PatternMatcher([(UPat(GroupOp.All, name="x"), track_visit)])
-    a = UOp.const(dtypes.int, 1)
-    b = UOp.const(dtypes.int, 2)
+    a = UOp.const(1)
+    b = UOp.const(2)
    graph_rewrite(a + b, pm, ctx=visited, bottom_up=True, walk=True)
    # bpm fires on each node before children: +, 1, 2
    self.assertEqual(visited, [Ops.ADD, 1, 2])
@@ -490,15 +490,15 @@ class TestWalkRewrite(unittest.TestCase):
    """Bidirectional walk: bpm fires pre-order, pm fires post-order."""
    visited = []
    def bpm_visit(ctx, x):
-      ctx.append((x.arg if x.op is Ops.CONST else x.op, "bpm"))
+      ctx.append((x.val if x.op is Ops.CONST else x.op, "bpm"))
      return None
    def pm_visit(ctx, x):
-      ctx.append((x.arg if x.op is Ops.CONST else x.op, "pm"))
+      ctx.append((x.val if x.op is Ops.CONST else x.op, "pm"))
      return None
    bpm = PatternMatcher([(UPat(GroupOp.All, name="x"), bpm_visit)])
    pm = PatternMatcher([(UPat(GroupOp.All, name="x"), pm_visit)])
-    a = UOp.const(dtypes.int, 1)
-    b = UOp.const(dtypes.int, 2)
+    a = UOp.const(1)
+    b = UOp.const(2)
    graph_rewrite(a + b, pm, ctx=visited, bpm=bpm, walk=True)
    # bpm fires pre-order, pm fires post-order
    self.assertEqual(visited, [
@@ -509,23 +509,23 @@ class TestWalkRewrite(unittest.TestCase):
    """If bpm matches, children are skipped and pm never fires on that node."""
    visited = []
    def bpm_match(ctx, x):
-      ctx.append((x.arg if x.op is Ops.CONST else x.op, "bpm"))
+      ctx.append((x.val if x.op is Ops.CONST else x.op, "bpm"))
      # rewrite const(1) -> const(10), short-circuiting its subtree
-      if x.op is Ops.CONST and x.arg == 1: return x.replace(arg=10)
+      if x.op is Ops.CONST and x.val == 1: return UOp.const(10, x.dtype)
      return None
    def pm_match(ctx, x):
-      ctx.append((x.arg if x.op is Ops.CONST else x.op, "pm"))
+      ctx.append((x.val if x.op is Ops.CONST else x.op, "pm"))
      return None
    bpm = PatternMatcher([(UPat(GroupOp.All, name="x"), bpm_match)])
    pm = PatternMatcher([(UPat(GroupOp.All, name="x"), pm_match)])
-    a = UOp.const(dtypes.int, 1)
-    b = UOp.const(dtypes.int, 2)
+    a = UOp.const(1)
+    b = UOp.const(2)
    ret = graph_rewrite(a + b, pm, ctx=visited, bpm=bpm, walk=True)
    # bpm matches const(1) and short-circuits it, so pm never fires on const(1)
    self.assertNotIn((1, "pm"), visited)
    # but pm still fires on const(2) and the rebuilt ADD
    self.assertIn((2, "pm"), visited)
-    self.assertIs(ret, UOp.const(dtypes.int, 10) + b)
+    self.assertIs(ret, UOp.const(10) + b)

 if __name__ == '__main__':
  unittest.main()
@@ -297,10 +297,10 @@ class TestPolyN(unittest.TestCase):
    from tinygrad.dtype import dtypes
    from tinygrad.uop.ops import UOp
    from test.helpers import eval_uop
-    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 1.0), [1.0, -2.0, 1.0])), 0.0)
-    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 2.0), [1.0, -2.0, 1.0])), 1.0)
-    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 3.0), [1.0, -2.0, 1.0])), 4.0)
-    np.testing.assert_allclose(eval_uop(polyN(UOp.const(dtypes.float, 4.0), [1.0, -2.0, 1.0])), 9.0)
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(1.0).cast(dtypes.float), [1.0, -2.0, 1.0])), 0.0)
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(2.0).cast(dtypes.float), [1.0, -2.0, 1.0])), 1.0)
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(3.0).cast(dtypes.float), [1.0, -2.0, 1.0])), 4.0)
+    np.testing.assert_allclose(eval_uop(polyN(UOp.const(4.0).cast(dtypes.float), [1.0, -2.0, 1.0])), 9.0)

 class TestTimeToStr(unittest.TestCase):
  def test_seconds(self):           self.assertEqual("   10.01s ", time_to_str(10.01))
@@ -8,15 +8,15 @@ from tinygrad.codegen import to_program
 class TestLinearizerFailures(unittest.TestCase):
  def test_fail_1(self):
    c0 = UOp.param(0, dtypes.float, (64,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 2), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 32), 2, AxisType.LOOP)
-    c3 = ((c1*UOp.const(dtypes.weakint, 32))+c2)
+    c1 = UOp.range(UOp.const(2), 1, AxisType.WEAK)
+    c2 = UOp.range(UOp.const(32), 2, AxisType.WEAK)
+    c3 = ((c1*UOp.const(32))+c2)
    c4 = UOp.param(1, dtypes.float, (163840,))
-    c5 = UOp.range(UOp.const(dtypes.weakint, 2560), 0, AxisType.REDUCE)
-    c6 = c4.index(((((((c5//UOp.const(dtypes.weakint, 8))%UOp.const(dtypes.weakint, 8))*UOp.const(dtypes.weakint, 8))+(c5%UOp.const(dtypes.weakint, 8)))+(((c2*UOp.const(dtypes.weakint, 40))+(c5//UOp.const(dtypes.weakint, 64)))*UOp.const(dtypes.weakint, 64)))+(c1*UOp.const(dtypes.weakint, 81920))))
+    c5 = UOp.range(UOp.const(2560), 0, AxisType.REDUCE)
+    c6 = c4.index(((((((c5//UOp.const(8))%UOp.const(8))*UOp.const(8))+(c5%UOp.const(8)))+(((c2*UOp.const(40))+(c5//UOp.const(64)))*UOp.const(64)))+(c1*UOp.const(81920))))
    c7 = UOp.param(2, dtypes.float, (64,))
    c8 = c7.index(c3)
-    c9 = ((((c6+(c8*UOp.const(dtypes.float, -1.0)))*(c6+(c8*UOp.const(dtypes.float, -1.0)))).reduce(c5, arg=Ops.ADD)*UOp.const(dtypes.float, 0.000390625))+UOp.const(dtypes.float, 1e-05)).sqrt().reciprocal()
+    c9 = ((((c6+(c8*UOp.const(-1.0)))*(c6+(c8*UOp.const(-1.0)))).reduce(c5, arg=Ops.ADD)*UOp.const(0.000390625))+UOp.const(1e-05)).sqrt().reciprocal()
    c10 = c0.index(c3).store(c9).end(c1, c2)
    ast = c10.sink(arg=KernelInfo())
    to_program(ast, renderer=Device[Device.DEFAULT].renderer)
@@ -109,6 +109,28 @@ class TestLLMServer(unittest.TestCase):

    self.assertGreater(len(contents), 0)

+  def test_interrupted_stream_logs_tokens(self):
+    with patch.object(self.mock_model, "generate", side_effect=lambda ids, **kwargs: iter([300, 301, 999])), \
+         patch("tinygrad.llm.serve.stderr_log") as log, patch("tinygrad.llm.serve.colored", side_effect=lambda text, color: text) as color:
+      stream = self.server.RequestHandlerClass.run_model(Mock(server=self.server), [200, 201, 202], "test")
+      next(stream)
+      next(stream)
+      stream.close()
+    interrupt = log.call_args.args[0]
+    self.assertFalse(interrupt.startswith("\n"))
+    self.assertTrue(interrupt.endswith("\n"))
+    self.assertIn("gen:", interrupt)
+    self.assertIn("out:    1", interrupt)
+    self.assertTrue(any(args[0].startswith("total:") and args[1] == "red" for args, _ in color.call_args_list))
+
+  def test_stream_disconnect_closes_source(self):
+    from tinygrad.llm.serve import Handler
+    source, handler = Mock(), Mock()
+    source.__iter__ = Mock(return_value=iter([{}]))
+    handler.wfile.write.side_effect = BrokenPipeError
+    Handler.stream_json(handler, source)
+    source.close.assert_called_once()
+
  def test_non_streaming(self):
    resp = self.client.chat.completions.create(
      model="test-model",
@@ -42,7 +42,7 @@ def check_assign(buffer_lists, copies=None):
  for orig_si, new_si in zip(linear.src, result.src):
    for orig, new in zip(orig_si.src[1:], new_si.src[1:]):
      if new.op is Ops.SLICE and id(orig) not in replace_map:
-        replace_map[id(orig)] = (new.src[0], new.src[1].arg * new.src[0].dtype.itemsize, new.arg * new.dtype.itemsize)
+        replace_map[id(orig)] = (new.src[0], new.src[1].val * new.src[0].dtype.itemsize, new.arg * new.dtype.itemsize)

  # verify pinned buffers are not planned
  for buf in held_bufs:
@@ -1,5 +1,5 @@
 import unittest, time
-from tinygrad import dtypes, Tensor, UOp, getenv
+from tinygrad import Tensor, UOp, getenv
 from tinygrad.helpers import Profiling

 PYPROFILE = getenv("PYPROFILE")
@@ -27,29 +27,29 @@ class TestBench(unittest.TestCase):
    print(f"{self._testMethodName:30s} {et*1e6/self.N:.2f} us")

  def test_uop_instant_creation(self):
-    for i in range(self.N): UOp.const(dtypes.int, 100+i)
+    for i in range(self.N): UOp.const(100+i)

  def test_uop_list_creation(self):
-    [UOp.const(dtypes.int, 100+i) for i in range(self.N)]
+    [UOp.const(100+i) for i in range(self.N)]

  def test_uop_add_2n(self):
-    a = UOp.const(dtypes.int, 2)
+    a = UOp.const(2)
    for _ in range(self.N): a = a + a

  def test_uop_toposort(self):
-    a = UOp.const(dtypes.int, 0)
-    for i in range(self.N): a = a + UOp.const(dtypes.int, 100+i)
+    a = UOp.const(0)
+    for i in range(self.N): a = a + UOp.const(100+i)
    self.start_time()
    self.assertEqual(len(a.toposort()), 2*self.N+1)

  def test_uop_toposort_2n(self):
-    a = UOp.const(dtypes.int, 0)
+    a = UOp.const(0)
    for _ in range(self.N): a = a + a
    self.start_time()
    self.assertEqual(len(a.toposort()), self.N+1)

  def test_uop_simplify(self):
-    a = UOp.const(dtypes.int, 2)
+    a = UOp.const(2)
    for _ in range(self.N): (a+a).simplify()

  def test_uop_simplify_complex(self):
@@ -68,7 +68,7 @@ class TestBench(unittest.TestCase):
    for _ in range(self.N): expr.simplify()

  def test_uop_chain_free(self):
-    a = UOp.const(dtypes.int, 2)
+    a = UOp.const(2)
    for _ in range(self.N): a = a + a
    self.start_time()
    del a
@@ -5,9 +5,9 @@ from tinygrad.uop.ops import PatternMatcher, UPat

 class TestPatternMatcher(unittest.TestCase):
  def test_simple_match(self):
-    matcher = PatternMatcher([(UPat(Ops.CONST, name="x", dtype=dtypes.float), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.int, 1)
+    matcher = PatternMatcher([(UPat(Ops.CONST, name="x", dtype=dtypes.weakfloat), lambda x: x.rtag())])
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(1)
    self.assertEqual(matcher.rewrite(c1), c1.rtag())
    self.assertEqual(matcher.rewrite(c2), None)

@@ -61,16 +61,16 @@ class TestPatternMatcher(unittest.TestCase):

  def test_uop(self):
    matcher = PatternMatcher([(UPat(Ops.CONST, name="x"), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
+    c1 = UOp.const(1.0)
    c2 = UOp(Ops.ADD, src=(c1, c1))
    self.assertEqual(matcher.rewrite(c1), c1.rtag())
    self.assertEqual(matcher.rewrite(c2), None)

  def test_uop_set(self):
    matcher = PatternMatcher([(UPat((Ops.CONST, Ops.CAST), name="x"), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.bool, False)
+    c1 = UOp.const(False)
    c2 = UOp(Ops.CAST, arg=dtypes.int, src=(c1,))
-    c3 = UOp.const(dtypes.float, 1.0)
+    c3 = UOp.const(1.0)
    c4 = UOp(Ops.ADD, src=(c3, c3))
    self.assertEqual(matcher.rewrite(c1), c1.rtag())
    self.assertEqual(matcher.rewrite(c2), c2.rtag())
@@ -82,11 +82,11 @@ class TestPatternMatcher(unittest.TestCase):
      (UPat(Ops.CONST, arg=False, name="x"), lambda x: x.rtag()),
      (UPat(Ops.MAX, name="x"), lambda x: x.rtag()),
    ])
-    c1 = UOp.const(dtypes.float, 0.0)
-    c2 = UOp.const(dtypes.bool, False)
+    c1 = UOp.const(0.0)
+    c2 = UOp.const(False)
    c3 = UOp(Ops.MAX, src=(c1, c1))
    c4 = UOp(Ops.MUL, src=(c1, c1))
-    c5 = UOp.const(dtypes.int, -1)
+    c5 = UOp.const(-1)
    self.assertEqual(matcher.rewrite(c1), c1.rtag())
    self.assertEqual(matcher.rewrite(c2), c2.rtag())
    self.assertEqual(matcher.rewrite(c3), c3.rtag())
@@ -96,11 +96,11 @@ class TestPatternMatcher(unittest.TestCase):
  def test_filter_arg(self):
    matcher = PatternMatcher([
      (UPat(Ops.MUL, src=[UPat(Ops.CONST, name="c"), UPat(Ops.CONST, arg=2)], name="x"),
-       lambda x,c: x.rtag() if c.arg in {1, -1} else None)
+       lambda x,c: x.rtag() if c.val in {1, -1} else None)
    ])
-    y1 = UOp.const(dtypes.int, 1)
-    y2 = UOp.const(dtypes.int, 2)
-    y3 = UOp.const(dtypes.int, -1)
+    y1 = UOp.const(1)
+    y2 = UOp.const(2)
+    y3 = UOp.const(-1)
    c1 = UOp(Ops.MUL, src=(y1, y2))
    c2 = UOp(Ops.MUL, src=(y2, y2))
    c3 = UOp(Ops.MUL, src=(y3, y2))
@@ -114,26 +114,27 @@ class TestPatternMatcher(unittest.TestCase):

  def test_dup_name(self):
    matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=(UPat(Ops.CONST, name="y"), UPat(Ops.CONST, name="y"))), lambda x, y: x.rtag())])
-    y1 = UOp.const(dtypes.float, 1.0)
-    y2 = UOp.const(dtypes.float, 1.0)
+    y1 = UOp.const(1.0)
+    y2 = UOp.const(1.0)
    c1 = UOp(Ops.ADD, src=(y1, y1))
    c2 = UOp(Ops.ADD, src=(y1, y2))
    self.assertEqual(matcher.rewrite(c1), c1.rtag())
    self.assertEqual(matcher.rewrite(c2), c1.rtag())

  def test_dtype(self):
-    matcher = PatternMatcher([(UPat(Ops.CONST, name="x", dtype=dtypes.float32), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float64, 1.0)
+    # a concrete const dtype lives on the pair's CAST
+    matcher = PatternMatcher([(UPat(Ops.CAST, name="x", dtype=dtypes.float32), lambda x: x.rtag())])
+    c1 = UOp.const(1.0).cast(dtypes.float32)
+    c2 = UOp.const(1.0).cast(dtypes.float64)
    self.assertEqual(matcher.rewrite(c1), c1.rtag())
    self.assertEqual(matcher.rewrite(c2), None)

  def test_dtype_set(self):
-    matcher = PatternMatcher([(UPat(Ops.CONST, name="x", dtype={dtypes.float32, dtypes.float64}), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float64, 1.0)
-    c3 = UOp.const(dtypes.float16, 1.0)
-    c4 = UOp.const(dtypes.int, 1)
+    matcher = PatternMatcher([(UPat(Ops.CAST, name="x", dtype={dtypes.float32, dtypes.float64}), lambda x: x.rtag())])
+    c1 = UOp.const(1.0).cast(dtypes.float32)
+    c2 = UOp.const(1.0).cast(dtypes.float64)
+    c3 = UOp.const(1.0).cast(dtypes.float16)
+    c4 = UOp.const(1).cast(dtypes.int)
    self.assertEqual(matcher.rewrite(c1), c1.rtag())
    self.assertEqual(matcher.rewrite(c2), c2.rtag())
    self.assertEqual(matcher.rewrite(c3), None)
@@ -141,8 +142,8 @@ class TestPatternMatcher(unittest.TestCase):

  def test_src_one(self):
    matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=(UPat(Ops.CONST), UPat(Ops.CONST))), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
    c3 = UOp(Ops.ADD, src=(c1,c2))
    self.assertEqual(matcher.rewrite(c3), c3.rtag())
    self.assertEqual(matcher.rewrite(c2), None)
@@ -158,8 +159,8 @@ class TestPatternMatcher(unittest.TestCase):

  def test_src_permutations(self):
    matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=[UPat(Ops.CONST), UPat(GroupOp.ALU)]), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
    c3 = UOp(Ops.ADD, src=(c1,c2))
    c4 = UOp(Ops.ADD, src=(c3,c2))
    c5 = UOp(Ops.ADD, src=(c2,c3))
@@ -171,8 +172,8 @@ class TestPatternMatcher(unittest.TestCase):

  def test_src_repeat(self):
    matcher = PatternMatcher([(UPat(GroupOp.ALU, name="x", src=UPat(Ops.CONST)), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
    c3 = UOp(Ops.ADD, src=(c1,c2))
    c4 = UOp(Ops.ADD, src=(c2,c3))
    self.assertEqual(matcher.rewrite(c3), c3.rtag())
@@ -180,9 +181,9 @@ class TestPatternMatcher(unittest.TestCase):

  def test_allow_len(self):
    matcher = PatternMatcher([(UPat(Ops.MULACC, name="x", src=(UPat(Ops.CONST),), allow_any_len=True), lambda x: x.rtag())])
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp.const(dtypes.float, 3.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
+    c3 = UOp.const(3.0)
    c4 = UOp(Ops.EXP2, src=(c1,))
    c5 = UOp(Ops.ADD, src=(c1,c2))
    c6 = UOp(Ops.MULACC, src=(c1,c2,c3))
@@ -191,8 +192,8 @@ class TestPatternMatcher(unittest.TestCase):
    self.assertEqual(matcher.rewrite(c6), c6.rtag())

  def test_deep_src_permutations(self):
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
    u1 = (c1 + c2) + c1
    u2 = (c2 + c1) + c1
    matcher = PatternMatcher([
@@ -5,7 +5,7 @@ from tinygrad.nn.state import get_parameters
 from tinygrad.engine.jit import TinyJit
 from tinygrad import Tensor, Device, GlobalCounters, dtypes, Variable
 from tinygrad.helpers import Context
-from test.helpers import slow, jit_cache_count
+from test.helpers import slow, jit_cache_count, KernelCountException
 from extra.lr_scheduler import OneCycleLR
 from test.helpers import derandomize_model

@@ -35,8 +35,9 @@ def helper_test(nm, gen, model, max_memory_allowed, max_kernels_allowed, all_jit
    assert mem_used < max_memory_allowed, f"{nm} used more than {max_memory_allowed:.3f} GB - {mem_used:.3} GB used"
    assert (max_memory_allowed - mem_used) / max_memory_allowed < 0.2, f"{max_memory_allowed:.3f} GB is too far from {mem_used:.3} GB used"
    if kernels_used:
-      assert kernels_used <= max_kernels_allowed, f"{nm} used more than {max_kernels_allowed} kernels, it used {kernels_used}"
-      assert (max_kernels_allowed - kernels_used) / max_kernels_allowed < 0.2, f"{max_kernels_allowed=} is too far from {kernels_used=} used"
+      if kernels_used > max_kernels_allowed: raise KernelCountException(max_kernels_allowed, kernels_used)
+      if (max_kernels_allowed - kernels_used) / max_kernels_allowed >= 0.2:
+        raise KernelCountException(max_kernels_allowed, kernels_used)
    if all_jitted:
      assert kernels_used > 0 and kernels_used == GlobalCounters.kernel_count or (kernels_used <= GlobalCounters.kernel_count and getattr(Device[Device.DEFAULT], "graph", None)), f"only {kernels_used} out of {GlobalCounters.kernel_count} were jitted"  # noqa: E501

@@ -2,32 +2,11 @@
 import gc, unittest, time
 from typing import cast
 from tinygrad import nn, dtypes, Device, Tensor, getenv
-from tinygrad.uop.ops import UOp, Ops, GroupOp, UPat, KernelInfo
-from tinygrad.helpers import DEBUG, GlobalCounters, Context
-from tinygrad.engine.realize import compile_linear, run_linear
-from tinygrad.codegen import to_program
-
-class KernelCountException(Exception): pass
-def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Tensor]|None=None, filter_sink=True):
-  if to_prerealize:
-    with Context(DEBUG=0, TRACK_MATCH_STATS=0): Tensor.realize(*to_prerealize)
-  if isinstance(t, Tensor): linear, var_vals = t.linear_with_vars()
-  elif isinstance(t, list) and isinstance(t[0], Tensor): linear, var_vals = Tensor.linear_with_vars(*t)
-  else:
-    assert isinstance(t, UOp), f"can't schedule {t}"
-    linear, var_vals = Tensor(t).linear_with_vars()
-  kernel_cnt = sum((len(call.device) if isinstance(call.device, tuple) else 1)
-                   for call in linear.src if call.src[0].op is Ops.SINK or not filter_sink)
-  if kernel_cnt != allowed:
-    print(f"SCHEDULE ISSUE, expecting {allowed} got {kernel_cnt}")
-    if DEBUG >= 3:
-      for i,call in enumerate(linear.src):
-        print("kernel", i+1)
-        print(call.src[0])
-    raise KernelCountException(f"{kernel_cnt} != {allowed}")
-  # test compiling the linear
-  compile_linear(linear)
-  return linear, var_vals
+from tinygrad.uop.ops import UOp, Ops, GroupOp, UPat, KernelInfo, AxisType
+from tinygrad.helpers import GlobalCounters, Context
+from tinygrad.engine.realize import run_linear, compile_linear
+from tinygrad.codegen import to_program, full_rewrite_to_sink
+from test.helpers import check_schedule, assert_kernel_count, KernelCountException

 def _realize_weights(m):
  for p in nn.state.get_parameters(m): p.realize()
@@ -143,7 +122,7 @@ class TestSimpleSchedule(unittest.TestCase):
    a = Tensor.empty(16,16).sum(axis=1)
    a1 = a.reshape(4,4)
    a2 = a.reshape(16,1,1)
-    self.assertEqual(len(Tensor.schedule_linear(a1, a2).src), 1)
+    check_schedule([a1, a2], 1)

 class TestSchedule(unittest.TestCase):
  def setUp(self):
@@ -155,8 +134,7 @@ class TestSchedule(unittest.TestCase):
  def test_arange_avgpool2d(self, kcount=1):
    x = Tensor.arange(25).reshape(1,1,5,5).cast(dtypes.float32)
    t = x.avg_pool2d(padding=1).clone()
-    linear, var_vals = t.linear_with_vars()
-    self.assertEqual(len(linear.src), kcount)
+    check_schedule(t, kcount)

  def test_arange_avgpool2d_fused_noopt(self):
    with Context(NOOPT=1): self.test_arange_avgpool2d(kcount=1)
@@ -224,7 +202,7 @@ class TestSchedule(unittest.TestCase):
    GlobalCounters.reset()
    expr = (a/b)/c
    expr.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertLessEqual(GlobalCounters.global_ops, 4*3)

  # NOTE: this is causing "LAZYCACHE=1 incorrectly reuses contiguous const" #4562
@@ -357,6 +335,11 @@ class TestSchedule(unittest.TestCase):
    out1 = a.sum() + b
    check_schedule([out0, out1], 2)

+  def test_reduce_broadcast_not_recomputed(self):
+    a = Tensor.empty(32, 16).realize()
+    out = a-a.mean(axis=0, keepdim=True)
+    check_schedule(out, 2)
+
  def test_scaled_dot_product_attention_multireduce_fusion(self):
    q = Tensor.empty(32,8,16,8).realize()
    k = Tensor.empty(32,8,16,8).realize()
@@ -609,9 +592,7 @@ class TestSchedule(unittest.TestCase):
    img = Tensor.randn(BS, CIN, 64, 64).realize()
    w = Tensor.uniform(16, CIN, 3, 3).realize()
    ret = Tensor.conv2d(img, w).relu().mean().backward()
-    linear, var_vals = Tensor.linear_with_vars(ret, img.grad, w.grad)
-    cnt = len([call for call in linear.src if call.src[0].op is Ops.SINK])
-    assert cnt == allowed, f"expected {allowed} kernels, got {cnt}"
+    check_schedule([ret, img.grad, w.grad], allowed)

  def test_conv2d_half(self): self.test_conv2d(4, dtype=dtypes.half)

@@ -632,7 +613,8 @@ class TestSchedule(unittest.TestCase):
        return len([call for call in linear.src if call.src[0].op is Ops.PROGRAM])

      with Context(IMAGE=1):
-        self.assertEqual(cnt(), 5)
+        got = cnt()
+        if got != 5: raise KernelCountException(5, got)

  def test_image_f16_residual_fusion(self):
    with Context(FLOAT16=1, OPENPILOT_HACKS=1):
@@ -647,7 +629,8 @@ class TestSchedule(unittest.TestCase):
        return len([call for call in linear.src if call.src[0].op is Ops.PROGRAM])

      with Context(IMAGE=1):
-        self.assertEqual(cnt(), 9)
+        got = cnt()
+        if got != 9: raise KernelCountException(9, got)

  def _test_fusion(self, shapes, f, cnt):
    with Context(DEBUG=0, TRACK_MATCH_STATS=0):
@@ -714,6 +697,19 @@ class TestSchedule(unittest.TestCase):
    xt = X[[Tensor([2]), Tensor([1])]]
    check_schedule(xt, 1)

+  def test_split_advanced_indexing_not_recomputed(self):
+    with Context(SPLIT_REDUCEOP=1):
+      X = Tensor.empty(32768, 4).realize()
+      idx = Tensor.randint(4, high=X.shape[0])
+      linear, _ = check_schedule(X[idx], 3, [Tensor._device_rng_counters[idx.device]])
+      # The split's final reduction remains, but the one-hot gather should collapse into a direct indexed load.
+      reduce_kernels = 0
+      for call in linear.src:
+        if call.src[0].op is not Ops.SINK: continue
+        sink = full_rewrite_to_sink(call.src[0], Device[call.device].renderer)
+        reduce_kernels += any(u.op is Ops.RANGE and u.arg[-1] is AxisType.REDUCE for u in sink.toposort())
+      self.assertEqual(reduce_kernels, 1)
+
  def test_push_through_reshape(self):
    x = Tensor.empty(10, 20).realize()
    out = x.argmax(1)
@@ -874,8 +870,7 @@ class TestSchedule(unittest.TestCase):
    t = Tensor.zeros((3, 3)).contiguous().realize()
    v = t[1]  # view - is_realized but not has_buffer_identity
    assert v.uop.is_realized
-    linear, _ = Tensor.linear_with_vars(v)
-    self.assertEqual(len(linear.src), 0)
+    check_schedule(v, 0)

  # NOTE: because empty does not have a lowered kernel if realize is called on a childless empty, it never gets allocated.
  def test_childless_empty_never_allocates(self):
@@ -1457,8 +1452,7 @@ class TestSchedule(unittest.TestCase):
    Tensor.manual_seed(0)
    x = Tensor.randn(4, 12, 64, 64, dtype=dtypes.half).realize()
    out = x.softmax(dtype=dtypes.float)
-    linear = out.schedule_linear()
-    self.assertEqual(len(linear.src), 3)
+    linear, _ = check_schedule(out, 3)
    # max reduction stays in input dtype (no numerical loss), upcast happens after subtracting max
    self.assertEqual(linear.src[0].src[1].dtype, dtypes.half)
    self.assertEqual(linear.src[1].src[1].dtype, dtypes.float)
@@ -1640,11 +1634,11 @@ class TestSchedule(unittest.TestCase):
    self.assertEqual(GlobalCounters.mem_used-base, 0)

  def test_const_schedule(self):
-    constv = Tensor.empty(2, 2).uop.const_like(10)
+    constv = Tensor.empty(2, 2).const_like(10).uop
    check_schedule(constv, 0)

  def test_const_schedule_contig(self):
-    constv = Tensor.empty(2, 2).uop.const_like(10).contiguous()
+    constv = Tensor.empty(2, 2).const_like(10).uop.contiguous()
    check_schedule(constv, 0)

  def test_advanced_simple_indexing_combined(self):
@@ -1873,8 +1867,7 @@ class TestFusionOp(unittest.TestCase):
    val = 1.0
    a = Tensor(val)
    for _ in range(24): a = Tensor.stack(a, a)[0]
-    linear = a.schedule_linear()
-    self.assertLessEqual(len(linear.src), 1)
+    check_schedule(a, 0)
    self.assertLess(time.perf_counter()-st, 2.0)

  def test_recursive_reshape(self):
@@ -1883,8 +1876,7 @@ class TestFusionOp(unittest.TestCase):
    b = Tensor.empty(16, 2).realize()
    r = a.sum(1)
    for _ in range(24): r = r.reshape(16, 2) + b
-    linear = r.schedule_linear()
-    self.assertEqual(len(linear.src), 1)
+    check_schedule(r, 1)
    self.assertLess(time.perf_counter()-st, 2.0)

 # NOTE: the NULL backend supports SLICE
@@ -2,7 +2,7 @@ import unittest, itertools

 from tinygrad.codegen.late.coalesce import indexing_simplify
 from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import UOp, Ops, graph_rewrite
+from tinygrad.uop.ops import UOp, Ops, graph_rewrite, pm_lower_index_dtype
 from tinygrad.uop.symbolic import simplify_valid, sym, pm_move_where_on_load
 from tinygrad.helpers import Context
 from test.helpers import full_rewrite
@@ -23,7 +23,7 @@ def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UO
    UOp.param(0, dtypes.float, image_shape).index(idx[1].valid(valid), idx[0].valid(valid)),
  ))

-def Special(expr, nmax): return UOp(Ops.SPECIAL, src=(UOp.const(dtypes.weakint, nmax),), arg=expr)
+def Special(expr, nmax): return UOp(Ops.SPECIAL, src=(UOp.const(nmax),), arg=expr)
 def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
 def Range(n, nmax): return UOp.range(nmax, n)

@@ -207,7 +207,7 @@ class TestImageSimplification(unittest.TestCase):
    if svalid is not None:
      check_uop_against_string(self, off.src[1].get_valid(), svalid)
    else:
-      self.assertEqual(off.src[1].get_valid(), UOp.const(dtypes.bool, True), "svalid is None but valid is not True")
+      self.assertEqual(off.src[1].get_valid(), UOp.const(True), "svalid is None but valid is not True")

  def test_idx_gt_c(self):
    # (idx1 < c+1).ne(True) ? (..., idx1-1+c) : 0 can drop the valid
@@ -455,7 +455,7 @@ class TestImageSimplification(unittest.TestCase):
    A1 = lidx0*32 + r0*32 + lidx1*4 - 99
    valid = ((lidx1 < 1).ne(True)) & ((lidx0 + r0) < 3).ne(True) & ((lidx0 + r0) < 19)
    alu0 = gidx0 + (A1 % 32)*32 + (A1 // 32 % 16)*1024
-    load = get_load_image_uop((1, 16384, 4), valid, (alu0, UOp.const(dtypes.weakint, 0)))
+    load = get_load_image_uop((1, 16384, 4), valid, (alu0, UOp.const(0)))
    try:
      self.check(load, None, "(gidx0+lidx0*1024+r0*1024+lidx1*128+-3168)", "0")
    except AssertionError:
@@ -474,7 +474,7 @@ class TestImageSimplification(unittest.TestCase):
    A1 = lidx0*16 + r0*16 + lidx1*4 - 51
    valid = ((lidx1 < 1).ne(True)) & ((lidx0 + r0) < 3).ne(True) & ((lidx0 + r0) < 11)
    alu0 = lidx2 + gidx0*4 + (A1 % 16)*64 + (A1 // 16 % 8)*1024
-    load = get_load_image_uop((1, 8192, 4), valid, (alu0, UOp.const(dtypes.weakint, 0)))
+    load = get_load_image_uop((1, 8192, 4), valid, (alu0, UOp.const(0)))
    try:
      self.check(load, None, "(lidx2+gidx0*4+lidx0*1024+r0*1024+lidx1*256+-3264)", "0")
    except AssertionError:
@@ -488,10 +488,20 @@ class TestImageSimplification(unittest.TestCase):
    gidx0 = Special("gidx0", 1064)
    r12 = Range(12, 3)
    valid = ((gidx0 < 645).ne(True)) & (gidx0 < 653)
-    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(dtypes.weakint, 0))
+    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(0))
    load = get_load_image_uop((1, 48, 4), valid, idx)
    self.check(load, None, "(r12*4+(gidx0+3)%4+(gidx0+3)//4*24+-3888)", "0")

+  def test_drop_gate_committed_in_the_index_pass(self):
+    # the fused index pass runs without symbolic, so committing a weak src must not leave a CAST that
+    # symbolic later folds inside the index only: the gate's copy of the expression has to stay the same node
+    f = UOp.variable("f", 0.0, 9.0, dtypes.float)
+    idx_y = (f + UOp.const(1.0)).cast(dtypes.int)
+    load = get_load_image_uop((10, 10, 4), (UOp.const(-1) < idx_y) & (idx_y < UOp.const(10)),
+                              (Special("gidx0", 10), idx_y))
+    off = graph_rewrite(load.sink(), pm_lower_index_dtype+indexing_simplify, ctx={}).src[0].src[0]
+    self.assertEqual(off.src[1].get_valid(), UOp.const(True))
+
 class TestDropTrueGate(unittest.TestCase):
  def test_drop_true_gate_on_index(self):
    # test that INDEX with a constant True valid gets simplified to drop the valid
@@ -499,8 +509,8 @@ class TestDropTrueGate(unittest.TestCase):
    from tinygrad.uop.ops import graph_rewrite
    from tinygrad.uop.symbolic import sym
    buf = UOp.param(0, dtypes.int, (1,))
-    idx = UOp.const(dtypes.weakint, 0)
-    true_gate = UOp.const(dtypes.bool, True)
+    idx = UOp.const(0)
+    true_gate = UOp.const(True)
    index_with_gate = UOp(Ops.INDEX, src=(buf, idx.valid(true_gate)))
    # apply the optimization
    result = graph_rewrite(index_with_gate, sym+indexing_simplify)
@@ -516,50 +526,50 @@ class TestRangeShrink(unittest.TestCase):
  def test_range_shrink_single_guard(self):
    # range 0..203 guarded by r < 4 everywhere -> shrink to 0..3
    r = Range(0, 204)
-    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
+    load = get_gated_load_uop(r < UOp.const(4), r)
    ranges = self.get_ranges(load.sink())
    self.assertEqual(len(ranges), 1)
-    self.assertEqual(ranges[0].src[0].arg, 4)
+    self.assertEqual(ranges[0].src[0].val, 4)

  def test_range_shrink_picks_max_guard(self):
    # two loads guard the same range with r < 4 and r < 8 -> shrink to max(4, 8) = 8
    r = Range(0, 204)
-    load1 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
-    load2 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 8), r)
+    load1 = get_gated_load_uop(r < UOp.const(4), r)
+    load2 = get_gated_load_uop(r < UOp.const(8), r)
    ranges = self.get_ranges(UOp.sink(load1, load2))
    self.assertEqual(len(ranges), 1)
-    self.assertEqual(ranges[0].src[0].arg, 8)
+    self.assertEqual(ranges[0].src[0].val, 8)

  def test_range_no_shrink_guard_ge_max(self):
    # guard r < 300 with range max 204 -> no shrink (guard doesn't constrain)
    r = Range(0, 204)
-    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 300), r)
+    load = get_gated_load_uop(r < UOp.const(300), r)
    ranges = self.get_ranges(load.sink())
    self.assertEqual(len(ranges), 1)
-    self.assertEqual(ranges[0].src[0].arg, 204)
+    self.assertEqual(ranges[0].src[0].val, 204)

  def test_range_no_shrink_when_unguarded_elsewhere(self):
    # one load guards r < 4, but another load uses r without a gate -> no shrink
    r = Range(0, 204)
-    load1 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
+    load1 = get_gated_load_uop(r < UOp.const(4), r)
    load2 = UOp(Ops.LOAD, src=(UOp.param(1, dtypes.float, (204,)).index(r),))
    ranges = self.get_ranges(UOp.sink(load1, load2))
    self.assertEqual(len(ranges), 1)
-    self.assertEqual(ranges[0].src[0].arg, 204)
+    self.assertEqual(ranges[0].src[0].val, 204)

  def test_range_no_shrink_when_used_in_reduce(self):
    # range used in both a gated load AND directly in the reduce expression -> no shrink
    r = Range(0, 204)
-    gated_load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
+    gated_load = get_gated_load_uop(r < UOp.const(4), r)
    red = (r.cast(dtypes.float) + gated_load).reduce(r, arg=Ops.ADD)
    ranges = self.get_ranges(red.sink())
    self.assertEqual(len(ranges), 1)
-    self.assertEqual(ranges[0].src[0].arg, 204)
+    self.assertEqual(ranges[0].src[0].val, 204)

  def test_range_shrink_to_single_iteration(self):
    # guard r < 1 shrinks range to 1 -> single iteration, range eliminated entirely
    r = Range(0, 204)
-    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 1), r)
+    load = get_gated_load_uop(r < UOp.const(1), r)
    ranges = self.get_ranges(load.sink())
    self.assertEqual(len(ranges), 0)

@@ -567,19 +577,19 @@ class TestRangeShrink(unittest.TestCase):
    # emulates mask.where(x.pad_to(mask.shape), Invalid): range should shrink accordingly
    from tinygrad.dtype import Invalid
    r = Range(0, 204)
-    x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
+    x = (r < 4).where(UOp.const(1.0), Invalid)
    ranges = self.get_ranges(UOp.param(0, dtypes.float, (204,)).index(r).store((r < 4).where(x, Invalid)).sink())
    self.assertEqual(len(ranges), 1)
-    self.assertEqual(ranges[0].src[0].arg, 4)
+    self.assertEqual(ranges[0].src[0].val, 4)

  def test_range_shrink_store_where_invalid_flipped(self):
    # above, but flipped
    from tinygrad.dtype import Invalid
    r = Range(0, 204)
-    x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
+    x = (r < 4).where(UOp.const(1.0), Invalid)
    ranges = self.get_ranges(UOp.param(0, dtypes.float, (204,)).index(r).store((r >= 4).where(Invalid, x)).sink())
    self.assertEqual(len(ranges), 1)
-    self.assertEqual(ranges[0].src[0].arg, 4)
+    self.assertEqual(ranges[0].src[0].val, 4)

 if __name__ == '__main__':
  unittest.main()
@@ -1,5 +1,6 @@
 import unittest
 from tinygrad import Variable
+from tinygrad.uop.ops import UOp


 class TestFuzzFailure(unittest.TestCase):
@@ -8,7 +9,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 2)
    v3=Variable('v3', 0, 1)
    expr = (((((((((((((((((((((((0//4)%2)//8)+-2)+-4)+-3)+v1)+-4)+v2)+-2)+v3)+v2)//3)%7)*1)//2)+v2)*-1)+2)+1)+0)+-3)+v3)
-    v1_val, v2_val, v3_val = v1.const_like(8), v2.const_like(0), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(8), UOp.const(0), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -19,7 +20,7 @@ class TestFuzzFailure(unittest.TestCase):
    v3=Variable('v3', 0, 3)
    expr = (((((((((((((((((((((((((0*4)//5)*2)*-1)*-2)+-4)*4)*2)*3)*4)+-4)*4)+v2)+v2)+v3)//3)+v2)+v1)//9)+3)+1)//1)+-4)//4)*2)
    expr = (((((v1+(v2+(((v3+(v2*2))+1)//3)))+4)//9)+-57)//(9*4))
-    v1_val, v2_val, v3_val = v1.const_like(6), v2.const_like(0), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(6), UOp.const(0), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -29,7 +30,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 1)
    v3=Variable('v3', 0, 2)
    expr = (((((((((((((((((((0//2)//3)+v3)+0)+-4)*-2)*-2)+-1)+2)+3)+v3)+0)//8)*-3)+0)*-2)*-4)*-2)//5)
-    v1_val, v2_val, v3_val = v1.const_like(0), v2.const_like(0), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(0), UOp.const(0), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -39,7 +40,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 3)
    v3=Variable('v3', 0, 4)
    expr = (((((((((((((((((((((((((((((0*-2)+0)*-1)//9)//6)//8)+v1)*-4)+v2)//4)//8)+4)*3)+v1)+v3)//8)//7)+4)+v3)*-4)+1)+v1)*3)+4)*2)//5)//2)//3)*-4)
-    v1_val, v2_val, v3_val = v1.const_like(2), v2.const_like(0), v3.const_like(2)
+    v1_val, v2_val, v3_val = UOp.const(2), UOp.const(0), UOp.const(2)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -49,7 +50,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 1)
    v3=Variable('v3', 0, 3)
    expr = ((((((((((((((0+v2)+v1)*0)+v2)//1)//7)+-2)+v2)+v1)*4)+-3)//5)+v2)+1)
-    v1_val, v2_val, v3_val = v1.const_like(0), v2.const_like(0), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(0), UOp.const(0), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -60,7 +61,7 @@ class TestFuzzFailure(unittest.TestCase):
    v3=Variable('v3', 0, 128)
    expr = (((((((((((((((((((((((((((((0//3)+4)+v1)//2)+-1)//1)*1)*-1)*4)//5)+v1)//6)+v1)*-1)+-4)+v2)+-2)*-3)+v3)+-4)+-2)*-1)//8)//4)*-4)+3)+v3)*
            -2)+v2)
-    v1_val, v2_val, v3_val = v1.const_like(8), v2.const_like(3), v3.const_like(2)
+    v1_val, v2_val, v3_val = UOp.const(8), UOp.const(3), UOp.const(2)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -70,7 +71,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 5)
    v3=Variable('v3', 0, 128)
    expr = (((((((((((((((((((((((((((((0+v2)*-4)+0)//9)+-4)*-2)*3)*4)//9)+v3)+v1)//4)+v1)+v3)+-1)*4)//4)+v2)//7)//3)+v1)+v2)+v3)+1)*2)//4)*3)+-1)*1)
-    v1_val, v2_val, v3_val = v1.const_like(0), v2.const_like(2), v3.const_like(65)
+    v1_val, v2_val, v3_val = UOp.const(0), UOp.const(2), UOp.const(65)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -80,7 +81,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 8)
    v3=Variable('v3', 0, 9)
    expr = (((((((0+-1)+2)+v1)*-2)//3)+v1)*-4)
-    v1_val, v2_val, v3_val = v1.const_like(0), v2.const_like(0), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(0), UOp.const(0), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -90,7 +91,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 1)
    v3=Variable('v3', 0, 8)
    expr = (((((((((((((((((((((((((((((0*-2)//1)+3)*-2)+-3)*-4)*1)+v1)+0)%2)%8)%9)+v2)%9)+-4)//4)+-1)*-2)+0)+v1)+v1)+3)+v1)+4)+-4)+0)*2)+-3)%6)
-    v1_val, v2_val, v3_val = v1.const_like(0), v2.const_like(1), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(0), UOp.const(1), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -101,7 +102,7 @@ class TestFuzzFailure(unittest.TestCase):
    v3=Variable("v3", 0, 32)
    x5 = (v1 <= 9).where(v1 * -4 - 4, v1 // 9) // 9
    expr = ((x5 >= -4).where(x5, (v2 % 3 + v2) // 5) * -1).maximum(((v1 * -2) % 6 + v3 % 1) * -1) * -1
-    v1_val, v2_val, v3_val = v1.const_like(9), v2.const_like(0), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(9), UOp.const(0), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -111,7 +112,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable("v2", 0, 128)
    v3=Variable("v3", 0, 5)
    expr = (((v2 * 0).maximum(8) - v2 * 2) % 5 + v1 // 6 + v1 + 5) % 5
-    v1_val, v2_val, v3_val = v1.const_like(0), v2.const_like(7), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(0), UOp.const(7), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    self.assertEqual(num, rn)
@@ -60,7 +60,7 @@ class TestTensorUOpClone(unittest.TestCase):
    t = _t(3, 4).float()
    self.assertIs(_strip_unique(t.clone().uop), _strip_unique(t.uop.clone()))
  def test_clone_deviceless_const(self):
-    u = UOp.const(dtypes.float, 2.0)
+    u = UOp.const(2.0)
    self.assertIs(_strip_unique(Tensor(u).clone().uop), _strip_unique(u.clone()))

 class TestTensorUOpGradient(unittest.TestCase):
@@ -382,7 +382,7 @@ class TestTensorUOpStack(unittest.TestCase):
    self.assertIs(_t(2, 3).uop.stack(w.uop).dtype, dtypes.float32)
  def test_stack_index_dtype(self):
    # index is outside the promotion lattice, equal dtypes bypass promotion
-    self.assertEqual(UOp.const(dtypes.weakint, 1).stack(UOp.const(dtypes.weakint, 2)).shape, (2,))
+    self.assertEqual(UOp.const(1).stack(UOp.const(2)).shape, (2,))

 class TestTensorUOpConv2d(unittest.TestCase):
  def test_conv2d_basic(self):
@@ -10,7 +10,7 @@ class TestTranscendentalFunctions(unittest.TestCase):
    # TODO: Test constant input when constant folding is fixed (or maybe test both variants)
    # Load input value from a buffer to prevent constant folding
    input_buf = UOp.param(1, dtypes.double, (1,))
-    loaded_value = input_buf.index(UOp.const(dtypes.int, 0)).load()
+    loaded_value = input_buf.index(UOp.const(0)).load()
    def eval_payne_hanek_reduction(v:float) -> tuple[float, int]:
      return tuple(eval_uop(u, [(dtypes.float64, [v])]) for u in payne_hanek_reduction(loaded_value))

@@ -27,48 +27,48 @@ class TestTranscendentalFunctions(unittest.TestCase):
    np.testing.assert_equal(q, 4)

  def test_cody_waite_reduction(self):
-    r, q = (eval_uop(u) for u in cody_waite_reduction(UOp.const(dtypes.float64, 12 * math.pi + 0.1)))
+    r, q = (eval_uop(u) for u in cody_waite_reduction(UOp.const(12 * math.pi + 0.1).cast(dtypes.float64)))
    np.testing.assert_allclose(r, 0.1)
    np.testing.assert_equal(q, 12)

  def test_frexp(self):
    for x in (1, -1):
-      mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(dtypes.float64, x)))
+      mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(float(x)).cast(dtypes.float64)))
      np.testing.assert_equal(mantissa, 0.5)
      np.testing.assert_equal(exponent, 1)

    for x in (2, -2):
-      mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(dtypes.float64, 2.0)))
+      mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(2.0).cast(dtypes.float64)))
      np.testing.assert_equal(mantissa, 0.5)
      np.testing.assert_equal(exponent, 2)

-    mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(dtypes.float64, 5.0)))
+    mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(5.0).cast(dtypes.float64)))
    np.testing.assert_equal(mantissa, 0.625)
    np.testing.assert_equal(exponent, 3)

-    mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(dtypes.float64, 1000.0)))
+    mantissa, exponent = (eval_uop(u) for u in frexp(UOp.const(1000.0).cast(dtypes.float64)))
    np.testing.assert_allclose(mantissa, 0.9765625)
    np.testing.assert_equal(exponent, 10)

  def test_rintk(self):
-    np.testing.assert_allclose(eval_uop(rintk(UOp.const(dtypes.float, 0.0))), 0)
-    np.testing.assert_allclose(eval_uop(rintk(UOp.const(dtypes.float, 5.0))), 5)
-    np.testing.assert_allclose(eval_uop(rintk(UOp.const(dtypes.float, 5.5))), 6)
-    np.testing.assert_allclose(eval_uop(rintk(UOp.const(dtypes.float, 5.999))), 6)
-    np.testing.assert_allclose(eval_uop(rintk(UOp.const(dtypes.float, -5.0))), -5)
-    np.testing.assert_allclose(eval_uop(rintk(UOp.const(dtypes.float, -5.5))), -6)
-    np.testing.assert_allclose(eval_uop(rintk(UOp.const(dtypes.float, -5.999))), -6)
+    np.testing.assert_allclose(eval_uop(rintk(UOp.const(0.0).cast(dtypes.float))), 0)
+    np.testing.assert_allclose(eval_uop(rintk(UOp.const(5.0).cast(dtypes.float))), 5)
+    np.testing.assert_allclose(eval_uop(rintk(UOp.const(5.5).cast(dtypes.float))), 6)
+    np.testing.assert_allclose(eval_uop(rintk(UOp.const(5.999).cast(dtypes.float))), 6)
+    np.testing.assert_allclose(eval_uop(rintk(UOp.const(-5.0).cast(dtypes.float))), -5)
+    np.testing.assert_allclose(eval_uop(rintk(UOp.const(-5.5).cast(dtypes.float))), -6)
+    np.testing.assert_allclose(eval_uop(rintk(UOp.const(-5.999).cast(dtypes.float))), -6)

  def test_pow2if(self):
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, 0), dtypes.float)), 1.0)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, 1), dtypes.float)), 2.0)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, 2), dtypes.float)), 4.0)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, 10), dtypes.float)), 1024.0)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, 63), dtypes.float)), 2**63)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, -1), dtypes.float)), 0.5)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, -2), dtypes.float)), 0.25)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, -10), dtypes.float)), 2**-10)
-    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, -63), dtypes.float)), 2**-63)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(0).cast(dtypes.int), dtypes.float)), 1.0)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(1).cast(dtypes.int), dtypes.float)), 2.0)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(2).cast(dtypes.int), dtypes.float)), 4.0)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(10).cast(dtypes.int), dtypes.float)), 1024.0)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(63).cast(dtypes.int), dtypes.float)), 2**63)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(-1).cast(dtypes.int), dtypes.float)), 0.5)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(-2).cast(dtypes.int), dtypes.float)), 0.25)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(-10).cast(dtypes.int), dtypes.float)), 2**-10)
+    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(-63).cast(dtypes.int), dtypes.float)), 2**-63)

 if __name__ == '__main__':
  unittest.main()
@@ -7,35 +7,35 @@ from tinygrad.uop.symbolic import sym
 from test.helpers import to_uops_list

 simple_pm = PatternMatcher([
-  (UPat.cvar('x', dtypes.int), lambda x: UOp.const(dtypes.float, 1.0) + UOp.const(dtypes.float, 2.0)),
-  (UPat.cvar('x') + UPat.cvar('y'), lambda x,y: UOp.const(dtypes.float, x.arg+y.arg)),
-  (UPat.cvar('x') * UPat.cvar('y') * UPat.cvar('z'), lambda x,y,z: UOp.const(dtypes.float, x.arg*y.arg*z.arg)),
-  ((UPat.var('x') + UPat.cvar('c1')) + UPat.cvar('c2'), lambda x,c1,c2: x + (c1.arg+c2.arg)),
+  (UPat.cvar('x', dtypes.weakint), lambda x: UOp.const(1.0) + UOp.const(2.0)),
+  (UPat.cvar('x') + UPat.cvar('y'), lambda x,y: UOp.const(x.val+y.val)),
+  (UPat.cvar('x') * UPat.cvar('y') * UPat.cvar('z'), lambda x,y,z: UOp.const(x.val*y.val*z.val)),
+  ((UPat.var('x') + UPat.cvar('c1')) + UPat.cvar('c2'), lambda x,c1,c2: x + (c1.val+c2.val)),
 ])

 def const_values(u:UOp):
-  if u.op is Ops.CONST: return (u.arg,) if not isinstance(u.arg, tuple) else u.arg
-  if u.op is Ops.STACK: return tuple(x.arg for x in u.src)
+  if u.op is Ops.CONST: return (u.val,)
+  if u.op is Ops.STACK: return tuple(x.val for x in u.src)
  raise AssertionError(f"expected const-like UOp, got {u.op}")

 class TestGraphRewriteConst(unittest.TestCase):
  def test_gep_const(self):
-    v1 = UOp.const(dtypes.int, (0,1,2))
+    v1 = UOp.const((0,1,2), dtypes.int)
    v2 = v1.index(1)
    ret = graph_rewrite(v2, sym)
    self.assertEqual(ret.dtype, dtypes.int)
-    self.assertEqual(ret.arg, 1)
+    self.assertEqual(ret.val, 1)

  def test_add_const(self):
-    v1 = UOp.const(dtypes.int, (0,1,2))
-    v2 = UOp.const(dtypes.int, (5,6,7))
+    v1 = UOp.const((0,1,2))
+    v2 = UOp.const((5,6,7))
    ret = graph_rewrite(v1+v2, sym)
    self.assertEqual(ret.op, Ops.STACK)
    self.assertEqual(const_values(ret), (5,7,9))

  def test_add_const_lose_v(self):
-    v1 = UOp.const(dtypes.int, (0,1,2))
-    v2 = UOp.const(dtypes.int, (2,1,0))
+    v1 = UOp.const((0,1,2))
+    v2 = UOp.const((2,1,0))
    ret = graph_rewrite(v1+v2, sym)
    self.assertEqual(ret.op, Ops.STACK)
    self.assertEqual(const_values(ret), (2,2,2))
@@ -49,47 +49,47 @@ class TestModularWraparound(unittest.TestCase):
    self.assertEqual(len(results), 2)  # +1 for SINK
    self.assertEqual(results[0].op, Ops.CONST)
    self.assertEqual(results[0].dtype, uop.dtype)
-    self.assertEqual(results[0].arg, expected)
+    self.assertEqual(results[0].val, expected)

  @xfail_broken_const_wraparound
  def test_cast(self):
    t = self._test
-    t(UOp.const(dtypes.uint, 0xABCD17D6).cast(dtypes.uint8), 0xD6)
-    t(UOp.const(dtypes.uint, 0xABCD17D6).cast(dtypes.uint8).cast(dtypes.uint), 0xD6)
+    t(UOp.const(0xABCD17D6, dtypes.uint).cast(dtypes.uint8), 0xD6)
+    t(UOp.const(0xABCD17D6, dtypes.uint).cast(dtypes.uint8).cast(dtypes.uint), 0xD6)

  @xfail_broken_const_wraparound
  def test_mul(self):
    t = self._test
-    t(UOp.const(dtypes.uint, 0xABCD17D6) * 0xAABBCCDD, 1147018174)
-    t(UOp.const(dtypes.int, 0xABCD17D6) * 10, -1241321892)
+    t(UOp.const(0xABCD17D6, dtypes.uint) * 0xAABBCCDD, 1147018174)
+    t(UOp.const(0xABCD17D6, dtypes.int) * 10, -1241321892)

  @xfail_broken_const_wraparound
  def test_div(self):
    t = self._test
-    t(UOp.const(dtypes.uint, 0xABCD17D6) * 0xAABBCCDD // 11, 104274379)
-    t(UOp.const(dtypes.int, 0xABCD17D6) * 10 // 11, -112847444)
+    t(UOp.const(0xABCD17D6, dtypes.uint) * 0xAABBCCDD // 11, 104274379)
+    t(UOp.const(0xABCD17D6, dtypes.int) * 10 // 11, -112847444)

  @xfail_broken_const_wraparound
  def test_neg(self):
    t = self._test
-    t(-UOp.const(dtypes.uint8, 1), 0xFF)
-    t(-UOp.const(dtypes.uint16, 1), 0xFFFF)
-    t(-UOp.const(dtypes.uint32, 1), 0xFFFFFFFF)
-    t(-UOp.const(dtypes.uint64, 1), 0xFFFFFFFFFFFFFFFF)
+    t(-UOp.const(1, dtypes.uint8), 0xFF)
+    t(-UOp.const(1, dtypes.uint16), 0xFFFF)
+    t(-UOp.const(1, dtypes.uint32), 0xFFFFFFFF)
+    t(-UOp.const(1, dtypes.uint64), 0xFFFFFFFFFFFFFFFF)

  @xfail_broken_const_wraparound
  def test_neg_min_int(self):
    t = self._test
-    t(-UOp.const(dtypes.int8, -2**7), -2**7)
-    t(-UOp.const(dtypes.int16, -2**15), -2**15)
-    t(-UOp.const(dtypes.int32, -2**31), -2**31)
-    t(-UOp.const(dtypes.int64, -2**63), -2**63)
+    t(-UOp.const(-2**7, dtypes.int8), -2**7)
+    t(-UOp.const(-2**15, dtypes.int16), -2**15)
+    t(-UOp.const(-2**31, dtypes.int32), -2**31)
+    t(-UOp.const(-2**63, dtypes.int64), -2**63)

  @xfail_broken_const_wraparound
  def test_payne_hanek_reduction_bug(self):
    t = self._test
-    a = (UOp.const(dtypes.uint, 43748177600).cast(dtypes.uint) | 36).cast(dtypes.ulong)
-    b = 2536655455 * a + 4294967296 * UOp.const(dtypes.ulong, 25366554550)
+    a = (UOp.const(43748177600, dtypes.uint).cast(dtypes.uint) | 36).cast(dtypes.ulong)
+    b = 2536655455 * a + 4294967296 * UOp.const(25366554550, dtypes.ulong)
    c = (b + 2261737165) // 4611686018427387904
    t(c, 0)

@@ -103,67 +103,67 @@ class TestGraphRewrite(unittest.TestCase):
  # NOTE: this shows why we can't have a UOp in arg
  @unittest.expectedFailure
  def test_no_dedup_args(self):
-    a1 = UOp.variable("a1", UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 11), dtypes.int)
-    a2 = UOp.variable("a2", UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 11), dtypes.int)
+    a1 = UOp.variable("a1", UOp.const(0), UOp.const(11), dtypes.int)
+    a2 = UOp.variable("a2", UOp.const(0), UOp.const(11), dtypes.int)
    sink = a1.sink(a2)
    variables = [x for x in graph_rewrite(sink, PatternMatcher([])).toposort() if x.op is Ops.PARAM and x.addrspace is AddrSpace.ALU]
    self.assertEqual(len(variables), 1)

  def test_simple(self):
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
    nout = graph_rewrite(c1+c2, simple_pm)
    self.assertEqual(nout.op, Ops.CONST)
-    self.assertEqual(nout.arg, 3.0)
+    self.assertEqual(nout.val, 3.0)

  def test_depth_2_late(self):
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp.const(dtypes.float, 3.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
+    c3 = UOp.const(3.0)
    nout = graph_rewrite(c1*c2*(c3+c3), simple_pm)
    self.assertEqual(nout.op, Ops.CONST)
-    self.assertEqual(nout.arg, 12.0)
+    self.assertEqual(nout.val, 12.0)

  def test_double(self):
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp.const(dtypes.float, 3.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
+    c3 = UOp.const(3.0)
    nout = graph_rewrite(c1+c2+c3, simple_pm)
    self.assertEqual(nout.op, Ops.CONST)
-    self.assertEqual(nout.arg, 6.0)
+    self.assertEqual(nout.val, 6.0)

  def test_triple(self):
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp.const(dtypes.float, 3.0)
-    c4 = UOp.const(dtypes.float, 4.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
+    c3 = UOp.const(3.0)
+    c4 = UOp.const(4.0)
    nout = graph_rewrite(c1+c2+c3+c4, simple_pm)
    self.assertEqual(nout.op, Ops.CONST)
-    self.assertEqual(nout.arg, 10.0)
+    self.assertEqual(nout.val, 10.0)

  def test_diamond(self):
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
-    c3 = UOp.const(dtypes.float, 3.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
+    c3 = UOp.const(3.0)
    nout = graph_rewrite((c1+c2)+(c1+c3), simple_pm)
    self.assertEqual(nout.op, Ops.CONST)
-    self.assertEqual(nout.arg, 7.0)
+    self.assertEqual(nout.val, 7.0)

  def test_magic_4(self):
-    c1 = UOp.const(dtypes.int, 4.0)
+    c1 = UOp.const(4)
    nout = graph_rewrite(c1, simple_pm)
    self.assertEqual(nout.op, Ops.CONST)
-    self.assertEqual(nout.arg, 3.0)
+    self.assertEqual(nout.val, 3.0)

  def test_depth_2_fold(self):
    v = UOp.variable("v", 0, 1, dtypes.float)
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    c1 = UOp.const(1.0)
+    c2 = UOp.const(2.0)
    nout = graph_rewrite(v+c1+c2, simple_pm)
    self.assertEqual(nout.op, Ops.ADD)
    self.assertEqual(nout.src[0].op, Ops.PARAM)
    self.assertEqual(nout.src[1].op, Ops.CONST)
-    self.assertEqual(nout.src[1].arg, 3.0)
+    self.assertEqual(nout.src[1].val, 3.0)

  def test_commutative_work(self):
    a = UOp.variable('a', 0, 1)
@@ -174,14 +174,14 @@ class TestGraphRewrite(unittest.TestCase):
    a = UOp.variable('a', 0, 1)
    tst = (2+a).simplify()
    self.assertIs(tst.src[0], a)
-    self.assertIs(tst.src[1], a.const_like(2))
+    self.assertIs(tst.src[1], UOp.const(2))

  def test_consts_go_last(self):
    a = UOp.variable('a', 0, 1)
    b = UOp.variable('b', 0, 1)
    c = UOp.variable('c', 0, 1)
    d = UOp.variable('d', 0, 1)
-    outs = [2+a, 2+a+d+3+b+c+4, a.const_like(2)+a, (4+d)+c+(2+a)+b]
+    outs = [2+a, 2+a+d+3+b+c+4, UOp.const(2)+a, (4+d)+c+(2+a)+b]
    for out in outs:
      sink = graph_rewrite(out, sym)
      print(sink.render())
@@ -191,67 +191,75 @@ class TestGraphRewrite(unittest.TestCase):

 class TestUOpGraph(unittest.TestCase):
  def test_add_constant_fold(self):
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    c1 = UOp.const(1.0, dtypes.float)
+    c2 = UOp.const(2.0, dtypes.float)
    out = c1+c2
    uops = to_uops_list([out])
    self.assertEqual(len(uops), 2)  # +1 for SINK
    out = uops[-2]
    self.assertEqual(out.op, Ops.CONST)
-    self.assertEqual(out.arg, 3.0)
+    self.assertEqual(out.val, 3.0)

  def test_where_same_fold(self):
    v = UOp.variable('tmp', 0, 1)
-    c0 = UOp.const(dtypes.weakint, 0)
+    c0 = UOp.const(0)
    vc = v != c0
-    c1 = UOp.const(dtypes.float, 1.0)
+    c1 = UOp.const(1.0, dtypes.float)
    out = vc.where(c1, c1)
    uops = to_uops_list([out])
    self.assertEqual(len(uops), 2)  # +1 for SINK
    out = uops[-2]
    self.assertEqual(out.op, Ops.CONST)
-    self.assertEqual(out.arg, 1.0)
+    self.assertEqual(out.val, 1.0)

  def test_where_const_fold(self):
-    bf = UOp.const(dtypes.bool, False)
-    c1 = UOp.const(dtypes.float, 1.0)
-    c2 = UOp.const(dtypes.float, 2.0)
+    bf = UOp.const(False)
+    c1 = UOp.const(1.0, dtypes.float)
+    c2 = UOp.const(2.0, dtypes.float)
    out = bf.where(c1, c2)
    uops = to_uops_list([out])
    self.assertEqual(len(uops), 2)  # +1 for SINK
    out = uops[-2]
    self.assertEqual(out.op, Ops.CONST)
-    self.assertEqual(out.arg, 2.0)
+    self.assertEqual(out.val, 2.0)

  def test_const_cast(self):
-    bf = UOp.const(dtypes.bool, False)
+    bf = UOp.const(False)
    out = bf.cast(dtypes.int)
    uops = to_uops_list([out])
    self.assertEqual(len(uops), 2)  # +1 for SINK
    out = uops[-2]
    self.assertEqual(out.op, Ops.CONST)
-    self.assertEqual(out.arg, 0)
+    self.assertEqual(out.val, 0)

  def test_const_bitcast(self):
-    bf = UOp.const(dtypes.float, 1.0)
+    bf = UOp.const(1.0, dtypes.float)
    out = bf.bitcast(dtypes.uint32)
    uops = to_uops_list([out])
    self.assertEqual(len(uops), 2)  # +1 for SINK
    out = uops[-2]
    self.assertEqual(out.op, Ops.CONST)
-    self.assertEqual(out.arg, 0x3F800000)
+    self.assertEqual(out.val, 0x3F800000)

  @unittest.expectedFailure
  def test_const_shape_change_bitcast(self):
-    bf = UOp.const(dtypes.uint8, 0x3F)
+    bf = UOp.const(0x3F).cast(dtypes.uint8)
    out = bf.bitcast(dtypes.half)
    uops = to_uops_list([out])
    self.assertEqual(len(uops), 2)  # +1 for SINK

+  def test_devectorize_derives_lane_dtype(self):
+    from tinygrad.codegen import do_devectorize
+    # an Invalid lane derives bool while the value lane derives float: the lane rebuild must derive, not inherit
+    lhs = UOp.stack(UOp.invalid(), UOp.const(1.0).cast(dtypes.float))
+    out = do_devectorize(lhs * lhs)
+    invalid_lane_mul = next(u for u in out.src[0].toposort() if u.op is Ops.MUL)
+    self.assertIs(invalid_lane_mul.dtype, dtypes.bool)
+
  @unittest.skip("this test isn't valid uops")
  def test_noop_vectorize_fold(self):
    d0 = UOp.param(0, dtypes.float, (1,))
-    idx = UOp.const(dtypes.int, 0)
+    idx = UOp.const(0)
    ld = d0.load(idx, dtype=dtypes.float)
    vec = UOp(Ops.STACK, dtypes.float, (ld,))
    x = vec.index(0)
@@ -265,7 +273,7 @@ class TestUOpGraph(unittest.TestCase):
    d0 = UOp.param(0, dtypes.float, (1,))
    d1 = UOp.param(1, dtypes.float, (1,))
    d2 = UOp.param(2, dtypes.float, (1,))
-    idx = UOp.const(dtypes.int, 0)
+    idx = UOp.const(0)
    def _test_vec(geps, count=4):
      vec = UOp(Ops.STACK, dtypes.float, geps)
      out = d0.index(idx).store(vec)
@@ -302,7 +310,7 @@ class TestUOpGraph(unittest.TestCase):

  def test_gep_vec_const_fold(self):
    for vec_size in [2, 4, 8]:
-      consts = [UOp.const(dtypes.float, float(i)) for i in range(vec_size)]
+      consts = [UOp.const(float(i), dtypes.float) for i in range(vec_size)]
      vec = UOp(Ops.STACK, src=tuple(consts))
      with Context(SPEC=0):
        uops = to_uops_list([vec.index(i) for i in range(vec_size)])
@@ -312,7 +320,7 @@ class TestUOpGraph(unittest.TestCase):
  def test_cast_alu_fold(self):
    d0 = UOp.param(0, dtypes.bool, (1,))
    d1 = UOp.param(1, dtypes.int, (1,))
-    idx = UOp.const(dtypes.int, 0)
+    idx = UOp.const(0)
    ld = d1.index(idx)
    alu = (ld<1).cast(dtypes.bool)
    out = d0.index(idx).store(alu)
@@ -322,7 +330,7 @@ class TestUOpGraph(unittest.TestCase):
  def test_double_cast_fold(self):
    d0 = UOp.param(0, dtypes.float, (1,))
    d1 = UOp.param(1, dtypes.int, (1,))
-    idx = UOp.const(dtypes.int, 0)
+    idx = UOp.const(0, dtypes.int)
    ld = d1.index(idx)
    alu = ld.cast(dtypes.float).cast(dtypes.float)
    out = d0.index(idx).store(alu)
@@ -331,8 +339,8 @@ class TestUOpGraph(unittest.TestCase):

  def test_depth_2_const_fold(self):
    v = UOp.variable("tmp", 0, 1, dtypes.int)
-    c2 = UOp.const(dtypes.int, 2)
-    c4 = UOp.const(dtypes.int, 4)
+    c2 = UOp.const(2, dtypes.int)
+    c4 = UOp.const(4, dtypes.int)
    vc = v+c2
    out = vc+c4
    uops = to_uops_list([out])
@@ -340,19 +348,19 @@ class TestUOpGraph(unittest.TestCase):
    out = uops[-2]  # -2 to skip SINK
    self.assertEqual(out.op, Ops.ADD)
    self.assertEqual(out.src[1].op, Ops.CONST)
-    self.assertEqual(out.src[1].arg, 6)
+    self.assertEqual(out.src[1].val, 6)

  def test_bitcast_to_same_dtype_fold(self):
    for dt in dtypes.ints + dtypes.floats + (dtypes.bool,):
      d0 = UOp.param(0, dt, (1,))
-      v = d0.index(UOp.const(dtypes.int, 0))
+      v = d0.index(UOp.const(0))
      uops = to_uops_list([v.bitcast(dt)])
      self.assertEqual(len([x for x in uops if x.op is Ops.BITCAST and x.dtype is dt]), 0, f"dtype = {dt}")

  def test_sub_with_cast_folds(self):
    a = Variable("a", 0, 5)
    uops = to_uops_list([a.cast(dtypes.int)+(-a).cast(dtypes.int)])
-    assert uops[0] == UOp.const(dtypes.int, 0)
+    assert uops[0] == UOp.const(0, dtypes.int)
    assert uops[-1].op == Ops.SINK

  def test_where_on_gated_load_fold(self):
@@ -364,7 +372,7 @@ class TestUOpGraph(unittest.TestCase):
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
      assert u.op is not Ops.WHERE
-      if u.op is Ops.LOAD and u.src[0].src[0].op is Ops.PARAM: assert u.src[1].arg==5
+      if u.op is Ops.LOAD and u.src[0].src[0].op is Ops.PARAM: assert u.src[1].val==5

  def test_where_on_gated_load_folds_swapped_branches(self):
    ridx0 = UOp.range(100, 0)
@@ -374,7 +382,7 @@ class TestUOpGraph(unittest.TestCase):
    uops = to_uops_list([w])
    for u in uops:
      assert u.op is not Ops.WHERE
-      if u.op is Ops.LOAD: assert u.src[1].arg==5
+      if u.op is Ops.LOAD: assert u.src[1].val==5

  def test_where_on_gated_load_with_cast(self):
    ridx0 = UOp.range(100, 0)
@@ -386,13 +394,13 @@ class TestUOpGraph(unittest.TestCase):
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
      assert u.op is not Ops.WHERE
-      if u.op is Ops.LOAD and u.src[0].src[0].op is Ops.PARAM: assert u.src[1].arg == 5
+      if u.op is Ops.LOAD and u.src[0].src[0].op is Ops.PARAM: assert u.src[1].val == 5

  def test_where_on_casted_gated_load_extra_cond(self):
    ridx0 = UOp.range(100, 0)
    d0 = UOp.param(0, dtypes.float, (100,))
    ld = d0.index(ridx0.valid(ridx0<50))
-    w = ((ridx0<50) & (ridx0>30)).where(ld, UOp.const(dtypes.float, 0)).cast(dtypes.half)
+    w = ((ridx0<50) & (ridx0>30)).where(ld, UOp.const(0.0)).cast(dtypes.half)
    out = UOp.param(1, dtypes.half, (100,))
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
@@ -402,7 +410,7 @@ class TestUOpGraph(unittest.TestCase):
    ridx0 = UOp.range(100, 0)
    d0 = UOp.param(0, dtypes.float, (100,))
    ld = d0.index(ridx0.valid(ridx0<50))
-    w = ((ridx0<50) & (ridx0>30)).where(UOp.const(dtypes.float, 0), ld).cast(dtypes.half)
+    w = ((ridx0<50) & (ridx0>30)).where(UOp.const(0.0), ld).cast(dtypes.half)
    out = UOp.param(1, dtypes.half, (100,))
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
@@ -418,22 +426,22 @@ class TestUOpGraph(unittest.TestCase):
    uops = to_uops_list([st])
    for u in uops:
      assert u.op is not Ops.WHERE
-      if u.op is Ops.STORE: assert u.src[1].arg==5
+      if u.op is Ops.STORE: assert u.src[1].val==5

  def test_load_idx_becomes_int(self):
    # mnist indexing with split reduceop
    # Make sure we are not doign math on the loaded index, which would promote it to long
    c0 = UOp.param(0, dtypes.uchar, (128000,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
+    c1 = UOp.range(UOp.const(512), 1, AxisType.WEAK)
+    c2 = UOp.range(UOp.const(250), 2, AxisType.WEAK)
    c3 = UOp.param(1, dtypes.int, (512,))
    c4 = c3.index(c1)
-    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
-    c6 = ((c2*UOp.const(dtypes.weakint, 240))+c5)
+    c5 = UOp.range(UOp.const(240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(240))+c5)
    c7 = UOp.param(2, dtypes.uchar, (60000,))
    c8 = c7.index(c6)
    c9 = ((c4<0).where((c4+60000), c4)!=c6.cast(dtypes.int)).where(0, c8.cast(dtypes.uint).cast(dtypes.uchar)).reduce(c5, arg=Ops.ADD)
-    c10 = c0.index(((c1*UOp.const(dtypes.weakint, 250))+c2)).store(c9).end(c1, c2)
+    c10 = c0.index(((c1*UOp.const(250))+c2)).store(c9).end(c1, c2)
    uops = to_uops_list([c10])
    for u in uops:
      self.assertNotEqual(u.dtype, dtypes.long)
@@ -441,19 +449,19 @@ class TestUOpGraph(unittest.TestCase):
  def test_load_idx_no_math_on_loaded(self):
    # test the (x+y)<c pattern where x has loads - we shouldn't do math on loaded indices
    c0 = UOp.param(0, dtypes.uchar, (128000,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
+    c1 = UOp.range(UOp.const(512), 1, AxisType.WEAK)
+    c2 = UOp.range(UOp.const(250), 2, AxisType.WEAK)
    c3 = UOp.param(1, dtypes.int, (512,))
    c4 = c3.index(c1)  # c4 is a load
-    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
-    c6 = ((c2*UOp.const(dtypes.weakint, 240))+c5)
+    c5 = UOp.range(UOp.const(240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(240))+c5)
    c7 = UOp.param(2, dtypes.uchar, (60000,))
    c8 = c7.index(c6)
    # (loaded + range) < const pattern - loaded value shouldn't be promoted to long
    loaded_idx = c4.cast(dtypes.weakint)
-    comparison = (loaded_idx + c5) < UOp.const(dtypes.weakint, 60000)
+    comparison = (loaded_idx + c5) < UOp.const(60000)
    c9 = comparison.where(c8.cast(dtypes.uint).cast(dtypes.uchar), 0).reduce(c5, arg=Ops.ADD)
-    c10 = c0.index(((c1*UOp.const(dtypes.weakint, 250))+c2)).store(c9).end(c1, c2)
+    c10 = c0.index(((c1*UOp.const(250))+c2)).store(c9).end(c1, c2)
    uops = to_uops_list([c10])
    for u in uops:
      self.assertNotEqual(u.dtype, dtypes.long)
@@ -462,9 +470,9 @@ class TestUOpGraph(unittest.TestCase):
    glbl0 = UOp.param(0, dtypes.int, (1,))
    glbl1 = UOp.param(1, dtypes.int, (1,))
    glbl2 = UOp.param(2, dtypes.int, (1,))
-    idx = UOp.const(dtypes.int, 0)
+    idx = UOp.const(0)
    ld0 = glbl1.index(UOp.invalid())
-    ld1 = glbl2.index(idx.valid(UOp.const(dtypes.bool, True)))
+    ld1 = glbl2.index(idx.valid(UOp.const(True)))
    uops = to_uops_list([glbl0.index(idx).store(ld1+ld0)])
    # the gate and invalid value are deleted from ld1
    self.assertEqual(len([u for u in uops if u.op is Ops.LOAD]), 1)
@@ -476,7 +484,7 @@ class TestUOpGraph(unittest.TestCase):
    st = smem.index(lidx).store(glbl0.index(lidx).load())
    barrier = st.barrier()
    ld0 = smem.after(barrier).index(UOp.invalid())
-    ld1 = smem.after(barrier).index((lidx+2).valid(UOp.const(dtypes.bool, True)))
+    ld1 = smem.after(barrier).index((lidx+2).valid(UOp.const(True)))
    uops = to_uops_list([glbl0.index(lidx).store(ld1+ld0)])

    # the gate and invalid value are deleted from ld1
@@ -484,10 +492,10 @@ class TestUOpGraph(unittest.TestCase):

  def test_fold_gated_store(self):
    glbl = UOp.param(0, dtypes.int, (1,))
-    idx0 = UOp.const(dtypes.int, 0)
-    val = UOp.const(dtypes.int, 42)
+    idx0 = UOp.const(0)
+    val = UOp.const(42)
    st0 = glbl.index(UOp.invalid()).store(val)
-    st1 = glbl.index(idx0.valid(UOp.const(dtypes.bool, True))).store(val)
+    st1 = glbl.index(idx0.valid(UOp.const(True))).store(val)
    uops = to_uops_list([st0, st1])
    # only the second store happens
    self.assertEqual(len([u for u in uops if u.op is Ops.STORE]), 1)
@@ -495,9 +503,9 @@ class TestUOpGraph(unittest.TestCase):
  @unittest.skip("this is a uop type error")
  def test_asserts_bad_gate(self):
    glbl0 = UOp.param(0, dtypes.int, (1,))
-    idx = UOp.const(dtypes.int, 0)
-    bad_gate = UOp.const(dtypes.int, 1)
-    with self.assertRaises(AssertionError): to_uops_list([UOp(Ops.STORE, src=(glbl0, idx, UOp.const(dtypes.int, 42), bad_gate))])
+    idx = UOp.const(0)
+    bad_gate = UOp.const(1)
+    with self.assertRaises(AssertionError): to_uops_list([UOp(Ops.STORE, src=(glbl0, idx, UOp.const(42), bad_gate))])

  def test_after_end(self):
    r = UOp.range(10, 0)
@@ -505,7 +513,7 @@ class TestUOpGraph(unittest.TestCase):
    c = r + 1
    self.assertIn(r, c.ranges)

-    e = UOp.const(dtypes.int, 1).end(r)
+    e = UOp.const(1).end(r)
    self.assertNotIn(r, e.ranges)

    a = c.after(e)
@@ -555,7 +563,7 @@ class TestConstBufferize(unittest.TestCase):
    CONST doesn't depend on ranges (constant is same value everywhere).
    """
    from tinygrad.schedule.rangeify import pm_const_buffer_folding, BufferizeOpts
-    c = UOp.const(dtypes.float, 42.0)
+    c = UOp.const(42.0)
    r1 = UOp.range(3, 0)
    bufferize_with_range = UOp(Ops.STAGE, src=(c, r1), arg=BufferizeOpts(device="CPU"))
    self.assertEqual(len(bufferize_with_range.src), 2)  # const + 1 range
@@ -563,13 +571,13 @@ class TestConstBufferize(unittest.TestCase):
    result = graph_rewrite(bufferize_with_range, pm_const_buffer_folding, name='test')
    # BUFFERIZE should be removed, result is const broadcast to shape
    self.assertNotEqual(result.op, Ops.STAGE)
-    const_vals = [u.arg for u in result.toposort() if u.op is Ops.CONST and u.dtype == dtypes.float]
+    const_vals = [u.val for u in result.toposort() if u.op is Ops.CONST and u.dtype is dtypes.weakfloat]
    self.assertIn(42.0, const_vals)

  def test_const_bufferize_with_multiple_ranges(self):
    """Test CONST.BUFFERIZE with multiple ranges is also folded."""
    from tinygrad.schedule.rangeify import pm_const_buffer_folding, BufferizeOpts
-    c = UOp.const(dtypes.float, 3.14)
+    c = UOp.const(3.14)
    r1 = UOp.range(3, 0)
    r2 = UOp.range(4, 1)
    bufferize_with_ranges = UOp(Ops.STAGE, src=(c, r1, r2), arg=BufferizeOpts(device="CPU"))
@@ -578,14 +586,14 @@ class TestConstBufferize(unittest.TestCase):
    result = graph_rewrite(bufferize_with_ranges, pm_const_buffer_folding, name='test')
    # BUFFERIZE should be removed
    self.assertNotEqual(result.op, Ops.STAGE)
-    const_vals = [u.arg for u in result.toposort() if u.op is Ops.CONST and u.dtype == dtypes.float]
+    const_vals = [u.val for u in result.toposort() if u.op is Ops.CONST and u.dtype is dtypes.weakfloat]
    self.assertIn(3.14, const_vals)

 class TestUOpTags(unittest.TestCase):
  def test_inc_by_one(self):
-    g = UOp.const(dtypes.int, 1) + UOp.const(dtypes.int, 1)
+    g = UOp.const(1) + UOp.const(1)
    assert g.ssimplify() == 2
-    pm_plus_1 = PatternMatcher([(UPat(Ops.CONST, name="x"), lambda x: x.replace(arg=x.arg+1, tag=1) if x.tag is None else None)])
+    pm_plus_1 = PatternMatcher([(UPat(Ops.CONST, name="x"), lambda x: UOp.const(x.val+1, x.dtype).rtag(1) if x.tag is None else None)])
    pm_strip_tags = PatternMatcher([(UPat(GroupOp.All, name="x"), lambda x: x.replace(tag=None) if x.tag is not None else None)])
    g = graph_rewrite(g, pm_plus_1)
    assert g.ssimplify() == 4
@@ -674,36 +682,36 @@ class TestUOpGetItem(unittest.TestCase):

 class TestUOpBroadcast(unittest.TestCase):
  def test_broadcast_row(self):
-    a = UOp.const(dtypes.float, 1, shape=(4, 8))
-    b = UOp.const(dtypes.float, 2, shape=(4, 1))
+    a = UOp.const(1.0).expand((4, 8))
+    b = UOp.const(2.0).expand((4, 1))
    c = a + b
    self.assertEqual(c.shape, (4, 8))
    self.assertEqual(c.op, Ops.ADD)

  def test_broadcast_col(self):
-    a = UOp.const(dtypes.float, 1, shape=(4, 8))
-    b = UOp.const(dtypes.float, 2, shape=(1, 8))
+    a = UOp.const(1.0).expand((4, 8))
+    b = UOp.const(2.0).expand((1, 8))
    c = a + b
    self.assertEqual(c.shape, (4, 8))
    self.assertEqual(c.op, Ops.ADD)

  def test_broadcast_lower_dim(self):
-    a = UOp.const(dtypes.float, 1, shape=(4, 8))
-    b = UOp.const(dtypes.float, 2, shape=(8,))
+    a = UOp.const(1.0).expand((4, 8))
+    b = UOp.const(2.0).expand((8,))
    c = a * b
    self.assertEqual(c.shape, (4, 8))
    self.assertEqual(c.op, Ops.MUL)

  def test_broadcast_scalar(self):
-    a = UOp.const(dtypes.float, 1, shape=(4, 8))
+    a = UOp.const(1.0).expand((4, 8))
    c = a * 2
    self.assertEqual(c.shape, (4, 8))
    self.assertEqual(c.op, Ops.MUL)

  def test_broadcast_symbolic_same_shape(self):
    t = Variable("t", 1, 10)
-    a = UOp.const(dtypes.float, 1, shape=(1, 1, t))
-    b = UOp.const(dtypes.float, 2, shape=(1, 1, t))
+    a = UOp.const(1.0).expand((1, 1, t))
+    b = UOp.const(2.0).expand((1, 1, t))
    c = a + b
    self.assertEqual(c.op, Ops.ADD)

@@ -1,34 +1,34 @@
 import unittest
-from tinygrad import UOp, dtypes
+from tinygrad import UOp

 class TestUOpRepr(unittest.TestCase):
  def test_simple_const(self):
-    a = UOp.const(dtypes.int, 42)
-    self.assertEqual(repr(a), "UOp(Ops.CONST, dtypes.int, arg=42, src=())")
+    a = UOp.const(42)
+    self.assertEqual(repr(a), "UOp(Ops.CONST, dtypes.weakint, arg=42, src=())")
  def test_different_consts(self):
-    a, b = UOp.const(dtypes.int, 42), UOp.const(dtypes.int, 3)
+    a, b = UOp.const(42), UOp.const(3)
    expected = (
-      "UOp(Ops.ADD, dtypes.int, arg=None, src=(\n" +
-      "  UOp(Ops.CONST, dtypes.int, arg=42, src=()),\n" +
-      "  UOp(Ops.CONST, dtypes.int, arg=3, src=()),))"
+      "UOp(Ops.ADD, dtypes.weakint, arg=None, src=(\n" +
+      "  UOp(Ops.CONST, dtypes.weakint, arg=42, src=()),\n" +
+      "  UOp(Ops.CONST, dtypes.weakint, arg=3, src=()),))"
    )
    self.assertEqual(repr(a+b), expected)
  def test_walrus_operator_indentation(self):
    # The reference should have the same indentation as the definition
-    a = UOp.const(dtypes.int, 42)
+    a = UOp.const(42)
    expected = (
-      "UOp(Ops.ADD, dtypes.int, arg=None, src=(\n" +
-      "  x0:=UOp(Ops.CONST, dtypes.int, arg=42, src=()),\n" +
+      "UOp(Ops.ADD, dtypes.weakint, arg=None, src=(\n" +
+      "  x0:=UOp(Ops.CONST, dtypes.weakint, arg=42, src=()),\n" +
      "  x0,))"
    )
    self.assertEqual(repr(a+a), expected)
  def test_nested_walrus_indentation(self):
    # Ensure indentation is consistent at multiple levels
-    b = (a:=UOp.const(dtypes.int, 1)) + a
+    b = (a:=UOp.const(1)) + a
    expected = (
-      "UOp(Ops.MUL, dtypes.int, arg=None, src=(\n" +
-      "  x0:=UOp(Ops.ADD, dtypes.int, arg=None, src=(\n" +
-      "    x1:=UOp(Ops.CONST, dtypes.int, arg=1, src=()),\n" +
+      "UOp(Ops.MUL, dtypes.weakint, arg=None, src=(\n" +
+      "  x0:=UOp(Ops.ADD, dtypes.weakint, arg=None, src=(\n" +
+      "    x1:=UOp(Ops.CONST, dtypes.weakint, arg=1, src=()),\n" +
      "    x1,)),\n" +
      "  x0,))"
    )
@@ -4,44 +4,48 @@ from tinygrad.uop.ops import UOp, resolve

 class TestUOpResolve(unittest.TestCase):
  def test_simple_int(self):
-    u = UOp.const(dtypes.int, 4)
+    u = UOp.const(4, dtypes.int)
    self.assertEqual(int(u), 4)

+  def test_weak_const(self):
+    self.assertEqual(int(UOp.const(5)), 5)
+    self.assertEqual(float(UOp.const(1.5)), 1.5)
+
  def test_int_add(self):
-    u = UOp.const(dtypes.int, 4) + 7
+    u = UOp.const(4, dtypes.int) + 7
    self.assertEqual(int(u), 11)

  def test_lt(self):
-    u = UOp.const(dtypes.int, 4) < 7
+    u = UOp.const(4) < 7
    self.assertTrue(u)

  def test_rfloordiv(self):
-    u = 8 // UOp.const(dtypes.int, 4)
+    u = 8 // UOp.const(4, dtypes.int)
    self.assertEqual(int(u), 2)

  def test_rtruediv(self):
-    u = 9 / UOp.const(dtypes.float, 4)
+    u = 9 / UOp.const(4, dtypes.float)
    self.assertEqual(float(u), 2.25)

  def test_leq(self):
-    u = UOp.const(dtypes.int, 4) <= 4
+    u = UOp.const(4) <= 4
    self.assertTrue(u)

  def test_ne(self):
-    u = UOp.const(dtypes.int, 4) != 7
+    u = UOp.const(4) != 7
    self.assertTrue(u)

  def test_ne_f(self):
-    u = UOp.const(dtypes.int, 4) != 4
+    u = UOp.const(4) != 4
    self.assertFalse(u)

  def test_ngt(self):
-    u = UOp.const(dtypes.int, 4) > 7
+    u = UOp.const(4) > 7
    self.assertFalse(u)

  def test_ssimplify(self):
-    self.assertEqual((8 % UOp.const(dtypes.int, 4)).ssimplify(), 0)
-    self.assertEqual((8 * UOp.const(dtypes.int, 4)).ssimplify(), 32)
+    self.assertEqual((8 % UOp.const(4)).ssimplify(), 0)
+    self.assertEqual((8 * UOp.const(4)).ssimplify(), 32)

  def test_ambiguous_less_than(self):
    u = UOp.variable("i", 1, 10)
@@ -52,7 +56,7 @@ class TestUOpResolve(unittest.TestCase):
    self.assertFalse(resolve(u < -1, True))

  def test_float_direct(self):
-    u = UOp.const(dtypes.float, 4.5) + 7
+    u = UOp.const(4.5, dtypes.float) + 7
    self.assertEqual(float(u), 11.5)

  def test_var_cmp_t(self):
@@ -12,13 +12,12 @@ from tinygrad.uop.validate import uops_to_z3
 def check_uop_against_string(self, v:UOp, s:str):
  sym_vars = {v.render():v for v in v.toposort() if v.op in (Ops.RANGE, Ops.SPECIAL, Ops.PARAM)}
  s_eval = eval(s, sym_vars)
-  if isinstance(s_eval, int) and v.dtype==dtypes.weakint: s_eval = UOp.const(dtypes.weakint, s_eval)
-  elif isinstance(s_eval, (bool, int, float)): s_eval = UOp.const(dtypes.from_py(s_eval), s_eval)
+  if isinstance(s_eval, (bool, int, float)): s_eval = UOp.const(s_eval)
  s_eval = graph_rewrite(s_eval, commutative, name="cannonicalize eval")
  self.assertIs(s_eval, v, f"eval did not match simplified: {s_eval} != {v.render()} for {s}")

 def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint): return UOp.variable(name,min_val,max_val,dtype)
-def uconst(val): return UOp.const(dtypes.weakint, val)
+def uconst(val): return UOp.const(val)
 def usum(ops): return functools.reduce(lambda x,y: x+y, ops)
 def uand(ops): return functools.reduce(lambda x,y: x*y, ops)

@@ -104,16 +103,16 @@ class TestSymbolic(unittest.TestCase):
    self.assertEqual(UOp.gcd(a, a*b, a*3).simplify(), a)
    self.assertEqual(UOp.gcd(a*a*a, a*b*a, a*3*a).simplify(), a*a)
    self.assertEqual(UOp.gcd(a*a*10, b*a*5, a*a*5).simplify(), a*5)
-    self.assertEqual(UOp.gcd(a*10, b*5, a*5).simplify(), a.const_like(5))
-    self.assertEqual(UOp.gcd(a, b*5, a*5).simplify(), a.const_like(1))
+    self.assertEqual(UOp.gcd(a*10, b*5, a*5).simplify(), uconst(5))
+    self.assertEqual(UOp.gcd(a, b*5, a*5).simplify(), uconst(1))

  def test_divides_exact(self):
    a = Variable("a", 1, 8)
    b = Variable("b", 1, 8)
    self.assertEqual((a*a*3).divide_exact(a).simplify(), a*3)
-    self.assertEqual((a*a*3).divide_exact(a*a*3).simplify(), a.const_like(1))
-    self.assertEqual((a*a*6).divide_exact(a*a*3).simplify(), a.const_like(2))
-    self.assertEqual((a*b*3).divide_exact(a.const_like(3)).simplify(), a*b)
+    self.assertEqual((a*a*3).divide_exact(a*a*3).simplify(), uconst(1))
+    self.assertEqual((a*a*6).divide_exact(a*a*3).simplify(), uconst(2))
+    self.assertEqual((a*b*3).divide_exact(uconst(3)).simplify(), a*b)
    self.assertEqual((a*a*3).divide_exact(a*(-3)).simplify(), a*-1)
    self.assertEqual((a*a*b*3).divide_exact(a*b).simplify(), a*3)
    self.assertEqual((a*3+a*b).divide_exact(a).simplify(), b+3)
@@ -347,6 +346,12 @@ class TestSymbolic(unittest.TestCase):
  def test_mul_lt(self):
    self.helper_test_variable(Variable("a", 0, 5)*4 < 13, 0, 1, "(a<4)")
    self.helper_test_variable(Variable("a", 0, 5)*4 < 16, 0, 1, "(a<4)")
+    self.helper_test_variable(Variable("a", -5, 5)*4 < -13, 0, 1, "(a<-3)")
+    self.helper_test_variable(Variable("a", -5, 5)*-4 < 13, 0, 1, "((a*-1)<4)")
+    c0, c1 = 2, 2**54+1
+    self.helper_test_variable(Variable("a", 0, c1)*c0 < c1, 0, 1, f"(a<{2**53+1})")
+    c0, c1 = -2, -(2**54-1)
+    self.helper_test_variable(Variable("a", 0, -c1)*c0 < c1, 0, 1, f"((a*-1)<{-(2**53-1)})")
    self.helper_test_variable(Variable("a", 0, 5)*(-2) < 0, 0, 1, "((a*-1)<0)")
    self.helper_test_variable(Variable("a", 0, 5)*4 >= 12, 0, 1, "((a<3)!=True)")
    self.helper_test_variable(Variable("a", 0, 5)*4 >= 13, 0, 1, "((a<4)!=True)")
@@ -877,7 +882,7 @@ class TestSymbolic(unittest.TestCase):
    idx = Variable("idx", 0, 24)
    self.helper_test_variable(idx//4, 0, 6, "(idx//4)")
    # TODO: simplify the true branch
-    self.helper_test_variable((idx<4).where(idx//4, idx.const_like(-1)), -1, 6, "(idx<4).where((idx//4), -1)")
+    self.helper_test_variable((idx<4).where(idx//4, uconst(-1)), -1, 6, "(idx<4).where((idx//4), -1)")

  def test_floordiv_lt(self):
    # x//d<c <=> x<c*d for d>0, and <=> c*d<x for d<0
@@ -936,7 +941,7 @@ class TestSymbolic(unittest.TestCase):

  def test_where_removal(self):
    cond = Variable("a", 0, 3) < 2
-    u1, u0 = cond.const_like(True), cond.const_like(False)
+    u1, u0 = UOp.const(True), UOp.const(False)
    self.helper_test_variable(cond, 0, 1, "(a<2)")
    self.helper_test_variable(cond.where(u1, u0), 0, 1, "(a<2)")
    self.helper_test_variable(cond.where(u1, u0).where(u1, u0), 0, 1, "(a<2)")
@@ -1011,12 +1016,12 @@ class TestSymbolic(unittest.TestCase):

    # TODO: copied from render, render does not support cast
    glbl = UOp.param(0, dtypes.int, (1,))
-    uops = get_uops(UOp(Ops.STORE, src=(glbl.index(UOp.const(dtypes.int, 0)), expr)).sink())
+    uops = get_uops(UOp(Ops.STORE, src=(glbl.index(UOp.const(0, dtypes.int)), expr)).sink())
    rewritten_uop = [uop for uop in uops if uop.op is Ops.STORE][0].src[1]

    # the vars are now scalar PARAMs
    pvar = {u.expr: u for u in rewritten_uop.toposort() if u.op is Ops.PARAM}
-    self.assertEqual(rewritten_uop, (pvar['s']<2).where(pvar['a'].cast(dtypes.half), pvar['b'].cast(dtypes.half)))
+    self.assertEqual(rewritten_uop, (pvar['s']<UOp.const(2, dtypes.int)).where(pvar['a'].cast(dtypes.half), pvar['b'].cast(dtypes.half)))

  def test_where_merge_branches(self):
    cond1 = Variable("s", 0, 10) < 6
@@ -1173,7 +1178,7 @@ class TestSymbolicVariables(unittest.TestCase):
    b = Variable("x", 1, 1).bind(1)
    s = b.simplify()
    self.assertEqual(s.op, Ops.CONST)
-    self.assertEqual(s.arg, 1)
+    self.assertEqual(s.val, 1)

 class TestSymInfer(unittest.TestCase):
  def test_sym_infer(self):
@@ -1197,9 +1202,13 @@ class TestSymInfer(unittest.TestCase):
    # floor: 1 % -1000 = -999, 1 // -1000 = -1
    assert sym_infer(a%b, var_vals) == -999
    assert sym_infer(a//b, var_vals) == -1
+  def test_sym_infer_with_cast(self):
+    a = Variable("a", 0, 100, dtypes.int)
+    assert sym_infer(a.cast(dtypes.long) + 1, {a.expr: 5}) == 6
+    assert sym_infer(a.cast(dtypes.float) * 0.5, {a.expr: 5}) == 2.5
  def test_sym_infer_with_bitcast(self):
    a = Variable("a", 1, 10, dtypes.int)
-    expr = ((a.bitcast(dtypes.uint) << UOp.const(dtypes.uint, 1)).bitcast(dtypes.int) + 2)
+    expr = ((a.bitcast(dtypes.uint) << UOp.const(1)).bitcast(dtypes.int) + 2)
    ret = sym_infer(expr, {a.expr: 2})
    assert isinstance(ret, int)
    assert ret == 6
@@ -1210,7 +1219,7 @@ class TestSymInfer(unittest.TestCase):
    c = Variable("c", 0, 0xFFFFFFFF, dtypes.uint)
    assert sym_infer(c.bitcast(dtypes.int), {c.expr: 0xFFFFFFFF}) == -1

-    assert sym_infer(UOp.const(dtypes.float, 1.5).bitcast(dtypes.uint), {}) == 1069547520
+    assert sym_infer(UOp.const(1.5).cast(dtypes.float).bitcast(dtypes.uint), {}) == 1069547520

  def test_sym_infer_deeply_nested(self):
    # build an expression that exceeds Python's nested parentheses limit for eval
@@ -1331,7 +1340,7 @@ class TestInvalidIndex(unittest.TestCase):
  def test_invalid_times_0(self):
    ridx = Variable("ridx", 0, 10)
    idx = (ridx<5).where(ridx, UOp.invalid())*0
-    self.assertIs(idx.simplify(), (ridx<5).where(UOp.const(dtypes.weakint, 0), UOp.invalid()),
+    self.assertIs(idx.simplify(), (ridx<5).where(uconst(0), UOp.invalid()),
                  "multiplying an index by 0 should preserve the invalid")

  def test_alu_moves_inside_invalid(self):
@@ -1356,24 +1365,24 @@ class TestInvalidIndex(unittest.TestCase):
    self.assertIs((UOp.invalid()<Variable("a",0,10)).simplify().dtype, dtypes.bool)

  def test_alu_invalid_vconst(self):
-    c1 = UOp.const(dtypes.weakint, (1, 1, Invalid, Invalid))
-    c2 = UOp.const(dtypes.weakint, (1, Invalid, 1, 1))
-    self.assertIs((c1+c2).simplify(), UOp.const(dtypes.weakint, (2, Invalid, Invalid, Invalid)))
+    c1 = UOp.const((1, 1, Invalid, Invalid))
+    c2 = UOp.const((1, Invalid, 1, 1))
+    self.assertIs((c1+c2).simplify(), UOp.const((2, Invalid, Invalid, Invalid)))

 class TestStoreLoadFolding(unittest.TestCase):
  """Tests for store(index, load(index)) -> NOOP rule. This rule matches patterns that EMERGE during simplification."""
  def test_store_load_folding(self):
    # store(idx, load(idx)) -> NOOP, including emergent patterns like store(idx, load(idx) + 0)
    buf = UOp.param(0, dtypes.int, (1,))
-    index = buf.index(UOp.const(dtypes.weakint, 0))
+    index = buf.index(UOp.const(0))
    # Direct: store(idx, load(idx)) -> NOOP
    self.assertEqual(graph_rewrite(index.store(index.load()), sym).op, Ops.NOOP)
    # Emergent: store(idx, load(idx) + 0) -> store(idx, load(idx)) -> NOOP
-    self.assertEqual(graph_rewrite(index.store(index.load() + UOp.const(dtypes.int, 0)), sym).op, Ops.NOOP)
+    self.assertEqual(graph_rewrite(index.store(index.load() + UOp.const(0)), sym).op, Ops.NOOP)
    # Emergent: store(idx, load(idx) * 1) -> store(idx, load(idx)) -> NOOP
-    self.assertEqual(graph_rewrite(index.store(index.load() * UOp.const(dtypes.int, 1)), sym).op, Ops.NOOP)
+    self.assertEqual(graph_rewrite(index.store(index.load() * UOp.const(1)), sym).op, Ops.NOOP)
    # Negative: store(idx, load(idx) + 1) should NOT fold
-    self.assertEqual(graph_rewrite(index.store(index.load() + UOp.const(dtypes.int, 1)), sym).op, Ops.STORE)
+    self.assertEqual(graph_rewrite(index.store(index.load() + UOp.const(1)), sym).op, Ops.STORE)

 class TestMoveWhereOnLoad(unittest.TestCase):
  def test_bool_index_preserves_dtype(self):
@@ -1384,7 +1393,7 @@ class TestMoveWhereOnLoad(unittest.TestCase):
    cond = (a < 4) & (r < 2)
    valid = (a < 2)  # pre-existing valid on the load (to pass can_move check for the r-only clause)
    idx = buf.index(a.valid(valid))
-    expr = cond.where(idx, idx.const_like(0))
+    expr = cond.where(idx, UOp.const(0))
    out = graph_rewrite(expr, pm_move_where_on_load)
    type_verify(out, spec_shared)  # Invalid matches any dtype

@@ -1469,7 +1478,7 @@ class TestFuzzFailure(unittest.TestCase):
    v2=Variable('v2', 0, 2)
    v3=Variable('v3', 0, 1)
    expr = (((((((((((((((((((((((0//4)%2)//8)+-2)+-4)+-3)+v1)+-4)+v2)+-2)+v3)+v2)//3)%7)*1)//2)+v2)*-1)+2)+1)+0)+-3)+v3)
-    v1_val, v2_val, v3_val = v1.const_like(8), v2.const_like(0), v3.const_like(0)
+    v1_val, v2_val, v3_val = UOp.const(8), UOp.const(0), UOp.const(0)
    num = expr.simplify().substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    rn = expr.substitute({v1:v1_val, v2:v2_val, v3:v3_val}).ssimplify()
    assert num==rn, f"{num} != {rn}"
@@ -5,12 +5,12 @@ from tinygrad.dtype import dtypes, Invalid
 class TestVminVmaxProperties(unittest.TestCase):
  def test_vmin_vmax_constant(self):
    # vmin and vmax for a constant
-    uop = UOp.const(dtypes.int32, 42)
+    uop = UOp.const(42)
    self.assertEqual(uop.vmin, 42)
    self.assertEqual(uop.vmax, 42)

  def test_vmin_vmax_cmpne(self):
-    uop = UOp.const(dtypes.int32, 42)
+    uop = UOp.const(42)
    def test_bool(u, x):
      self.assertEqual(u.vmin, x)
      self.assertEqual(u.vmax, x)
@@ -81,8 +81,8 @@ class TestVminVmaxProperties(unittest.TestCase):

  def test_vmin_vmax_multiplication_0_inf(self):
    # vmin and vmax for multiplication with a variable
-    x = UOp.const(dtypes.float, 0.0)
-    y = UOp.load(UOp.param(0, dtypes.float, (1,)), UOp.const(dtypes.int, 0), dtype=dtypes.float)
+    x = UOp.const(0.0)
+    y = UOp.load(UOp.param(0, dtypes.float, (1,)), UOp.const(0), dtype=dtypes.float)
    uop = x * y
    # TODO: these should be 0, but definitely should not be nan
    self.assertEqual(uop.vmin, -math.inf)
@@ -162,12 +162,18 @@ class TestVminVmaxProperties(unittest.TestCase):
    self.assertEqual(x_uint.vmin, dtypes.uint.min)
    self.assertEqual(x_uint.vmax, dtypes.uint.max)

+  def test_vmin_vmax_cast_float_to_int(self):
+    self.assertEqual(UOp.variable('x', -4.5, 4.5, dtypes.float).cast(dtypes.int)._min_max, (-4, 4))
+    self.assertEqual(UOp.const(4.5).cast(dtypes.float).cast(dtypes.int)._min_max, (4, 4))
+    x = UOp.const(4.5).cast(dtypes.float)
+    self.assertIs(x.ne(x.cast(dtypes.int).cast(dtypes.float)).simplify().arg, True)
+
  def test_vmin_vmax_invalid(self):
    i = UOp.invalid()
    self.assertNotEqual(i.vmin, i.vmax)

  def test_vmin_vmax_invalid_vconst(self):
-    x = UOp.const(dtypes.weakint, (0, 4, Invalid, Invalid))
+    x = UOp.const((0, 4, Invalid, Invalid))
    self.assertEqual((x.vmin, x.vmax), (0, 4))

 class TestVminVmaxDivMod(unittest.TestCase):
@@ -198,14 +204,14 @@ class TestVminVmaxDivMod(unittest.TestCase):

  def test_vmin_vmax_floordiv_floormod(self):
    x = UOp.variable('x', -7, 7)
-    floordiv = x.alu(Ops.FLOORDIV, x.const_like(3))
+    floordiv = x.alu(Ops.FLOORDIV, UOp.const(3))
    self.assertEqual(floordiv.vmin, -3)
    self.assertEqual(floordiv.vmax, 2)
-    floormod = x.alu(Ops.FLOORMOD, x.const_like(3))
+    floormod = x.alu(Ops.FLOORMOD, UOp.const(3))
    self.assertEqual(floormod.vmin, 0)
    self.assertEqual(floormod.vmax, 2)
    # negative const divisor: floormod range is [c+1, 0]
-    floormod_neg = x.alu(Ops.FLOORMOD, x.const_like(-3))
+    floormod_neg = x.alu(Ops.FLOORMOD, UOp.const(-3))
    self.assertEqual(floormod_neg.vmin, -2)
    self.assertEqual(floormod_neg.vmax, 0)

@@ -286,44 +292,44 @@ class TestVminVmaxDivMod(unittest.TestCase):
 class TestVminVmaxVConst(unittest.TestCase):
  def test_vmin_vmax_vconst_single_element(self):
    # vmin and vmax for a single-element vector constant
-    uop = UOp.const(dtypes.int32, (42,))
+    uop = UOp.const((42,))
    self.assertEqual(uop.vmin, 42)
    self.assertEqual(uop.vmax, 42)

  def test_vmin_vmax_vconst_multiple_elements(self):
    # vmin and vmax for a multi-element vector constant
-    uop = UOp.const(dtypes.int32, (10, 20, -5, 7))
+    uop = UOp.const((10, 20, -5, 7))
    self.assertEqual(uop.vmin, -5)
    self.assertEqual(uop.vmax, 20)

  def test_vmin_vmax_vconst_all_equal(self):
    # vmin and vmax for a vector where all elements are equal
-    uop = UOp.const(dtypes.int32, (7, 7, 7))
+    uop = UOp.const((7, 7, 7))
    self.assertEqual(uop.vmin, 7)
    self.assertEqual(uop.vmax, 7)

  def test_vmin_vmax_vconst_with_negative_values(self):
    # vmin and vmax for a vector constant containing negative values
-    uop = UOp.const(dtypes.int32, (-10, -20, -5, -15))
+    uop = UOp.const((-10, -20, -5, -15))
    self.assertEqual(uop.vmin, -20)
    self.assertEqual(uop.vmax, -5)

  def test_vmin_vmax_vconst_with_floats(self):
    # vmin and vmax for a vector constant of float values
-    uop = UOp.const(dtypes.float32, (1.5, -3.2, 0.0))
+    uop = UOp.const((1.5, -3.2, 0.0))
    self.assertEqual(uop.vmin, -3.2)
    self.assertEqual(uop.vmax, 1.5)

  def test_vmin_vmax_vconst_with_bools(self):
    # vmin and vmax for a vector constant of bool values
-    uop = UOp.const(dtypes.bool, (True, False, False))
+    uop = UOp.const((True, False, False))
    self.assertIs(uop.vmin, False)
    self.assertIs(uop.vmax, True)

  def test_vmin_vmax_vector_with_gep(self):
    # vmin and vmax for a vector constant of bool values
    d1 = UOp.param(1, dtypes.int, (1,))
-    idx = UOp.const(dtypes.int, 0)
+    idx = UOp.const(0)
    val = UOp(Ops.LOAD, src=(d1.index(idx),))
    uop = (val // 32)
    self.assertEqual(uop.vmin, -67108864)
@@ -332,17 +338,17 @@ class TestVminVmaxVConst(unittest.TestCase):
 class TestConstFactor(unittest.TestCase):
  def test_const_factor_constant(self):
    # const_factor for a constant
-    uop = UOp.const(dtypes.int32, 42)
+    uop = UOp.const(42)
    self.assertEqual(uop.const_factor(), 42)

  def test_const_factor_addition(self):
    # const_factor for an addition of constants
-    uop = UOp.const(dtypes.int32, 30) + UOp.const(dtypes.int32, 12)
+    uop = UOp.const(30) + UOp.const(12)
    self.assertEqual(uop.const_factor(), 6)  # GCD(30, 12) = 6

  def test_const_factor_multiplication(self):
    # const_factor for a multiplication of constants
-    uop = UOp.const(dtypes.int32, 5) * UOp.const(dtypes.int32, 7)
+    uop = UOp.const(5) * UOp.const(7)
    self.assertEqual(uop.const_factor(), 5)  # For multiplication, it's one of the factors

  def test_const_factor_with_variable(self):
@@ -377,14 +383,14 @@ class TestConstFactor(unittest.TestCase):
 class TestDivides(unittest.TestCase):
  def test_divides_constant_exact(self):
    # Divides a constant by an exact divisor
-    uop = UOp.const(dtypes.int32, 42)
+    uop = UOp.const(42)
    result = uop.divides(7)
    self.assertIsNotNone(result)
    self.assertEqual(result.const_factor(), 6)  # 42 / 7 = 6

  def test_divides_constant_inexact(self):
    # Try to divide a constant by a non-exact divisor
-    uop = UOp.const(dtypes.int32, 42)
+    uop = UOp.const(42)
    result = uop.divides(5)
    self.assertIsNone(result)  # 42 is not divisible by 5

@@ -5,23 +5,24 @@ from tinygrad.tensor import Tensor
 from tinygrad.helpers import Timing, Context, cdiv
 from tinygrad.dtype import dtypes, AddrSpace, ConstFloat, Invalid  # noqa: F401
 from tinygrad.device import Device
-from tinygrad.uop.ops import Ops, ParamArg, UOp, UPat, dtype_from_uop, exec_alu, graph_rewrite, pm_lower_index_dtype  # noqa: F401  # ParamArg used by eval(str(uop)) roundtrip tests
+from tinygrad.uop.ops import Ops, ParamArg, PatternMatcher, UOp, UPat, dtype_from_uop, exec_alu, graph_rewrite, pm_lower_index_dtype  # noqa: F401  # ParamArg used by eval(str(uop)) roundtrip tests
 from tinygrad.uop.spec import spec_program, spec_shared, type_verify
 from tinygrad.uop.symbolic import sym, pm_remove_invalid
 from test.helpers import eval_uop, to_uops_list

 class TestDTypeFromUOp(unittest.TestCase):
  def test_broadcastable_promotion(self):
-    self.assertEqual(dtype_from_uop(Ops.ADD, (UOp.const(dtypes.float32, 1.0), UOp.const(dtypes.float16, 1.0)), None), dtypes.float32)
-    self.assertEqual(dtype_from_uop(Ops.MUL, (UOp.const(dtypes.int8, 1), UOp.const(dtypes.int32, 1)), None), dtypes.int32)
+    self.assertEqual(dtype_from_uop(Ops.ADD, (UOp.const(1.0).cast(dtypes.float32), UOp.const(1.0).cast(dtypes.float16)), None), dtypes.float32)
+    self.assertEqual(dtype_from_uop(Ops.MUL, (UOp.const(1).cast(dtypes.int8), UOp.const(1).cast(dtypes.int32)), None), dtypes.int32)

  def test_same_dtype_fast_path(self):
-    src = (UOp.const(dtypes.weakint, 1), UOp.const(dtypes.weakint, 2))
+    src = (UOp.const(1), UOp.const(2))
    self.assertEqual(dtype_from_uop(Ops.ADD, src, None), dtypes.weakint)

  def test_where_promotion(self):
-    cond = UOp.const(dtypes.bool, True)
-    self.assertEqual(dtype_from_uop(Ops.WHERE, (cond, UOp.const(dtypes.float32, 1.0), UOp.const(dtypes.float16, 1.0)), None), dtypes.float32)
+    cond = UOp.const(True)
+    srcs = (cond, UOp.const(1.0).cast(dtypes.float32), UOp.const(1.0).cast(dtypes.float16))
+    self.assertEqual(dtype_from_uop(Ops.WHERE, srcs, None), dtypes.float32)
    idx = UOp.range(4, 0)
    self.assertEqual(idx.valid(idx < 4).dtype, dtypes.weakint)

@@ -37,33 +38,39 @@ class TestDTypeFromUOp(unittest.TestCase):
    self.assertEqual(UOp(Ops.CONST, arg=True).dtype, dtypes.bool)
    self.assertEqual(UOp(Ops.CONST, arg=Invalid).dtype, dtypes.bool)
    # an explicit (strong) const dtype is legal until the field is removed
-    self.assertEqual(UOp.const(dtypes.int32, 3).dtype, dtypes.int32)
+    self.assertEqual(UOp.const(3, dtypes.int32).dtype, dtypes.int32)

  def test_weak_dtype_rejected_by_program_spec(self):
    for weak, concrete, value in ((dtypes.weakint, dtypes.int32, 1), (dtypes.weakfloat, dtypes.float32, 1.0)):
-      with self.assertRaises(RuntimeError): type_verify(UOp.const(weak, value).sink(), spec_program)
-      type_verify(UOp.const(concrete, value).sink(), spec_program)
+      with self.assertRaises(RuntimeError): type_verify(UOp.const(value, weak).sink(), spec_program)
+      type_verify(UOp.const(value, concrete).sink(), spec_program)
+
+  def test_invalid_stated_dtype(self):
+    # UOp.const normalizes a stated dtype away (const_like/full pass their position's); the core constructor does not,
+    # and the spec is what rejects a non-bool Invalid
+    self.assertIs(UOp.const(Invalid, dtypes.float32), UOp.invalid())
+    with self.assertRaises(RuntimeError): type_verify(UOp(Ops.CONST, dtypes.float32, arg=Invalid), spec_shared)

  def test_invalid_dtype_and_consumers(self):
    invalid = UOp.invalid()
    self.assertIs(invalid.dtype, dtypes.bool)
-    self.assertIs(UOp.const(dtypes.float32, Invalid), invalid)
-    self.assertIs((moved:=invalid.reshape((1,))).cast(dtypes.float32), moved)
+    self.assertIs(UOp.const(Invalid, dtypes.float32), invalid)
    scratch = Tensor.invalids(4, dtype=dtypes.float32)
    self.assertEqual((scratch.dtype, next(u.dtype for u in scratch.uop.toposort() if u.op is Ops.BUFFER), next(u.dtype for u in scratch.uop.toposort()
-      if u.arg is Invalid)), (dtypes.float32, dtypes.float32, dtypes.bool))
-    invalid, value = UOp.invalid(), UOp.const(dtypes.float32, 1)
+      if u.is_invalid)), (dtypes.float32, dtypes.float32, dtypes.bool))
+    invalid, value = UOp.invalid(), UOp.const(1, dtypes.float32)
+    for u in (UOp.param(0, dtypes.bool, ()).where(value, invalid), value+invalid, UOp.stack(value, invalid)): self.assertIs(u.src[-1], invalid)
    for u in (UOp(Ops.STACK, dtypes.float32, src=(value, invalid)), UOp(Ops.ADD, dtypes.float32, src=(value, invalid)),
-              UOp.const(dtypes.bool, True).where(value, invalid), UOp(Ops.CMPLT, src=(invalid, value)), UOp(Ops.CMPLT, src=(value, invalid)),
+              UOp.const(True).where(value, invalid), UOp(Ops.CMPLT, src=(invalid, value)), UOp(Ops.CMPLT, src=(value, invalid)),
              UOp.param(0, dtypes.float32, (4,)).index(invalid)): type_verify(u, spec_shared)
    gate, value = UOp.param(0, dtypes.bool, ()), UOp.param(1, dtypes.float, ())
-    self.assertIs((out:=graph_rewrite(gate.where(value, UOp.invalid()), pm_remove_invalid)).src[2], UOp.const(dtypes.float, 0))
+    self.assertIs((out:=graph_rewrite(gate.where(value, UOp.invalid()), pm_remove_invalid)).src[2], UOp.const(0, dtypes.float))
    type_verify(out.sink(), spec_program)

  def test_remove_invalid_stack_lanes(self):
-    stack = UOp(Ops.STACK, dtypes.half, (UOp.const(dtypes.half, 1), UOp.invalid()))
+    stack = UOp(Ops.STACK, dtypes.half, (UOp.const(1, dtypes.half), UOp.invalid()))
    out = graph_rewrite(stack, pm_remove_invalid)
-    self.assertEqual(out.src, (UOp.const(dtypes.half, 1), UOp.const(dtypes.half, 0)))
+    self.assertEqual(out.src, (UOp.const(1, dtypes.half), UOp.const(0, dtypes.half)))
    type_verify(out.sink(), spec_program)

 class TestLowerIndexDtype(unittest.TestCase):
@@ -72,7 +79,7 @@ class TestLowerIndexDtype(unittest.TestCase):
    # width the offset bounds select (this one needs long)
    buf = UOp.param(0, dtypes.float, (2**31+64,))
    i = UOp.variable("i", 0, 2**28)
-    shrink = UOp(Ops.SHRINK, src=(buf, (i*24).valid(i < 2**28), UOp.const(dtypes.weakint, 4)))
+    shrink = UOp(Ops.SHRINK, src=(buf, (i*24).valid(i < 2**28), UOp.const(4)))
    lowered = graph_rewrite(shrink.sink(), pm_lower_index_dtype)
    self.assertTrue(all(u.dtype != dtypes.weakint for u in lowered.backward_slice_with_self), "lowering must resolve all weakint")
    sh = next(u for u in lowered.backward_slice_with_self if u.op is Ops.SHRINK)
@@ -111,6 +118,26 @@ class TestSafeCast(unittest.TestCase):
    self.assertEqual(a.cast(dtypes.int8).cast(dtypes.int64).simplify(), a.cast(dtypes.int64))
    self.assertEqual(a.cast(dtypes.int8).cast(dtypes.float).simplify(), a.cast(dtypes.float))

+class TestConstFloatEq(unittest.TestCase):
+  def test_nan_eq_ne_agree(self):
+    nan = dtypes.float32.const(math.nan)
+    self.assertTrue(nan == math.nan)
+    self.assertFalse(nan != math.nan)  # float.__ne__ would say True here
+    self.assertFalse(nan == Invalid)
+    self.assertTrue(nan != Invalid)    # __ne__ must defer to the reflected eq, not swallow NotImplemented
+
+  def test_invalid_eq_defers_to_reflected(self):
+    class HoldsInvalid:  # a carrier that knows it holds Invalid. returning False for foreign types would silence its eq
+      def __eq__(self, other): return other is Invalid
+    self.assertTrue(Invalid == HoldsInvalid())
+    self.assertFalse(Invalid != HoldsInvalid())
+
+  def test_matchers_agree_on_nan(self):
+    n = UOp.const(math.nan, dtypes.float32)
+    for compiled in (False, True):
+      pm = PatternMatcher([(UPat(Ops.CONST, arg=math.nan), lambda: True)], compiled=compiled)
+      self.assertTrue(pm.rewrite(n), f"{compiled=}")
+
 class TestExecALU(unittest.TestCase):
  def test_sqrt(self):
    self.assertEqual(exec_alu(Ops.SQRT, dtypes.float, (0.0,)), 0.0)
@@ -198,9 +225,9 @@ class TestGatedStoreRewrite(unittest.TestCase):
  def test_tiny_gate_store(self):
    gmem = UOp.param(0, dtypes.float, (8,))
    gidx0 = UOp.special(4, 'gidx0')
-    gate = gidx0<UOp.const(dtypes.weakint, 1)
-    idx = UOp(Ops.INDEX, src=(gmem, (gidx0 * UOp.const(dtypes.weakint, 2)).valid(gate)))
-    val = UOp.const(dtypes.float, 42.0)
+    gate = gidx0<UOp.const(1)
+    idx = UOp(Ops.INDEX, src=(gmem, (gidx0 * UOp.const(2)).valid(gate)))
+    val = UOp.const(42.0).cast(dtypes.float)
    store = UOp(Ops.STORE, src=(idx, val))
    uops = to_uops_list([store])
    if_uop = next(u for u in uops if u.op is Ops.IF)
@@ -215,10 +242,10 @@ class TestGatedStoreRewrite(unittest.TestCase):
    gmem0 = UOp.param(0, dtypes.float, (8,))
    gmem1 = UOp.param(1, dtypes.float, (8,))
    gidx0 = UOp.special(4, 'gidx0')
-    idx = gidx0 * UOp.const(dtypes.weakint, 2)
-    idx0 = UOp(Ops.INDEX, src=(gmem0, idx.valid(gidx0<UOp.const(dtypes.weakint, 1))))
+    idx = gidx0 * UOp.const(2)
+    idx0 = UOp(Ops.INDEX, src=(gmem0, idx.valid(gidx0<UOp.const(1))))
    idx1 = UOp(Ops.INDEX, src=(gmem1, idx))
-    val = UOp.const(dtypes.float, 42.0)
+    val = UOp.const(42.0).cast(dtypes.float)
    stores = [UOp.store(idx0, val), UOp.store(idx1, val)]
    uops = to_uops_list(stores)
    if_uop = next(u for u in uops if u.op is Ops.IF)
@@ -235,11 +262,11 @@ class TestGatedStoreRewrite(unittest.TestCase):
    gmem0 = UOp.param(0, dtypes.float, (8,))
    gmem1 = UOp.param(1, dtypes.float, (8,))
    gidx0 = UOp.special(4, 'gidx0')
-    idx = gidx0*UOp.const(dtypes.weakint, 2)
-    gate = gidx0<UOp.const(dtypes.weakint, 1)
+    idx = gidx0*UOp.const(2)
+    gate = gidx0<UOp.const(1)
    idx0 = UOp(Ops.INDEX, src=(gmem0, idx.valid(gate)))
    idx1 = UOp(Ops.INDEX, src=(gmem1, idx.valid(gate)))
-    val = UOp.const(dtypes.float, 42.0)
+    val = UOp.const(42.0).cast(dtypes.float)
    stores = [UOp.store(idx0, val), UOp.store(idx1, val)]
    uops = to_uops_list(stores)
    ifs = [u for u in uops if u.op is Ops.IF]
@@ -257,7 +284,7 @@ class TestFastIdiv(unittest.TestCase):
  def test_division_power_of_two(self):
    for dt in (dtypes.int32, dtypes.uint32):
      g = UOp.param(0, dt, (3,))
-      c = UOp.const(dt, 2)
+      c = UOp.const(2).cast(dt)
      l = g.index(c)
      a = UOp(Ops.CDIV, dt, (l, c))
      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
@@ -270,7 +297,7 @@ class TestFastIdiv(unittest.TestCase):
    # FLOORMOD by a power of two lowers to AND (correct floor mod for any sign in two's complement)
    for dt in (dtypes.int32, dtypes.uint32):
      g = UOp.param(0, dt, (9,))
-      c = UOp.const(dt, 8)
+      c = UOp.const(8).cast(dt)
      a = UOp(Ops.FLOORMOD, dt, (g.index(c), c))
      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
      ops = [x.op for x in uops]
@@ -282,7 +309,7 @@ class TestFastIdiv(unittest.TestCase):
    # uint FLOORDIV by a power of two lowers to a shift, leaving no IDIV/FLOORDIV in the kernel
    for dt in (dtypes.uint32, dtypes.uint64):
      g = UOp.param(0, dt, (3,))
-      c = UOp.const(dt, 2)
+      c = UOp.const(2).cast(dt)
      a = UOp(Ops.FLOORDIV, dt, (g.index(c), c))
      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
      ops = [x.op for x in uops]
@@ -294,7 +321,7 @@ class TestFastIdiv(unittest.TestCase):
  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "WEBGPU doesn't support long")
  def test_fast_idiv_and_mod(self):
    g = UOp.param(0, dtypes.uint32, (4,))
-    c = UOp.const(dtypes.uint, 3)
+    c = UOp.const(3).cast(dtypes.uint)
    l = g.index(c)
    a = UOp(Ops.CDIV, src=(l, c))
    uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
@@ -314,7 +341,7 @@ class TestFastIdiv(unittest.TestCase):
  def test_fast_idiv_bounded_numerator_zero(self):
    x = UOp.variable("x", 0, 1, dtype=dtypes.int32)
    for val in range(2):
-      self.assertEqual(eval_uop(x.alu(Ops.CDIV, x.const_like(3)), vals=(val,)), cdiv(val, 3))
+      self.assertEqual(eval_uop(x.alu(Ops.CDIV, UOp.const(3).cast(x.dtype)), vals=(val,)), cdiv(val, 3))

  @Context(DISABLE_FAST_IDIV=0)
  def test_fast_idiv_remove_powers_of_two(self):
@@ -329,7 +356,7 @@ class TestFastIdiv(unittest.TestCase):
  def test_fast_idiv_overflow(self):
    # This will be possible with a slightly different method for fast_idiv
    g = UOp.param(0, dtypes.uint32, (8,))
-    c = UOp.const(dtypes.uint, 7)
+    c = UOp.const(7).cast(dtypes.uint)
    l = UOp(Ops.LOAD, src=(g.index(c),))
    a = UOp(Ops.CDIV, src=(l, c))
    uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
@@ -340,7 +367,7 @@ class TestFastIdiv(unittest.TestCase):

  def test_disable_fast_idiv(self):
    g = UOp.param(0, dtypes.uint32, (4,))
-    c = UOp.const(dtypes.uint, 3)
+    c = UOp.const(3).cast(dtypes.uint)
    l = g.index(c)
    a = UOp(Ops.CDIV, src=(l, c))
    with Context(DISABLE_FAST_IDIV=1):
@@ -352,8 +379,8 @@ class TestFastIdiv(unittest.TestCase):
 class TestUOpMethod(unittest.TestCase):
  @unittest.skip("uops lt no longer ordered")
  def test_compare_alu_same_src_different_arg(self):
-    a = UOp.const(dtypes.float, 2.0)
-    b = UOp.const(dtypes.float, 3.0)
+    a = UOp.const(2.0)
+    b = UOp.const(3.0)

    add = UOp(Ops.ADD, src=(a, b))
    mul = UOp(Ops.MUL, src=(a, b))
@@ -368,8 +395,8 @@ class TestUOpMethod(unittest.TestCase):
    self.assertEqual(list(var_vals)[0], a.expr)

  def test_const_factor(self):
-    gidx0 = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 8),), arg='gidx0')
-    self.assertEqual(UOp.const(dtypes.int, 17).const_factor(), 17)
+    gidx0 = UOp(Ops.SPECIAL, src=(UOp.const(8),), arg='gidx0')
+    self.assertEqual(UOp.const(17).const_factor(), 17)
    self.assertEqual(gidx0.const_factor(), 1)
    self.assertEqual((gidx0*3).const_factor(), 3)
    self.assertEqual((gidx0*3+6).const_factor(), 3)
@@ -388,26 +415,26 @@ class TestUOpMethod(unittest.TestCase):

  def test_const_zero_neg_zero_different(self):
    # -0.0 and 0.0 must be different UOps (for IEEE754 correctness, e.g. 1/-0.0 = -inf)
-    pos_zero = UOp.const(dtypes.float, 0.0)
-    neg_zero = UOp.const(dtypes.float, -0.0)
+    pos_zero = UOp.const(0.0)
+    neg_zero = UOp.const(-0.0)
    self.assertIsNot(pos_zero, neg_zero)
    self.assertNotEqual(hash(pos_zero.arg), hash(neg_zero.arg))

  def test_const_nan_same(self):
    # nan constants should be deduplicated
-    nan1 = UOp.const(dtypes.float, float('nan'))
-    nan2 = UOp.const(dtypes.float, float('nan'))
+    nan1 = UOp.const(float('nan'))
+    nan2 = UOp.const(float('nan'))
    self.assertIs(nan1, nan2)

 class TestUOpStr(unittest.TestCase):
  def test_uop_str(self):
-    a = UOp.const(dtypes.float, 2.0) + UOp.const(dtypes.float, 3.0)
+    a = UOp.const(2.0) + UOp.const(3.0)
    for _ in range(20): a = a + a
    assert len(str(a)) < 10_000, "exponential string growth"
    assert str(eval(str(a))) == str(a)

  def test_vectorized_str(self):
-    vec = UOp(Ops.STACK, src=tuple(UOp.const(dtypes.int, x) for x in range(4)))
+    vec = UOp(Ops.STACK, src=tuple(UOp.const(x) for x in range(4)))
    assert str(eval(str(vec))) == str(vec)

  def test_reduceop_arg(self):
@@ -426,7 +453,7 @@ class TestUPatHelpers(unittest.TestCase):

 class TestUopsObject(unittest.TestCase):
  def test_timing(self):
-    with Timing("create 10k uops:"): ret = [UOp(Ops.CONST, dtypes.int, arg=10000000+i) for i in range(10000)]
+    with Timing("create 10k uops:"): ret = [UOp.const(10000000+i, dtypes.int) for i in range(10000)]
    assert len(ret) == 10000

  def test_nested(self):
@@ -442,16 +469,16 @@ class TestUOpRender(unittest.TestCase):
    u = UOp(Ops.STACK, dtype=dtypes.void, src=())
    self.assertEqual(u.render(), "{}")
  def test_render_vectorize_same(self):
-    u = UOp(Ops.STACK, dtype=dtypes.int, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))
+    u = UOp(Ops.STACK, src=(UOp.const(0),)*3)
    self.assertEqual(u.render(simplify=False), "{0,0,0}")
  def test_render_vectorize_different(self):
-    u = UOp(Ops.STACK, dtype=dtypes.int, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
+    u = UOp(Ops.STACK, src=tuple(UOp.const(i) for i in range(3)))
    self.assertEqual(u.render(simplify=False), "{0,1,2}")
  def test_render_vectorize_same_simplified(self):
-    u = UOp(Ops.STACK, dtype=dtypes.int, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))
+    u = UOp(Ops.STACK, src=(UOp.const(0),)*3)
    self.assertEqual(u.render(), "{0,0,0}")
  def test_render_vectorize_different_simplified(self):
-    u = UOp(Ops.STACK, dtype=dtypes.int, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
+    u = UOp(Ops.STACK, src=tuple(UOp.const(i) for i in range(3)))
    self.assertEqual(u.render(), "{0,1,2}")

 class TestContiguousViewOffset(unittest.TestCase):
@@ -147,21 +147,21 @@ class TestUOpsStats(unittest.TestCase):
  #MULACC should have the same stats as MUL + ADD
  def test_mulacc(self):
    globl = UOp.param(0, dtypes.int, (3,))
-    o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
-    o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
+    o1 = UOp.const(1, dtypes.int)
+    o2 = UOp.const(2, dtypes.int)
    u1 = globl.index(o1)
    u2 = globl.index(o2)
-    u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
+    u3 = UOp.const(3, dtypes.int)
    u4 = UOp(Ops.MUL, src=(u1,u2))
    u5 = UOp(Ops.ADD, src=(u4,u3))
    uops = tuple(u5.toposort())

    globl = UOp.param(0, dtypes.int, (3,))
-    o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
-    o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
+    o1 = UOp.const(1, dtypes.int)
+    o2 = UOp.const(2, dtypes.int)
    u1 = globl.index(o1)
    u2 = globl.index(o2)
-    u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
+    u3 = UOp.const(3, dtypes.int)
    u4 = UOp(Ops.MULACC, src=(u1,u2,u3))
    uops_fma = tuple(u4.toposort())

@@ -12,12 +12,12 @@ class TestValidateOOB(unittest.TestCase):
  def test_const_index(self):
    with Context(CHECK_OOB=1, SPEC=2):
      buf = UOp.param(0, dtypes.int, (16,))
-      to_uops_list([buf.index(UOp.const(dtypes.int, 0)).load(dtype=dtypes.int)])  # valid
-      to_uops_list([buf.index(UOp.const(dtypes.int, 15)).load(dtype=dtypes.int)])  # valid (last element)
+      to_uops_list([buf.index(UOp.const(0)).load(dtype=dtypes.int)])  # valid
+      to_uops_list([buf.index(UOp.const(15)).load(dtype=dtypes.int)])  # valid (last element)
      with self.assertRaises(RuntimeError):
-        to_uops_list([buf.index(UOp.const(dtypes.int, 16)).load(dtype=dtypes.int)])  # off by one
+        to_uops_list([buf.index(UOp.const(16)).load(dtype=dtypes.int)])  # off by one
      with self.assertRaises(RuntimeError):
-        to_uops_list([buf.index(UOp.const(dtypes.int, 42)).load(dtype=dtypes.int)])  # way out
+        to_uops_list([buf.index(UOp.const(42)).load(dtype=dtypes.int)])  # way out

  def test_variable_index(self):
    with Context(CHECK_OOB=1, SPEC=2):
@@ -138,11 +138,19 @@ class TestValidateOOB(unittest.TestCase):
      with self.assertRaises(RuntimeError):
        to_uops_list([buf1.index((ld0 * 2).valid((ld0 >= 0) & (ld0 < 64))).load(dtype=dtypes.int)])  # oob

+  def test_load_from_shrink_as_index(self):
+    with Context(CHECK_OOB=1, SPEC=2):
+      buf0 = UOp.param(0, dtypes.int, (16,))
+      buf1 = UOp.param(1, dtypes.int, (64,))
+      shrink = UOp(Ops.SHRINK, src=(buf0, UOp.const(0, dtypes.int), UOp.const(4)))
+      ld0 = shrink.load(dtype=dtypes.int).index(0)
+      to_uops_list([buf1.index(ld0.valid((ld0 >= 0) & (ld0 < 64))).load(dtype=dtypes.int)])
+
  def test_load_bool_as_mask(self):
    with Context(CHECK_OOB=1, SPEC=2):
      buf_bool = UOp.param(0, dtypes.bool, (16,))
      buf_int = UOp.param(1, dtypes.int, (8,))
-      gidx = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.weakint, 16),), arg="gidx0")
+      gidx = UOp(Ops.SPECIAL, src=(UOp.const(16),), arg="gidx0")
      ld_bool = buf_bool.index(gidx).load()
      with self.assertRaises(RuntimeError):
        to_uops_list([buf_int.index(gidx.valid(ld_bool)).load()])  # gidx 0..15, buf_int size 8
@@ -156,12 +164,12 @@ class TestValidateOOB(unittest.TestCase):
      sbuf = UOp.placeholder((8,), dtypes.uint, slot=0, addrspace=AddrSpace.LOCAL)

      # Define indices, valids and barrier
-      gidx = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 416),), arg="gidx0")
-      lidx = UOp(Ops.SPECIAL, src=(UOp.const(dtypes.int, 10),), arg="lidx0")
+      gidx = UOp(Ops.SPECIAL, src=(UOp.const(416),), arg="gidx0")
+      lidx = UOp(Ops.SPECIAL, src=(UOp.const(10),), arg="lidx0")

      gate = (gidx<400) & (lidx<8)

-      local_store = sbuf.index(lidx.valid(lidx<8)).store(UOp.const(dtypes.uint, 1))
+      local_store = sbuf.index(lidx.valid(lidx<8)).store(UOp.const(1))

      barrier = UOp(Ops.BARRIER, src=(local_store,))
      if_barrier = UOp(Ops.IF, src=(gate, barrier))
@@ -179,7 +187,7 @@ class TestValidateOOB(unittest.TestCase):
      glbl0 = UOp.param(0, dtypes.int, (16,))
      mask = UOp.param(0, dtypes.bool, (16,))
      ridx = UOp.range(20, 0)
-      ld0 = UOp(Ops.LOAD, src=(glbl0.index(UOp.const(ridx, ridx<16&mask))))
+      ld0 = UOp(Ops.LOAD, src=(glbl0.index(UOp.const(ridx<16&mask, ridx))))
      to_uops_list([ld0])

 if __name__ == "__main__":
@@ -96,10 +96,10 @@ class TestViz(unittest.TestCase):
  def test_exceptions(self):
    # VIZ tracks rewrites up to and including the error
    def count_3(x:UOp):
-      assert x.arg <= 3
-      return x.replace(arg=x.arg+1)
+      assert x.val <= 3
+      return UOp.const(x.val+1, x.dtype)
    err_pm = PatternMatcher([(UPat.cvar("x"), count_3),])
-    a = UOp.const(dtypes.int, 1)
+    a = UOp.const(1)
    with save_viz() as viz:
      with self.assertRaises(AssertionError): exec_rewrite(a, [err_pm])
    lst = viz.list_items()
@@ -199,11 +199,11 @@ class TestViz(unittest.TestCase):
    self.assertEqual(ansistrip(a2["label"]), "CUSTOM\nx\nyzww\nw")

  def test_inf_loop(self):
-    a = UOp.const(dtypes.int, 3)
-    b = UOp.const(dtypes.int, 4)
+    a = UOp.const(3)
+    b = UOp.const(4)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
    ])
    with save_viz() as viz:
      # use smaller stack limit for faster test (default is 250000)
@@ -224,9 +224,9 @@ class TestViz(unittest.TestCase):
    list(viz.get_details(0, 0))

  def test_enter_calls_rewrite(self):
-    pm = PatternMatcher([(UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4))])
+    pm = PatternMatcher([(UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype))])
    with save_viz() as viz:
-      inner = UOp.const(dtypes.int, 3)
+      inner = UOp.const(3)
      call = UOp(Ops.CALL, src=(UOp(Ops.SINK, src=(inner,)),))
      func = UOp(Ops.FUNCTION, src=(UOp(Ops.TUPLE, src=(call,)),))
      graph_rewrite(func, TrackedPatternMatcher(pm.patterns), enter_calls=True)
@@ -236,8 +236,8 @@ class TestViz(unittest.TestCase):
  def test_const_node_visibility(self):
    with save_viz() as viz:
      a = UOp.variable("a", 0, 10, dtype=dtypes.int)
-      z = UOp.const(a.dtype, 0)
-      y = UOp.const(dtypes.float, math.pi)
+      z = UOp.const(0, a.dtype)
+      y = UOp.const(math.pi, dtypes.float)
      alu = a*z
      ret = exec_rewrite(sink:=UOp.sink(alu, y), [sym])
    lst = viz.list_items()
@@ -253,7 +253,7 @@ class TestViz(unittest.TestCase):

  def test_const_reshape_expand_folded(self):
    # CONST->EXPAND should be folded into the ALU node, not shown as separate EXPAND nodes
-    c = UOp.const(dtypes.float, 1.0, shape=(3,4))  # creates CONST->EXPAND chain
+    c = UOp.const(1.0).expand((3,4))  # creates CONST->EXPAND chain
    a = UOp.variable("a", 0.0, 10.0, dtypes.float)
    alu = a + c
    with save_viz() as viz:
@@ -267,13 +267,13 @@ class TestViz(unittest.TestCase):

  def test_stack_movement_not_folded_unless_all_const(self):
    a = UOp.variable("a", 0, 10, dtype=dtypes.int)
-    c = UOp.const(dtypes.int, 1)
+    c = UOp.const(1)
    stack = a.stack(c)
    reshaped = stack.reshape((1, 2))
    graph = uop_to_json(VizData(), reshaped)
    self.assertFalse(graph[id(stack)]["exclude"])

-    const_stack = c.stack(UOp.const(dtypes.int, 2))
+    const_stack = c.stack(UOp.const(2))
    const_reshaped = const_stack.reshape((1, 2))
    const_graph = uop_to_json(VizData(), const_reshaped)
    self.assertTrue(const_graph[id(const_stack)]["exclude"])
@@ -401,7 +401,7 @@ class TestVizIntegration(unittest.TestCase):
    with save_viz() as viz:
      def test(root):
        return graph_rewrite(root, sym)
-      test(c:=UOp.const(dtypes.int, 1))
+      test(c:=UOp.const(1))
      test(c+1)
    ls = viz.list_items()
    self.assertEqual(len(ls), 1)
@@ -414,7 +414,7 @@ class TestVizIntegration(unittest.TestCase):
      @track_rewrites()
      def test(root):
        return graph_rewrite(root, sym)
-      test(c:=UOp.const(dtypes.int, 1))
+      test(c:=UOp.const(1))
      test(c+1)
    ls = viz.list_items()
    self.assertEqual(len(ls), 2)
@@ -425,21 +425,22 @@ class TestVizIntegration(unittest.TestCase):
    with save_viz() as viz:
      def default_test(root): return graph_rewrite(root, sym)
      tracked_test = track_rewrites()(default_test)
-      c = UOp.const(dtypes.int, 1)
+      c = UOp.const(1)
      default_test(c+1) # goes to the default group
      tracked_test(c)   # all rewrites after this go inside the second group.
      default_test(c+2)
    ls = viz.list_items()
    self.assertEqual(len(ls), 2)
    graph = next(viz.get_details(0, 0))["graph"]
+    # both operands of c+1 are the same bare weak CONST, so the graph has two nodes
    self.assertEqual(list(graph), [id(c), id(c+1)])
    self.assertTrue(graph[id(c)]["exclude"])
    self.assertFalse(graph[id(c+1)]["exclude"])
    self.assertEqual(list(next(viz.get_details(1, 0))["graph"]), [id(c)])
    graph = next(viz.get_details(1, 1))["graph"]
-    self.assertEqual(list(graph), [id(c), id(c.const_like(2)), id(c+2)])
+    self.assertEqual(list(graph), [id(c), id((c+2).src[1]), id(c+2)])
    self.assertTrue(graph[id(c)]["exclude"])
-    self.assertTrue(graph[id(c.const_like(2))]["exclude"])
+    self.assertTrue(graph[id((c+2).src[1])]["exclude"])
    self.assertFalse(graph[id(c+2)]["exclude"])

  def test_recurse(self):
@@ -1,6 +1,7 @@
 import unittest, sys
 from tinygrad import Tensor, GlobalCounters, dtypes, Context
 from tinygrad.helpers import WINO
+from test.helpers import check_schedule

@unittest.skipIf(sys.platform.startswith("win"), "flaky on Windows")
 class TestWinograd(unittest.TestCase):
@@ -13,7 +14,7 @@ class TestWinograd(unittest.TestCase):
  def test_forward_kernels(self):
    x,w = Tensor.rand(1,4,9,9).realize(), Tensor.rand(4,4,3,3).realize()
    out = Tensor.conv2d(x,w)
-    self.assertEqual(len(out.schedule_linear().src), 4)
+    check_schedule(out, 4)

  def test_backward_counters(self):
    # contiguous_backward on the pooled input keeps the input-transform adjoint out of the overlap accumulation, so
@@ -81,6 +81,13 @@ class TestTensorCores(unittest.TestCase):
    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
      helper_tc_allclose(tc.dims[0], tc.dims[1], tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)

+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  def test_tensor_cores_nested_reduce(self):
+    tc = Device[Device.DEFAULT].renderer.tensor_cores[0]
+    a, b = Tensor.empty(tc.dims[1]*2, tc.dims[2], dtype=tc.dtype_in), Tensor.empty(tc.dims[2], tc.dims[0], dtype=tc.dtype_in)
+    ast = replace_opts(a.matmul(b, dtype=tc.dtype_out).sum(0).schedule_linear().src[-1].src[0], [Opt(OptOps.TC, 0, (-1, 0, 1))])
+    with self.assertRaises(KernelOptError): to_program(ast, Device[Device.DEFAULT].renderer)
+
  @Context(ALLOW_TF32=1)
  @unittest.skipIf(Device.DEFAULT == "PYTHON", "not generated on EMULATED device")
  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
@@ -39,6 +39,10 @@ class TestTiny(unittest.TestCase):
    out = Tensor.ones(N).contiguous().sum()
    self.assertEqual(out.item(), N)

+  def test_eye(self):
+    out = Tensor.eye(3).flatten()
+    self.assertListEqual(out.tolist(), [1.0,0.0,0.0, 0.0,1.0,0.0, 0.0,0.0,1.0])
+
  def test_gemm(self, N=getenv("GEMM_N", 64), dtype=dtypes.float):
    a = Tensor.ones(N,N, dtype=dtype).contiguous()
    b = Tensor.eye(N, dtype=dtype).clone()
@@ -1,5 +1,5 @@
 import unittest
-from tinygrad import Tensor, dtypes
+from tinygrad import Tensor, UOp, dtypes
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import Ops

@@ -43,6 +43,13 @@ class TestRingAllReduce(unittest.TestCase):
    self.assertEqual(len(sinks), 2)
    self.assertTrue(all(dst != src for dst, src in pairs))

+  def test_symbolic_shape(self):
+    rows = UOp.variable("rows", 1, 4).bind(3)
+    t = Tensor.ones(4, 4).shard(("CPU:0", "CPU:1"), axis=1).realize()
+    out = t[:rows].sum(1).realize()
+    self.assertEqual(out.shape, (rows,))
+    self.assertTrue((out == 4).all().item())
+
  def test_correct_ring(self):
    with Context(RING=2):
      N = 4
@@ -4,6 +4,7 @@ import numpy as np
 from tinygrad import dtypes, Tensor, TinyJit, GlobalCounters, Variable
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.helpers import temp, DEV, Context
+from test.helpers import assert_kernel_count

 N = 200  # has to be bigger than the cache to fail

@@ -42,7 +43,7 @@ class TestAssign(unittest.TestCase):
    # it should copy into the empty buffer
    GlobalCounters.reset()
    c.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)

  def test_assign_slice(self):
    X = Tensor([1,2,3,4]).realize()
@@ -50,7 +51,7 @@ class TestAssign(unittest.TestCase):
    xs.assign(xs+1)
    GlobalCounters.reset()
    self.assertListEqual(X.tolist(), [1,2,4,5])
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)

  def test_assign_slice_alt(self):
    X = Tensor([1,2,3,4]).realize()
@@ -58,7 +59,7 @@ class TestAssign(unittest.TestCase):
    xs1.assign(xs2+1)
    GlobalCounters.reset()
    self.assertListEqual(X.tolist(), [1,4,5,4])
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)

  def test_assign_flip(self):
    ref = np.arange(16, dtype=np.float32)
@@ -68,7 +69,7 @@ class TestAssign(unittest.TestCase):
    xs.assign(xs + X)
    ref = ref + ref[::-1]
    np.testing.assert_allclose(X.numpy(), ref)
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)

  def test_assign_add(self):
    for T in (1, 2, 10):#, 100): # this crashes in CI, not sure why
@@ -331,14 +332,14 @@ class TestAssign(unittest.TestCase):
    a = (Tensor.arange(16).reshape(4,4).clone().realize() + 1)
    GlobalCounters.reset()
    b.assign(a.contiguous()).realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)

  def test_assign_contiguous_permute(self):
    b = Tensor.arange(16).reshape(4,4).clone().realize()
    a = (Tensor.arange(16).reshape(4,4).clone().realize() + 1).permute((1,0))
    GlobalCounters.reset()
    b.assign(a.contiguous()).realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)

  def test_permuted_assignment(self):
    a = Tensor(np.arange(N*N, dtype=np.float32)).reshape(N,N)
@@ -413,7 +414,7 @@ class TestAssign(unittest.TestCase):

    GlobalCounters.reset()
    Tensor.realize(b, c, d)
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    np.testing.assert_allclose(b.numpy(), a.sum(1).numpy()+1)
    np.testing.assert_allclose(c.numpy(), a.sum(1).numpy()+2)
    np.testing.assert_allclose(d.numpy(), a.sum(1).numpy()+3)
@@ -461,7 +462,7 @@ class TestAssign(unittest.TestCase):
    b.assign(r + b)
    c.assign(r + b_perm.contiguous())
    Tensor.realize(b, c)
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)
    np.testing.assert_equal(b.numpy(), a.numpy().sum(1) + np.arange(32 * 32).reshape(32, 32))
    np.testing.assert_equal(c.numpy(), a.numpy().sum(1) + np.arange(32 * 32).reshape(32, 32).transpose(1, 0))

@@ -471,7 +472,7 @@ class TestAssign(unittest.TestCase):
    a.assign(a + b)
    GlobalCounters.reset()
    a.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    np.testing.assert_equal(a.numpy(), np.ones((4, 4))+np.pad(np.ones((4, 4))[:, 0:2], ((0, 0), (0, 2)), constant_values=2))

  def test_permuted_assignment_masked_view_not_contiguous(self):
@@ -510,7 +511,7 @@ class TestAssign(unittest.TestCase):
    expected[0:10] = expected[50:60].copy()
    GlobalCounters.reset()
    a[0:10].assign(a[50:60]).realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)  # currently conservative, forces contiguous
+    assert_kernel_count(2)  # currently conservative, forces contiguous
    np.testing.assert_allclose(a.numpy(), expected)

  def test_setitem_half(self):
@@ -630,7 +631,7 @@ class TestAssign(unittest.TestCase):
    GlobalCounters.reset()
    x.realize()
    # N assigns (1 kernel each) producing N kernels total
-    self.assertEqual(GlobalCounters.kernel_count, N)
+    assert_kernel_count(N)

  def test_shared_computation_assign_kernel_count(self):
    """When a .contiguous() is shared between an assign value and the next layer's input (like QKV projection in LLM),
@@ -648,7 +649,7 @@ class TestAssign(unittest.TestCase):
    GlobalCounters.reset()
    caches[-1][:1].contiguous().realize()
    # N matmuls + N assigns + 1 final read = 2*N+1 (AFTER embedding allows full graph scheduling with shared contiguous reuse)
-    self.assertEqual(GlobalCounters.kernel_count, 2*N+1)
+    assert_kernel_count(2*N+1)

  def test_double_assign_from_const(self):
    a = Tensor.empty(2)
@@ -656,12 +657,12 @@ class TestAssign(unittest.TestCase):
    a.assign(Tensor.ones(2, buffer=False))
    GlobalCounters.reset()
    a.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(a.tolist(), [1.,1.])

  def test_assign_deviceless_const(self):
    s = Tensor.empty(4, device="CPU:1", dtype=dtypes.float)
-    s.assign(Tensor(UOp.const(dtypes.float, 2.0)))
+    s.assign(Tensor(UOp.const(2.0).cast(dtypes.float)))
    np.testing.assert_equal(s.numpy(), [2, 2, 2, 2])

  def test_nested_after_contiguous_store(self):
@@ -672,7 +673,7 @@ class TestAssign(unittest.TestCase):
    contig.assign(Tensor([1, 4, 3], dtype=dtypes.int64))
    GlobalCounters.reset()
    base.assign(contig).realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)  # TODO: first copy is dead, could be 1
+    assert_kernel_count(2)  # TODO: first copy is dead, could be 1
    self.assertEqual(base.tolist(), [1,4,3])

  def test_nested_after_contiguous_store_no_init(self):
@@ -682,7 +683,7 @@ class TestAssign(unittest.TestCase):
    contig.assign(Tensor([1, 4, 3], dtype=dtypes.int64))
    GlobalCounters.reset()
    base.assign(contig).realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(base.tolist(), [1,4,3])

 class TestAssignOrdering(unittest.TestCase):
@@ -47,7 +47,7 @@ class TestGatedDeltaNetBlock(unittest.TestCase):
  def _make_config(self, **kwargs):
    return TransformerConfig(**({"num_blocks":1, "dim":4, "hidden_dim":8, "n_heads":1, "n_kv_heads":1,
                                 "norm_eps":1e-5, "vocab_size":32, "head_dim":4, "rope_theta":10000.0,
-                                 "rope_dim":4, "v_head_dim":4, "max_context":4, "full_attention_interval":2,
+                                 "rope_dim":4, "v_head_dim":4, "max_context":4, "ssm_layers":(True,),
                                 "ssm":SSMConfig(conv_kernel=2, state_size=2, group_count=1, time_step_rank=1, inner_size=2)} | kwargs))

  def _make_block(self, config:TransformerConfig) -> GatedDeltaNetBlock:
@@ -176,6 +176,20 @@ class TestGatedDeltaNetBlock(unittest.TestCase):
      np.testing.assert_allclose(recurrent_state, expected_recurrent[step], rtol=1e-3, atol=1e-3,
                                 err_msg=f"GatedDeltaNet reset recurrent cache mismatch at step {step}")

+  def test_kda_channel_decay(self):
+    config = self._make_config(n_heads=2, ssm=SSMConfig(conv_kernel=2, state_size=2, group_count=2, time_step_rank=2, inner_size=4, kda=True))
+    block, x = GatedDeltaNetBlock(config, config.ssm), Tensor([[[1., 2., 0., 0.]]])
+    # f_b(f_a(x)) = [1, 2, 3, 4]
+    block.ssm_f_a.weight = Tensor([[1., 0., 0., 0.], [0., 1., 0., 0.]])
+    block.ssm_f_b.weight = Tensor([[1., 0.], [0., 1.], [1., 1.], [2., 1.]])
+    block._init_state(x)
+    initial_state = Tensor.arange(8, dtype=dtypes.float32).reshape(1, 2, 2, 2)
+    block.recurrent_state.assign(initial_state).realize()
+    block.ssm_a = Tensor([[-1.], [-1.]])
+    block._attention(x, 0).realize()
+    alpha = np.exp(-self._softplus_np(np.arange(1, 5)).reshape(1, 2, 1, 2))
+    np.testing.assert_allclose(block.recurrent_state.numpy(), initial_state.numpy() * alpha, rtol=1e-5, atol=1e-5)
+
 class TestPairwiseTopk(unittest.TestCase):
  def test_basic_topk(self):
    x = Tensor([[[1.0, 3.0, 2.0, 5.0, 4.0]]])
@@ -3,7 +3,8 @@ import tempfile, unittest, math
 from tinygrad import Tensor, dtypes, TinyJit
 from tinygrad.helpers import Context
 from tinygrad.dtype import least_upper_float
-from tinygrad.uop.ops import UOp, Ops, dtype_from_uop
+from tinygrad.uop.ops import UOp, Ops, dtype_from_uop, graph_rewrite, pm_lower_index_dtype, pm_commit_weak
+from tinygrad.uop.symbolic import symbolic_simple
 from tinygrad.uop.spec import spec_shared, type_verify
 from tinygrad.engine.jit import JitError

@@ -11,18 +12,18 @@ from tinygrad.engine.jit import JitError
 class TestWeakPromotion(unittest.TestCase):
  def test_rand_requires_concrete(self):
    with self.assertRaises(ValueError): Tensor.rand(2, dtype=dtypes.weakfloat)
-    with self.assertRaises(ValueError): Tensor.const(dtypes.weakfloat, 1.0).rand_like()
-    with self.assertRaises(ValueError): Tensor.const(dtypes.weakfloat, 1.0).randn_like()
+    with self.assertRaises(ValueError): Tensor.const(1.0).rand_like()
+    with self.assertRaises(ValueError): Tensor.const(1.0).randn_like()

  def test_reduce_strips_weakness(self):
    for weak, value, strong in ((dtypes.weakint, 1, dtypes.default_int), (dtypes.weakfloat, 1.0, dtypes.default_float)):
-      t = Tensor.const(weak, value).expand(3)
+      t = Tensor.const(value, weak).expand(3)
      for out in (t.sum(), t.max(), t.prod(), t.cumsum(0), t.cummax(0)[0]): self.assertEqual(out.dtype, strong)
-    self.assertEqual((Tensor.const(dtypes.weakfloat, 1.0).expand(3).sum() + Tensor([1], dtype=dtypes.float16)).dtype, dtypes.float32)
+    self.assertEqual((Tensor.const(1.0).expand(3).sum() + Tensor([1], dtype=dtypes.float16)).dtype, dtypes.float32)

  def test_materialize_at_default_dtype(self):
    for weak, value, strong in ((dtypes.weakfloat, 0.5, dtypes.default_float),):
-      t = Tensor.const(weak, value)
+      t = Tensor.const(value, weak)
      self.assertEqual(t.dtype, weak)
      self.assertEqual(t.data().itemsize, strong.itemsize)
      self.assertEqual(t.numpy().dtype.itemsize, strong.itemsize)
@@ -32,7 +33,7 @@ class TestWeakPromotion(unittest.TestCase):
      self.assertEqual(t.contiguous().dtype, weak)

  def test_assign_into_weak_commits(self):
-    t = Tensor.const(dtypes.weakfloat, 0.5)
+    t = Tensor.const(0.5)
    t.assign(Tensor(1.0, dtype=dtypes.default_float))
    self.assertEqual((t.dtype, t.item()), (dtypes.default_float, 1.0))

@@ -43,10 +44,63 @@ class TestWeakPromotion(unittest.TestCase):
    r = Tensor([2], dtype=dtypes.uint8, device="CPU").copysign(Tensor([1], dtype=dtypes.uint32, device="CPU"))
    self.assertEqual((r.dtype, r.tolist()), (dtypes.uint32, [2]))

-  def test_uop_scalar_const_unchanged(self):
-    for dtype, value in ((dtypes.weakint, 1), (dtypes.int32, 1), (dtypes.float32, 0.5)):
+  def test_minimum_reflects_weak_operand(self):
+    r = Tensor(1).minimum(Tensor([2], dtype=dtypes.uint8, device="CPU"))
+    self.assertEqual((r.dtype, r.tolist()), (dtypes.uint8, [1]))
+    for dt in dtypes.uints:
+      r = Tensor([dt.max], dtype=dt, device="CPU").minimum(1)
+      self.assertEqual((r.dtype, r.tolist()), (dt, [1]))
+      self.assertNotIn(Ops.CAST, [u.op for u in r._uop.toposort()])
+
+  def test_broadcasted_keeps_const_weak(self):
+    # a python scalar stays a bare weak CONST through _broadcasted, lifted only to the KIND of the lub
+    x, y = Tensor([1], dtype=dtypes.int8)._broadcasted(3)
+    self.assertEqual((y._uop.base.op, y.dtype, x.dtype), (Ops.CONST, dtypes.weakint, dtypes.int8))
+    x, y = Tensor([1], dtype=dtypes.int8)._broadcasted(0.5)
+    self.assertEqual((y._uop.base.op, y.dtype, x.dtype), (Ops.CONST, dtypes.weakfloat, dtypes.weakfloat))
+    x, y = Tensor.const(1).reshape(1)._broadcasted(Tensor([1.0], dtype=dtypes.float32))
+    self.assertEqual((x._uop.base.op, x._uop.base.val, x.dtype, x.shape, y.dtype),
+                     (Ops.CONST, 1, dtypes.weakfloat, (1,), dtypes.float32))
+
+  def test_uop_scalar_const_lifts_kind(self):
+    for dtype, value, out_dtype, const_dtype in ((dtypes.weakint, 1, dtypes.weakint, dtypes.weakint),
+                                                 (dtypes.int32, 1, dtypes.int32, dtypes.weakint),
+                                                 (dtypes.int32, 0.5, dtypes.weakfloat, dtypes.weakfloat),
+                                                 (dtypes.float32, 1, dtypes.float32, dtypes.weakfloat)):
      out = UOp.variable("x", 0.0 if dtype == dtypes.float32 else 0, 10.0 if dtype == dtypes.float32 else 10, dtype) + value
-      self.assertEqual((out.dtype, out.src[1].dtype), (dtype, dtype))
+      self.assertEqual((out.dtype, out.src[1].op, out.src[1].dtype), (out_dtype, Ops.CONST, const_dtype))
+    # the kind lift converts the VALUE too (the arg is the only dtype carrier once UOp.const loses its dtype arg),
+    # and a bare weak const UOp is the same spelling as the python scalar: both lift to the same node
+    x = UOp.variable("x", 0.0, 1.0, dtypes.float32)
+    self.assertIsInstance((x + 2).src[1].val, float)
+    self.assertIs(x + UOp.const(2), x + 2)
+
+  def test_index_dtype_ignores_weakness(self):
+    with Context(SPEC=2):
+      idx = UOp.const(0).cast(dtypes.int32)
+      weak = UOp.const(1.0).expand((1,))
+      self.assertEqual(UOp(Ops.INDEX, dtypes.float32, (weak, idx)).dtype, dtypes.float32)
+      with self.assertRaisesRegex(RuntimeError, "bad dtype"): UOp(Ops.INDEX, dtypes.int32, (weak, idx))
+
+  def test_store_weak_value_uses_destination_dtype(self):
+    with Context(DEFAULT_FLOAT=dtypes.float16):
+      dst = UOp.param(0, dtypes.bfloat16, (1,)).index(UOp.const(0).cast(dtypes.int32))
+      gate = UOp.const(True)
+      out = graph_rewrite(dst.store(UOp.const(5.0), gate), pm_lower_index_dtype, ctx={})
+    # a bare weak CONST commits directly: the pass runs without symbolic, so a CAST here would survive it
+    self.assertEqual((out.src[1], out.src[2]), (UOp.const(5.0, dtypes.bfloat16), gate))
+
+  def test_weak_srcs_commit_only_at_a_concrete_lub(self):
+    weak_lub = UOp(Ops.ADD, src=(UOp.const(1), UOp.const(1.0)))
+    self.assertIs(graph_rewrite(weak_lub, pm_lower_index_dtype, ctx={}), weak_lub)
+    concrete = UOp.const(2.0).cast(dtypes.float16)
+    where = graph_rewrite(UOp(Ops.WHERE, src=(UOp.const(True), concrete, UOp.const(1.0))), pm_lower_index_dtype, ctx={})
+    self.assertEqual(tuple(x.dtype for x in where.src), (dtypes.bool, dtypes.float16, dtypes.float16))
+
+  def test_weak_shift_lhs_commits_the_node(self):
+    # a shift derives its lhs's dtype, so committing the lhs restates the root (WGSL's packed store writes `mask << shift_am`)
+    shl = graph_rewrite(UOp.const(0xFFFF) << UOp.variable("x", 0, 16, dtypes.uint), symbolic_simple+pm_commit_weak)
+    self.assertEqual((shl.dtype, shl.src[0]), (dtypes.uint, UOp.const(0xFFFF, dtypes.uint)))

  @unittest.expectedFailure  # TODO: a weak const defers to its consumer (JAX): these dtypes change once python scalars are weak consts
  def test_changed_rows(self):
@@ -79,17 +133,17 @@ class TestWeakPromotion(unittest.TestCase):
  def test_weak_int_binop(self):
    v = UOp.variable("i", 0, 10, dtypes.weakint)
    self.assertEqual((v << 1).dtype, dtypes.weakint)
-    self.assertEqual(dtype_from_uop(Ops.SHL, (UOp.const(dtypes.int8, 1), UOp.const(dtypes.uint32, 1)), None), dtypes.int64)
-    self.assertEqual(UOp.const(dtypes.weakint, 1).alu(Ops.SHL, UOp.const(dtypes.uint8, 1)).dtype, dtypes.uint8)
+    self.assertEqual(dtype_from_uop(Ops.SHL, (UOp.const(1, dtypes.int8), UOp.const(1, dtypes.uint32)), None), dtypes.int8)
+    self.assertEqual(UOp.const(1).alu(Ops.SHL, UOp.const(1, dtypes.uint)).dtype, dtypes.weakint)
    self.assertEqual((v & 3).dtype, dtypes.weakint)
-    with self.assertRaises(RuntimeError): Tensor.const(dtypes.weakfloat, 1.0) << Tensor.const(dtypes.weakfloat, 1.0)
-    with self.assertRaises(RuntimeError): UOp.const(dtypes.int32, 1).alu(Ops.SHL, UOp.const(dtypes.float64, 1))
+    with self.assertRaises(RuntimeError): Tensor.const(1.0) << Tensor.const(1.0)
+    with self.assertRaises(RuntimeError): UOp.const(1, dtypes.int32).alu(Ops.SHL, UOp.const(1, dtypes.float64))
    for op in (Ops.SHL, Ops.SHR):
      with self.assertRaises(RuntimeError):
-        UOp.const(dtypes.float32, 1).alu(op, UOp.const(dtypes.int32, 1))
+        UOp.const(1, dtypes.float32).alu(op, UOp.const(1, dtypes.int32))
    # float bitwise builds, the spec rejects it
    with Context(SPEC=1):
-      f32, wf = UOp.const(dtypes.float32, 1.0), UOp.const(dtypes.weakfloat, 1.0)
+      f32, wf = UOp.const(1.0, dtypes.float32), UOp.const(1.0)
      for bad in (f32.alu(Ops.AND, f32), UOp(Ops.AND, dtypes.float32, (f32, f32)), UOp(Ops.AND, dtypes.int32, (wf, wf))):
        with self.assertRaises(RuntimeError): type_verify([bad], spec_shared)

@@ -117,11 +171,18 @@ class TestWeakPromotion(unittest.TestCase):
      t.realize()
      self.assertNotIn(t.uop.buffer.dtype, dtypes.weaks)

+  def test_computed_float_index_lowers(self):
+    # a half-pixel nearest index resolves its float-scaled range before the gather
+    idx = (Tensor.arange(8) + 0.5) / 4 - 0.5
+    idx = (idx.clip(0, 1) - 0.5).ceil().int()
+    out = Tensor([0, 1], device="NULL")[idx].contiguous().realize()
+    self.assertNotIn(out.uop.buffer.dtype, dtypes.weaks)
+

 class TestWeakStorageBoundary(unittest.TestCase):
  # weak has no storage: a weak assignment source casts when it defers to the destination, everything else raises
  def test_weak_source(self):
-    w05 = Tensor.const(dtypes.weakfloat, 0.5).reshape(1)
+    w05 = Tensor.const(0.5).reshape(1)
    dst = Tensor.zeros(2, dtype=dtypes.int8, device="CPU").contiguous().realize()
    with self.assertRaises(RuntimeError): dst.assign(w05.expand(2))                   # weakfloat into int does not defer
    with self.assertRaises(RuntimeError): dst[0:1] = w05
@@ -142,7 +203,7 @@ class TestWeakMaterializationEntries(unittest.TestCase):
  def test_reads_commit_storage_raises(self):
    for weak, value, strong in ((dtypes.weakfloat, 0.5, dtypes.default_float),):
      def weak_val():
-        return Tensor([True], device="CPU").where(Tensor.const(weak, value), Tensor.const(weak, value))
+        return Tensor([True], device="CPU").where(Tensor.const(value, weak), Tensor.const(value, weak))
      self.assertEqual(weak_val().dtype, weak)
      self.assertEqual(weak_val().to("CPU").dtype, weak)
      self.assertEqual(weak_val().data().format, strong.fmt)
@@ -158,7 +219,7 @@ class TestWeakMaterializationEntries(unittest.TestCase):
  def test_weak_is_virtual(self):
    # NOTE: int64 lub uint64 is weakfloat, so this is device-ful weak from promotion, never from a cast to weak
    devful = Tensor([1], dtype=dtypes.int64, device="CPU") + Tensor([1], dtype=dtypes.uint64, device="CPU")
-    for t in (Tensor.const(dtypes.weakfloat, 0.5), devful):
+    for t in (Tensor.const(0.5), devful):
      self.assertTrue(t.uop.is_virtual)
      # realize is a no-op, so a weak input can never become the real buffer TinyJit needs
      with self.assertRaises(JitError): TinyJit(lambda x: (x+1).realize())(t)
@@ -169,7 +230,7 @@ class TestWeakMaterializationEntries(unittest.TestCase):

  def test_empty_reads_commit(self):
    for weak, strong in ((dtypes.weakfloat, dtypes.default_float),):
-      empty = Tensor.const(weak, 0).reshape(1).shrink(((0, 0),))
+      empty = Tensor.const(0, weak).reshape(1).shrink(((0, 0),))
      self.assertEqual(empty.data().format, strong.fmt)
      self.assertEqual(empty.numpy().dtype.itemsize, strong.itemsize)
      self.assertEqual(empty.tolist(), [])
@@ -2,8 +2,9 @@ import numpy as np
 import unittest
 from tinygrad.function import function
 from tinygrad import Tensor, GlobalCounters, Device
-from tinygrad.dtype import dtypes, Invalid
+from tinygrad.dtype import Invalid
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, ProgramInfo
+from test.helpers import assert_kernel_count

 class TestFunction(unittest.TestCase):
  def test_simple(self):
@@ -593,7 +594,7 @@ class TestFunctionTuple(unittest.TestCase):
    state = Tensor([10., 20., 30., 40.], device="CPU").contiguous().realize()
    @function(precompile=True, allow_implicit=True)
    def f(a:Tensor):
-      after = state.uop.after(state.uop.shrink(((0, 2),)).store(UOp.const(dtypes.float32, Invalid, shape=(2,))))
+      after = state.uop.after(state.uop.shrink(((0, 2),)).store(Invalid))
      return Tensor(after).contiguous() + a
    out = f(Tensor([1., 1., 1., 1.], device="CPU").contiguous().realize())
    np.testing.assert_allclose(out.numpy(), [11., 21., 31., 41.])
@@ -618,7 +619,7 @@ class TestFunctionTuple(unittest.TestCase):
    out = f(a)
    GlobalCounters.reset()
    out.realize()
-    self.assertEqual(GlobalCounters.kernel_count, kernel_count)
+    assert_kernel_count(kernel_count)
    np.testing.assert_allclose(out.numpy(), [3., 5., 7., 9.])

  def test_custom_kernel_precompile_further_compute_multi(self): self.test_custom_kernel_precompile_further_compute(multi=True, kernel_count=4)
@@ -134,14 +134,14 @@ class TestInvalidTensor(unittest.TestCase):
    self._invalid_test_helper(out, [1.0, 2.0, None, None])

  def test_uop_where_keeps_invalid_bare(self):
-    cond = UOp.const(dtypes.weakint, 0) < UOp.const(dtypes.weakint, 1)
-    idx = UOp(Ops.STACK, src=tuple(UOp.const(dtypes.weakint, x) for x in range(3)))
+    cond = UOp.const(0) < UOp.const(1)
+    idx = UOp(Ops.STACK, src=tuple(UOp.const(x) for x in range(3)))
    out = cond.where(idx, UOp.invalid())
    self.assertIs(cond.op, Ops.CMPLT)
    self.assertIs(idx.op, Ops.STACK)
    self.assertIs(out.op, Ops.WHERE)
    self.assertIs(out.src[2].op, Ops.CONST)
-    self.assertIs(out.src[2].arg, Invalid)
+    self.assertTrue(out.src[2].is_invalid)

 if __name__ == '__main__':
  unittest.main()
@@ -351,13 +351,13 @@ class TestJit(unittest.TestCase):
    @TinyJit
    def f(x:Tensor) -> Tensor: return (x + 1).realize()
    with self.assertRaises(JitError):
-      f(Tensor(UOp.const(dtypes.float, 2.0))).item()
+      f(Tensor(UOp.const(2.0).cast(dtypes.float))).item()

  def test_jit_deviceless_compute_input(self):
    @TinyJit
    def f(x:Tensor) -> Tensor: return (x + 1).realize()
    with self.assertRaises(JitError):
-      f(Tensor(UOp.const(dtypes.float, 2.0) + UOp.const(dtypes.float, 1.0))).item()
+      f(Tensor(UOp.const(2.0).cast(dtypes.float) + UOp.const(1.0).cast(dtypes.float))).item()

  def test_jit_init_empty_alt(self):
    @TinyJit
@@ -3,18 +3,53 @@ from unittest.mock import patch
 from tinygrad import Tensor, UOp
 from tinygrad.schedule import schedule_cache
 from tinygrad.llm.model import Transformer, TransformerConfig
+from tinygrad.llm.serve import StreamRouter

 TEST_CONFIG = TransformerConfig(num_blocks=1, dim=64, hidden_dim=128, n_heads=2, n_kv_heads=2,
                           norm_eps=1e-5, vocab_size=100, head_dim=32, rope_theta=10000.0, rope_dim=32, v_head_dim=32, max_context=32)
+V_START_POS = UOp.variable("start_pos", 0, TEST_CONFIG.max_context-1)
+V_TOKS = UOp.variable("toks", 1, 32)  # 32 is the default chunk_size in generate

 class TestTransformerGenerate(unittest.TestCase):
+  def test_warmup(self):
+    model, calls = Transformer(TEST_CONFIG), []
+    def generate(tokens):
+      calls.append(tokens)
+      yield from (1, 2)
+    with patch.object(model, "generate", generate): model.warmup()
+    self.assertEqual(calls, [[0], [0]])
+
+  def test_first_recurrent_generate_before_state_init(self):
+    model = Transformer(TEST_CONFIG)
+    model.has_recurrent_block = True
+    with patch.object(Transformer, '__call__', return_value=Tensor([[42]])):
+      self.assertEqual(next(model.generate([0])), 42)
+
+  def test_recurrent_live_state_reuse(self):
+    model = Transformer(TEST_CONFIG)
+    model.has_recurrent_block = True
+    model._cached_tokens = [1, 2, 3, 4, 5]
+    self.assertEqual(model.get_start_pos([1, 2, 3, 4, 5, 42, 10]), 5)
+    calls = []
+    def mock_call(self, tokens, start_pos, temperature, **kwargs):
+      calls.append((tokens.shape, start_pos))
+      return Tensor([[42]])
+    with patch.object(Transformer, '__call__', mock_call):
+      next(model.generate([1, 2, 3, 4, 5, 42, 10]))
+    self.assertEqual(calls, [((1, 1), V_START_POS.bind(5)), ((1, 1), V_START_POS.bind(6))])
+
+  def test_template_starts_reasoning(self):
+    router = StreamRouter(reasoning=True)
+    self.assertEqual(list(router.route("reasoning</think>answer")),
+                     [("reasoning_content", "reasoning"), ("content", "answer")])
+
  def test_kv_cache_reuse(self):
    """Test that generate reuses the KV cache when tokens extend the cached prefix."""
    model = Transformer(TEST_CONFIG)

    captured_inputs = []
-    def mock_call(self, tokens, start_pos, temperature):
-      captured_inputs.append((tokens.shape, start_pos if isinstance(start_pos, int) else start_pos.val))
+    def mock_call(self, tokens, start_pos, temperature, **kwargs):
+      captured_inputs.append((tokens.shape, start_pos))
      return Tensor([[42]])

    with patch.object(Transformer, '__call__', mock_call):
@@ -31,17 +66,15 @@ class TestTransformerGenerate(unittest.TestCase):
      next(gen)

    # should process tokens[6:] = [42, 10, 11, 12] since first 6 have cached k/v
-    toks_shape = captured_inputs[0][0][-1]
-    self.assertEqual(toks_shape.val if isinstance(toks_shape, UOp) else toks_shape, 4)
-    self.assertEqual(captured_inputs[0][1], 6)
+    self.assertEqual(captured_inputs, [((1, V_TOKS.bind(4)), V_START_POS.bind(6))])

  def test_kv_cache_invalidation(self):
    """Test that generate invalidates the KV cache when tokens diverge from the cached prefix."""
    model = Transformer(TEST_CONFIG)

    captured_inputs = []
-    def mock_call(self, tokens, start_pos, temperature):
-      captured_inputs.append((tokens.shape, start_pos if isinstance(start_pos, int) else start_pos.val))
+    def mock_call(self, tokens, start_pos, temperature, **kwargs):
+      captured_inputs.append((tokens.shape, start_pos))
      return Tensor([[42]])

    with patch.object(Transformer, '__call__', mock_call):
@@ -55,9 +88,7 @@ class TestTransformerGenerate(unittest.TestCase):
      next(gen)

    # should process all 3 tokens from start
-    toks_shape = captured_inputs[0][0][-1]
-    self.assertEqual(toks_shape.val if isinstance(toks_shape, UOp) else toks_shape, 3)
-    self.assertEqual(captured_inputs[0][1], 0)
+    self.assertEqual(captured_inputs, [((1, V_TOKS.bind(3)), V_START_POS.bind(0))])

  def test_two_prompts_schedule_cache(self):
    """Third prompt should hit the schedule cache, not miss (first two warm up both jits: prefill + decode)."""
@@ -90,7 +121,7 @@ class TestTransformerGenerate(unittest.TestCase):

    def get_prefill_flags(tokens, chunk_size):
      is_prefill = []
-      def mock_call(self, tokens, start_pos, temperature):
+      def mock_call(self, tokens, start_pos, temperature, **kwargs):
        is_prefill.append(resolve(tokens.shape[1] != 1))
        return Tensor([[42]])
      with patch.object(Transformer, '__call__', mock_call):
@@ -151,7 +182,7 @@ class TestTransformerGenerate(unittest.TestCase):
    """Temperature from generate should be passed through to __call__."""
    model = Transformer(TEST_CONFIG)
    captured_temps = []
-    def mock_call(self, tokens, start_pos, temperature):
+    def mock_call(self, tokens, start_pos, temperature, **kwargs):
      captured_temps.append(float(temperature.item()))
      return Tensor([[42]])
    with patch.object(Transformer, '__call__', mock_call):
@@ -17,7 +17,7 @@ class TestMetalGraph(unittest.TestCase):
      buf.op = Ops.SLICE
      src = MagicMock()
      src.dtype = dtypes.uint8
-      buf.src = (src, UOp.const(dtypes.weakint, offset))
+      buf.src = (src, UOp.const(offset))
      buf.dtype = dtypes.uint8
    else:
      buf.op = Ops.BUFFER
@@ -2,7 +2,7 @@ import unittest, numpy as np
 from tinygrad import Tensor, Variable, Context, Device, TinyJit, GlobalCounters, dtypes, UOp, nn, getenv
 from tinygrad.nn.state import get_parameters, get_state_dict
 from tinygrad.uop.ops import Ops
-from test.helpers import not_support_multi_device, needs_second_gpu, slow
+from test.helpers import not_support_multi_device, needs_second_gpu, slow, assert_kernel_count, KernelCountException
 from hypothesis import given, strategies as strat, settings

 settings.register_profile("my_profile", max_examples=200, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
@@ -60,8 +60,8 @@ class TestMultiTensor(unittest.TestCase):
  def test_shard_elementwise(self): self._test_shard_op(lambda t:(t+t).reshape(2, 2), [[2.,2.],[2.,2.]])
  def test_alu_deviceless_const(self):
    s = Tensor([1.0, 2, 3, 4]).shard((f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1"), axis=0)
-    np.testing.assert_equal((s + Tensor(UOp.const(dtypes.float, 1.0))).numpy(), [2, 3, 4, 5])
-    np.testing.assert_equal((s + Tensor(UOp.const(dtypes.float, 1.0)).reshape((1,)).expand((4,))).numpy(), [2, 3, 4, 5])
+    np.testing.assert_equal((s + Tensor(UOp.const(1.0).cast(dtypes.float))).numpy(), [2, 3, 4, 5])
+    np.testing.assert_equal((s + Tensor(UOp.const(1.0).cast(dtypes.float)).reshape((1,)).expand((4,))).numpy(), [2, 3, 4, 5])

  def test_add_rank_expand_shard(self):
    # a sharded src keeps its own rank under implicit broadcast, its shard axis right-aligns into the output
@@ -143,9 +143,8 @@ class TestMultiTensor(unittest.TestCase):
    GlobalCounters.reset()
    with Context(ALLREDUCE_CAST=1, RING=0, ALL2ALL=0):
      tst.realize()
-    kernel_count = GlobalCounters.kernel_count
+    assert_kernel_count(kernel_count)
    np.testing.assert_allclose(tst.numpy(), (a_src.numpy()+b_src.numpy()).sum(0))
-    self.assertEqual(kernel_count, kernel_count)

  def test_allreduce_cast_half_assign(self): self.test_allreduce_cast_half(assign=True, kernel_count=10)

@@ -584,7 +583,7 @@ class TestMultiTensor(unittest.TestCase):
      zeros = Tensor.zeros(3).realize()
    b = a.to(devices_2)*zeros.to(devices_2)
    sched = b.schedule_linear().src
-    self.assertEqual(len(sched), 0)
+    if len(sched) != 0: raise KernelCountException(0, len(sched))
    self.assertListEqual(b.tolist(), [0, 0, 0])

@unittest.skipIf(not_support_multi_device(), "no multi")
@@ -597,7 +596,7 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):
    t = Tensor.arange(64).reshape(8, 8).clone().realize()
    t.shard_([f"{Device.DEFAULT}:{i}" for i in range(4)], axis=0)

-    with self.assertRaises(AssertionError):
+    with self.assertRaises(RuntimeError):
      # sharded axis shrink on non-device boundry is not allowed
      a = t.shrink(((0, 3), (0, 8))).contiguous()
      a.schedule_linear()
@@ -1,15 +1,16 @@
 import unittest
 from tinygrad import Tensor, dtypes, GlobalCounters
+from test.helpers import assert_kernel_count

 class TestSetitemInto(unittest.TestCase):
  def test_setitem_into_unrealized(self):
    GlobalCounters.reset()
    t = Tensor.arange(4, dtype=dtypes.int32).reshape(2, 2)
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(t.tolist(), [[0, 1], [5, 5]])

@@ -18,11 +19,11 @@ class TestSetitemInto(unittest.TestCase):
    GlobalCounters.reset()
    a = Tensor.arange(8, dtype=dtypes.int32).reshape(2, 4)
    w = a[0] + a[1]  # unrealized ADD with SHRINK in graph: [4, 6, 8, 10]
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    w[1] = 99
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    w.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(w.tolist(), [4, 99, 8, 10])

@@ -30,61 +31,61 @@ class TestSetitemInto(unittest.TestCase):
    GlobalCounters.reset()
    t = Tensor.empty(4, dtype=dtypes.int32)
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(GlobalCounters.global_mem, 4)
    t[1].realize()
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(t[1].item(), 5)

  def test_setitem_into_empty_alu(self):
    GlobalCounters.reset()
    t = Tensor.empty(4, dtype=dtypes.int32) + 1
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertLessEqual(GlobalCounters.global_mem, 32)
    t[1].realize()
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(t[1].item(), 5)

  def test_setitem_into_tensor(self):
    t = Tensor([1, 2, 3, 4], dtype=dtypes.int32).realize()
    GlobalCounters.reset()
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t[1].realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(GlobalCounters.global_mem, 4)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertListEqual(t.tolist(), [1, 5, 3, 4])

  def test_setitem_into_tensor_alu(self):
    t = Tensor([1, 2, 3, 4], dtype=dtypes.int32).realize() + 1
    GlobalCounters.reset()
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t[1].realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertLessEqual(GlobalCounters.global_mem, 32)
    t[1].realize()
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertListEqual(t.tolist(), [2, 5, 4, 5])

  def test_setitem_into_const(self):
    GlobalCounters.reset()
    t = Tensor.ones(4, dtype=dtypes.int32, buffer=False)
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(t.tolist(), [1, 5, 1, 1])

@@ -92,9 +93,9 @@ class TestSetitemInto(unittest.TestCase):
    GlobalCounters.reset()
    t = Tensor.ones(4, dtype=dtypes.int32, buffer=False) + 1
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(t.tolist(), [2, 5, 2, 2])

@@ -105,18 +106,18 @@ class TestSetitemInto(unittest.TestCase):
    t = Tensor.arange(4, dtype=dtypes.int32)
    self.assertIs(other.uop, t.uop)
    t[1] = 5
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    self.assertListEqual(t.tolist(), [0, 5, 2, 3])

  def test_setitem_slice_const(self):
    t = Tensor.zeros(100, dtype=dtypes.int32).contiguous().realize()
    GlobalCounters.reset()
    t[20:50] = 3
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(GlobalCounters.global_mem, 30*4)  # 30 elements written

  def test_setitem_slice_tensor(self):
@@ -124,18 +125,18 @@ class TestSetitemInto(unittest.TestCase):
    v = Tensor.zeros(30, dtype=dtypes.int32).contiguous().realize()
    GlobalCounters.reset()
    t[20:50] = v
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(GlobalCounters.global_mem, 30*4*2)  # 30 read + 30 written

  def test_setitem_full(self):
    t = Tensor.zeros(100, dtype=dtypes.int32).contiguous().realize()
    GlobalCounters.reset()
    t[:] = 3
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    t.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertEqual(GlobalCounters.global_mem, 100*4)  # full buffer written

 if __name__ == '__main__':
@@ -66,8 +66,8 @@ class TestTensorData(unittest.TestCase):
    assert dat.shape == ()

  def test_const_dtype_for_uop(self):
-    self.assertEqual(Tensor.const(dtypes.int8, UOp.const(dtypes.float32, 1.0)).dtype, dtypes.int8)
-    self.assertEqual(Tensor.const(dtypes.int32, UOp.variable("x", 1, 10).bind(5)).item(), 5)
+    self.assertEqual(Tensor.const(UOp.const(1.0).cast(dtypes.float32), dtypes.int8).dtype, dtypes.int8)
+    self.assertEqual(Tensor.const(UOp.variable("x", 1, 10).bind(5), dtypes.int32).item(), 5)

  def test_data_float32(self):
    a = Tensor([[1,2.5],[3,4]], dtype=dtypes.float32)
@@ -57,11 +57,11 @@ def _make_buffer_view(src:UOp) -> UOp|None:
  if (offset := src.contiguous_view_offset()) is None: return None
  buf = src.base
  if buf.op is Ops.SLICE:
-    byte_offset = buf.src[1].arg * buf.src[0].dtype.itemsize + offset * src.dtype.itemsize
+    byte_offset = buf.src[1].val * buf.src[0].dtype.itemsize + offset * src.dtype.itemsize
    buf = buf.src[0]
    if byte_offset % buf.dtype.itemsize != 0: return None
    offset = byte_offset // buf.dtype.itemsize
-  return UOp(Ops.SLICE, src.dtype, (buf, UOp.const(None, offset)), src.numel())
+  return UOp(Ops.SLICE, src.dtype, (buf, UOp.const(offset)), src.numel())

 def contiguous_mops_to_view(c:UOp, src:UOp):
  """MOPS(BUFFER) → SLICE when movement ops collapse to a contiguous range."""
@@ -83,7 +83,7 @@ def contiguous_mops_to_view(c:UOp, src:UOp):
    resolved = graph_rewrite(src, multi_pm, name="multi_buffer_view")
    if resolved.op is not Ops.UNSHARD: return None
    if (view := _make_buffer_view(resolved.src[0])) is None: return None
-    return view.reshape(resolved.src[0].shape).unshard(resolved.arg, resolved.src[1]).contiguous(tag=c.tag)
+    return view.reshape(resolved.src[0].shape).unshard(resolved.arg, resolved.src[1:]).contiguous(tag=c.tag)

  return None

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	1796f6bb5b	more KernelCountException	2026-08-04 21:37:31 -07:00
sirhcmandGitHub	de57be1f26	kill nvidia pids at benchmarks start (#17406 )	2026-08-04 23:45:49 -04:00
chenyuandGitHub	9b508dfafc	remove invalid special case in cast [PR] (#17405 )	2026-08-04 23:08:07 -04:00
George HotzandGitHub	e1f42681fa	add new schedule tests + format better (#17402 ) * add new schedule tests + format better * assert_kernel_count	2026-08-04 18:46:38 -07:00
chenyuandGitHub	3eab809e06	update minimum to not create strong type const [PR] (#17401 )	2026-08-04 21:29:24 -04:00
George HotzandGitHub	6122b3c98f	use check_schedule in tests where possible (#17400 )	2026-08-04 18:17:29 -07:00
chenyuandGitHub	f295f9fc99	use weak 0 in convert_pad_to_where_to_keep_behavior_local [pr] (#17398 )	2026-08-04 19:44:34 -04:00
chenyuandGitHub	d79772f057	fix pow on extreme inputs (#17397 ) * fix pow on extreme inputs * WEBGPU	2026-08-04 19:30:43 -04:00
chenyuandGitHub	c1a10e0726	fix _min_max for CAST from float to int [pr] (#17396 ) * fix _min_max for CAST from float to int [pr] * fix	2026-08-04 18:06:50 -04:00
chenyuandGitHub	7b6d2ddf23	more weak const without cast in const_like [PR] (#17395 ) * more weak const without cast in const_like [PR] * that?	2026-08-04 15:10:43 -04:00
George HotzandGitHub	f489f4bc36	add test_eye + color INDEX (#17394 )	2026-08-04 11:08:08 -07:00
wozeparrotandGitHub	80d2073a11	fa: fix dq hazard with D=64 (#17393 )	2026-08-04 09:38:01 -07:00
chenyuandGitHub	85e9440446	keep weak consts weak in symbolic [PR] (#17392 )	2026-08-04 12:16:53 -04:00
chenyuandGitHub	0db63e13cf	update dtype_from_uop for IMAGE INDEX [PR] (#17391 )	2026-08-04 11:50:24 -04:00
chenyuandGitHub	6eedca57ce	dtype_from_uop for CUSTOM, CUSTOMI, PYLITERAL [PR] (#17390 ) these do not change dtype	2026-08-04 11:02:13 -04:00
b1tgandGitHub	0796853845	support symbolic shapes in allreduce (#17364 )	2026-08-04 07:53:12 -07:00
geohot	568bfb6a37	Revert "usb: wait for PCIe link after power on (#17380 )" This reverts commit `104ee90ccf`.	2026-08-04 06:42:26 -07:00
qazalandGitHub	f993228b73	llama: accurate mxfp4 mfu (#17388 ) * llama: accurate mxfp4 mfu * train_llama3 import	2026-08-04 18:41:57 +09:00
qazalandGitHub	c9cd44b751	more custom kernel contig input edge case tests (#17387 )	2026-08-04 18:20:50 +09:00
chenyuandGitHub	13fff4f258	const_like cleanups [PR] (#17386 )	2026-08-03 23:57:20 -04:00
chenyuandGitHub	0170a3075a	use python bitcast in fold_bitcast [PR] (#17385 )	2026-08-03 23:09:53 -04:00
George HotzandGitHub	c21a552f3d	llm: bugfixes + warmup (#17384 )	2026-08-03 18:23:14 -07:00
chenyuandGitHub	67dc02d7e7	bitcast in python for _bits_to_rand [PR] (#17383 ) * bitcast in python for _bits_to_rand [PR] const in mixin would be weak only without width, so not bitcast	2026-08-03 21:14:58 -04:00
George HotzandGitHub	3cb786f447	llm: update test_llm_server tests (#17382 )	2026-08-03 16:06:17 -07:00
geohot	87289a7410	hotfix: revert test_scalar_alu_index, violates spec	2026-08-03 15:25:51 -07:00
George HotzandGitHub	c2625c78cb	scalar ALU index fix + llm: preserve_thinking (#17381 ) * cstyle: scalar ALU index fix, serve: preserve_thinking, test: fix Handler import - cstyle.py: return scalar directly when ALU buffer has 1 element - cli.py: add preserve_thinking param to FallbackTemplate.render - serve.py: pass preserve_thinking=True when rendering chat completions - test_llm_server.py: fix import to use Handler from llm.serve * real fix	2026-08-03 15:10:21 -07:00
chenyuandGitHub	33755a3465	improve threefry codegen [pr] (#17379 ) decomp uint64 can handle part of it	2026-08-03 17:24:25 -04:00
YassineYousfiandGitHub	104ee90ccf	usb: wait for PCIe link after power on (#17380 )	2026-08-03 14:19:05 -07:00
chenyuandGitHub	a2385ae21d	MAX_LINE_COUNT=26000 (#17378 ) oh well	2026-08-03 15:37:58 -04:00
wozeparrotandGitHub	3331944547	gptoss: fix moe routing (#17377 )	2026-08-03 11:16:44 -07:00
nimlgenandGitHub	7c1ce50f63	hcq2: epoch (#17376 ) * hcq2: epoch * x * minor	2026-08-03 16:37:25 +03:00
qazalandGitHub	be5f62d269	llama: refactor amax stuff and skip in fp4 (#17375 )	2026-08-03 20:05:52 +09:00
nimlgenandGitHub	e22935c758	hcq2: inputs table (#17374 ) * revert this * x * simpler * fix	2026-08-03 13:40:43 +03:00
chenyuandGitHub	314df72b5f	deflake test_hcq with MOCKGPU (#17370 ) for MOCKGPU we compare with e2e wall time which would be device agnostic	2026-08-03 13:22:59 +03:00
chenyuandGitHub	23c7813f44	no hard coded dtype int for rangeify debuf [PR] (#17373 )	2026-08-02 20:43:04 -04:00
George HotzandGitHub	05bc7c6994	fix llm reasoning and Linear import (#17372 )	2026-08-02 12:31:52 -07:00
chenyuandGitHub	09dabfe05e	minor pm_float_decomp cleanup [PR] (#17371 ) make the rule order independently correct	2026-08-02 15:17:33 -04:00
chenyuandGitHub	59df317b12	use UOp.const to create new consts [PR] (#17368 ) replace arg won't work with ConstArg	2026-08-02 13:31:55 -04:00
wozeparrotandGitHub	e14cadb1fb	gptoss: set ASM_GEMM (#17363 )	2026-08-02 06:52:36 -07:00
chenyuandGitHub	0258c7fefc	minor argstr and alloc stack cleanup [PR] (#17361 )	2026-08-01 22:37:06 -04:00
George HotzandGitHub	fb607fb990	faster devectorizer with one line (#17360 )	2026-08-01 13:06:35 -07:00
George HotzandGitHub	15c936db01	merge devectorize + indexing (#17354 ) * external benchmark schedule in 5 sec (codex slop) * prune * real? * delete	2026-08-01 11:40:15 -07:00
wozeparrotandGitHub	98b700bad1	gptoss optim fixes (#17356 )	2026-08-01 09:10:06 -07:00
chenyuandGitHub	8e524ca467	CONST related cleanups [pr] (#17352 ) ConstFloat(nan) != nan should be False, and some Invalid bool cleanups	2026-08-01 09:54:28 -04:00
nimlgenandGitHub	665822ab34	hcq2: faster replace (#17353 ) * hcq2: use runtime device for submit params * hcq2: parameterize buffers in linear time	2026-08-01 16:32:04 +03:00
chenyuandGitHub	6c0ec39279	UOp.is_invalid [PR] (#17351 ) helper to prep ConstArg	2026-08-01 02:36:32 -04:00
chenyuandGitHub	b502fc1367	more const arg -> val (#17350 )	2026-08-01 02:18:10 -04:00
qazalandGitHub	161783d8f7	add _device_num back to ast.variables (kimi) (#17327 )	2026-08-01 14:56:50 +09:00
George HotzandGitHub	5a1c641f79	more arg -> val (#17349 ) * more arg -> val * kimi * more	2026-07-31 22:44:24 -07:00
Adeeb ShihadehandGitHub	20b8ecff50	support new USB vendor ID (#17348 )	2026-07-31 21:45:06 -07:00
George HotzandGitHub	9082ecef5d	use .val to access the value of Ops.CONST (#17347 )	2026-07-31 19:51:42 -07:00
George HotzandGitHub	099d69ff7d	ci: split macos unit test into metal and mock runners (#17346 )	2026-07-31 19:41:18 -07:00
George HotzandGitHub	a88f832f0c	remove UOp.val (#17345 )	2026-07-31 18:58:38 -07:00
chenyuandGitHub	850989115d	__int__ and __float__ work for weak (#17342 )	2026-07-31 19:56:37 -04:00
sirhcmandGitHub	15d515299e	heuristics: try multiple TC axes (#17341 )	2026-07-31 19:25:18 -04:00
sirhcmandGitHub	85ced44db6	tc: don't allow reduce over output dims (#17340 )	2026-07-31 16:48:51 -04:00
chenyuandGitHub	277433259e	fix sym_infer for CAST (#17338 )	2026-07-31 14:38:38 -04:00
chenyuandGitHub	8dc225e28e	dtype_from_uop(INS) is None [PR] (#17337 )	2026-07-31 14:33:15 -04:00
chenyuandGitHub	ad24750487	const(value, dtype) -> const(value).cast(dtype) in tests (#17335 )	2026-07-31 13:24:31 -04:00
qazalandGitHub	a11ee26bb8	viz: prep for faster cli DEBUG=3 (#17334 ) * move data * split	2026-08-01 02:16:30 +09:00
sirhcmandGitHub	b95bd5b2a5	don't reset chestnut in benchmark (#17333 )	2026-07-31 12:39:04 -04:00
nimlgenandGitHub	1095bbe409	hcq2: fix ib reuse (#17330 ) * hcq2: initialize IB reuse counters at link * x	2026-07-31 18:43:23 +03:00
chenyuandGitHub	7f4dbb8090	remove shape= from UOp.const [PR] (#17331 ) inlined to const_like	2026-07-31 11:18:59 -04:00
chenyuandGitHub	8e2f175542	const(dtype, b) -> const(b, dtype) [PR] (#17328 ) prep for dtype removal	2026-07-31 09:46:37 -04:00
nimlgenandGitHub	155b84ee80	hcq2 faster schedule (#17324 ) * avoid quadratic STACK dtype promotion * build HCQ patch stacks directly * pack HCQ command buffers linearly * remove HCQ command buffer simplification	2026-07-31 16:08:18 +03:00
wozeparrotandGitHub	4f5cadd15d	gptoss ci (#17325 )	2026-07-31 05:51:58 -07:00
nimlgenandGitHub	0c4bfaeb48	coalesce ints (#17323 ) * merge ints * fix z3 validation of coalesced loads * fix uint vector names in CUDA and Metal	2026-07-31 15:49:56 +03:00
qazalandGitHub	6d2700f0b7	failing test for unbound _device_num err in BEAM (#17326 ) * min failing test * switch to cpu * err	2026-07-31 20:24:44 +09:00
wozeparrotandGitHub	93b74c75fc	gptoss: grouped moe (#17322 )	2026-07-31 03:25:32 -07:00
qazalandGitHub	f7964acb64	llama with MXFP4 (#17321 ) * mxfp4 in llama * less * name	2026-07-31 18:25:25 +09:00
qazalandGitHub	0a3325f9c2	add mxfp4 quantize and layout kernels (#17320 )	2026-07-31 14:50:18 +09:00
sirhcmandGitHub	13452b3775	benchmark comma big model (#17312 )	2026-07-31 01:37:34 -04:00
qazalandGitHub	a8c1e89500	fp4 asm gemm 6+ pflops (#17315 ) * fp4 gemm * better kernargs structure * move to .s files * work * work * p2 * style * use .py * move to dsl * cleanup * add MFMA_SCALE_X2_ENCODING * cleanup mfma * fma docs * gemm_mxfp4 * more cleanup * move * move to cdna_asm_gemm * change * rm * change * mx	2026-07-31 14:25:23 +09:00
George HotzandGitHub	d65ea465ed	cleanup gemm fragment + add store unshard (#17313 ) * cleanup gemm fragment + add store unshard * multi * fix	2026-07-30 20:53:28 -07:00
chenyuandGitHub	6608b9d8e2	adjust lower weak order [pr] (#17314 ) * adjust lower weak order [pr] * no symbolic_simple	2026-07-30 23:36:12 -04:00
Noah SchiroandGitHub	f2c2f4456b	Add softmin (#17292 ) * Add softmin * Remove extra tests	2026-07-30 20:05:11 -04:00
chenyuandGitHub	fe8ece7efa	fix weak for image gate fusion [pr] (#17310 )	2026-07-30 19:47:47 -04:00
chenyuandGitHub	ce500c1946	broadcast doesn't cast CONST [pr] (#17308 ) * broadcast doesn't cast CONST [pr] it might change arg from python int to python float, but not cast the CONST, so we don't need or_casted in symbolic. we lower the weak dtype right before decomp * fix	2026-07-30 18:51:59 -04:00
sirhcmandGitHub	b290372121	ftdi reset chestnut before running comma benchmark (#17306 )	2026-07-30 14:28:45 -04:00
sirhcmandGitHub	da15c43e51	update openpilot benchmarks (#17304 )	2026-07-30 14:22:20 -04:00
chenyuandGitHub	b488cc7df2	update dtype_from_uop for SHR/SHL (#17307 ) certain backend cast the distance, and we should not re-broadcast it	2026-07-30 13:57:58 -04:00
chenyuandGitHub	341c4ed4f5	rewrite writes new dtype with dtype_from_uop try 2 [pr] (#17305 ) with NIR fix which is INDEX related some Ops currently depends on explicitly set dtype and not dtype_from_uop	2026-07-30 13:47:21 -04:00
chenyuandGitHub	d05a3e6c0b	Revert "rewrite writes new dtype with dtype_from_uop [pr] (#17302 )" (#17303 ) This reverts commit `417245ab26`.	2026-07-30 13:08:26 -04:00
chenyuandGitHub	417245ab26	rewrite writes new dtype with dtype_from_uop [pr] (#17302 ) some Ops currently depends on explicitly set dtype and not dtype_from_uop	2026-07-30 12:46:24 -04:00
b1tgandGitHub	e25bf77ce9	kimi delta attention (#17281 ) * kimi delta attention * config * test	2026-07-30 08:23:56 -07:00
wozeparrotandGitHub	b5a2a5666a	gptoss moe routing (#17284 )	2026-07-30 07:49:14 -07:00
nimlgenandGitHub	3f6f0a19ab	hcq2: less graph_rewrites (#17300 )	2026-07-30 16:44:45 +03:00
qazalandGitHub	7b7b9c1ae0	spec tests for movement ops before custom_kernel (#17299 ) * spec * fixing * xfail * remove * update tests * add back * change to max_kernels * split and cleanup	2026-07-30 19:20:24 +09:00
qazalandGitHub	aba5ba4b40	update rangeify comment for mop after AFTER (#17298 )	2026-07-30 17:09:32 +09:00
chenyuandGitHub	d52ef3077c	fix do_devectorize dtype (#17295 ) each one derives its dtype from src, not shared	2026-07-29 23:04:25 -04:00
sirhcmandGitHub	bfc9fc6e0a	nicer TinyJit decorator (#17293 )	2026-07-29 21:48:00 -04:00
sirhcmandGitHub	060f447db6	qcom: match cl for SP_CS_INSTR_SIZE (#17289 )	2026-07-29 18:47:02 -04:00
chenyuandGitHub	fd912b348c	generic c0*x<c1 [pr] (#17291 )	2026-07-29 18:45:23 -04:00
George HotzandGitHub	138676ab81	improve fragment example + index unshard (kimi) (#17288 ) * fix dtypes in fragment example * match tilelang * flip locals * fix index on unshard * test fixes * kimi needs more taste	2026-07-29 15:38:38 -07:00
chenyuandGitHub	027907a544	fix c0*x<c1 symbolic [pr] (#17290 ) use `//` and not math.ceil	2026-07-29 18:32:31 -04:00
nimlgenandGitHub	d4ba8b6e0f	hcq2: use stack (#17286 )	2026-07-29 22:09:34 +03:00
George HotzandGitHub	b30c7e00d4	support 2d on UNSHARD (kimi) (#17285 ) * support 2d on UNSHARD * fixes * Fix test and spec * single barrier * 2d sharding works for devices too * cleanups * no _rewrap	2026-07-29 12:01:59 -07:00
sirhcmandGitHub	aab51fb7b6	64-bit UOp.variable support, try 2 (#17256 )	2026-07-29 14:39:36 -04:00
chenyuandGitHub	dd86a30798	test case for weakfloat cast to weakint INDEX (#17287 )	2026-07-29 14:15:52 -04:00
George HotzandGitHub	52c9e5a99e	rename LOOP -> WEAK and STRONGLOOP -> LOOP (#17283 )	2026-07-29 10:38:36 -07:00
chenyuandGitHub	6c2b9fac08	weak const cleanups [PR] (#17282 )	2026-07-29 13:14:20 -04:00
George HotzandGitHub	bd296a7359	enable alloc_fragment support with UNSHARD (kimi) (#17272 ) * enable alloc_fragment support with UNSHARD (kimi) * cleaner with implicit barrier * cleanups * cleaner * strongloop * dcount cleanups	2026-07-29 09:46:45 -07:00
Noah SchiroandGitHub	3df1b07c86	Patch half precision ops for older gpu archs (#17274 )	2026-07-29 09:42:24 -07:00
chenyuandGitHub	3803f1583b	fix add_raw_barrier [pr] (#17278 ) * fix add_raw_barrier [pr] a kernel might need multiple barriers * smaller?	2026-07-29 09:30:00 -07:00
qazalandGitHub	6ea7d366fa	test permuted input in custom Ops.PROGRAM test (#17279 ) * test permuted input in custom Ops.PROGRAM test * dtype	2026-07-29 18:29:05 +09:00
George HotzandGitHub	527e57300c	fix smu reset for kernel >= 7 (#17277 ) * fix smu reset for kernel >7 * match kernel	2026-07-28 23:11:21 -07:00