ptx also

simplifies _one_hot_along_dim

.github/workflows/benchmark.yml

@@ -521,6 +521,8 @@ jobs:
       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: Test copy speeds
+      run: SIZE=64e6 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3 test/external/external_test_usb_asm24.py TestDevCopySpeeds
 
   driverbenchmarks:
     name: PCI Driver Benchmark (DEV=${{ matrix.dev }})

.github/workflows/szdiff.yml

@@ -14,12 +14,15 @@ jobs:
     outputs:
       branchstat: ${{ steps.brstat.outputs.stat}}
     steps:
-      - name: Check code from PR branch 
+      - name: Check code from PR branch
         uses: actions/checkout@v6
         with:
           repository: ${{ github.event.pull_request.head.repo.full_name }}
           ref: ${{ github.event.pull_request.head.sha }}
           fetch-depth: 0
+          # PR code is only inspected with git rev-list, never executed
+          allow-unsafe-pr-checkout: true
+          persist-credentials: false
       - name: Check whether branch is up-to-date
         id: brstat
         run: |
@@ -51,6 +54,9 @@ jobs:
           repository: ${{ github.event.pull_request.head.repo.full_name }}
           ref: ${{ github.event.pull_request.head.sha }}
           path: pr
+          # PR code is only line-counted by master's sz.py, never executed
+          allow-unsafe-pr-checkout: true
+          persist-credentials: false
         # the base default to tinygrad master and cannot be other fork branch for security purpose
       - name: Checkout code from tinygrad master
         uses: actions/checkout@v6

.github/workflows/test.yml

@@ -171,7 +171,7 @@ jobs:
         llvm: 'true'
         amd: 'true'
     - name: Run NULL backend tests
-      run: DEV=NULL python -m pytest -n=auto test/null/ --durations=20
+      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
     - name: Run targeted tests on NULL backend
       run: |
         DEV=NULL python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
@@ -291,7 +291,7 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1391 ALLOWED_GATED_READ_IMAGE=58 FLOAT16=1 DEV="CL::IMAGE_PITCH_ALIGNMENT=64" IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=54 FLOAT16=1 DEV="CL::IMAGE_PITCH_ALIGNMENT=64" IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
           # IMAGE_PITCH_ALIGNMENT=64 matches adreno 630
       - name: Test openpilot CL compile fp32 (test correctness)
         run: |
@@ -500,6 +500,29 @@ jobs:
       - name: Run LLVM test
         run: DEV=MOCKKFD+AMD:LLVM python test/device/test_amd_llvm.py
 
+  hcq2:
+    name: hcq2
+    runs-on: *linux
+    timeout-minutes: 5
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v6
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: hcq2
+          deps: testing_unit
+          amd: 'true'
+      - name: Run HCQ2 tests
+        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/test_tiny.py
+      - name: Run HCQ2 multi-device tests
+        run: |
+          HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_multitensor.py \
+            TestMultiTensor.test_simple_add TestMultiTensor.test_shard_reduce \
+            TestMultiTensor.test_backward_sum TestMultiTensor.test_matmul_shard_0_0
+      - name: Run HCQ2 JIT tests
+        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_jit.py
+
   testmockam:
     name: Linux (am)
     runs-on: *linux
@@ -620,7 +643,7 @@ jobs:
     - name: Run unit tests
       run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
     - name: Run NULL backend tests
-      run: DEV=NULL python -m pytest -n=auto test/null/ --durations=20
+      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)

conftest.py

@@ -0,0 +1,10 @@
+import os, pytest, signal, threading
+
+@pytest.hookimpl(wrapper=True)
+def pytest_runtest_call(item):
+  t = threading.Timer(int(os.getenv("TEST_TIMEOUT", 300)), os.kill, args=(os.getpid(), signal.SIGABRT))
+  t.start()
+  try: yield
+  finally:
+    t.cancel()
+    t.join()

examples/mlperf/model_train.py

@@ -1462,6 +1462,8 @@ 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)
     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)
@@ -1753,11 +1755,12 @@ def train_gptoss():
 
   scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)
 
-  # realize everything here
-  if optim.master_params: Tensor.realize(*optim.master_params)
+  if optim.master_params:
+    for m in optim.master_params: m.realize()
   Tensor.realize(*optim.params, *fp8_inv_scales)
 
   @TinyJit
+  @Context(TRAINING=1)
   def minibatch(tokens:Tensor):
     if is_dp: tokens = tokens.to(None).shard(device, 0)
     if not is_sharding: tokens = tokens.to(None)

examples/mlperf/models/flat_llama.py

@@ -37,8 +37,8 @@ def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
   return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal(), new_amax
 
 def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
-           x_fp8:Tensor|None=None, x_new_amax:Tensor|None=None,
-           grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None, x_prequant_mx:tuple|None=None) -> tuple[Tensor,...]:
+           x_fp8:Tensor|None=None, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None, x_prequant_mx:tuple|None=None,
+           next_amax_x:Tensor|None=None) -> tuple[Tensor,...]:
   if not fp8:
     if ASM_GEMM:
       from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
@@ -56,13 +56,14 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
     else:
       x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*l_shape, x_q.shape[-1])
       out = x_phys @ (w.cast(dtypes.bfloat16) * _mx_block_scale(w_inv_scale)).T
-    return out, (amax_x.detach() if amax_x is not None else None), x_q
+    return out, x_q
   if x_fp8 is None:
-    if FUSED_INPUT_QUANTIZE and amax_x is not None:
+    if FUSED_INPUT_QUANTIZE:
       from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed
-      x_fp8, _, x_new_amax, _ = quantize_fp8_delayed(x, amax_x, FP8_DTYPE)
+      x_fp8, _ = quantize_fp8_delayed(x, amax_x, next_amax_x, FP8_DTYPE)
     else:
-      x_fp8, _, x_new_amax = quantize_fp8(x, amax_state=amax_x)
+      x_fp8, _, new_amax_x = quantize_fp8(x, amax_state=amax_x)
+      next_amax_x.assign(new_amax_x)
   if ASM_GEMM:
     from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
     if can_use_asm_gemm(x_fp8, w.T):
@@ -73,51 +74,51 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
       else:
         out = asm_gemm(x_fp8, w.T, x_scale=amax_x, w_scale=w_inv_scale, grad_amax_state=grad_amax_state,
                        next_grad_amax_state=next_grad_amax_state)
-      return out, x_new_amax, 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_new_amax, x_fp8
+      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,
-                         grad_amax_state:Tensor, next_grad_amax_state:Tensor):
+                         next_amax_x:Tensor, grad_amax_state:Tensor, next_grad_amax_state:Tensor):
   if FUSED_ADD_NORM_MUL_QUANTIZE:
     from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_rmsnorm_mul_quantize_fp8
-    x_fp8, new_amax, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE)
-    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x, x_new_amax=new_amax,
+    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,
                        grad_amax_state=grad_amax_state, next_grad_amax_state=next_grad_amax_state)
     return out, x_normed, rrms, ret
   x_normed, rrms = rmsnorm(x, eps)
   out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale, grad_amax_state=grad_amax_state,
-                     next_grad_amax_state=next_grad_amax_state)
+                     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,
-                             grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
+                             next_amax_x:Tensor, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
   if FUSED_ADD_NORM_MUL_QUANTIZE:
     from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
-    x_fp8, new_amax, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE)
-    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x, x_new_amax=new_amax,
+    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,
                        grad_amax_state=grad_amax_state, next_grad_amax_state=next_grad_amax_state)
     return out, h, x_normed, rrms, ret
   h = x + residual
   x_normed, rrms = rmsnorm(h, eps)
   out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale, grad_amax_state=grad_amax_state,
-                     next_grad_amax_state=next_grad_amax_state)
+                     next_grad_amax_state=next_grad_amax_state, next_amax_x=next_amax_x)
   return out, h, x_normed, rrms, ret
 
 def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
-                             amax_x2: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:
     from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
-    x2_fp8, new_amax_x2 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13,
-                                                 next_grad_amax_state=next_grad_amax_xw13)
-    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, amax_x=amax_x2, x_new_amax=new_amax_x2,
+    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)
+    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, amax_x=amax_x2,
                        grad_amax_state=grad_amax_xout, next_grad_amax_state=next_grad_amax_xout)
     return out, ret
   hidden = x_w13.shape[-1] // 2
   x_w1, x_w3 = x_w13[..., :hidden], x_w13[..., hidden:]
   out, *ret = matmul(x_w1.silu() * x_w3, w2, amax_x=amax_x2, w_inv_scale=s_2, grad_amax_state=grad_amax_xout,
-                     next_grad_amax_state=next_grad_amax_xout)
+                     next_grad_amax_state=next_grad_amax_xout, next_amax_x=next_amax_x2)
   return out, ret
 
 class FlatTransformer:
@@ -154,7 +155,7 @@ class FlatTransformer:
     self.tok_embeddings = nn.Embedding(vocab_size, dim)
     self.tok_embeddings.weight = Tensor.normal(vocab_size, dim, mean=0.0, std=0.02, dtype=dtypes.bfloat16)
     self.output = Tensor.normal(1, vocab_size, dim, mean=0.0, std=0.02, dtype=dtypes.bfloat16)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().is_param_(False)
+    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)
     names = ["xqkv", "xo", "x2"]
@@ -186,89 +187,87 @@ class FlatTransformer:
 
   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):
     bsz, seqlen, _ = x.shape
-    amaxs, saves = [], []
+    saves = []
 
-    xqkv, x_normed, rrms, (new_amax, *s) = norm_quantize_matmul(x, attention_norm, wqkv, s_qkv, self.norm_eps,
+    xqkv, x_normed, rrms, s = norm_quantize_matmul(x, attention_norm, wqkv, s_qkv, self.norm_eps,
                                                                   amax_x=amax_xqkv, grad_amax_state=grad_amax_xqkv,
-                                                                  next_grad_amax_state=next_grad_amax_xqkv)
-    amaxs.append(new_amax)
+                                                                  next_grad_amax_state=next_grad_amax_xqkv, next_amax_x=next_amax_xqkv)
     saves.extend([x_normed, rrms, *s, xqkv])
-    xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
-    xq = xqkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
-    xk = xqkv[:, :, :, self.n_rep].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-    xv = xqkv[:, :, :, self.n_rep+1].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)
     if getenv("HK_FLASH_ATTENTION"):
-      from extra.thunder.amd.fa import flash_attention
+      from extra.thunder.amd.fa import flash_attention, fused_qkv_rope
+      xq, xk, xv = fused_qkv_rope(xqkv, freqs_cis, self.n_heads, self.n_kv_heads, self.head_dim)
       attn, *save = flash_attention(xq, xk, xv, is_causal=True, write_flat=True)
       saves.extend(save)
     else:
+      xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
+      xq = xqkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
+      xk = xqkv[:, :, :, self.n_rep].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
+      xv = xqkv[:, :, :, self.n_rep+1].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
+      xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
+      xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)
       xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
       attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
     attn = attn.reshape(bsz, seqlen, -1)
 
-    out, new_amax, *s = matmul(attn, wo, amax_x=amax_xo, w_inv_scale=s_o, grad_amax_state=grad_amax_xo,
-                               next_grad_amax_state=next_grad_amax_xo)
-    amaxs.append(new_amax)
+    out, *s = matmul(attn, wo, amax_x=amax_xo, w_inv_scale=s_o, grad_amax_state=grad_amax_xo,
+                               next_grad_amax_state=next_grad_amax_xo, next_amax_x=next_amax_xo)
     saves.extend([*s, out])
-    return out, amaxs, saves
+    return out, saves
 
   def feed_forward(self, x:Tensor, residual:Tensor, **kwargs):
-    amaxs, saves = [], []
+    saves = []
 
     if SPLIT_W13:
       h = x + residual
       x_normed, rrms = rmsnorm(h, self.norm_eps)
       saves.extend([x_normed, rrms])
       inp = x_normed * kwargs["ffn_norm"]
-      x_w1, new_amax, *s = matmul(inp, kwargs["w1"], amax_x=kwargs["amax_x1"], w_inv_scale=kwargs["s_1"],
-                                  grad_amax_state=kwargs["grad_amax_xw1"], next_grad_amax_state=kwargs["next_grad_amax_xw1"])
-      amaxs.append(new_amax)
+      x_w1, *s = matmul(inp, kwargs["w1"], amax_x=kwargs["amax_x1"], w_inv_scale=kwargs["s_1"],
+                                  grad_amax_state=kwargs["grad_amax_xw1"], next_grad_amax_state=kwargs["next_grad_amax_xw1"],
+                                  next_amax_x=kwargs["next_amax_x1"])
       saves.extend([*s, x_w1])
-      x_w3, new_amax, *s = matmul(inp, kwargs["w3"], amax_x=kwargs["amax_x3"], w_inv_scale=kwargs["s_3"],
-                                  grad_amax_state=kwargs["grad_amax_xw3"], next_grad_amax_state=kwargs["next_grad_amax_xw3"])
-      amaxs.append(new_amax)
+      x_w3, *s = matmul(inp, kwargs["w3"], amax_x=kwargs["amax_x3"], w_inv_scale=kwargs["s_3"],
+                                  grad_amax_state=kwargs["grad_amax_xw3"], next_grad_amax_state=kwargs["next_grad_amax_xw3"],
+                                  next_amax_x=kwargs["next_amax_x3"])
       saves.extend([*s, x_w3])
       if FUSED_SILU_W13 and MXFP8:
         from extra.llama_kernels.fused_silu_mul_quantize_mxfp8 import fused_silu_mul_quantize_mxfp8
         aq, ae8, asi = fused_silu_mul_quantize_mxfp8(x_w1.reshape(-1, x_w1.shape[-1]), x_w3.reshape(-1, x_w3.shape[-1]))
-        out, new_amax, *s = matmul(None, kwargs["w2"], x_prequant_mx=(aq, ae8, asi), amax_x=kwargs["amax_x2"],
-                                   w_inv_scale=kwargs["s_2"], grad_amax_state=kwargs["grad_amax_xout"],
-                                   next_grad_amax_state=kwargs["next_grad_amax_xout"])
+        out, *s = matmul(None, kwargs["w2"], x_prequant_mx=(aq, ae8, asi), amax_x=kwargs["amax_x2"],
+                         w_inv_scale=kwargs["s_2"], grad_amax_state=kwargs["grad_amax_xout"],
+                         next_grad_amax_state=kwargs["next_grad_amax_xout"], next_amax_x=kwargs["next_amax_x2"])
         out = out.reshape(*x_w1.shape[:-1], kwargs["w2"].shape[0])
       else:
-        out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
-                                   grad_amax_state=kwargs["grad_amax_xout"], next_grad_amax_state=kwargs["next_grad_amax_xout"])
-      amaxs.append(new_amax)
+        out, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
+                         grad_amax_state=kwargs["grad_amax_xout"], next_grad_amax_state=kwargs["next_grad_amax_xout"],
+                         next_amax_x=kwargs["next_amax_x2"])
       saves.extend([*s, out])
     else:
-      x_w13, h, x_normed, rrms, (new_amax, *s) = add_norm_quantize_matmul(x, residual, kwargs["ffn_norm"], kwargs["w13"], kwargs["s_13"],
+      x_w13, h, x_normed, rrms, s = add_norm_quantize_matmul(x, residual, kwargs["ffn_norm"], kwargs["w13"], kwargs["s_13"],
                                                                           self.norm_eps, amax_x=kwargs["amax_x13"],
+                                                                          next_amax_x=kwargs["next_amax_x13"],
                                                                           grad_amax_state=kwargs["grad_amax_xw13"],
                                                                           next_grad_amax_state=kwargs["next_grad_amax_xw13"])
-      amaxs.append(new_amax)
       saves.extend([x_normed, rrms, *s, x_w13])
-      out, (new_amax, *s) = silu_w13_quantize_matmul(x_w13, kwargs["w2"], kwargs["s_2"], amax_x2=kwargs["amax_x2"],
+      out, s = silu_w13_quantize_matmul(x_w13, kwargs["w2"], kwargs["s_2"], amax_x2=kwargs["amax_x2"],
+                                                     next_amax_x2=kwargs["next_amax_x2"],
                                                      grad_amax_xw13=kwargs["grad_amax_xw13"],
                                                      next_grad_amax_xw13=kwargs["next_grad_amax_xw13"],
                                                      grad_amax_xout=kwargs["grad_amax_xout"],
                                                      next_grad_amax_xout=kwargs["next_grad_amax_xout"])
-      amaxs.append(new_amax)
       saves.extend([*s, out])
-    return out, h, amaxs, saves
+    return out, h, saves
 
   @function(precompile=True, precompile_backward=True)
   def run_layer(self, x:Tensor, freqs_cis:Tensor, attn_kwargs:dict, ffn_kwargs:dict, save:bool=True):
-    attn, attn_amaxs, attn_saves = self.attention(x, freqs_cis, **attn_kwargs)
-    ffn, h, ffn_amaxs, ffn_saves = self.feed_forward(x, attn, **ffn_kwargs)
+    attn, attn_saves = self.attention(x, freqs_cis, **attn_kwargs)
+    ffn, h, ffn_saves = self.feed_forward(x, attn, **ffn_kwargs)
     h = h + ffn
-    amaxs = tuple(a.detach() for a in (*attn_amaxs, *ffn_amaxs))
-    if save: return (h, *amaxs, *attn_saves, *ffn_saves)
-    else: return (h, *amaxs)
+    if save: return (h, *attn_saves, *ffn_saves)
+    else: return (h,)
 
   def shard(self, device:tuple[str, ...], mp:bool=False):
     from tinygrad.nn.state import get_parameters
@@ -313,26 +312,27 @@ class FlatTransformer:
 
   def __call__(self, tokens:Tensor, save:bool=True):
     h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
+    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
     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])
+                        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])
       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])
       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])
-      h, *ret = self.run_layer(h, freqs_cis, attn_kwargs, ffn_kwargs, save=save)
-      amax_names = ["xqkv", "xo"] + (["x1", "x3"] if SPLIT_W13 else ["x13"]) + ["x2"]
-      for name, new_val in zip(amax_names, ret[:len(amax_names)]):
-        na[name][i].assign(new_val)
+                          next_grad_amax_xw13=nga["xw13"][i], next_amax_x13=na["x13"][i])
+      h, *_ = self.run_layer(h, freqs_cis, attn_kwargs, ffn_kwargs, save=save)
 
     logits = matmul(self.norm(h), self.output[0], fp8=False)[0]
     return logits
@@ -415,6 +415,9 @@ 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)
       logits = model(tokens[:, :-1], save=llama_size=="8B")
       loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
     with Timing("python backward: "):

examples/mlperf/models/gpt_oss.py

@@ -10,9 +10,9 @@ if __name__ == "__main__":
 from tinygrad import Tensor, nn, function, getenv, dtypes, TinyJit
 from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
 from tinygrad.uop.ops import Ops, UOp
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
+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, quantize_mxfp8
+from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8
 
 FP8_DTYPE = dtypes.fp8e4m3
 FP8_MAX = 448.0
@@ -39,8 +39,11 @@ def quant_dequant_mx(x:Tensor) -> Tensor:
   fxn = _quant_dequant_fwd_fxn(x.as_param(0).uop, x.device)
   return Tensor(UOp.maketuple(fxn.uop).call(x.uop, grad_fxn=_quant_dequant_bwd).gettuple(0))
 
+def _mx_scale(e8:Tensor) -> Tensor:
+  return _mx_block_scale(e8) if e8.ndim == 2 else _mx_block_scale_3d(e8)
+
 def _dequant_fwd(w_q:Tensor, w_scale:Tensor) -> Tensor:
-  return w_q.cast(dtypes.bfloat16) * _mx_block_scale(w_scale)
+  return w_q.cast(dtypes.bfloat16) * _mx_scale(w_scale)
 
 @functools.cache
 def _dequant_fwd_fxn(wq_p, ws_p, device):
@@ -48,7 +51,7 @@ def _dequant_fwd_fxn(wq_p, ws_p, device):
 
 def _dequant_bwd(grad:UOp, call:UOp) -> tuple:
   w_scale = Tensor(call.src[2])
-  return ((Tensor(grad).cast(dtypes.bfloat16) * _mx_block_scale(w_scale).cast(dtypes.bfloat16)).uop, None)
+  return ((Tensor(grad).cast(dtypes.bfloat16) * _mx_scale(w_scale).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)
@@ -67,6 +70,11 @@ def swiglu(x:Tensor, limit:float=7.0, alpha:float=1.702) -> Tensor:
   x_linear = x_linear.clamp(-limit, limit)
   return (x_glu * (alpha * x_glu).sigmoid()) * (x_linear + 1)
 
+def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0) -> Tensor:
+  freqs = 1.0 / (theta ** (Tensor.arange(0, dim, 2, dtype=dtypes.float32)[:(dim // 2)] / dim))
+  freqs = Tensor.arange(end, dtype=dtypes.float32).unsqueeze(dim=1) * freqs.unsqueeze(dim=0)
+  return Tensor.stack(freqs.cos(), freqs.sin(), dim=-1).cast(dtypes.default_float).reshape(1, end, 1, dim//2, 2)
+
 class GPTOSS:
   def __init__(self, dim:int, n_layers:int, n_heads:int, n_kv_heads:int, head_dim:int, n_experts:int, experts_per_tok:int,
                intermediate_size:int, vocab_size:int, norm_eps:float=1e-5, rope_theta:int=150000, sliding_window:int=128,
@@ -109,14 +117,30 @@ class GPTOSS:
     w_q, w_e8, _ = quantize_mxfp8(w)
     return w_q, w_e8.is_param_(False)
 
-  def _attn_mask(self, seqlen:int, sliding:bool, dtype) -> Tensor:
+  def _attn_mask(self, seqlen:int, dtype) -> Tensor:
     i, j = Tensor.arange(seqlen).reshape(seqlen, 1), Tensor.arange(seqlen).reshape(1, seqlen)
-    allowed = j <= i
-    if sliding: allowed = allowed & (i - j < self.sliding_window)
-    return allowed.where(0.0, -1e30).cast(dtype).contiguous()
+    return (j <= i).where(0.0, -1e30).cast(dtype).contiguous()
 
-  def attention(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, *, attention_norm:Tensor, wqkv:Tensor, wqkv_scale:Tensor,
-                wqkv_bias:Tensor, wo:Tensor, wo_scale:Tensor, wo_bias:Tensor, sinks:Tensor):
+  def _sliding_attention(self, xq:Tensor, xk:Tensor, xv:Tensor, sinks:Tensor) -> Tensor:
+    bsz, seqlen, H, hd = xq.shape
+    KV, R, W = self.n_kv_heads, self.n_rep, self.sliding_window
+    assert seqlen % W == 0, f"seqlen {seqlen} must be a multiple of sliding_window {W} for banded attention"
+    nb = seqlen // W
+    q = xq.reshape(bsz, seqlen, KV, R, hd).permute(0, 2, 3, 1, 4).reshape(bsz, KV, R, nb, W, hd).float()
+    k, v = (x.permute(0, 2, 1, 3).reshape(bsz, KV, 1, nb, W, hd).float() for x in (xk, xv))
+    kk, vv = (x.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb].cat(x, dim=-2) for x in (k, v))
+    sc = (q @ kk.transpose(-1, -2)) * self.sm_scale   # (B,KV,R,nb,W,2W)
+    i, j, pv = Tensor.arange(W).reshape(W, 1), Tensor.arange(2 * W).reshape(1, 2 * W), Tensor.arange(nb).reshape(nb, 1, 1) >= 1
+    sc = ((j > i) & (j <= i + W) & (pv | (j >= W))).where(sc, -float("inf"))
+    sink = sinks.reshape(1, KV, R, 1, 1, 1).float()
+    m = sc.max(-1, keepdim=True).maximum(sink)
+    e = (sc - m).exp()
+    p = (e / (e.sum(-1, keepdim=True) + (sink - m).exp())).cast(dtypes.bfloat16)
+    attn = p @ vv.cast(dtypes.bfloat16)
+    return attn.reshape(bsz, KV, R, seqlen, hd).permute(0, 3, 1, 2, 4).reshape(bsz, seqlen, H * hd)
+
+  def attention(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, sliding:bool, *, attention_norm:Tensor, wqkv:Tensor,
+                wqkv_scale:Tensor, wqkv_bias:Tensor, wo:Tensor, wo_scale:Tensor, wo_bias:Tensor, sinks:Tensor):
     bsz, seqlen, _ = x.shape
     x_normed, rrms = rmsnorm(x, self.norm_eps)
     qkv = matmul_mx(x_normed * attention_norm, wqkv, wqkv_scale) + wqkv_bias
@@ -124,16 +148,23 @@ class GPTOSS:
     xq = qkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
     xk, xv = qkv[:, :, :, self.n_rep], qkv[:, :, :, self.n_rep + 1]
     xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
+    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)  # (B,N,H,D)/(B,N,KV,D)
 
-    xq = xq.cast(dtypes.bfloat16).reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
-    xk = xk.cast(dtypes.bfloat16).permute(0, 2, 1, 3).unsqueeze(2)
-    xv = xv.cast(dtypes.bfloat16).permute(0, 2, 1, 3).unsqueeze(2)
-    scores = (xq @ xk.transpose(-2, -1)).float() * self.sm_scale + mask
-    sink = sinks.reshape(1, self.n_kv_heads, self.n_rep, 1, 1).float()
-    m = scores.max(-1, keepdim=True).maximum(sink)
-    e = (scores - m).exp()
-    w = (e / (e.sum(-1, keepdim=True) + (sink - m).exp())).cast(dtypes.bfloat16)
-    attn = (w @ xv).permute(0, 3, 1, 2, 4).reshape(bsz, seqlen, self.n_heads * self.head_dim)
+    if sliding:
+      attn = self._sliding_attention(xq, xk, xv, sinks)
+    elif getenv("HK_FLASH_ATTENTION"):
+      from extra.thunder.amd.fa import flash_attention
+      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks)
+      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
+    else:
+      xqm = xq.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
+      xkm, xvm = xk.permute(0, 2, 1, 3).unsqueeze(2), xv.permute(0, 2, 1, 3).unsqueeze(2)
+      scores = (xqm @ xkm.transpose(-2, -1)).float() * self.sm_scale + mask
+      sink = sinks.reshape(1, self.n_kv_heads, self.n_rep, 1, 1).float()
+      m = scores.max(-1, keepdim=True).maximum(sink)
+      e = (scores - m).exp()
+      w = (e / (e.sum(-1, keepdim=True) + (sink - m).exp())).cast(dtypes.bfloat16)
+      attn = (w @ xvm).permute(0, 3, 1, 2, 4).reshape(bsz, seqlen, self.n_heads * self.head_dim)
 
     out = matmul_mx(attn, wo, wo_scale) + wo_bias
     return out, [x_normed, rrms, attn]
@@ -157,8 +188,8 @@ class GPTOSS:
     return out, [x_normed, rrms]
 
   @function(precompile=True, precompile_backward=True)
-  def run_layer(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, attn_kwargs:dict, ffn_kwargs:dict, save:bool=True):
-    attn, attn_saves = self.attention(x, freqs_cis, mask, **attn_kwargs)
+  def run_layer(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, sliding:bool, attn_kwargs:dict, ffn_kwargs:dict, save:bool=True):
+    attn, attn_saves = self.attention(x, freqs_cis, mask, sliding, **attn_kwargs)
     h = x + attn
     ffn, ffn_saves = self.feed_forward(h, **ffn_kwargs)
     h = h + ffn
@@ -175,8 +206,7 @@ class GPTOSS:
     h = self.tok_embeddings(tokens)
     bsz, seqlen = tokens.shape
     freqs_cis = self.freqs_cis.cast(h.dtype)[:, :seqlen, :, :, :]
-    mask_full = self._attn_mask(seqlen, False, dtypes.float32)
-    mask_sliding = self._attn_mask(seqlen, True, dtypes.float32)
+    mask_full = None if getenv("HK_FLASH_ATTENTION") else self._attn_mask(seqlen, dtypes.float32)
     for i in range(self.n_layers):
       attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wqkv_scale=self.wqkv_scale[i],
                          wqkv_bias=self.wqkv_bias[i], wo=self.wo[i], wo_scale=self.wo_scale[i], wo_bias=self.wo_bias[i],
@@ -184,8 +214,7 @@ class GPTOSS:
       ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], gate=self.gate[i], gate_bias=self.gate_bias[i],
                         w_gate_up=self.w_gate_up[i], w_gate_up_scale=self.w_gate_up_scale[i], w_gate_up_bias=self.w_gate_up_bias[i],
                         w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
-      mask = mask_sliding if i % 2 == 0 else mask_full
-      h, *_ = self.run_layer(h, freqs_cis, mask, attn_kwargs, ffn_kwargs, save=save)
+      h, *_ = self.run_layer(h, freqs_cis, mask_full, i % 2 == 0, attn_kwargs, ffn_kwargs, save=save)
 
     logits = self.norm(h) @ self.output.T
     return logits

examples/mlperf/optim.py

@@ -96,7 +96,7 @@ class GradAccClipAdamW(Optimizer):
     up = up.float().shard_like(w) + self.lr.to(w.device) * wd * w.detach()
     new_w = w.detach() - up
     if master is not None: master.assign(new_w)
-    if self.zero: new_w = self._zero_gather(new_w)
+    if self.zero and not (MXFP8 and t.dtype in dtypes.fp8s): new_w = self._zero_gather(new_w)
     # when master is offloaded to a different device than the param, results are resharded back onto the param's (sharded) device
     offloaded = master is not None and master.device != t.device
     if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16:
@@ -106,6 +106,7 @@ class GradAccClipAdamW(Optimizer):
       if MXFP8:
         from extra.gemm.cdna_asm_gemm import quantize_mxfp8
         w_q, w_e8, _ = quantize_mxfp8(new_w.reshape(-1, new_w.shape[-1]))
+        if self.zero: w_q, w_e8 = self._zero_gather(w_q), self._zero_gather(w_e8)
         new_e8 = w_e8.reshape(t._inv_scale.shape)
         t._inv_scale.assign(new_e8.shard_like(t._inv_scale) if offloaded else new_e8)
         ret = w_q.reshape(new_w.shape)

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -20,7 +20,7 @@ export ZERO_OPTIM=${ZERO_OPTIM:-1}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
 export GBS=$((BS * GRADIENT_ACC_STEPS))
 
 export MODEL="gptoss"
@@ -34,7 +34,7 @@ export SEED=${SEED:-5760}
 export DATA_SEED=${DATA_SEED:-5760}
 
 export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
 
 export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh

@@ -20,7 +20,7 @@ export ZERO_OPTIM=${ZERO_OPTIM:-1}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
 export GBS=$((BS * GRADIENT_ACC_STEPS))
 
 export MODEL="gptoss"
@@ -34,6 +34,6 @@ export SEED=${SEED:-$RANDOM}
 export DATA_SEED=${DATA_SEED:-5760}
 
 export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
 
 python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/profile.sh

@@ -2,5 +2,4 @@
 export BENCHMARK=${BENCHMARK:-5}
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
-SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
-[ "$BENCHMARK" -le 3 ] || python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
+[ "$BENCHMARK" -le 3 ] || [[ $DEV == NULL* ]] || python -m tinygrad.viz.cli -s AMD -t --interval "train @ 2" "train @ 3"

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export MXFP8=${MXFP8:-1}
+export ZERO_OPTIM=${ZERO_OPTIM:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="gptoss"
+export BASEDIR="/raid/datasets/c4-8b/"
+export EVAL_TARGET=3.34 EVAL_FREQ=12288
+export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
+if [ -z "$FULL_LAYERS" ]; then
+  export LAYERS=${LAYERS:-2}
+fi
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export MXFP8=${MXFP8:-1}
+export ZERO_OPTIM=${ZERO_OPTIM:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="gptoss"
+export BASEDIR="/raid/datasets/c4-8b/"
+export EVAL_TARGET=3.34 EVAL_FREQ=12288
+export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export FAST_CE=${FAST_CE:-1}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export SPLIT_W13=${SPLIT_W13:-0}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
+if [ -z "$FULL_LAYERS" ]; then
+  export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
+fi
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam_mp.sh

@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export FAST_CE=${FAST_CE:-0}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-0}
+export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-0}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-0}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-0}
+export SPLIT_W13=${SPLIT_W13:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
+if [ -z "$FULL_LAYERS" ]; then
+  export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
+fi
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export FAST_CE=${FAST_CE:-1}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export SPLIT_W13=${SPLIT_W13:-0}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run_mp.sh

@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=${DEV:-AMD}
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
+export USE_ATOMICS=${USE_ATOMICS:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export FAST_CE=${FAST_CE:-0}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-0}
+export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-0}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-0}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-0}
+export SPLIT_W13=${SPLIT_W13:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-32}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+python3 examples/mlperf/model_train.py

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/profile.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+export BENCHMARK=${BENCHMARK:-5}
+export EVAL_BS=0
+VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
+[ "$BENCHMARK" -le 3 ] || [[ $DEV == NULL* ]] || python -m tinygrad.viz.cli -s AMD -t --interval "train @ 2" "train @ 3"

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/run_and_time.sh

@@ -0,0 +1,58 @@
+#!/usr/bin/env bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="."
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
+export DEV=AMD
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export HK_FLASH_ATTENTION=1
+export ALL2ALL=1
+export LATE_ALLREDUCE=0
+export USE_ATOMICS=1
+export ASM_GEMM=1
+export WQKV=1
+export MASTER_WEIGHTS=1
+export FP8=1
+export ALLREDUCE_CAST=1
+export FAST_CE=1
+export FUSED_INPUT_QUANTIZE=1
+export FUSED_GRAD_QUANTIZE=1
+export FUSED_ADD_NORM_MUL_QUANTIZE=1
+export FUSED_SILU_W13=1
+export SPLIT_W13=0
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=8 MP=1 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=2
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=8B
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=8192
+
+export SEED=$RANDOM
+export DATA_SEED=$SEED
+
+export JITBEAM=3
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export LOGMLPERF=1
+
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="llama31_8b_8xMI350x_${DATETIME}_${SEED}.log"
+
+# beam
+FAKEDATA=1 BENCHMARK=10 INITMLPERF=1 LLAMA_LAYERS=2 python3 examples/mlperf/model_train.py | tee "$LOGFILE"
+
+# run
+RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a "$LOGFILE"

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/startup_walltime.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+export BENCHMARK=5
+export EVAL_BS=0
+export FAKEDATA=1
+export NULL_ALLOW_COPYOUT=1
+export HIP_VISIBLE_DEVICES=""
+export DEV=NULL:HIP:gfx950
+export JITBEAM=0
+export LLAMA_LAYERS=${LLAMA_LAYERS:-"2"}
+time examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh

examples/mlperf/training_submission_v6.1/tinycorp/systems/tinybox_8xMI300X.json

@@ -0,0 +1,38 @@
+{
+    "submitter": "tinycorp",
+    "division": "closed",
+    "status": "Available on-premise",
+    "system_name": "tinybox 8xMI300X",
+    "number_of_nodes": "1",
+    "host_processors_per_node": "2",
+    "host_processor_model_name": "AMD EPYC 9354",
+    "host_processor_core_count": "32",
+    "host_processor_vcpu_count": "64",
+    "host_processor_frequency": "",
+    "host_processor_caches": "",
+    "host_processor_interconnect": "",
+    "host_memory_capacity": "2304GB",
+    "host_storage_type": "NVMe SSD",
+    "host_storage_capacity": "3x 4TB raid array",
+    "host_networking": "",
+    "host_networking_topology": "",
+    "host_memory_configuration": "24x 96GB DDR5",
+    "accelerators_per_node": "8",
+    "accelerator_model_name": "AMD Instinct MI300X 192GB HBM3",
+    "accelerator_host_interconnect": "PCIe 5.0 x16",
+    "accelerator_frequency": "",
+    "accelerator_on-chip_memories": "",
+    "accelerator_memory_configuration": "HBM3",
+    "accelerator_memory_capacity": "192GB",
+    "accelerator_interconnect": "",
+    "accelerator_interconnect_topology": "",
+    "cooling": "air",
+    "hw_notes": "",
+    "framework": "tinygrad, branch mlperf_training_v5.0",
+    "other_software_stack": {
+        "python": "3.10.16",
+        "ROCm": "3.0.0+94441cb"
+    },
+    "operating_system": "Ubuntu 24.04.1 LTS",
+    "sw_notes": ""
+  }

examples/mlperf/training_submission_v6.1/tinycorp/systems/tinybox_8xMI350X.json

@@ -0,0 +1,38 @@
+{
+    "submitter": "tinycorp",
+    "division": "closed",
+    "status": "Available on-premise",
+    "system_name": "tinybox 8xMI350X",
+    "number_of_nodes": "1",
+    "host_processors_per_node": "2",
+    "host_processor_model_name": "AMD EPYC 9575F",
+    "host_processor_core_count": "32",
+    "host_processor_vcpu_count": "64",
+    "host_processor_frequency": "",
+    "host_processor_caches": "",
+    "host_processor_interconnect": "",
+    "host_memory_capacity": "3072 GiB",
+    "host_storage_type": "NVMe SSD",
+    "host_storage_capacity": "4TB",
+    "host_networking": "",
+    "host_networking_topology": "",
+    "host_memory_configuration": "24x 128GB DDR5",
+    "accelerators_per_node": "8",
+    "accelerator_model_name": "AMD Instinct MI350X 288GB HBM3e",
+    "accelerator_host_interconnect": "PCIe 5.0 x16",
+    "accelerator_frequency": "",
+    "accelerator_on-chip_memories": "",
+    "accelerator_memory_configuration": "HBM3",
+    "accelerator_memory_capacity": "288GB",
+    "accelerator_interconnect": "",
+    "accelerator_interconnect_topology": "",
+    "cooling": "air",
+    "hw_notes": "",
+    "framework": "tinygrad, branch mlperf_training_v6.0",
+    "other_software_stack": {
+        "python": "3.12.3",
+        "ROCm": "7.1.1"
+    },
+    "operating_system": "Ubuntu 24.04.3 LTS",
+    "sw_notes": ""
+  }

examples/mlperf/training_submission_v6.1/tinycorp/systems/tinybox_green.json

@@ -0,0 +1,38 @@
+{
+  "submitter": "tinycorp",
+  "division": "closed",
+  "status": "Available on-premise",
+  "system_name": "tinybox green",
+  "number_of_nodes": "1",
+  "host_processors_per_node": "1",
+  "host_processor_model_name": "AMD EPYC 7532",
+  "host_processor_core_count": "32",
+  "host_processor_vcpu_count": "64",
+  "host_processor_frequency": "",
+  "host_processor_caches": "",
+  "host_processor_interconnect": "",
+  "host_memory_capacity": "128GB",
+  "host_storage_type": "NVMe SSD",
+  "host_storage_capacity": "4 TB raid array + 1 TB boot",
+  "host_networking": "",
+  "host_networking_topology": "",
+  "host_memory_configuration": "8x 16GB DDR4",
+  "accelerators_per_node": "6",
+  "accelerator_model_name": "NVIDIA GeForce RTX 4090",
+  "accelerator_host_interconnect": "PCIe 4.0 x16",
+  "accelerator_frequency": "",
+  "accelerator_on-chip_memories": "",
+  "accelerator_memory_configuration": "GDDR6X",
+  "accelerator_memory_capacity": "24GB",
+  "accelerator_interconnect": "",
+  "accelerator_interconnect_topology": "",
+  "cooling": "air",
+  "hw_notes": "",
+  "framework": "tinygrad, branch mlperf_training_v5.0",
+  "other_software_stack": {
+    "python": "3.10.12",
+    "CUDA": "12.4"
+  },
+  "operating_system": "Ubuntu 22.04.4",
+  "sw_notes": ""
+}

examples/mlperf/training_submission_v6.1/tinycorp/systems/tinybox_red.json

@@ -0,0 +1,37 @@
+{
+  "submitter": "tinycorp",
+  "division": "closed",
+  "status": "Available on-premise",
+  "system_name": "tinybox red",
+  "number_of_nodes": "1",
+  "host_processors_per_node": "1",
+  "host_processor_model_name": "AMD EPYC 7532",
+  "host_processor_core_count": "32",
+  "host_processor_vcpu_count": "64",
+  "host_processor_frequency": "",
+  "host_processor_caches": "",
+  "host_processor_interconnect": "",
+  "host_memory_capacity": "128GB",
+  "host_storage_type": "NVMe SSD",
+  "host_storage_capacity": "4 TB raid array + 1 TB boot",
+  "host_networking": "",
+  "host_networking_topology": "",
+  "host_memory_configuration": "8x 16GB DDR4",
+  "accelerators_per_node": "6",
+  "accelerator_model_name": "AMD Radeon RX 7900 XTX",
+  "accelerator_host_interconnect": "PCIe 4.0 x16",
+  "accelerator_frequency": "",
+  "accelerator_on-chip_memories": "",
+  "accelerator_memory_configuration": "GDDR6",
+  "accelerator_memory_capacity": "24GB",
+  "accelerator_interconnect": "",
+  "accelerator_interconnect_topology": "",
+  "cooling": "air",
+  "hw_notes": "",
+  "framework": "tinygrad, branch mlperf_training_v5.0",
+  "other_software_stack": {
+    "python": "3.10.12"
+  },
+  "operating_system": "Ubuntu 22.04.4",
+  "sw_notes": ""
+}

extra/f16_decompress.py

@@ -5,10 +5,10 @@ def bit_extract(x: Tensor, e: int, s: int) -> Tensor:
   return (x >> s) & mask
 
 def u16_to_f16(x: Tensor) -> Tensor:
-  sign = bit_extract(x, 15, 15).float()
+  sign = bit_extract(x, 15, 15).bool()
   exponent = bit_extract(x, 14, 10).float()
   fraction = bit_extract(x, 9, 0).float()
-  return sign.where(-1, 1) * exponent.where((exponent - 15.0).exp2() * (1 + fraction / 1024.0), 6.103515625e-5 * (fraction / 1024.0))
+  return sign.where(-1, 1) * exponent.bool().where((exponent - 15.0).exp2() * (1 + fraction / 1024.0), 6.103515625e-5 * (fraction / 1024.0))
 
 def u32_to_f16(oo: Tensor) -> Tensor:
   f1 = u16_to_f16(oo>>16)

extra/gemm/amd_copy_matmul.py

@@ -80,7 +80,7 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
       # NOTE: since this is part of K, these 2 can be anywhere in the frags and long as a and b match
       a_frag = a_frag.reshape(2, 8)[lane_m, :]
       b_frag = b_frag.reshape(2, 8)[lane_m, :]
-    wmma = UOp.wmma(a_frag, b_frag, acc_frag.after(k), ((16, 16, 16), 'AMD', 32))
+    wmma = UOp.wmma(a_frag, b_frag, acc_frag.after(k), (16, 16, 16), 'AMD', 32)
     acc_store = acc_frag.store(wmma).end(tile_m, tile_n)
   else:
     # registers for LOCAL -> REG

extra/gemm/amd_flash_attention.py

@@ -13,7 +13,7 @@ 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)
+WMMA_ARG = (WMMA_M, WMMA_N, WMMA_K), 'AMD', 32
 LOG2E = math.log2(math.e)
 
 def warp_shfl_xor(val, offset, lane):
@@ -97,7 +97,7 @@ def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
   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 = 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)
 
@@ -158,7 +158,7 @@ def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
   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)
+  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).barrier().end(n_tile)

extra/gemm/cdna_asm_gemm.py

@@ -128,6 +128,11 @@ def _mx_block_scale(e8:Tensor) -> Tensor:
   rows, scale_K = e8.shape
   return (e8.cast(dtypes.float32) - 127.0).exp2().reshape(rows, scale_K, 1).expand(rows, scale_K, 32).reshape(rows, scale_K*32)
 
+def _mx_block_scale_3d(e8:Tensor) -> Tensor:
+  # batched (E, rows, scale_K) dequant scale 2^(e8-127) broadcast to (E, rows, scale_K*32)
+  E, rows, scale_K = e8.shape
+  return (e8.cast(dtypes.float32) - 127.0).exp2().reshape(E, rows, scale_K, 1).expand(E, rows, scale_K, 32).reshape(E, rows, scale_K*32)
+
 counters = {"used":0, "todos":[]}
 def todo(msg:str) -> bool: counters["todos"].append(msg); return False
 def _asm_gemm_report():
@@ -169,10 +174,10 @@ def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
   m = UOp.range(M, 1, AxisType.LOOP)
   n = UOp.range(N, 2, AxisType.LOOP)
   k = UOp.range(K, 0, AxisType.REDUCE)
-  mul = (A.flatten().index((m*UOp.const(dtypes.index, K)+k))*
-         B.flatten().index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
+  mul = (A.flatten().index((m*UOp.const(dtypes.weakint, K)+k))*
+         B.flatten().index((k*UOp.const(dtypes.weakint, 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.index, N)+n)).store(red).end(m, n)
+  store = C.flatten().index((m*UOp.const(dtypes.weakint, 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
@@ -275,10 +280,7 @@ def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=
       elif getenv("FUSED_GRAD_QUANTIZE", 0):
         grad_amax_t = Tensor(grad_amax_state, device=a.device)
         g_amax = grad_amax_t
-        g_fp8, _, new_grad_amax, _ = quantize_fp8_delayed(g_t, g_amax)
-        store_effect = next_grad_amax_state.store(new_grad_amax.uop)
-        assert g_fp8.uop.op is Ops.AFTER, f"expected AFTER, got {g_fp8.uop.op}"
-        g_fp8 = Tensor(g_fp8.uop.replace(src=g_fp8.uop.src + (store_effect,)), device=a.device)
+        g_fp8, _ = quantize_fp8_delayed(g_t, g_amax, Tensor(next_grad_amax_state, device=a.device))
       else:
         grad_amax_t = Tensor(grad_amax_state, device=a.device)
         g_amax = grad_amax_t

extra/gemm/metal_uop_matmul.py

@@ -1,5 +1,5 @@
 from tinygrad import UOp, dtypes
-from tinygrad.uop.ops import AxisType, Ops, KernelInfo, AddrSpace
+from tinygrad.uop.ops import AxisType, KernelInfo, AddrSpace
 from extra.gemm.amd_uop_matmul import test_matmul
 
 N = 2048
@@ -27,11 +27,8 @@ def hand_spec_tc_cores():
   acc = acc[0].set(0.0)
   acc = acc[1].set(0.0)
 
-  # TODO: make this simple
-  wmma_arg = ('WMMA_8_8_8_float_float', (8, 8, 8), dtypes.float, dtypes.float, 'METAL', 32, (((3, 2),), ((3, 2),), ((3, 2),)), ())
-
   acc_load = UOp.stack(acc.after(gk)[0], acc.after(gk)[1])
-  out = UOp(Ops.WMMA, dtypes.float, (a_tc, b_tc, acc_load), arg=wmma_arg)
+  out = UOp.wmma(a_tc, b_tc, acc_load, (8, 8, 8), 'METAL', 32)
 
   end_loop = UOp.group(*[acc[i].store(out.index(i)) for i in range(2)]).end(gk)
 

extra/gemm/mi350x_uop_matmul.py

@@ -6,7 +6,7 @@ os.environ["AMD_LLVM"] = "0"
 from tinygrad import Tensor, Context, dtypes, UOp, GlobalCounters
 from tinygrad.helpers import DEBUG, getenv
 from tinygrad.dtype import AddrSpace
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
+from tinygrad.uop.ops import AxisType, KernelInfo
 
 WARP_SIZE = 64
 
@@ -137,8 +137,7 @@ def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
     acc_load = acc_after[N_inner_loop, M_inner_loop]
 
     # do WMMA
-    wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
-    out = UOp(Ops.WMMA, dtypes.float, (Ar[M_inner_loop], Br[N_inner_loop], acc_load), arg=wmma_arg)
+    out = UOp.wmma(Ar[M_inner_loop], Br[N_inner_loop], acc_load, (16, 16, 32), 'AMD', 64)
 
     # store back the acc
     acc_store = acc[N_inner_loop, M_inner_loop].store(out)
@@ -193,8 +192,7 @@ acc = acc[init_l:=UOp.range(4, 1)].set(0.0, end=init_l)
 
 # do the wmma
 acc_load = UOp.stack(*[acc.after(K_loop)[i] for i in range(4)])
-wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
-out = UOp(Ops.WMMA, dtypes.float, (A_in, B_in, acc_load), arg=wmma_arg)
+out = UOp.wmma(A_in, B_in, acc_load, (16, 16, 32), 'AMD', 64)
 
 # store back the acc
 acc = acc.after(UOp.group(*[acc[i].store(out.index(i)) for i in range(4)]).end(K_loop))

extra/gemm/mi350x_uop_matmul_2.py

@@ -6,7 +6,7 @@ os.environ["AMD_LLVM"] = "0"
 from tinygrad import Tensor, Context, dtypes, UOp, GlobalCounters
 from tinygrad.helpers import DEBUG, getenv
 from tinygrad.dtype import AddrSpace
-from tinygrad.uop.ops import sint, AxisType, KernelInfo, Ops
+from tinygrad.uop.ops import AxisType, KernelInfo
 
 WARP_SIZE = 64
 
@@ -60,8 +60,7 @@ def compute_on_locals(acc:UOp, Asl:UOp, Bsl:UOp, rng:int, afters:tuple[UOp, ...]
   acc_load = acc_after[N_inner_loop, M_inner_loop]
 
   # do WMMA
-  wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
-  out = UOp(Ops.WMMA, dtypes.float, (Ar[M_inner_loop], Br[N_inner_loop], acc_load), arg=wmma_arg)
+  out = UOp.wmma(Ar[M_inner_loop], Br[N_inner_loop], acc_load, (16, 16, 32), 'AMD', 64)
 
   # store back the acc
   acc_store = acc[N_inner_loop, M_inner_loop].store(out)

extra/hcq2/hcq2.py

@@ -2,8 +2,8 @@ from __future__ import annotations
 from typing import cast, Callable, TypeVar, Generic, Any
 import struct, functools, time, collections, itertools
 from dataclasses import replace, dataclass
-from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, dedup, pluralize
-from tinygrad.helpers import to_tuple, round_up, partition, data64_le, panic
+from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, dedup, pluralize, JIT_BATCH_SIZE, unwrap
+from tinygrad.helpers import to_tuple, round_up, partition, data64_le, panic, ContextVar
 from tinygrad.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer
 from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, track_rewrites, GroupOp
 from tinygrad.uop.symbolic import symbolic
@@ -19,6 +19,8 @@ HCQDeviceType = TypeVar('HCQDeviceType', bound='HCQ2Compiled')
 # *****************
 # 0. helpers
 
+HCQ_RUNTIME_DEV = ContextVar("HCQ_RUNTIME_DEV", "CPU")
+
 HCQ_DEVS = frozenset(("AMD",))
 HCQ_P2P_DEVS = HCQ_DEVS | frozenset(("CPU",))
 HCQ_CACHE_TAGS = frozenset(("program", "systems", "template"))
@@ -47,7 +49,7 @@ def make_getaddr(u, device=None):
   return UOp(Ops.GETADDR, dtypes.uint64, src=(u,), arg=device or to_tuple(u.device)[0])
 
 def make_ins(op, *srcs):
-  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)
+  return UOp(Ops.INS, arg=op, src=tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs))
 
 def make_placeholder(devs, size:int, dtype, name=None, unique=True) -> UOp:
   return UOp.param(next(UOp.unique_num) if unique else 0, dtype, shape=(size,), device=devs).rtag(name or "temp")
@@ -56,7 +58,7 @@ def make_patch(buf:UOp, off:sint, val:UOp, dtype=None) -> UOp:
   return buf.index(UOp.const(dtypes.int, off // buf.dtype.itemsize)).store(val.simplify().cast(dtype or buf.dtype))
 
 def make_binary_patch(buf:UOp, blob:bytes) -> UOp:
-  data = UOp(Ops.BITCAST, buf.dtype, (UOp(Ops.BINARY, dtypes.uint8, src=(), arg=blob),))
+  data = UOp(Ops.BINARY, src=(), arg=blob).bitcast(buf.dtype)
   r = UOp.range(len(blob) // buf.dtype.itemsize, 0, dtype=dtypes.int, src=(buf, data))
   return buf.index(r).store(data.index(r).load()).end(r)
 
@@ -100,159 +102,139 @@ def stage_copy(dst:UOp, src:UOp) -> UOp|None:
 pm_insert_copy_staging = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src"))), stage_copy)])
 
 # *****************
-# 2.1. tag hcq calls
-
-def tag_hcq_call(ctx:itertools.count, call:UOp) -> UOp:
-  if (hcq_devs:=next((b.device for b in call.src[1:] if all_devices_in(b.device, HCQ_DEVS)), None)) is None: return call
-
-  queue = "COMPUTE:0" if call.src[0].op is Ops.PROGRAM else "COPY:0"
-  info = HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), to_tuple(hcq_devs), queue)
-  return call.replace(arg=replace(call.arg, aux=info)).rtag(next(ctx))
-pm_tag_hcq_calls = PatternMatcher([(UPat(Ops.LINEAR, name="l"), lambda ctx, l: l.replace(src=tuple(tag_hcq_call(ctx, s) for s in l.src)))])
-
-# *****************
-# 2.2. deps tracking
-# device.timeline_signal/value are the per-device schedule epoch. Before a schedule queue accesses memory owned by device N for the first time,
-# it waits for device[N].timeline_signal >= device[N].timeline_value - 1. This orders the schedule after all prior schedules that touched device N.
-#
-# queue.timeline_signal/value are per-queue progress counters used only inside a schedule.
-# Only the owner queue signals its queue.timeline_signal. Values are monotonic.
-#
-# At schedule end, one finalizer queue per touched device[N] waits for every active queue on device[N] to reach its schedule-local
-# final queue.timeline value, then signals device[N].timeline_signal with the schedule's reserved device epoch. After that, buffers/transients
-# for device N from this schedule are safe for the next schedule
-#
-# C programs reserve and bump timeline values, then patch command buffers with the concrete wait/signal values.
+# 2. deps
 
 class HCQDepsTracker(DepsTracker):
   @staticmethod
   def _key(buf:Any) -> tuple[Any, int, int]:
     return (buf.arg.slot, 0, buf.max_numel() * buf.dtype.itemsize) if isinstance(buf, UOp) else DepsTracker._key(buf)
 
-def make_deps(u:UOp, dep_lanes:list[tuple[UOp, int, int]], nlanes:int) -> UOp:
-  deps:dict[UOp, list[int|None]] = collections.defaultdict(lambda: [None]*nlanes)
-  for dep, dlane, lane in dep_lanes: deps[dep][lane] = dlane
-  return u.after(*deps, arg=tuple(tuple(v) for v in deps.values()))
-
-def sched_sync(ctx:DepsTracker, call:UOp) -> UOp|None:
-  if not isinstance(call.arg.aux, HCQInfo): return None
-
+def _get_call_bufs_by_lane(call:UOp, devices:tuple[str, ...]) -> list[list[Any]]:
   refs = get_call_arg_uops(call)
-  outs, _ = get_call_outs_ins(call)
-  devices, queue = call.arg.aux.device, call.arg.aux.queue
+  return [[b if b.op is Ops.PARAM else mb.bufs[lane] if isinstance(mb:=b.buffer, MultiBuffer) else mb for b in refs] for lane in range(len(devices))]
 
-  dep_lanes:list[tuple[UOp, int, int]] = []
-  for lane, d in enumerate(devices):
-    lane_refs = [b if b.op is Ops.PARAM else mb.bufs[lane] if isinstance(mb:=b.buffer, MultiBuffer) else mb for b in refs]
-    for dep, dlane in ctx.access_resources(lane_refs, outs, (call, lane)): dep_lanes.append((dep, dlane, lane))
+def _get_deps(ctx:DepsTracker, bufs_by_lane:list[list[Any]], write, key:tuple[tuple[str, ...], str, int]) -> list[tuple[tuple, int, int]]:
+  dep_lanes:list[tuple[tuple, int, int]] = []
+  for lane, bufs in enumerate(bufs_by_lane):
+    dep_lanes += [(dep, dlane, lane) for dep, dlane in ctx.access_resources(bufs, write if write is not None else range(len(bufs)), (key, lane))]
+  return dep_lanes
 
+def _build_wait_cmds(dep_lanes:list[tuple[tuple, int, int]], devices:tuple[str, ...], queue:str) -> tuple[list[UOp], set[int]]:
+  # opt1: same-queue ops are fifo-ordered
   if devices[0].split(":")[0] in {"AMD", "QCOM"} or queue.startswith("COPY"):
-    dep_lanes = [(dep, dlane, lane) for dep, dlane, lane in dep_lanes if (dep.arg.aux.device[dlane], dep.arg.aux.queue) != (devices[lane], queue)]
+    dep_lanes = [(dep, dlane, lane) for dep, dlane, lane in dep_lanes if (dep[0][dlane], dep[1]) != (devices[lane], queue)]
 
-  # keep latest dep per (dep device, queue, cur lane)
-  latest = {((dep.arg.aux.device[dlane], dep.arg.aux.queue), lane): (dep, dlane) for dep, dlane, lane in sorted(dep_lanes, key=lambda x: x[0].tag)}
-  return make_deps(call, [(dep, dlane, lane) for (_, lane), (dep, dlane) in latest.items()], len(devices))
-pm_sched_sync = PatternMatcher([(UPat(Ops.CALL, name="call"), sched_sync)])
+  # opt2: keep latest dep per (dep device, queue, cur lane)
+  latest = {((dep[0][dlane], dep[1]), lane): (dep, dlane) for dep, dlane, lane in sorted(dep_lanes, key=lambda x: x[0][2])}
+  deps:dict[tuple, list[int|None]] = collections.defaultdict(lambda: [None]*len(devices))
+  for (_, lane), (dep, dlane) in latest.items(): deps[dep][lane] = dlane
+
+  waits = []
+  for (ddevs, dqueue, dtag), lanes in deps.items():
+    sig = make_mstack([make_signal(d if dl is None else ddevs[dl], queue=dqueue, sentinel=dl is None) for dl, d in zip(lanes, devices)])
+    val = make_mstack([make_signal_value(d if dl is None else ddevs[dl], queue=dqueue) for dl, d in zip(lanes, devices)])
+    waits.append(UOp(Ops.INS, arg="wait", src=(sig, val.index(UOp.const(dtypes.int, 0)) + dtag)))
+  return waits, {dtag for _, _, dtag in deps}
+
+def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[tuple[tuple[str, ...], str]],
+                      tracker:HCQDepsTracker) -> tuple[list[UOp], set[int]]:
+  # collect all buffers which belong to devices
+  dev_bufs:dict[str, dict[int, Any]] = collections.defaultdict(dict)
+  for call, devices in batch:
+    for b in itertools.chain.from_iterable(_get_call_bufs_by_lane(call, devices)):
+      for bd in to_tuple(b.device): dev_bufs[bd][id(b)] = b
+
+  zero, n, finalizers, waited = UOp.const(dtypes.int, 0), len(batch_info), [], set()
+  for _, devgroup in itertools.groupby(sorted(dedup([d for devs, _ in batch_info for d in devs])), key=lambda d: d.split(":")[0]):
+    devs = tuple(devgroup)
+
+    # to finalize the batch, sync all accesses from other devices to buffers that belong to this device
+    fin_deps = [dl for dl in _get_deps(tracker, [list(dev_bufs[d].values()) for d in devs], None, key=(devs, "COMPUTE:0", n)) if dl[0][2] < n]
+    waits, cur_waited = _build_wait_cmds(fin_deps, devs, "COMPUTE:0")
+    waited |= cur_waited
+
+    # wait the syncs, store the device epoch; value bumps are a separate call: no lane may bump until every lane has patched its waits
+    store = UOp(Ops.INS, arg="store", src=(make_signal(devs), (tl:=make_signal_value(devs)).index(zero)))
+    submit = make_submit(*waits, store, devs=devs, queue="COMPUTE:0")
+    upd = [(tl, 1)] + [(make_signal_value(devs, queue=qn), n) for qn in dedup([qn for bdevs, qn in batch_info if set(bdevs) & set(devs)])]
+    bump = UOp.barrier(*[s.index(zero, dtype=s.dtype).store(s.index(zero) + inc) for s, inc in upd])
+    finalizers += [UOp.custom_function("hcq", b.sink()).call(aux=HCQInfo("hcq_finalizer", Estimates(), devs, "COMPUTE:0")) for b in (submit, bump)]
+  return finalizers, waited
+
+def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]]) -> list[UOp]:
+  batch_info = [(devices, "COMPUTE:0" if call.src[0].op is Ops.PROGRAM else "COPY:0") for call, devices in batch]
+
+  # schedule deps
+  waited:set[int] = set()
+  deps_tracker = HCQDepsTracker()
+  call_waits:list[list[UOp]] = []
+  for tag, ((call, _), (devices, queue)) in enumerate(zip(batch, batch_info)):
+    deps = _get_deps(deps_tracker, _get_call_bufs_by_lane(call, devices), get_call_outs_ins(call)[0], key=(devices, queue, tag))
+    cmds, cur_waited = _build_wait_cmds(deps, devices, queue)
+    call_waits.append(cmds)
+    waited |= cur_waited
+
+  # build finalizers
+  finalizers, finalizer_waited = _build_finalizers(batch, batch_info, deps_tracker)
+  waited |= finalizer_waited
+
+  src = []
+  for tag, ((call, _), (devices, queue), cmds) in enumerate(zip(batch, batch_info, call_waits)):
+    # first queue use, sync prior device work with main signal
+    if batch_info.index((devices, queue)) == tag:
+      epoch = UOp(Ops.INS, arg="wait", src=(make_signal(devices), make_signal_value(devices).index(0) - 1))
+      cmds = [UOp(Ops.INS, arg="barrier", src=()), epoch] + cmds
+
+    # and make hcq call
+    info = HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), devices, queue)
+    cmds = [*cmds, call.replace(arg=replace(call.arg, aux=info))]
+
+    # signal queue timeline if someone waits for us
+    if tag in waited: cmds += [UOp(Ops.INS, arg="store", src=(make_signal(devices, queue), make_signal_value(devices, queue).index(0) + tag))]
+    src.append(UOp.custom_function("hcq", make_submit(*cmds, devs=devices, queue=queue).sink()).call(name="hcq", aux=info))
+  return src + finalizers
+
+def sched_hcq_batches(l:UOp) -> UOp:
+  srcs:list[UOp] = []
+  batch:list[tuple[UOp, tuple[str, ...]]] = []
+  for call in l.src:
+    if (devs:=next((b.device for b in call.src[1:] if all_devices_in(b.device, HCQ_DEVS)), None)) is not None: batch.append((call, to_tuple(devs)))
+    else: srcs, batch = srcs + _finalize_batch(batch) + [call], []
+  return l.replace(src=tuple(srcs + _finalize_batch(batch)))
+pm_sched_hcq_batches = PatternMatcher([(UPat(Ops.LINEAR, name="l"), sched_hcq_batches)])
 
 # *****************
-# 2.3. merge into queues
+# 3. merge into queues
 
-def _merged_hcq_call(calls:list[UOp]):
-  info = replace(unwrap_after(calls[0]).arg.aux, estimates=sum((unwrap_after(c).arg.aux.estimates for c in calls), start=Estimates()))
-  cmdbuf = make_submit(*calls, devs=info.device, queue=info.queue)
-  return UOp.custom_function("hcq", cmdbuf.sink()).call(name="hcq", aux=info)
+def _merged_hcq_call(calls:list[UOp]) -> UOp: # TODO: simplify?
+  if len(calls) == 1: return calls[0]
+  info = replace(calls[0].arg.aux, name=f"submit {calls[0].arg.aux.queue} ({len(calls)})",
+                 estimates=sum((c.arg.aux.estimates for c in calls), start=Estimates()))
+  cmds = [cmd for c in calls for cmd in get_submit(c).src[0].src]
+  return UOp.custom_function("hcq", make_submit(*cmds, devs=info.device, queue=info.queue).sink()).call(name="hcq", aux=info)
 
 def merge_queues(linear:UOp) -> UOp:
   new_src:list[UOp] = []
-  opened_qs:dict[tuple[tuple[str, ...], str], list[UOp]] = {} # (devs, queue) -> list of calls, kept in submit order
+  opened_qs:dict[tuple[tuple[str, ...], str], list[UOp]] = {} # (devs, queue) -> list of hcq calls, kept in submit order
+  limits = collections.defaultdict(lambda: JIT_BATCH_SIZE.value)
 
   for call in linear.src:
-    if not isinstance(unwrap_after(call).arg.aux, HCQInfo):
+    if not isinstance(info:=call.arg.aux, HCQInfo) or info.name == "hcq_finalizer": # non-hcq call or finalizer: close all open queues
       new_src += [_merged_hcq_call(opened_qs.pop(k)) for k in list(opened_qs)] + [call]
       continue
 
-    devices, queue = unwrap_after(call).arg.aux.device, unwrap_after(call).arg.aux.queue
-
-    if (old:=opened_qs.pop((devices, queue), None)) is not None: new_rec = old + [call]
+    if (old:=opened_qs.pop(key:=(info.device, info.queue), None)) is not None:
+      if limits[key] and len(old) >= limits[key]: new_src, old, limits[key] = new_src + [_merged_hcq_call(old)], [], limits[key] * 2
+      new_rec = old + [call]
     else:
       # no such queue opened: close every open submit on this queue that shares a device, so submit order is kept
-      closing = [k for k in opened_qs if k[1] == queue and set(k[0]) & set(devices)]
+      closing = [k for k in opened_qs if k[1] == info.queue and set(k[0]) & set(info.device)]
       new_src += [_merged_hcq_call(opened_qs.pop(k)) for k in closing]
       new_rec = [call]
-    opened_qs[(devices, queue)] = new_rec
+    opened_qs[(info.device, info.queue)] = new_rec
   return linear.replace(src=tuple(new_src + [_merged_hcq_call(c) for c in opened_qs.values()]))
 pm_merge_queues = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), merge_queues)])
 
-# *****************
-# 2.4. finalizer
-
-def add_finalizer(ctx:itertools.count, linear:UOp) -> UOp:
-  # collect by device type
-  parts:dict[str, list[UOp]] = collections.defaultdict(list)
-  for call in linear.src:
-    if (c:=unwrap_after(call)).src[0].op is not Ops.CUSTOM_FUNCTION or c.src[0].arg != "hcq": continue
-    parts[c.arg.aux.device[0].split(':')[0]].append(unwrap_after(get_submit(call).src[0].src[0]))
-
-  nbump = next(ctx)
-  finalizers = []
-  for calls in parts.values():
-    devs = tuple(dedup(d for call in calls for d in unwrap_after(call).arg.aux.device))
-    zero = UOp.const(dtypes.int, 0)
-    tl = make_signal_value(devs)
-
-    # split each (multi-device) call into per-device deps, then store the device timeline value into the device signal after them
-    dep_lanes = [(call, dlane, devs.index(d)) for call in calls for dlane, d in enumerate(unwrap_after(call).arg.aux.device)]
-    store = make_deps(make_signal(devs).store(tl.index(zero)), dep_lanes, len(devs))
-    submit = make_submit(store, devs=devs, queue="COMPUTE:0")
-
-    upd = [(tl, 1)] + [(make_signal_value(devs, queue=qn), nbump) for qn in dedup([unwrap_after(call).arg.aux.queue for call in calls])]
-    patches = [s.after(submit).index(zero, dtype=s.dtype).store(s.index(zero) + inc) for s, inc in upd]
-    finalizers.append(UOp.custom_function("hcq", UOp.barrier(*patches).sink()).call(aux=HCQInfo("hcq finalizer", Estimates(), devs, "COMPUTE:0")))
-  return linear.replace(src=linear.src + tuple(finalizers))
-pm_add_finalizer = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), add_finalizer)])
-
-# *****************
-# 2.5. global sync
-
-def add_global_sync(ctx:set[tuple[str, ...]], submit:UOp, q:UOp) -> UOp|None:
-  if (devs:=q.arg[0]) in ctx: return None
-  ctx.add(devs)
-
-  # some devices from a command buffer might be used for the first time this schedule, so we wait for their global timeline epoch.
-  wait = make_signal(devs).wait(make_signal_value(devs).index(UOp.const(dtypes.int, 0)) - 1)
-  return submit.replace(src=(q.replace(src=(UOp(Ops.BARRIER, dtypes.void), wait, *q.src)),))
-pm_add_global_sync = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_global_sync)])
-
-# *****************
-# 3.1. lower loads/stores
-
-def add_loads(ctx:set[int], submit:UOp, q:UOp) -> UOp|None:
-  cur_devs = q.arg[0]
-  new_src:list[UOp] = []
-  for s in q.src:
-    if s.op is Ops.AFTER:
-      for lanes, dep in zip(s.arg, s.src[1:]):
-        devs, queue = dep.arg.aux.device, dep.arg.aux.queue
-        ctx.add(dep.tag) # mark op to update signal.
-
-        sig = make_mstack([make_signal(d if dl is None else devs[dl], queue=queue, sentinel=dl is None) for dl, d in zip(lanes, cur_devs)])
-        val = make_mstack([make_signal_value(d if dl is None else devs[dl], queue=queue) for dl, d in zip(lanes, cur_devs)]).index(UOp.const(dtypes.int, 0))
-        new_src.append(sig.wait(val + dep.tag))
-      s = s.src[0]
-    new_src.append(s)
-  return submit.replace(src=(q.replace(src=tuple(new_src)),))
-pm_add_inner_loads = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_loads)])
-
-def add_stores(ctx:set[int], submit:UOp, q:UOp) -> UOp|None:
-  devs, queue = q.arg
-  new_src:list[UOp] = []
-  for op in q.src:
-    new_src.append(op)
-    if (sigval:=unwrap_after(op).tag) in ctx:
-      new_src.append(make_signal(devs, queue=queue).store(make_signal_value(devs, queue=queue).index(UOp.const(dtypes.int, 0)) + sigval))
-  return submit.replace(src=(q.replace(src=tuple(new_src)),))
-pm_add_inner_stores = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_stores)])
-
 # *****************
 # 4.1. hcq lowering: programs
 
@@ -277,7 +259,7 @@ def is_value_known_at_link(val:UOp) -> bool:
   addressed_bufs = [b for g in val.toposort() if g.op is Ops.GETADDR for b in unwrap_mstack(g.buf_uop)]
 
   # addr of input params is not known at link time
-  return not runtime_reads and all(b.op is not Ops.PARAM or b.tag is not None for b in addressed_bufs)
+  return not val.variables() and not runtime_reads and all(b.op is not Ops.PARAM or b.tag is not None for b in addressed_bufs)
 
 def is_link_patch(p:UOp, jit:bool) -> bool:
   store = p.src[0] if (is_binary_patch:=p.op is Ops.END) else p
@@ -303,35 +285,32 @@ pm_split_patches = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION
 
 # *****************
 
-def make_addr_table(call:UOp, gaddrs:list[UOp], name:str):
+def make_addr_table(call:UOp, gaddrs:list[UOp], name:str) -> tuple[dict[UOp, UOp], tuple[UOp, ...]]:
   bare = {g: g.replace(src=(unwrap_after(g.src[0]),)) for g in gaddrs}
 
-  order = sorted(dedup(bare.values()), key=lambda g: ((b:=unwrap_mstack(g.buf_uop)[0]).arg.slot, to_tuple(b.tag)))
+  order = sorted(dedup(bare.values()), key=lambda g: ((b:=unwrap_mstack(g.buf_uop)[0]).arg.slot, repr(b.tag)))
   slots, table = {g:i for i,g in enumerate(order)}, make_placeholder(call.arg.aux.device, len(order), dtypes.uint64, name)
 
   reads = {g: table.after(*g.src[0].src[1:] if g.src[0].op is Ops.AFTER else ()).index(UOp.const(dtypes.int, slots[bare[g]])).load() for g in gaddrs}
   return reads, (table.after(*[make_patch(table, i * table.dtype.itemsize, addr) for addr, i in slots.items()]),) if slots else ()
 
-def rm_rt_getaddrs(call:UOp) -> UOp|None:
-  if not (gaddrs:=[u for u in call.src[0].toposort() if u.op is Ops.GETADDR]): return None
+def make_blob_bufs(call:UOp, blobs:list[UOp]) -> tuple[dict[UOp, UOp], tuple[UOp, ...]]:
+  bufs = {b: make_placeholder(call.arg.aux.device, b.max_numel(), b.dtype, "template") for b in blobs}
+  return bufs, tuple(buf.after(make_binary_patch(buf, b.src[0].arg)) for b,buf in bufs.items())
+
+def rm_rt_uops(call:UOp) -> UOp|None:
+  if not (rt_uops:=[u for u in call.src[0].toposort() if u.op is Ops.GETADDR or (u.op is Ops.BITCAST and u.src[0].op is Ops.BINARY)]): return None
+  gaddrs, blobs = partition(rt_uops, lambda u: u.op is Ops.GETADDR)
   inputs, internals = partition(gaddrs, lambda g: all(x.op is Ops.PARAM and x.tag is None for x in unwrap_mstack(g.buf_uop)))
   runtimes, systems = partition(internals, lambda g: any(x.tag in {"program", "kernargs", "cmdbuf"} for x in unwrap_mstack(g.buf_uop)))
 
   # exec fills the inputs table with the input addresses every run, so it has no fill patches
-  (input_reads, _), (rt_reads, rt_fills), (sys_reads, sys_fills) = (make_addr_table(call, gs, name) for gs, name in
-    ((inputs, "inputs"), (runtimes, "runtime"), (systems, "systems")))
-  return call.replace(src=(call.src[0].substitute(input_reads | rt_reads | sys_reads), *call.src[1:], *rt_fills, *sys_fills),
+  (reads, _), *tables = [make_addr_table(call, gs, n) for gs,n in ((inputs, "inputs"), (runtimes, "runtime"), (systems, "systems"))] + \
+                        [make_blob_bufs(call, blobs)]
+  reads, fills = reads | {k:v for r,_ in tables for k,v in r.items()}, [f for _,fs in tables for f in fs]
+  return call.replace(src=(call.src[0].substitute(reads), *call.src[1:], *fills),
                       arg=replace(call.arg, aux=replace(call.arg.aux, input_idxs=tuple(sorted(dedup(g.buf_uop.arg.slot for g in inputs))))))
-pm_rm_rt_getaddrs = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), rm_rt_getaddrs)])
-
-# *****************
-
-def rm_rt_binaries(call:UOp) -> UOp|None:
-  if not (blobs:=[u for u in call.src[0].toposort() if u.op is Ops.BITCAST and u.src[0].op is Ops.BINARY]): return None
-  blob_bufs = {blob: make_placeholder(call.arg.aux.device, blob.max_numel(), blob.dtype, "template") for blob in blobs}
-  fills = [buf.after(make_binary_patch(buf, blob.src[0].arg)) for blob, buf in blob_bufs.items()]
-  return call.replace(src=(call.src[0].substitute(blob_bufs), *call.src[1:], *fills))
-pm_rm_rt_binaries = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), rm_rt_binaries)])
+pm_rm_rt_uops = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), rm_rt_uops)])
 
 # *****************
 
@@ -375,7 +354,7 @@ def pack_hcq_placeholders(call:UOp) -> UOp|None:
       sizes[b.tag] = offs[b] + b.max_numel()
   counts = collections.Counter(b.tag for b in bufs)
   bases = {b.tag:make_placeholder(b.device, sizes[b.tag], b.dtype, b.tag) for b in bufs if counts[b.tag] > 1}
-  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.index, offs.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
+  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.weakint, offs.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
   return call.replace(src=(call.src[0].substitute(subs, walk=True), *call.src[1:])) if subs else None
 pm_pack_placeholders = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), pack_hcq_placeholders)])
 
@@ -383,7 +362,7 @@ pm_pack_placeholders = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNC
 # 8. callify hcq programs
 
 pm_callify_hcq = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="hcq", src=(UPat(Ops.SINK),), name="cf"),
-  lambda cf: cf.replace(src=(to_program(cf.src[0].replace(arg=KernelInfo("hcq_submit"), tag=1), Device["CPU"].renderer),)))])
+  lambda cf: cf.replace(src=(to_program(cf.src[0].replace(arg=KernelInfo("hcq_submit"), tag=1), Device[HCQ_RUNTIME_DEV.value].renderer),)))])
 
 hcq_compile_cache:dict[tuple[bytes, bool], UOp] = {}
 
@@ -392,27 +371,23 @@ def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None, jit=False) -> UOp:
   if input_uops is not None: linear = graph_rewrite(linear, pm_replace_buffers, ctx=input_uops, walk=True, enter_calls=True, name="replace buffer")
 
   if (final_linear:=(hcq_compile_cache.get(cache_key:=(linear.key, jit)))) is None:
-    # schedule
+    # prep
     linear = linear.substitute(back_map:={s.param_like(i): s for i,s in enumerate(input_uops)} if input_uops is not None else {}, walk=True)
     linear = graph_rewrite(linear, pm_insert_copy_staging + pm_flatten_linear, name="insert copy staging")
-    linear = graph_rewrite(linear, pm_tag_hcq_calls, ctx=(enumerator:=itertools.count(0)), walk=True, name="tag hcq calls")
-    linear = graph_rewrite(linear, pm_sched_sync, ctx=HCQDepsTracker(), walk=True, name="schedule sync")
-    linear = linear.substitute({s: p for p, s in back_map.items()}, walk=True)
+
+    # schedule
+    linear = graph_rewrite(linear, pm_sched_hcq_batches, walk=True, name="schedule hcq batches")
+    linear = linear.substitute({s: p for p, s in back_map.items()}, walk=True, enter_calls=True)
     linear = graph_rewrite(linear, pm_merge_queues, walk=True, name="merge queues")
-    linear = graph_rewrite(linear, pm_add_finalizer, ctx=enumerator, walk=True, name="add finalizer")
-    linear = graph_rewrite(linear, pm_add_global_sync, ctx=set(), walk=True, name="add global sync", enter_calls=True)
 
     # lowering to hcq ir
-    linear = graph_rewrite(linear, pm_add_inner_loads, ctx=(waited:=set()), walk=True, name="add loads", enter_calls=True)
-    linear = graph_rewrite(linear, pm_add_inner_stores, ctx=waited, walk=True, name="add stores", enter_calls=True)
     linear = graph_rewrite(linear, pm_encode_cmdbufs, walk=True, name="encode cmdbufs", enter_calls=True)
     linear = graph_rewrite(linear, pm_pack_placeholders, walk=True, name="pack placeholders")
 
     # pie
     linear = graph_rewrite(linear, pm_split_patches, ctx=jit, walk=True, name="split rt/lt patches")
     linear = graph_rewrite(linear, pm_early_simplify + symbolic, bottom_up=False, name="simplify packed placeholders", enter_calls=True)
-    linear = graph_rewrite(linear, pm_rm_rt_getaddrs, walk=True, name="replace rt getaddrs")
-    linear = graph_rewrite(linear, pm_rm_rt_binaries, walk=True, name="replace rt binaries")
+    linear = graph_rewrite(linear, pm_rm_rt_uops, walk=True, name="replace rt uops")
     linear = graph_rewrite(linear, pm_replace_params, walk=True, name="replace with args")
 
     # and compile it
@@ -425,7 +400,7 @@ def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None, jit=False) -> UOp:
 
 def bufferize_buf(ctx:bool, buf:UOp) -> UOp|None:
   if buf.tag is None: return None
-  return make_mstack(tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=(dv, ctx)), "CPU") for dev in to_tuple(buf.device)))
+  return make_mstack(tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=(dv, ctx)), HCQ_RUNTIME_DEV.value) for dev in to_tuple(buf.device)))
 pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, name="buf"), bufferize_buf)])
 
 # *****************
@@ -435,12 +410,14 @@ def push_stack(op, s): return UOp(Ops.STACK, op.dtype.scalar(),
   tuple(op.replace(dtype=op.dtype.scalar(), src=tuple(x if y is s else y for y in op.src)) for x in s.src))
 
 def fold_binary(buf:UOp, blob:UOp) -> UOp:
-  for b in (m.bufs if isinstance(m:=buf.buffer, MultiBuffer) else (m,)): b.ensure_allocated()._buf.cpu_view().view(fmt='B')[:len(blob.arg)] = blob.arg
+  for b in (m.bufs if isinstance(m:=buf.buffer, MultiBuffer) else (m,)):
+    b.ensure_allocated().as_memoryview(force_zero_copy=True, no_sync=True).cast('B')[:len(blob.arg)] = blob.arg
   return UOp(Ops.NOOP)
 
 def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
   for b, v in zip((bs:=mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)), val.src if val.op is Ops.STACK else (val,)*len(bs)):
-    struct.pack_into(f'<{v.dtype.fmt}', b.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * buf.dtype.itemsize, truncate[v.dtype](v.arg))
+    data = struct.pack(f'<{v.dtype.fmt}', truncate[v.dtype](v.arg))
+    b.ensure_allocated().as_memoryview(force_zero_copy=True, no_sync=True).cast('B')[(byte_off:=off.arg*buf.dtype.itemsize):byte_off+len(data)] = data
   return UOp(Ops.NOOP)
 
 def resolve_getaddr(buf:UOp, g:UOp) -> UOp:
@@ -509,25 +486,26 @@ class HCQ2Compiled(Compiled):
     self.rt_allocator = BumpAllocator(64 << 20, wrap=False)
 
   def new_buffer(self, b:UOp, jit:bool) -> Buffer:
-    if jit or b.tag in HCQ_CACHE_TAGS: return Buffer(self.device, b.max_numel(), b.dtype, options=BufferSpec(cpu_access=True, nolru=True))
+    if jit or b.tag in HCQ_CACHE_TAGS:
+      return Buffer(self.device, b.max_numel(), b.dtype, options=BufferSpec(uncached=True, cpu_access=True, nolru=True))
     return self.rt_buffer.view(b.max_numel(), b.dtype, self.rt_allocator.alloc(b.max_numel() * b.dtype.itemsize, alignment=128))
 
   @functools.cache
   def timeline_signal(self, queue:str|None=None, init_value:int=0) -> Buffer:
     buf = Buffer(self.device, 1, dtypes.uint64, options=BufferSpec(host=True, uncached=True, cpu_access=True), preallocate=True)
-    buf._buf.cpu_view().mv.cast('Q')[0] = init_value
+    buf.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = init_value
     return buf
 
   @functools.cache
   def timeline_value(self, queue:str|None=None, init_value:int=1) -> Buffer:
     buf = Buffer("CPU", 1, dtypes.uint64, preallocate=True)
-    buf.as_memoryview(force_zero_copy=True).cast('Q')[0] = init_value
+    buf.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = init_value
     return buf
 
   def synchronize(self, timeout:int|None=None):
     if not hasattr(self, 'iface'): return
-    sig = self.timeline_signal()._buf.cpu_view().mv.cast('Q')
-    tl = self.timeline_value().as_memoryview(force_zero_copy=True).cast('Q')
+    sig = self.timeline_signal().as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
+    tl = self.timeline_value().as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
     st = time.perf_counter()
     while sig[0] < tl[0] - 1:
       if time.perf_counter() - st > (timeout or 3000) / 1000: self.on_device_hang()
@@ -555,22 +533,19 @@ class HCQ2Compiled(Compiled):
     # if the device has an interface, call device_fini to clean up resources
     if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()
 
+@dataclass
 class HCQ2Buffer:
-  def __init__(self, va_addr:sint, size:int, meta:Any=None, _base:HCQ2Buffer|None=None, view:MMIOInterface|None=None, owner:HCQ2Compiled|None=None):
-    self.va_addr, self.size, self.meta, self._base, self.view, self.owner = va_addr, size, meta, _base, view, owner
+  va_addr:sint
+  meta:Any=None
+  view:MMIOInterface|None=None
 
-  def offset(self, offset:int=0, size:int|None=None) -> HCQ2Buffer:
-    return HCQ2Buffer(self.va_addr+offset, size or (self.size - offset), owner=self.owner, meta=self.meta,
-      _base=self._base or self, view=(self.view.view(offset=offset, size=size) if self.view is not None else None))
-
-  def cpu_view(self) -> MMIOInterface:
-    assert self.view is not None, "buffer has no cpu_view"
-    return self.view
-
-  @property
-  def base(self) -> HCQ2Buffer: return self._base or self
+  def offset(self, offset:int, size:int) -> HCQ2Buffer:
+    return HCQ2Buffer(self.va_addr+offset, meta=self.meta, view=(self.view.view(offset=offset, size=size) if self.view is not None else None))
 
 class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
+  def _as_buffer(self, buf:HCQ2Buffer) -> memoryview:
+    return unwrap(buf.view).mv
+
   def _map(self, buf:HCQ2Buffer) -> HCQ2Buffer:
     if not hasattr(self, '_do_map'): raise NotImplementedError("map failed: no method implemented")
     return self._do_map(buf)
@@ -586,24 +561,3 @@ class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
     self.dev.iface.free(mb)
 
   def _offset(self, buf, size:int, offset:int) -> HCQ2Buffer: return buf.offset(offset=offset, size=size)
-
-  def _wrap(self, dev:str, sz:int, opaque:HCQ2Buffer) -> Buffer:
-    return Buffer(dev, sz, dtypes.uint8, opaque=opaque, options=BufferSpec(external_ptr=1))
-
-  def _copy(self, dst:Buffer, src:Buffer):
-    from tinygrad.engine.realize import run_linear
-    du, su = UOp.from_buffer(dst), UOp.from_buffer(src)
-    run_linear(UOp(Ops.LINEAR, src=(su.param_like(1).copy_to_device(dst.device).call(du, su),)), update_stats=True)
-
-  def _copyin(self, dest:HCQ2Buffer, src:memoryview):
-    s = Buffer(self.dev.device, len(src), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
-    s._buf.cpu_view()[:len(src)] = src
-    self._copy(self._wrap(self.dev.device, len(src), dest), s)
-
-  def _copyout(self, dest:memoryview, src:HCQ2Buffer):
-    d = Buffer(self.dev.device, len(dest), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
-    self._copy(d, self._wrap(self.dev.device, len(dest), src))
-    self.dev.synchronize()
-    dest[:] = d._buf.cpu_view()[:len(dest)]
-
-  # def _as_buffer(self, buf): return buf.cpu_view().mv

extra/hcq2/ops_amd2.py

@@ -138,10 +138,10 @@ def pm4_program(ctx, call, prg):
 pm_pm4_opsel = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="prg"),), name="call", allow_any_len=True), pm4_program),
 
-  (UPat(Ops.WAIT, src=(UPat(name="dst"), UPat(name="val"))), pm4_wait),
-  (UPat(Ops.BARRIER), pm4_barrier),
-  (UPat(Ops.CUSTOM_FUNCTION, arg="timestamp", src=(UPat(name="dst"),)), pm4_timestamp),
-  (UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), pm4_store),
+  (UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))), pm4_wait),
+  (UPat(Ops.INS, arg="barrier"), pm4_barrier),
+  (UPat(Ops.INS, arg="timestamp", src=(UPat(name="dst"),)), pm4_timestamp),
+  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), pm4_store),
 ])
 
 def pm4_submit(cmdbuf, devs):
@@ -202,10 +202,10 @@ def sdma_timestamp(ctx, dst):
 pm_sdma_opsel = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.COPY),), name="call", allow_any_len=True), sdma_copy),
 
-  (UPat(Ops.BARRIER), lambda: UOp(Ops.NOOP, dtypes.void, ())),
-  (UPat(Ops.WAIT, src=(UPat(name="dst"), UPat(name="val"))), sdma_wait),
-  (UPat(Ops.CUSTOM_FUNCTION, arg="timestamp", src=(UPat(name="dst"),)), sdma_timestamp),
-  (UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), sdma_store),
+  (UPat(Ops.INS, arg="barrier"), lambda: UOp(Ops.NOOP, dtypes.void, ())),
+  (UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))), sdma_wait),
+  (UPat(Ops.INS, arg="timestamp", src=(UPat(name="dst"),)), sdma_timestamp),
+  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), sdma_store),
 ])
 
 def sdma_submit(cmdbuf, devs):
@@ -536,7 +536,7 @@ class AMDDevice(HCQ2Compiled):
 
   timestamp_divider = 100.0  # AMD GPU clock: ticks/us
 
-  ifaces = [KFDIface, PCIIface]
+  ifaces = [KFDIface, PCIIface, _mock(KFDIface, "MOCKIface"), _mock(KFDIface), _mock(PCIIface)]
 
   def is_am(self) -> bool: return isinstance(self.iface, (PCIIface,))
   def is_usb(self) -> bool: return False

extra/llama_kernels/__init__.py

@@ -20,11 +20,6 @@ def local_abs_max(x:Tensor) -> Tensor:
   fxn = _local_abs_max_fxn(param.uop, x.device)
   return Tensor(fxn[0].uop.call(x.uop).gettuple(0))
 
-def scalar_amax(amax_buf:Tensor) -> Tensor:
-  if isinstance(amax_buf.device, tuple):
-    return local_abs_max(amax_buf).detach()
-  return amax_buf.max().detach()
-
 def shard_shape(shape:tuple, axis:int, ndev:int) -> list:
   s = list(shape)
   s[axis] //= ndev

extra/llama_kernels/cast_amax/__init__.py

@@ -3,7 +3,7 @@ import functools, pathlib
 from tinygrad import Tensor, dtypes
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.renderer import Estimates
-from extra.llama_kernels import NUM_WG, THREADS_PER_WG, compile_cpp, alloc_like, alloc_local, scalar_amax, dname_of
+from extra.llama_kernels import NUM_WG, THREADS_PER_WG, compile_cpp, alloc_like, dname_of
 
 # module-level mailbox: grad_xw13 UOp -> (grad_xw13_fp8 UOp, delayed amax UOp)
 # lets cdna_asm_gemm's bwd reuse the fp8 companion produced by the fused silu_mul bwd kernel
@@ -11,13 +11,13 @@ from extra.llama_kernels import NUM_WG, THREADS_PER_WG, compile_cpp, alloc_like,
 _grad_fp8_mailbox:dict[UOp, tuple[UOp, UOp]] = {}
 
 @functools.cache
-def _custom_fused_bwd_w13(grad_xw13_fp8:UOp, grad_amax_buf:UOp, grad_amax:UOp,
+def _custom_fused_bwd_w13(grad_xw13_fp8:UOp, grad_amax_next:UOp, grad_amax:UOp,
                           xw13:UOp, grad_x2:UOp, amax_state:UOp, grad_amax_state:UOp, dname:str) -> UOp:
   hidden = xw13.shape[2] // 2
   n_elems = xw13.shape[0] * xw13.shape[1] * hidden
   threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
-  mem = n_elems * 2 * 3 + n_elems * 2 + NUM_WG * 4 + 4
-  sink = UOp.sink(grad_xw13_fp8.base, grad_amax_buf.base, grad_amax.base,
+  mem = n_elems * 2 * 3 + n_elems * 2 + 4 + 4
+  sink = UOp.sink(grad_xw13_fp8.base, grad_amax_next.base, grad_amax.base,
                   xw13.base, grad_x2.base, amax_state.base, grad_amax_state.base, threads, workgroups,
                   arg=KernelInfo(f"fused_silu_mul_bwd_w13_{n_elems}", estimates=Estimates(ops=10*n_elems, mem=mem)))
   src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_bwd_w13.cpp", n_elems, hidden)
@@ -25,14 +25,14 @@ def _custom_fused_bwd_w13(grad_xw13_fp8:UOp, grad_amax_buf:UOp, grad_amax:UOp,
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
 
 @functools.cache
-def _custom_fused_cast_amax_w13(fp8_out:UOp, amax_buf:UOp, xw13:UOp, amax_state:UOp, grad_amax_state:UOp,
+def _custom_fused_cast_amax_w13(fp8_out:UOp, amax_out:UOp, xw13:UOp, amax_state:UOp, grad_amax_state:UOp,
                                 next_grad_amax_state:UOp, dname:str) -> UOp:
   # NOTE: grad_amax_state is plumbed through as an unused fwd input so the bwd kernel can read it via kernel.src
   hidden = xw13.shape[2] // 2
   n_elems = xw13.shape[0] * xw13.shape[1] * hidden
   threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
-  mem = n_elems * 2 * 2 + n_elems + NUM_WG * 4
-  sink = UOp.sink(fp8_out.base, amax_buf.base, xw13.base, amax_state.base, threads, workgroups,
+  mem = n_elems * 2 * 2 + n_elems + 4
+  sink = UOp.sink(fp8_out.base, amax_out.base, xw13.base, amax_state.base, threads, workgroups,
                   arg=KernelInfo(f"fused_silu_mul_cast_amax_w13_{n_elems}", estimates=Estimates(ops=5*n_elems, mem=mem)))
   src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_fwd_w13.cpp", n_elems, hidden)
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
@@ -43,26 +43,23 @@ def _fused_quantize_bwd_w13(gradient:UOp, kernel:UOp):
   device = xw13.device
   axis = xw13.axis if isinstance(device, tuple) else None
   grad_xw13_fp8 = alloc_like(xw13.shape, dtypes.fp8e4m3,  device, axis)
-  grad_amax_buf = alloc_local((NUM_WG,), dtypes.float32,  device, axis)
+  grad_amax_next = Tensor(next_grad_amax_state, device=device)
   grad_amax_state_t = Tensor(grad_amax_state, device=device)
   fxn = functools.partial(_custom_fused_bwd_w13, dname=dname_of(device))
   grad_amax = grad_amax_state_t.empty_like()
-  grad_xw13_fp8, grad_amax_buf, grad_amax, *_ = Tensor.custom_kernel(
-    grad_xw13_fp8, grad_amax_buf, grad_amax,
+  grad_xw13_fp8, grad_amax_next, grad_amax, *_ = Tensor.custom_kernel(
+    grad_xw13_fp8, grad_amax_next, grad_amax,
     Tensor(xw13, device=device), Tensor(gradient, device=device).cast(dtypes.bfloat16),
     Tensor(amax_state, device=device), grad_amax_state_t, fxn=fxn)
   grad_xw13_uop = grad_xw13_fp8.uop.cast(dtypes.bfloat16)
-  new_grad_amax = scalar_amax(grad_amax_buf)
-  store_effect = next_grad_amax_state.store(new_grad_amax.uop)
   assert grad_xw13_fp8.uop.op is Ops.AFTER, f"expected AFTER, got {grad_xw13_fp8.uop.op}"
-  grad_xw13_fp8_uop = grad_xw13_fp8.uop.replace(src=grad_xw13_fp8.uop.src + (store_effect,))
   # Stash fp8 companion for cdna_asm_gemm's bwd to attach to grad_a.
-  _grad_fp8_mailbox[grad_xw13_uop] = (grad_xw13_fp8_uop, grad_amax_state_t.uop)
+  _grad_fp8_mailbox[grad_xw13_uop] = (grad_xw13_fp8.uop, grad_amax_state_t.uop)
   return (None, None, grad_xw13_uop, None, None, None)
 
 def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_state:Tensor,
-                           next_grad_amax_state:Tensor) -> tuple[Tensor, Tensor]:
-  # NOTE: silu(xw1)*xw3 -> fp8 + amax over fused xw13 layout. Returns (fp8, new_amax)
+                           next_grad_amax_state:Tensor, amax_out:Tensor) -> Tensor:
+  # NOTE: silu(xw1)*xw3 -> fp8 + amax over fused xw13 layout. Returns fp8.
   # grad_amax_state: delayed amax for grad_xw13 fp8 quantization in the backward.
   assert xw13.dtype == dtypes.bfloat16, f"expected bf16, got {xw13.dtype}"
   MBS, SEQ, H2 = xw13.shape
@@ -70,8 +67,7 @@ def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_
   HIDDEN = H2 // 2
   axis = xw13.uop.axis if isinstance(xw13.device, tuple) else None
   fp8_out  = alloc_like((MBS, SEQ, HIDDEN), fp8_dtype,      xw13.device, axis)
-  amax_buf = alloc_local((NUM_WG,),         dtypes.float32, xw13.device, axis)
   fxn = functools.partial(_custom_fused_cast_amax_w13, dname=dname_of(xw13.device))
-  fp8_out, amax_buf, *_ = Tensor.custom_kernel(fp8_out, amax_buf, xw13, amax_state, grad_amax_state, next_grad_amax_state,
+  fp8_out, amax_out, *_ = Tensor.custom_kernel(fp8_out, amax_out, xw13, amax_state, grad_amax_state, next_grad_amax_state,
                                                 fxn=fxn, grad_fxn=_fused_quantize_bwd_w13)
-  return fp8_out, scalar_amax(amax_buf)
+  return fp8_out

extra/llama_kernels/cast_amax/cast_amax_bwd_w13.cpp

@@ -23,14 +23,14 @@ static_assert(HIDDEN % VEC == 0, "HIDDEN must be divisible by VEC");
 
 // fused silu*mul backward, three outputs in a single HBM pass:
 //   1) fp8  grad_xw13_fp8  — delayed-scale quantize using grad_amax_state (mailbox to matmul bwd)
-//   2) fp32 grad_amax_buf  — per-WG partial |grad_xw13|, reduced into next step's grad_amax_state
+//   2) fp32 grad_amax_next — scalar |grad_xw13| via global atomic max
 //   3) fp32 grad_amax_out  — delayed grad amax used for quantize/GEMM epilogue scale
 // grad_amax_state is read for the fp8 scale. The store of new_grad_amax into grad_amax_state's
 // buffer is built in Python as a separate effect and threaded into grad_a via .after(store).
 extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
 fused_silu_mul_bwd_w13(
     __hip_fp8_storage_t*  __restrict__ grad_xw13_fp8_out,    // fp8,  2*N_ELEMS
-    float*                __restrict__ grad_amax_buf,        // fp32, NUM_WG per-WG partials
+    float*                __restrict__ grad_amax_next,       // fp32 scalar, initialized to 0 before launch
     float*                __restrict__ grad_amax_out,        // fp32 scalar delayed grad amax
     const __hip_bfloat16* __restrict__ xw13,                 // bf16, 2*N_ELEMS
     const __hip_bfloat16* __restrict__ grad_x2,              // bf16, N_ELEMS
@@ -92,5 +92,6 @@ fused_silu_mul_bwd_w13(
     if (tid < s) sdata[tid] = fmaxf(sdata[tid], sdata[tid + s]);
     __syncthreads();
   }
-  if (tid == 0) grad_amax_buf[wg] = sdata[0];
+  if (tid == 0 && sdata[0] > *grad_amax_next)
+    atomicMax(reinterpret_cast<int32_t*>(grad_amax_next), __float_as_int(sdata[0]));
 }

extra/llama_kernels/cast_amax/cast_amax_fwd_w13.cpp

@@ -24,7 +24,7 @@ static_assert(HIDDEN % VEC == 0, "HIDDEN must be divisible by VEC (so VEC loads
 extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
 fused_silu_mul_cast_amax_w13(
     __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, N_ELEMS
-    float*                __restrict__ amax_buf,        // fp32, NUM_WG (per-WG amaxes)
+    float*                __restrict__ amax_out,        // fp32 scalar, initialized to 0 before launch
     const __hip_bfloat16* __restrict__ xw13,            // bf16, 2*N_ELEMS
     const float*          __restrict__ amax_state)      // fp32 scalar
 {
@@ -67,7 +67,7 @@ fused_silu_mul_cast_amax_w13(
     *reinterpret_cast<uint64_t*>(&fp8_out[base]) = *reinterpret_cast<uint64_t*>(out);
   }
 
-  // LDS tree reduction: per-workgroup amax
+  // LDS tree reduction: per-workgroup amax, then global atomic into the scalar.
   sdata[tid] = local_max;
   __syncthreads();
   for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
@@ -75,5 +75,5 @@ fused_silu_mul_cast_amax_w13(
     __syncthreads();
   }
 
-  if (tid == 0) amax_buf[wg] = sdata[0];
+  if (tid == 0 && sdata[0] > *amax_out) atomicMax(reinterpret_cast<int32_t*>(amax_out), __float_as_int(sdata[0]));
 }

extra/llama_kernels/fused_ce/__init__.py

@@ -16,7 +16,7 @@ def _custom_fused_ce_loss_fwd(loss_out:UOp, max_out:UOp, lse_out:UOp, logits:UOp
   row_lse = (logits[b, s, v_lse].cast(dtypes.float) - row_max).exp().reduce(v_lse, arg=Ops.ADD).log() + row_max
 
   v_smooth = UOp.range(vocab, 3, axis_type=AxisType.REDUCE)
-  target = logits[b, s, targets[row].cast(dtypes.index)].cast(dtypes.float)
+  target = logits[b, s, targets[row].cast(dtypes.weakint)].cast(dtypes.float)
   mean_logits = logits[b, s, v_smooth].cast(dtypes.float).reduce(v_smooth, arg=Ops.ADD) / vocab
   loss = row_lse - (1.0 - label_smoothing) * target - label_smoothing * mean_logits
   stores = UOp.group(loss_out[row].store(loss), max_out[row].store(row_max), lse_out[row].store(row_lse))
@@ -32,7 +32,7 @@ def _custom_fused_ce_loss_bwd(d_logits:UOp, logits:UOp, lse:UOp, targets:UOp, sc
   s = row % seq
 
   prob = (logits[b, s, v].cast(dtypes.float) - lse[row]).exp()
-  target = v.eq(targets[row].cast(dtypes.index)).where(1.0 - label_smoothing, 0.0)
+  target = v.eq(targets[row].cast(dtypes.weakint)).where(1.0 - label_smoothing, 0.0)
   smooth = label_smoothing / vocab
   grad = (prob - target - smooth) * scale[0]
 

extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/__init__.py

@@ -3,19 +3,19 @@ import functools, pathlib
 from tinygrad import Tensor, dtypes
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.renderer import Estimates
-from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, alloc_like, alloc_local, scalar_amax, dname_of, compile_hip
+from extra.llama_kernels import NUM_WG, THREADS_PER_WG, alloc_like, alloc_local, dname_of, compile_hip
 
 def _src() -> str: return (pathlib.Path(__file__).parent/"fused_rmsnorm_mul_quantize_fp8.cpp").read_text()
 def _src_bwd() -> str: return (pathlib.Path(__file__).parent/"fused_rmsnorm_mul_quantize_fp8_bwd.cpp").read_text()
 
 @functools.cache
-def _custom_fwd(fp8_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_buf:UOp,
+def _custom_fwd(fp8_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_out:UOp,
                 x:UOp, weight:UOp, amax_state:UOp, dname:str, eps_val:float) -> UOp:
   MBS, SEQ, HIDDEN = x.shape
   n_elems = MBS * SEQ * HIDDEN
   threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
-  mem = n_elems * 2 + n_elems + MBS * SEQ * 4 + n_elems + HIDDEN * 2 + NUM_WG * 4 + 4
-  sink = UOp.sink(fp8_out.base, x_normed_out.base, rrms_out.base, amax_buf.base,
+  mem = n_elems * 2 + n_elems + MBS * SEQ * 4 + n_elems + HIDDEN * 2 + 4 + 4
+  sink = UOp.sink(fp8_out.base, x_normed_out.base, rrms_out.base, amax_out.base,
                   x.base, weight.base, amax_state.base, threads, workgroups,
                   arg=KernelInfo(f"fused_rmsnorm_mul_quantize_fp8_{n_elems}_h{HIDDEN}_eps{eps_val:.0e}",
                                  estimates=Estimates(ops=6*n_elems, mem=mem)))
@@ -26,13 +26,13 @@ def _custom_fwd(fp8_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_buf:UOp,
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
 
 @functools.cache
-def _custom_fwd_add(fp8_out:UOp, h_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_buf:UOp,
+def _custom_fwd_add(fp8_out:UOp, h_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_out:UOp,
                     x:UOp, residual:UOp, weight:UOp, amax_state:UOp, dname:str, eps_val:float) -> UOp:
   MBS, SEQ, HIDDEN = x.shape
   n_elems = MBS * SEQ * HIDDEN
   threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
-  mem = n_elems * 2 * 4 + MBS * SEQ * 4 + HIDDEN * 2 + NUM_WG * 4 + 4
-  sink = UOp.sink(fp8_out.base, h_out.base, x_normed_out.base, rrms_out.base, amax_buf.base,
+  mem = n_elems * 2 * 4 + MBS * SEQ * 4 + HIDDEN * 2 + 4 + 4
+  sink = UOp.sink(fp8_out.base, h_out.base, x_normed_out.base, rrms_out.base, amax_out.base,
                   x.base, residual.base, weight.base, amax_state.base, threads, workgroups,
                   arg=KernelInfo(f"fused_add_rmsnorm_mul_quantize_fp8_{n_elems}_h{HIDDEN}_eps{eps_val:.0e}",
                                  estimates=Estimates(ops=7*n_elems, mem=mem)))
@@ -85,7 +85,7 @@ def _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_st
   return grad_total.uop, grad_weight_uop
 
 def _fused_bwd(gradient:UOp, kernel:UOp):
-  # NOTE: fwd inputs (fp8_out, x_normed_out, rrms_out, amax_buf, x, weight, amax_state)
+  # NOTE: fwd inputs (fp8_out, x_normed_out, rrms_out, amax_out, x, weight, amax_state)
   _, x_normed_u, rrms_u, _, x_u, weight_u, amax_state_u = kernel.src[1:]
   grad_x, grad_w = _bwd_common(gradient, None, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel)
   return (None, None, None, None, grad_x, grad_w, None)
@@ -112,8 +112,9 @@ def _fused_add_bwd(*args, **kwargs):
   grad_h, grad_w = _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel)
   return (None, None, None, None, None, grad_h, grad_h, grad_w, None)
 
-def fused_rmsnorm_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor]:
-  # NOTE: rmsnorm(x) * weight -> fp8 + amax. Returns (fp8, new_amax, x_normed, rrms).
+def fused_rmsnorm_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, eps:float, fp8_dtype,
+                                   amax_out:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  # NOTE: rmsnorm(x) * weight -> fp8 + amax. Returns (fp8, x_normed, rrms).
   # x_normed + rrms are saved for the rmsnorm backward (also recomputed here from x regs).
   assert x.dtype == dtypes.bfloat16 and weight.dtype == dtypes.bfloat16
   assert x.shape[-1] == weight.shape[-1], f"HIDDEN mismatch: x={x.shape}, weight={weight.shape}"
@@ -123,16 +124,15 @@ def fused_rmsnorm_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, e
   fp8_out      = alloc_like((MBS, SEQ, HIDDEN), fp8_dtype,       x.device, axis)
   x_normed_out = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, x.device, axis)
   rrms_out     = alloc_like((MBS, SEQ),         dtypes.float32,  x.device, axis)
-  amax_buf     = alloc_local((NUM_WG,),         dtypes.float32,  x.device, axis)
   fxn = functools.partial(_custom_fwd, dname=dname_of(x.device), eps_val=eps)
-  fp8_out, x_normed_out, rrms_out, amax_buf, *_ = Tensor.custom_kernel(
-    fp8_out, x_normed_out, rrms_out, amax_buf, x, weight, amax_state, fxn=fxn, grad_fxn=_fused_bwd)
-  return fp8_out, scalar_amax(amax_buf), x_normed_out, rrms_out
+  fp8_out, x_normed_out, rrms_out, amax_out, *_ = Tensor.custom_kernel(
+    fp8_out, x_normed_out, rrms_out, amax_out, x, weight, amax_state, fxn=fxn, grad_fxn=_fused_bwd)
+  return fp8_out, x_normed_out, rrms_out
 
 def fused_add_rmsnorm_mul_quantize_fp8(x:Tensor, residual:Tensor, weight:Tensor, amax_state:Tensor,
-                                       eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor, Tensor]:
+                                       eps:float, fp8_dtype, amax_out:Tensor) -> tuple[Tensor, Tensor, Tensor, Tensor]:
   # NOTE: h = x + residual; y_normed = rmsnorm(h); fp8 = quantize(y_normed * weight).
-  # Returns (fp8, new_amax, h, x_normed, rrms). h is also written so downstream can
+  # Returns (fp8, h, x_normed, rrms). h is also written so downstream can
   # reuse it without recomputing x+residual — eliminates the separate residual-add kernel.
   assert x.dtype == dtypes.bfloat16 and residual.dtype == dtypes.bfloat16 and weight.dtype == dtypes.bfloat16
   assert x.shape == residual.shape
@@ -143,9 +143,8 @@ def fused_add_rmsnorm_mul_quantize_fp8(x:Tensor, residual:Tensor, weight:Tensor,
   h_out        = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, x.device, axis)
   x_normed_out = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, x.device, axis)
   rrms_out     = alloc_like((MBS, SEQ),         dtypes.float32,  x.device, axis)
-  amax_buf     = alloc_local((NUM_WG,),         dtypes.float32,  x.device, axis)
   fxn = functools.partial(_custom_fwd_add, dname=dname_of(x.device), eps_val=eps)
-  fp8_out, h_out, x_normed_out, rrms_out, amax_buf, *_ = Tensor.custom_kernel(
-    fp8_out, h_out, x_normed_out, rrms_out, amax_buf, x, residual, weight, amax_state,
+  fp8_out, h_out, x_normed_out, rrms_out, amax_out, *_ = Tensor.custom_kernel(
+    fp8_out, h_out, x_normed_out, rrms_out, amax_out, x, residual, weight, amax_state,
     fxn=fxn, grad_fxn=_fused_add_bwd)
-  return fp8_out, scalar_amax(amax_buf), h_out, x_normed_out, rrms_out
+  return fp8_out, h_out, x_normed_out, rrms_out

extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/fused_rmsnorm_mul_quantize_fp8.cpp

@@ -7,7 +7,7 @@
 //   fp8 = fp8_sat(y * (FP8_MAX / amax_state))
 // Also writes:
 //   rrms[row]        — saved for the rmsnorm backward
-//   amax_buf[wg]     — per-WG |y| partials, reduced later to update amax_state
+//   amax_out         — scalar |y| via global atomic max
 //
 // Layout: one WG per row, ROWS_PER_WG rows per WG via grid-stride (ROWS = N_ELEMS / HIDDEN).
 // Each thread handles HIDDEN / THREADS_PER_WG elements per row.
@@ -48,7 +48,7 @@ fused_add_rmsnorm_mul_quantize_fp8(
     __hip_bfloat16*       __restrict__ h_out,           // bf16, ROWS*HIDDEN — x + residual (saved for downstream)
     __hip_bfloat16*       __restrict__ x_normed_out,    // bf16, ROWS*HIDDEN
     float*                __restrict__ rrms_out,        // fp32, ROWS
-    float*                __restrict__ amax_buf,        // fp32, NUM_WG
+    float*                __restrict__ amax_out,        // fp32 scalar, initialized to 0 before launch
     const __hip_bfloat16* __restrict__ x,               // bf16, ROWS*HIDDEN
     const __hip_bfloat16* __restrict__ residual,        // bf16, ROWS*HIDDEN — added into x before rmsnorm
     const __hip_bfloat16* __restrict__ weight,          // bf16, HIDDEN
@@ -60,7 +60,7 @@ fused_rmsnorm_mul_quantize_fp8(
     __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, ROWS*HIDDEN
     __hip_bfloat16*       __restrict__ x_normed_out,    // bf16, ROWS*HIDDEN (saved for rmsnorm bwd)
     float*                __restrict__ rrms_out,        // fp32, ROWS (fp32 to match rmsnorm_bwd.cpp expectation)
-    float*                __restrict__ amax_buf,        // fp32, NUM_WG per-WG partials
+    float*                __restrict__ amax_out,        // fp32 scalar, initialized to 0 before launch
     const __hip_bfloat16* __restrict__ x,               // bf16, ROWS*HIDDEN
     const __hip_bfloat16* __restrict__ weight,          // bf16, HIDDEN (per-hidden scale)
     const float*          __restrict__ amax_state)      // fp32 scalar
@@ -144,12 +144,12 @@ fused_rmsnorm_mul_quantize_fp8(
     __syncthreads();  // before next row's sum_sq reduce reuses sdata
   }
 
-  // Final per-WG amax reduce.
+  // Final per-WG amax reduce, then global atomic into the scalar.
   sdata[tid] = local_max;
   __syncthreads();
   for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
     if (tid < s) sdata[tid] = fmaxf(sdata[tid], sdata[tid + s]);
     __syncthreads();
   }
-  if (tid == 0) amax_buf[wg] = sdata[0];
+  if (tid == 0 && sdata[0] > *amax_out) atomicMax(reinterpret_cast<int32_t*>(amax_out), __float_as_int(sdata[0]));
 }

extra/llama_kernels/quantize_fp8_delayed/__init__.py

@@ -3,14 +3,13 @@ from tinygrad import Tensor, dtypes
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import prod
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
-from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, alloc_like, alloc_local, scalar_amax
+from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, alloc_like
 
 @functools.cache
-def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_state:UOp) -> UOp:
+def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_out:UOp, x:UOp, amax_state:UOp, device=None) -> UOp:
   VEC = 8
   n_elems = prod(x.shape)
   assert n_elems % (NUM_WG * THREADS_PER_WG * VEC) == 0
-  assert amax_partial.shape[0] == NUM_WG
 
   x = x.reshape(n_elems)
   fp8_out = fp8_out.reshape(n_elems)
@@ -46,8 +45,13 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_st
     lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)).barrier())
     step //= 2
 
-  amax_store = amax_partial[tid.eq(0).where(wg, UOp.invalid())].store(lds[0])
-  return amax_store.end(tid, wg).sink(arg=KernelInfo(f"quantize_fp8_with_amax_{n_elems}", opts_to_apply=()))
+  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))
+  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=()))
 
 @functools.cache
 def _custom_quantize_fp8_scalar(fp8_out:UOp, x:UOp, amax_state:UOp) -> UOp:
@@ -69,25 +73,19 @@ def _quantize_fp8_delayed_bwd(gradient:UOp, kernel:UOp):
   grad_x = (Tensor(gradient, device=device).float() * scale).cast(dtypes.bfloat16)
   return (None, None, grad_x.uop, None)
 
-def quantize_fp8_delayed(x:Tensor, amax_state:Tensor, fp8_dtype=dtypes.fp8e4m3) -> tuple[Tensor, Tensor, Tensor, UOp]:
-  # NOTE: one-pass bf16 -> fp8 quantize with delayed scaling. Returns (fp8, inv_scale, new_amax, store_effect).
-  # Fused kernel reads x once and writes fp8 + per-WG |x| partials (then a small reduce produces scalar new_amax).
-  # store_effect writes new_amax into amax_state's buffer — the caller must thread it into a realized
-  # output via `.after(store_effect)`. Calling `amax_state.assign(new_amax)` inside a grad_fxn does
-  # NOT work because .assign mutates only the temp Tensor's .uop, not the original layer-owned buffer.
+def quantize_fp8_delayed(x:Tensor, amax_state:Tensor, amax_out:Tensor, fp8_dtype=dtypes.fp8e4m3) -> tuple[Tensor, Tensor]:
+  # NOTE: one-pass bf16 -> fp8 quantize with delayed scaling.
+  # Fused kernel reads x once and writes fp8 + scalar amax via global atomic max.
   assert x.dtype == dtypes.bfloat16, f"expected bf16, got {x.dtype}"
   axis = x.uop.axis if isinstance(x.device, tuple) else None
   fp8_out      = alloc_like(x.shape,  fp8_dtype,      x.device, axis)
   n_elems = prod(x.uop.shard_shape)
   assert n_elems % NUM_WG == 0, f"{n_elems=} must divide over {NUM_WG=}"
-  amax_partial = alloc_local((NUM_WG,), dtypes.float32, x.device, axis)
-  fxn = _custom_quantize_fp8_with_amax
-  fp8_out, amax_partial, *_ = Tensor.custom_kernel(fp8_out, amax_partial, x, amax_state,
-                                                    fxn=fxn, grad_fxn=_quantize_fp8_delayed_bwd)
-  new_amax = scalar_amax(amax_partial)
+  fxn = functools.partial(_custom_quantize_fp8_with_amax, device=x.device)
+  fp8_out, amax_out, *_ = Tensor.custom_kernel(fp8_out, amax_out, x, amax_state,
+                                                fxn=fxn, grad_fxn=_quantize_fp8_delayed_bwd)
   inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
-  store_effect = amax_state.uop.store(new_amax.uop)
-  return fp8_out, inv_scale, new_amax, store_effect
+  return fp8_out, inv_scale
 
 def quantize_fp8_scalar(x:Tensor, amax_state:Tensor, fp8_dtype=dtypes.fp8e4m3) -> Tensor:
   # NOTE: pure one-pass bf16 -> fp8 quantize with delayed scalar scale. No amax computation.

extra/thunder/amd/fa.py

@@ -16,6 +16,96 @@ def _sharded_empty(shape:Tensor, ref:Tensor, axis:int|None, dtype:DTypeLike|None
   axis = ref.uop.axis if axis is None else axis
   return Tensor(Tensor.invalids(*shape, dtype=dtype, device=ref.device).uop.multi(axis), dtype=dtype, device=ref.device)
 
+@functools.cache
+def custom_fused_qkv_rope_forward(q:UOp, k:UOp, v:UOp, xqkv:UOp, freqs_cis:UOp,
+                                  device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+  code = (pathlib.Path(__file__).parent / "fused_qkv_rope.cpp").read_text()
+  threads = 256
+  thread_idx = UOp.special(threads, "lidx0")
+  block_idx_x, block_idx_y = UOp.special(B, "gidx0"), UOp.special(N, "gidx1")
+  sink = UOp.sink(q.base, k.base, v.base, xqkv.base, freqs_cis.base, thread_idx, block_idx_x, block_idx_y,
+                  arg=KernelInfo(name="fused_qkv_rope_forward"))
+  compile_args = ["-std=c++20", "-ffast-math", f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}",
+                  f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DTHREADS_PER_BLOCK={threads}"]
+  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+
+@functools.cache
+def custom_fused_qkv_rope_backward(dxqkv:UOp, dq:UOp, dk:UOp, dv:UOp, freqs_cis:UOp,
+                                   device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+  assert (B, N, H, H_KV, D) == (2, 8192, 32, 8, 128)
+  code = (pathlib.Path(__file__).parent / "fused_qkv_rope_bwd.cpp").read_text()
+  threads = 256
+  thread_idx = UOp.special(threads, "lidx0")
+  gsz = (B, N // 64, H + 2 * H_KV)
+  block_idx_x, block_idx_y, block_idx_z = (UOp.special(x, f"gidx{i}") for i, x in enumerate(gsz))
+  sink = UOp.sink(dxqkv.base, dq.base, dk.base, dv.base, freqs_cis.base, thread_idx, block_idx_x, block_idx_y, block_idx_z,
+                  arg=KernelInfo(name="fused_qkv_rope_backward"))
+  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math", f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}",
+                  f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DTHREADS_PER_BLOCK={threads}"]
+  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+
+def _fa_native_grads(dq:UOp, dk:UOp, dv:UOp) -> tuple[UOp, UOp, UOp]|None:
+  def unwrap_partial(x:UOp) -> UOp|None:
+    expected = (Ops.CAST, Ops.REDUCE, Ops.PERMUTE, Ops.CAST, Ops.RESHAPE, Ops.AFTER)
+    for op in expected:
+      if x.op is not op: return None
+      if op is not Ops.AFTER: x = x.src[0]
+    return x
+  dq_native, dk_partial, dv_partial = dq.base, unwrap_partial(dk), unwrap_partial(dv)
+  if dq_native.op is not Ops.AFTER or dk_partial is None or dv_partial is None: return None
+  B, N, H, D, H_KV = dq.shape[0], dq.shape[1], dq.shape[2], dq.shape[3], dk.shape[2]
+  heads_per_wg = 2 if D == 128 and (H // H_KV) % 2 == 0 else 1
+  partials = (H // H_KV) // heads_per_wg
+  if dq_native.shape != (B, H, N, D) or dk_partial.shape != (B * partials, N, H_KV, D) or dv_partial.shape != dk_partial.shape: return None
+  return dq_native, dk_partial, dv_partial
+
+def _fused_qkv_rope_grad(dq_u:UOp, dk_u:UOp, dv_u:UOp, call:UOp) -> tuple[None, None, None, UOp, None]:
+  dq, dk, dv = Tensor(dq_u, device=dq_u.device), Tensor(dk_u, device=dk_u.device), Tensor(dv_u, device=dv_u.device)
+  xqkv_u, freqs_u = call.src[4], call.src[5]
+  xqkv, freqs_cis = Tensor(xqkv_u, device=xqkv_u.device), Tensor(freqs_u, device=freqs_u.device)
+  B, N, _ = xqkv.shape
+  H, H_KV, D = dq.shape[2], dk.shape[2], dq.shape[3]
+  num_devices = len(xqkv.device) if isinstance(xqkv.device, tuple) else 1
+  is_dp, is_mp = xqkv.uop.axis == 0, xqkv.uop.axis == 2
+  B_local = B // num_devices if is_dp else B
+  H_local = H // num_devices if is_mp else H
+  H_KV_local = H_KV // num_devices if is_mp else H_KV
+  single_device = xqkv.device[0] if isinstance(xqkv.device, tuple) else xqkv.device
+  arch = Device[single_device].renderer.target.arch
+  fa_native = _fa_native_grads(dq_u, dk_u, dv_u)
+  assert fa_native is not None, "fused QKV RoPE backward requires native Flash Attention gradients"
+  dq, dk, dv = (Tensor(x, device=x.device) for x in fa_native)
+  dxqkv = _sharded_empty_like(xqkv, axis=xqkv.uop.axis if isinstance(xqkv.device, tuple) else None)
+  fxn = functools.partial(custom_fused_qkv_rope_backward, device=single_device, arch=arch,
+                          B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D)
+  dxqkv = Tensor.custom_kernel(dxqkv, dq, dk, dv, freqs_cis, fxn=fxn)[0]
+  return None, None, None, dxqkv.uop, None
+
+def fused_qkv_rope(xqkv:Tensor, freqs_cis:Tensor, n_heads:int, n_kv_heads:int, head_dim:int) -> tuple[Tensor, Tensor, Tensor]:
+  B, N, packed_dim = xqkv.shape
+  assert packed_dim == n_kv_heads * (n_heads // n_kv_heads + 2) * head_dim
+  assert freqs_cis.dtype == dtypes.bfloat16, f"fused QKV RoPE requires bfloat16 frequencies, got {freqs_cis.dtype}"
+  assert freqs_cis.shape == (1, freqs_cis.shape[1], 1, head_dim // 2, 2) and freqs_cis.shape[1] >= N, \
+    f"invalid RoPE frequency shape {freqs_cis.shape} for sequence length {N} and head dimension {head_dim}"
+  num_devices = len(xqkv.device) if isinstance(xqkv.device, tuple) else 1
+  is_dp, is_mp = xqkv.uop.axis == 0, xqkv.uop.axis == 2
+  B_local = B // num_devices if is_dp else B
+  H_local = n_heads // num_devices if is_mp else n_heads
+  H_KV_local = n_kv_heads // num_devices if is_mp else n_kv_heads
+  assert H_local % H_KV_local == 0 and head_dim % 2 == 0 and head_dim <= 512
+  single_device = xqkv.device[0] if isinstance(xqkv.device, tuple) else xqkv.device
+  arch = Device[single_device].renderer.target.arch
+  axis = 0 if is_dp else 2 if is_mp else None
+  q = _sharded_empty((B, N, n_heads, head_dim), xqkv, axis=axis, dtype=dtypes.bfloat16)
+  k = _sharded_empty((B, N, n_kv_heads, head_dim), xqkv, axis=axis, dtype=dtypes.bfloat16)
+  v = _sharded_empty((B, N, n_kv_heads, head_dim), xqkv, axis=axis, dtype=dtypes.bfloat16)
+  fxn = functools.partial(custom_fused_qkv_rope_forward, device=single_device, arch=arch,
+                          B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=head_dim)
+  q, k, v, *_ = Tensor.custom_kernel(q, k, v, xqkv, freqs_cis, fxn=fxn, grad_fxn=_fused_qkv_rope_grad)
+  return q, k, v
+
 def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
   return _sharded_empty(ref.shape, ref, axis)
 
@@ -31,8 +121,9 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
 
     dq = _sharded_empty((B, H, N, D), xq, axis=shard_axis_t)
     GROUP_SIZE = H_local // H_KV_local
-    dk_partial = _sharded_empty((B * GROUP_SIZE, N, H_KV, D), xk, axis=shard_axis)
-    dv_partial = _sharded_empty((B * GROUP_SIZE, N, H_KV, D), xv, axis=shard_axis)
+    HEADS_PER_WG = 2 if D == 128 and GROUP_SIZE % 2 == 0 else 1
+    dk_partial = _sharded_empty((B * GROUP_SIZE // HEADS_PER_WG, N, H_KV, D), xk, axis=shard_axis)
+    dv_partial = _sharded_empty((B * GROUP_SIZE // HEADS_PER_WG, N, H_KV, D), xv, axis=shard_axis)
 
     # delta_vec = (do * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()
     delta_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
@@ -46,8 +137,8 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
       dq = dq.reshape(B, H, N//16, 4, 2, 2, D//32, 4, 4, 2).permute(0, 1, 2, 7, 8, 3, 4, 6, 5, 9).reshape(B, H, N, D).transpose(1, 2)
 
     # reduce partial dK/dV across GROUP_SIZE query heads
-    dk = dk_partial.reshape(B, GROUP_SIZE, N, H_KV, D).sum(1)
-    dv = dv_partial.reshape(B, GROUP_SIZE, N, H_KV, D).sum(1)
+    dk = dk_partial.reshape(B, GROUP_SIZE // HEADS_PER_WG, N, H_KV, D).sum(1)
+    dv = dv_partial.reshape(B, GROUP_SIZE // HEADS_PER_WG, N, H_KV, D).sum(1)
 
     if not has_sink: return None, None, dq.uop, dk.uop, dv.uop
     sinks = Tensor(ker.src[6], device=ker.src[6].device)
@@ -160,9 +251,11 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
                   f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}"]
 
   BLOCK_SIZE_KV = 256
+  GROUP_SIZE = H // H_KV
+  HEADS_PER_WG = 2 if D == 128 and GROUP_SIZE % 2 == 0 else 1
   NUM_WARPS = 4
   NUM_THREADS = 64 * NUM_WARPS
-  gsz = (H, N // BLOCK_SIZE_KV, B)
+  gsz = (H // HEADS_PER_WG, N // BLOCK_SIZE_KV, B)
   lsz = (NUM_THREADS, 1, 1)
   threadIdx_x = UOp.special(lsz[0], "lidx0")
   blockIdx_x, blockIdx_y, blockIdx_z = UOp.special(gsz[0], "gidx0"), UOp.special(gsz[1], "gidx1"), UOp.special(gsz[2], "gidx2")

extra/thunder/amd/fa_bwd_causal.cpp

@@ -28,6 +28,7 @@ constexpr int ATTN_H_KV = 8; // number of key/value heads (for GQA)
 #endif
 
 constexpr int GROUP_SIZE = ATTN_H / ATTN_H_KV; // queries per KV head group
+constexpr int HEADS_PER_WG = (ATTN_D == 128 && GROUP_SIZE % 2 == 0) ? 2 : 1;
 
 #ifndef ATTN_N
 constexpr int ATTN_N = 1024; // sequence length
@@ -53,7 +54,7 @@ using namespace kittens;
 using _gl_QdO  = gl<bf16, ATTN_B, ATTN_N, ATTN_H, ATTN_D>;
 using _gl_KV   = gl<bf16, ATTN_B, ATTN_N, ATTN_H_KV, ATTN_D>;
 using _gl_dQ   = gl<bf16, ATTN_B, ATTN_H, ATTN_N, ATTN_D>;
-using _gl_dKV  = gl<bf16, ATTN_B * GROUP_SIZE, ATTN_N, ATTN_H_KV, ATTN_D>;
+using _gl_dKV  = gl<bf16, ATTN_B * (GROUP_SIZE / HEADS_PER_WG), ATTN_N, ATTN_H_KV, ATTN_D>;
 using _gl_Lvec = gl<float, ATTN_B, ATTN_H, 1, ATTN_N>;
 
 template<int D> struct attn_bwd_combined_globals {
@@ -63,7 +64,7 @@ template<int D> struct attn_bwd_combined_globals {
   _gl_dQ dQg;
   _gl_dKV dKg, dVg;
   _gl_Lvec L_vec, delta_vec;
-  dim3 grid() { return dim3(ATTN_H, (ATTN_N / BLOCK_SIZE_KV), ATTN_B); }
+  dim3 grid() { return dim3(ATTN_H / HEADS_PER_WG, (ATTN_N / BLOCK_SIZE_KV), ATTN_B); }
   dim3 block() { return dim3(NUM_THREADS); }
   size_t dynamic_shared_memory() { return MAX_SHARED_MEMORY; }
 };
@@ -71,7 +72,7 @@ template<int D> struct attn_bwd_combined_globals {
 template<int D> __launch_bounds__(NUM_THREADS, 1)
 __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr, bf16 *dO_ptr, bf16 *Q_ptr, bf16 *K_ptr, bf16 *V_ptr, float *L_vec_ptr, float *delta_vec_ptr) {
 
-  const int q_head_idx_fixed = blockIdx.x;  // This is the query head index [0, ATTN_H)
+  const int q_head_idx_fixed = blockIdx.x * HEADS_PER_WG;  // First query head handled by this workgroup.
   const int kv_head_idx = q_head_idx_fixed / GROUP_SIZE;
   const int q_head_in_group = q_head_idx_fixed % GROUP_SIZE;
   const int seq_idx = blockIdx.y;
@@ -88,7 +89,7 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
   // first Q step that can overlap this K_span:
   const int first_step = max(0, k_start_min / STEP_QO);
   const int num_steps_per_head = total_steps_per_head - first_step;
-  const int num_steps = num_steps_per_head;
+  const int num_steps = num_steps_per_head * HEADS_PER_WG;
   const int k_pos = j * WARP_SIZE_KV;
 
   constexpr float L_SCALE_FACTOR = 1.44269504089f;
@@ -270,12 +271,12 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
   if (num_steps > 1) {
     // Prologue
     {
-      const int q_head_idx = (0) / num_steps_per_head + first_q_head;
-      const int q_seq_idx = ((0) % num_steps_per_head) + first_step;
+      const int q_head_idx = first_q_head;
+      const int q_seq_idx = first_step;
       const int q_pos = q_seq_idx * STEP_QO;
 
-      const int next_q_head_idx = (0 + 1) / num_steps_per_head + first_q_head;
-      const int next_q_seq_idx = ((0 + 1) % num_steps_per_head) + first_step;
+      const int next_q_head_idx = first_q_head;
+      const int next_q_seq_idx = first_step + 1;
 
       // dot slice 0
       {
@@ -1332,15 +1333,18 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
 
     // 9. for 1 <= i <= T_r (1024 / 32 = 32)  
     for (int i = 1; i < num_steps - 1; ++i, tic ^= 1, toc ^= 1) {
-      const int last_q_head_idx = (i - 1) / num_steps_per_head + first_q_head;
-      const int last_q_seq_idx = ((i - 1) % num_steps_per_head) + first_step;
+      const int last_head_offset = (i - 1) >= num_steps_per_head;
+      const int last_q_head_idx = last_head_offset + first_q_head;
+      const int last_q_seq_idx = i - 1 - last_head_offset * num_steps_per_head + first_step;
 
-      const int q_head_idx = i / num_steps_per_head + first_q_head;
-      const int q_seq_idx = (i % num_steps_per_head) + first_step;
+      const int head_offset = i >= num_steps_per_head;
+      const int q_head_idx = head_offset + first_q_head;
+      const int q_seq_idx = i - head_offset * num_steps_per_head + first_step;
       const int q_pos = q_seq_idx * STEP_QO;
 
-      const int next_q_head_idx = (i + 1) / num_steps_per_head + first_q_head;
-      const int next_q_seq_idx = ((i + 1) % num_steps_per_head) + first_step;
+      const int next_head_offset = (i + 1) >= num_steps_per_head;
+      const int next_q_head_idx = next_head_offset + first_q_head;
+      const int next_q_seq_idx = i + 1 - next_head_offset * num_steps_per_head + first_step;
 
       // dot slice 0
       {
@@ -2378,11 +2382,11 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
     }
   }
 
-  const int last_q_head_idx = (num_steps - 2) / num_steps_per_head + first_q_head;
-  const int last_q_seq_idx = ((num_steps - 2) % num_steps_per_head) + first_step;
+  const int last_q_head_idx = first_q_head + HEADS_PER_WG - 1;
+  const int last_q_seq_idx = first_step + num_steps_per_head - 2;
 
-  const int q_head_idx = (num_steps - 1) / num_steps_per_head + first_q_head;
-  const int q_seq_idx = ((num_steps - 1) % num_steps_per_head) + first_step;
+  const int q_head_idx = first_q_head + HEADS_PER_WG - 1;
+  const int q_seq_idx = first_step + num_steps_per_head - 1;
   const int q_pos = q_seq_idx * STEP_QO;
   // Epilogue
   {
@@ -3407,14 +3411,14 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
     }
   }
 
-  store<1>(g.dVg, dV_j, {batch_idx * GROUP_SIZE + q_head_in_group, 0, kv_head_idx, 0}, {0, j, 0, 0});
+  store<1>(g.dVg, dV_j, {batch_idx * (GROUP_SIZE / HEADS_PER_WG) + q_head_in_group / HEADS_PER_WG, 0, kv_head_idx, 0}, {0, j, 0, 0});
   __builtin_amdgcn_s_waitcnt(0);
   __builtin_amdgcn_s_barrier();
 
   // We first copy dV_j_T from accumulator GPRs to vector GPRs and then perform the store
   accvgpr_read(dV_j_T, dK_j_T);
   mul(dV_j_T, dV_j_T, dP_SCALE_FACTOR);
-  store<1>(g.dKg, dV_j, {batch_idx * GROUP_SIZE + q_head_in_group, 0, kv_head_idx, 0}, {0, j, 0, 0});
+  store<1>(g.dKg, dV_j, {batch_idx * (GROUP_SIZE / HEADS_PER_WG) + q_head_in_group / HEADS_PER_WG, 0, kv_head_idx, 0}, {0, j, 0, 0});
 
   // Write out final dQ_i slice
   mul(dQ_i_T, dQ_i_T, dP_SCALE_FACTOR);
@@ -3422,6 +3426,3 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
 }
 
 template __global__ void attend_bwd_combined_ker<ATTN_D>(bf16*, bf16*, bf16*, bf16*, bf16*, bf16*, bf16*, float*, float*);
-
-
-

extra/thunder/amd/fused_qkv_rope.cpp

@@ -0,0 +1,69 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+
+#ifndef ATTN_B
+#define ATTN_B 2
+#endif
+#ifndef ATTN_N
+#define ATTN_N 8192
+#endif
+#ifndef ATTN_H
+#define ATTN_H 32
+#endif
+#ifndef ATTN_H_KV
+#define ATTN_H_KV 8
+#endif
+#ifndef ATTN_D
+#define ATTN_D 128
+#endif
+#ifndef THREADS_PER_BLOCK
+#define THREADS_PER_BLOCK 256
+#endif
+
+constexpr int GROUP_SIZE = ATTN_H / ATTN_H_KV;
+constexpr int HALF_D = ATTN_D / 2;
+constexpr int PACKED_D = (GROUP_SIZE + 2) * ATTN_D;
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_BLOCK) void
+fused_qkv_rope_forward(
+    __hip_bfloat16*       __restrict__ q,
+    __hip_bfloat16*       __restrict__ k,
+    __hip_bfloat16*       __restrict__ v,
+    const __hip_bfloat16* __restrict__ xqkv,
+    const __hip_bfloat16* __restrict__ freqs_cis) {
+  const int b = blockIdx.x;
+  const int n = blockIdx.y;
+  const int bn = b * ATTN_N + n;
+  const int packed_bn = bn * ATTN_H_KV * PACKED_D;
+  const int q_bn = bn * ATTN_H * ATTN_D;
+  const int kv_bn = bn * ATTN_H_KV * ATTN_D;
+
+  if (threadIdx.x < HALF_D) {
+    const int pair = threadIdx.x;
+    const int even = pair << 1;
+    const float c = static_cast<float>(freqs_cis[((n * HALF_D + pair) * 2) + 0]);
+    const float s = static_cast<float>(freqs_cis[((n * HALF_D + pair) * 2) + 1]);
+
+    for (int kvh = 0; kvh < ATTN_H_KV; kvh++) {
+      const int base = packed_bn + kvh * PACKED_D;
+
+      for (int rep = 0; rep < GROUP_SIZE; rep++) {
+        const int qbase = base + rep * ATTN_D;
+        const int h = kvh * GROUP_SIZE + rep;
+        const float a = static_cast<float>(xqkv[qbase + even]);
+        const float bb = static_cast<float>(xqkv[qbase + even + 1]);
+        const int out = q_bn + h * ATTN_D + even;
+        q[out] = static_cast<__hip_bfloat16>(a * c - bb * s);
+        q[out + 1] = static_cast<__hip_bfloat16>(a * s + bb * c);
+      }
+
+      const float a = static_cast<float>(xqkv[base + GROUP_SIZE * ATTN_D + even]);
+      const float bb = static_cast<float>(xqkv[base + GROUP_SIZE * ATTN_D + even + 1]);
+      const int out = kv_bn + kvh * ATTN_D + even;
+      k[out] = static_cast<__hip_bfloat16>(a * c - bb * s);
+      k[out + 1] = static_cast<__hip_bfloat16>(a * s + bb * c);
+      v[out] = xqkv[base + (GROUP_SIZE + 1) * ATTN_D + even];
+      v[out + 1] = xqkv[base + (GROUP_SIZE + 1) * ATTN_D + even + 1];
+    }
+  }
+}

extra/thunder/amd/fused_qkv_rope_bwd.cpp

@@ -0,0 +1,161 @@
+#include "kittens.cuh"
+
+using namespace kittens;
+
+#ifndef ATTN_B
+#define ATTN_B 2
+#endif
+#ifndef ATTN_N
+#define ATTN_N 8192
+#endif
+#ifndef ATTN_H
+#define ATTN_H 32
+#endif
+#ifndef ATTN_H_KV
+#define ATTN_H_KV 8
+#endif
+#ifndef ATTN_D
+#define ATTN_D 128
+#endif
+#ifndef THREADS_PER_BLOCK
+#define THREADS_PER_BLOCK 256
+#endif
+constexpr int GROUP_SIZE = ATTN_H / ATTN_H_KV;
+constexpr int HALF_D = ATTN_D / 2;
+constexpr int PACKED_H = ATTN_H_KV * (GROUP_SIZE + 2);
+constexpr int HEADS_PER_WG = ATTN_D == 128 && GROUP_SIZE % 2 == 0 ? 2 : 1;
+constexpr int KV_PARTIALS = GROUP_SIZE / HEADS_PER_WG;
+constexpr int NUM_WARPS = 4;
+constexpr int TILE_N = 16;
+
+template<typename T> using grad_tile = rt<T, TILE_N, ATTN_D, row_l, rt_16x32_s>;
+
+template<int axis, ducks::rt::row_layout RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<RT>>
+__device__ __forceinline__ void load_fa_shuffled(RT &dst, const GL &src, const COORD &idx) {
+  using U = typename GL::dtype;
+  using U2 = base_types::packing<U>::packed_type;
+  U *src_ptr = (U*)&src[(idx.template unit_coord<axis, 3>())];
+  const int row_stride = src.template stride<axis>();
+  const int lane = kittens::laneid();
+  const int tile_row_stride = row_stride * dst.base_tile_rows;
+  const int tile_stride = dst.base_tile_rows * dst.base_tile_cols;
+  const uint32_t buffer_size = src.batch() * src.depth() * src.rows() * src.cols() * sizeof(U);
+  const buffer_resource br = make_buffer_resource(reinterpret_cast<uintptr_t>(src_ptr), buffer_size, 0x00020000);
+
+  #pragma unroll
+  for (int i = 0; i < dst.height; i++) {
+    #pragma unroll
+    for (int j = 0; j < dst.width; j++) {
+      const float4 loaded = std::bit_cast<float4>(llvm_amdgcn_raw_buffer_load_b128(
+        std::bit_cast<i32x4>(br), (i * tile_row_stride + j * tile_stride + lane * 8) * sizeof(U), 0, 0));
+      const U2 *packed = reinterpret_cast<const U2*>(&loaded);
+      #pragma unroll
+      for (int k = 0; k < dst.packed_per_base_tile; k++) dst.tiles[i][j].data[k] = packed[k];
+    }
+  }
+}
+
+template<int axis, ducks::rt::row_layout RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<RT>>
+__device__ __forceinline__ void store_fa_shuffled(const GL &dst, const RT &src, const COORD &idx) {
+  using U = typename GL::dtype;
+  U *dst_ptr = (U*)&dst[(idx.template unit_coord<axis, 3>())];
+  const int row_stride = dst.template stride<axis>();
+  const int lane = kittens::laneid();
+  const int row_offset = (lane % 4) * 4;
+  const int col_offset = ((lane / 32) * 16) + (((lane % 32) / 16) * 2) + (((lane % 16) / 4) * 4);
+  const uint32_t buffer_size = dst.batch() * dst.depth() * dst.rows() * dst.cols() * sizeof(U);
+  const buffer_resource br = make_buffer_resource(reinterpret_cast<uintptr_t>(dst_ptr), buffer_size, 0x00020000);
+
+  #pragma unroll
+  for (int i = 0; i < src.height; i++) {
+    const int row = src.base_tile_rows * i + row_offset;
+    #pragma unroll
+    for (int j = 0; j < src.width; j++) {
+      const int col = src.base_tile_cols * j + col_offset;
+      #pragma unroll
+      for (int k = 0; k < src.packed_per_base_tile; k++) llvm_amdgcn_raw_buffer_store_b32(
+        *reinterpret_cast<const uint32_t*>(&src.tiles[i][j].data[k]), std::bit_cast<i32x4>(br),
+        ((row + k) * row_stride + col) * sizeof(U), 0, 0);
+    }
+  }
+}
+
+template<ducks::rt::row_layout RT>
+__device__ __forceinline__ void inverse_rope_fa(RT &tile, const bf16_2 *freqs, const int n_base) {
+  const int lane = kittens::laneid();
+  const int row_offset = (lane % 4) * 4;
+  const int col_offset = ((lane / 32) * 16) + (((lane % 32) / 16) * 2) + (((lane % 16) / 4) * 4);
+  #pragma unroll
+  for (int i = 0; i < tile.height; i++) {
+    #pragma unroll
+    for (int j = 0; j < tile.width; j++) {
+      const int col = tile.base_tile_cols * j + col_offset;
+      #pragma unroll
+      for (int k = 0; k < tile.packed_per_base_tile; k++) {
+        const int row = tile.base_tile_rows * i + row_offset + k;
+        const float2 cs = __bfloat1622float2(freqs[(n_base + row) * HALF_D + col / 2]);
+        const float2 g = __bfloat1622float2(tile.tiles[i][j].data[k]);
+        tile.tiles[i][j].data[k] = __float22bfloat162_rn(make_float2(g.x * cs.x + g.y * cs.y, -g.x * cs.y + g.y * cs.x));
+      }
+    }
+  }
+}
+
+template<ducks::rt::row_layout RT>
+__device__ __forceinline__ void inverse_rope(RT &tile, const bf16_2 *freqs, const int n_base) {
+  const int lane = kittens::laneid();
+  #pragma unroll
+  for (int i = 0; i < tile.height; i++) {
+    const int row = tile.base_tile_rows * i + lane % tile.base_tile_rows;
+    #pragma unroll
+    for (int j = 0; j < tile.width; j++) {
+      #pragma unroll
+      for (int k = 0; k < tile.packed_per_base_tile; k++) {
+        const int col = tile.base_tile_cols * j + tile.base_tile_stride * (lane / tile.base_tile_rows) + 2 * k;
+        const float2 cs = __bfloat1622float2(freqs[(n_base + row) * HALF_D + col / 2]);
+        const float2 g = __bfloat1622float2(tile.tiles[i][j].data[k]);
+        tile.tiles[i][j].data[k] = __float22bfloat162_rn(make_float2(g.x * cs.x + g.y * cs.y, -g.x * cs.y + g.y * cs.x));
+      }
+    }
+  }
+}
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_BLOCK) void
+fused_qkv_rope_backward(
+    bf16*       __restrict__ dxqkv,
+    const bf16* __restrict__ dq,
+    const bf16* __restrict__ dk,
+    const bf16* __restrict__ dv,
+    const bf16* __restrict__ freqs_cis) {
+  gl<bf16, -1, -1, -1, -1> out{dxqkv, ATTN_B, ATTN_N, PACKED_H, ATTN_D};
+  gl<bf16, -1, -1, -1, -1> dqg{const_cast<bf16*>(dq), ATTN_B, ATTN_H, ATTN_N, ATTN_D};
+  gl<bf16, -1, -1, -1, -1> dkg{const_cast<bf16*>(dk), ATTN_B * KV_PARTIALS, ATTN_N, ATTN_H_KV, ATTN_D};
+  gl<bf16, -1, -1, -1, -1> dvg{const_cast<bf16*>(dv), ATTN_B * KV_PARTIALS, ATTN_N, ATTN_H_KV, ATTN_D};
+  const int b = blockIdx.x, n_tile = blockIdx.y * NUM_WARPS + kittens::warpid(), n_base = n_tile * TILE_N;
+  const int field = blockIdx.z;
+
+  if (field < ATTN_H) {
+    grad_tile<bf16> tile;
+    load_fa_shuffled<2>(tile, dqg, {b, field, n_tile, 0});
+    inverse_rope_fa(tile, reinterpret_cast<const bf16_2*>(freqs_cis), n_base);
+    const int out_head = (field / GROUP_SIZE) * (GROUP_SIZE + 2) + field % GROUP_SIZE;
+    store_fa_shuffled<1>(out, tile, {b, n_tile, out_head, 0});
+  } else {
+    const bool is_k = field < ATTN_H + ATTN_H_KV;
+    const int kvh = field - ATTN_H - (is_k ? 0 : ATTN_H_KV);
+    const auto &src = is_k ? dkg : dvg;
+    grad_tile<bf16> partial, tile;
+    grad_tile<float> partial_f, sum;
+    zero(sum);
+    #pragma unroll
+    for (int p = 0; p < KV_PARTIALS; p++) {
+      load<1>(partial, src, {b * KV_PARTIALS + p, n_tile, kvh, 0});
+      copy(partial_f, partial);
+      add(sum, sum, partial_f);
+    }
+    copy(tile, sum);
+    if (is_k) inverse_rope(tile, reinterpret_cast<const bf16_2*>(freqs_cis), n_base);
+    const int out_head = kvh * (GROUP_SIZE + 2) + GROUP_SIZE + !is_k;
+    store<1>(out, tile, {b, n_tile, out_head, 0});
+  }
+}

extra/thunder/amd/gemm_fp8.cpp

@@ -148,12 +148,28 @@ __global__ __launch_bounds__(512, 2) void hk_fp8_gemm(bf16 *C_ptr, fp8e4m3 *A_pt
     RT_C cC;
     RT_C cD;
 
-    // Calculate which block this threadblock should work on
-    int global_block_id = blockIdx.x;
-
-    // Convert linear block ID to 2D coordinates
-    int block_row = global_block_id / blocks_per_col;
-    int block_col = global_block_id % blocks_per_col;
+    int block_row, block_col;
+    if constexpr (N > M) {
+        // Wide outputs repeatedly consume the same A rows. Keep a short strip
+        // resident on each XCD while walking N to improve local cache reuse.
+        int wgid = chiplet_transform_chunked(int(blockIdx.x), total_blocks_needed, NUM_XCDS, 64);
+        constexpr int WGM = 3;
+        const int num_wgid_in_group = WGM * blocks_per_col;
+        const int group_id = wgid / num_wgid_in_group;
+        const int first_block_row = group_id * WGM;
+        const int group_size_m = min(blocks_per_row - first_block_row, WGM);
+        block_row = first_block_row + ((wgid % num_wgid_in_group) % group_size_m);
+        block_col = (wgid % num_wgid_in_group) / group_size_m;
+    } else {
+        int wgid = chiplet_transform_chunked(int(blockIdx.x), total_blocks_needed, NUM_XCDS, 64);
+        constexpr int WGM = 8;
+        const int num_wgid_in_group = WGM * blocks_per_col;
+        const int group_id = wgid / num_wgid_in_group;
+        const int first_block_row = group_id * WGM;
+        const int group_size_m = min(blocks_per_row - first_block_row, WGM);
+        block_row = first_block_row + ((wgid % num_wgid_in_group) % group_size_m);
+        block_col = (wgid % num_wgid_in_group) / group_size_m;
+    }
     int block_m = block_row * BLOCK_SIZE_ROW;
     int block_n = block_col * BLOCK_SIZE_COL;
 

extra/thunder/amd/gemm_fp8_atb.cpp

@@ -134,7 +134,7 @@ __global__ __launch_bounds__(512, 2) void hk_fp8_atb_gemm(bf16 *C_ptr, fp8e4m3 *
 
     int wgid = blockIdx.x;
     const int WGM = 8;
-    wgid = chiplet_transform_chunked(wgid, total_blocks_needed, NUM_XCDS, 64);
+    wgid = chiplet_transform_chunked(wgid, total_blocks_needed, NUM_XCDS, 32);
 
     const int num_wgid_in_group = WGM * blocks_per_col;
     int group_id = wgid / num_wgid_in_group;

extra/thunder/tiny/tk/group.py

@@ -1,7 +1,6 @@
 import math
 from typing import cast, Callable
-from tinygrad import dtypes
-from tinygrad.uop.ops import AxisType, UOp, Ops
+from tinygrad.uop.ops import AxisType, UOp
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import prod
 
@@ -75,9 +74,9 @@ class Group:
 
     a_base_shape = cast(RT, a).base_shape
     if a_base_shape.cols == 16:
-      wmma_arg = ('WMMA_16_16_16___bf16_float', (16, 16, 16), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2)), ((4, 2), (3, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 16)
     elif a_base_shape.cols == 32:
-      wmma_arg = ('WMMA_16_16_32___bf16_float', (16, 16, 32), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 32)
     else: raise NotImplementedError(f"mma_AB not implemented for {a_base_shape.cols=}")
 
     for height in self.ker.range(c.shape[-3], track=False):
@@ -92,7 +91,7 @@ class Group:
           else: raise NotImplementedError(f"mma_AB not implemented for {a_base_shape.cols=}")
           d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
-          out = UOp(Ops.WMMA, dtypes.float32, (a_in, b_in, d_in), arg=wmma_arg)
+          out = UOp.wmma(a_in, b_in, d_in, wmma_dims, 'AMD', 64)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
           c_store = UOp.group(*c_i).end(height, width, inner)
 
@@ -105,9 +104,9 @@ class Group:
 
     a_base_shape = cast(RT, a).base_shape
     if a_base_shape.cols == 16:
-      wmma_arg = ('WMMA_16_16_16___bf16_float', (16, 16, 16), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2)), ((4, 2), (3, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 16)
     elif a_base_shape.cols == 32:
-      wmma_arg = ('WMMA_16_16_32___bf16_float', (16, 16, 32), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 32)
     else: raise NotImplementedError(f"mma_ABt not implemented for {a_base_shape.cols=}")
 
     for height in self.ker.range(c.shape[-3], track=False):
@@ -122,7 +121,7 @@ class Group:
           else: raise NotImplementedError(f"mma_ABt not implemented for {a_base_shape.cols=}")
           d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
-          out = UOp(Ops.WMMA, dtypes.float32, (a_in, b_in, d_in), arg=wmma_arg)
+          out = UOp.wmma(a_in, b_in, d_in, wmma_dims, 'AMD', 64)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
           c_store = UOp.group(*c_i).end(height, width, inner)
 
@@ -135,9 +134,9 @@ class Group:
 
     a_base_shape = cast(RT, a).base_shape
     if a_base_shape.cols == 16:
-      wmma_arg = ('WMMA_16_16_16___bf16_float', (16, 16, 16), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2)), ((4, 2), (3, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 16)
     elif a_base_shape.cols == 32:
-      wmma_arg = ('WMMA_16_16_32___bf16_float', (16, 16, 32), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 32)
     else: raise NotImplementedError(f"mma_AtB not implemented for {a_base_shape.cols=}")
 
     for height in self.ker.range(c.shape[-3], track=False):
@@ -152,7 +151,7 @@ class Group:
           else: raise NotImplementedError(f"mma_AtB not implemented for {a_base_shape.cols=}")
           d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
-          out = UOp(Ops.WMMA, dtypes.float32, (a_in, b_in, d_in), arg=wmma_arg)
+          out = UOp.wmma(a_in, b_in, d_in, wmma_dims, 'AMD', 64)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
           c_store = UOp.group(*c_i).end(height, width, inner)
 
@@ -165,9 +164,9 @@ class Group:
 
     a_base_shape = cast(RT, a).base_shape
     if a_base_shape.cols == 16:
-      wmma_arg = ('WMMA_16_16_16___bf16_float', (16, 16, 16), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2)), ((4, 2), (3, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 16)
     elif a_base_shape.cols == 32:
-      wmma_arg = ('WMMA_16_16_32___bf16_float', (16, 16, 32), dtypes.bfloat16, dtypes.float, 'AMD', 64, (((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2), (9, 2)), ((4, 2), (3, 2))), ()) # type: ignore
+      wmma_dims = (16, 16, 32)
     else: raise NotImplementedError(f"mma_AtBt not implemented for {a_base_shape.cols=}")
 
     for height in self.ker.range(c.shape[-3], track=False):
@@ -182,7 +181,7 @@ class Group:
           else: raise NotImplementedError(f"mma_AtBt not implemented for {a_base_shape.cols=}")
           d_in = UOp.stack(*[c[height, width, i] for i in range(4)])
 
-          out = UOp(Ops.WMMA, dtypes.float32, (a_in, b_in, d_in), arg=wmma_arg)
+          out = UOp.wmma(a_in, b_in, d_in, wmma_dims, 'AMD', 64)
           c_i = [c[height, width, i].store(out.index(i)) for i in range(4)]
           c_store = UOp.group(*c_i).end(height, width, inner)
 

extra/torch_backend/backend.py

@@ -565,8 +565,10 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
   "aten.floor_divide": lambda x,y: x//y,
   "aten.floor_divide_.Tensor": lambda x,y: x//y,
   "aten.__lshift__.Scalar": lambda x,y: x<<y,
+  "aten.__lshift__.Tensor": lambda x,y: x<<y,
   "aten.__ilshift__.Scalar": lambda x,y: x<<y,
   "aten.__rshift__.Scalar": lambda x,y: x>>y,
+  "aten.__rshift__.Tensor": lambda x,y: x>>y,
   "aten.__irshift__.Scalar": lambda x,y: x>>y,
   # inplace ops using replace for fusion
   "aten.zero_": lambda x: x.const_like(0),

pyproject.toml

@@ -74,7 +74,6 @@ testing_minimal = [
   "torch==2.9.1",
   "pytest",
   "pytest-xdist",
-  "pytest-timeout",
   "pytest-split",
   "hypothesis>=6.148.9",
   "z3-solver<4.15.4",  # 4.15.4 has a segfault when creating many z3.Context()
@@ -160,9 +159,6 @@ norecursedirs = [
   ".hypothesis",
   ".git",
 ]
-timeout = 300
-timeout_method = "thread"
-timeout_func_only = true
 testpaths = ["test"]
 filterwarnings = [
   # Ignore SWIG warnings from importlib

test/amd/test_asm_kernel.py

@@ -36,7 +36,7 @@ def custom_add_var(A:UOp, B:UOp) -> UOp:
   A,B = A.flatten(), B.flatten()
   assert A.dtype == dtypes.uint32, f"buffer dtype must be uint32, got {A.dtype}"
   threads = UOp.special(A.numel(), "lidx0")
-  var = UOp.param(2, dtypes.index, vmin_vmax=(0, 10), name="var", addrspace=AddrSpace.ALU)
+  var = UOp.param(2, dtypes.weakint, vmin_vmax=(0, 10), name="var", addrspace=AddrSpace.ALU)
   insts = [
     s_load_b128(s[4:7], s[0:1]),
     s_load_b32(s[8], s[0:1], offset=0x10), # all threads load the same variable

test/amd/test_roundtrip.py

@@ -121,7 +121,8 @@ class TestTinygradKernelRoundtrip(unittest.TestCase):
     arch = self.arch
 
     from tinygrad.runtime.support.elf import elf_loader
-    from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler
+    from tinygrad.runtime.support.compiler_amd import HIPCompiler
+    from tinygrad.runtime.support.compiler_llvm import AMDLLVMCompiler
     from tinygrad.helpers import DEV
 
     kernels, _, _ = get_kernels_from_tinygrad(op_fn)

test/backend/test_call.py

@@ -0,0 +1,39 @@
+import unittest, ctypes
+from tinygrad import Tensor, UOp
+from tinygrad.device import Device
+from tinygrad.dtype import dtypes
+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)
+  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)
+  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")
+class TestCall(unittest.TestCase):
+  def test_call_out_param(self):
+    called = []
+    @ctypes.CFUNCTYPE(None, ctypes.c_int, ctypes.POINTER(ctypes.c_int))
+    def fxn(n, out):
+      called.append(n)
+      out[0] = n * 2
+    f = Tensor([ctypes.cast(fxn, ctypes.c_void_p).value], dtype=dtypes.uint64, device="CPU")
+    c = Tensor.empty(2, dtype=dtypes.int, device="CPU")
+    c = Tensor.custom_kernel(f, c, fxn=call_out_kernel)[1]
+    self.assertEqual(c.tolist(), [6, 7])
+    self.assertEqual(called, [3])
+
+  def test_call_ret(self):
+    @ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_int)
+    def fxn(n): return n + 1
+    f = Tensor([ctypes.cast(fxn, ctypes.c_void_p).value], dtype=dtypes.uint64, device="CPU")
+    c = Tensor.empty(1, dtype=dtypes.int, device="CPU")
+    c = Tensor.custom_kernel(f, c, fxn=call_ret_kernel)[1]
+    c.realize()
+    self.assertEqual(c.item(), 44)
+
+if __name__ == "__main__": unittest.main()

test/backend/test_custom_kernel.py

@@ -199,8 +199,7 @@ class TestCustomKernel(unittest.TestCase):
     c = Tensor.empty(N, N)
 
     tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm)[0]
-    err = (tst - (a@b)).square().max()
-    self.assertLess(err.item(), 1e-6)
+    self.assertTrue(tst.allclose(a@b, atol=1e-3).item())
 
   def test_gemm_multi(self):
     devs = ("CPU:0", "CPU:1")
@@ -209,8 +208,7 @@ class TestCustomKernel(unittest.TestCase):
     b = Tensor.randn(N, N).to(devs)
     c = Tensor(Tensor.empty(N//2, N, device=devs).uop.multi(0), device=devs)
     tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm)[0]
-    err = (tst - (a@b)).square().max()
-    self.assertLess(err.item(), 1e-6)
+    self.assertTrue(tst.allclose(a@b, atol=1e-3).item())
 
   def test_gemm_backward_custom(self): self.test_gemm_backward(True)
   # NOTE: grad_fxn doesn't work with pyrender
@@ -233,14 +231,9 @@ class TestCustomKernel(unittest.TestCase):
     real_grad_a, real_grad_b = a.grad, b.grad
     Tensor.realize(ref, real_grad_a, real_grad_b)
 
-    err = (tst - ref).square().max()
-    self.assertLess(err.item(), 1e-6)
-
-    err = (grad_a - real_grad_a).square().max()
-    self.assertLess(err.item(), 1e-6)
-
-    err = (grad_b - real_grad_b).square().max()
-    self.assertLess(err.item(), 1e-6)
+    self.assertTrue(tst.allclose(ref, atol=1e-3).item())
+    self.assertTrue(grad_a.allclose(real_grad_a, atol=1e-3).item())
+    self.assertTrue(grad_b.allclose(real_grad_b, atol=1e-3).item())
 
   def test_simple_qkv(self):
     N, d = 8, 4
@@ -253,8 +246,7 @@ class TestCustomKernel(unittest.TestCase):
     O_ref = ((Q @ K.T) / (d ** 0.5)) @ V
 
     Tensor.realize(O_custom, O_ref)
-    err = (O_custom - O_ref).square().max()
-    self.assertLess(err.item(), 1e-6)
+    self.assertTrue(O_custom.allclose(O_ref, atol=1e-3).item())
 
   def test_gemm_qkv(self):
     B, N, K_DIM, H_KV, REP, D = 2, 7, 6, 2, 2, 6

test/backend/test_dtype.py

@@ -8,7 +8,7 @@ from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
 from tinygrad import Context, Device, Tensor, dtypes
 from hypothesis import given, settings, strategies as strat
-from test.helpers import rand_for_dtype
+from test.helpers import rand_for_dtype, min_normal
 from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX, FP8E4M3FNUZ_MAX, FP8E5M2FNUZ_MAX
 import pytest
 pytestmark = pytest.mark.filterwarnings("ignore")
@@ -25,10 +25,10 @@ def get_available_cast_dtypes(dtype: DType) -> List[DType]:
   if dtype not in supported_dtypes and dtype not in dtypes.fp8s+(dtypes.half,dtypes.bfloat16): return []
   return dts
 
-def _to_torch_storage_type(dtype:DType):
-  if dtype == dtypes.bfloat16: return torch.float32
-  if dtype in dtypes.fp8s: return torch.float32
-  return _to_torch_dtype(dtype)
+def _to_torch_storage(a:Tensor) -> torch.Tensor:
+  # tolist() of an fp8 Tensor gives floats, so convert and store in uint8
+  if a.dtype in dtypes.fp8s: return torch.tensor([float_to_fp8(x, a.dtype) for x in a.flatten().tolist()], dtype=torch.uint8).reshape(a.shape)
+  return torch.tensor(a.tolist(), dtype=_to_torch_dtype(a.dtype))
 
 def _test_to_np(a:Tensor, np_dtype, target):
   if DEBUG >= 2: print(a)
@@ -46,12 +46,15 @@ def _test_cast(a:Tensor, target_dtype:DType):
   if a.is_floating_point() and dtypes.is_unsigned(target_dtype):
     # converting negative float to unsigned integer is undefined
     a = a.abs()
+  if a.is_floating_point() and dtypes.is_float(target_dtype) and (mn:=min_normal(target_dtype)) >= min_normal(a.dtype):
+    # subnormals are zero, so an input below the target's min normal casts to 0
+    a = (a.abs() < mn).where(0, a)
 
   expected = list(a.numpy().astype(_to_np_dtype(target_dtype)))
   if target_dtype in dtypes.fp8s: expected = [truncate[target_dtype](x) for x in expected]
   _test_op(lambda: a.cast(target_dtype), target_dtype, expected)
 def _test_bitcast(a:Tensor, target_dtype:DType, target=None):
-  expected = torch.tensor(a.tolist(), dtype=_to_torch_storage_type(a.dtype)).view(_to_torch_dtype(target_dtype)).tolist()
+  expected = _to_torch_storage(a).view(_to_torch_dtype(target_dtype)).tolist()
   if target_dtype in dtypes.fp8s: expected = [fp8_to_float(x, target_dtype) for x in expected]
   _test_op(lambda: a.bitcast(target_dtype), target_dtype, target or expected)
 
@@ -61,10 +64,12 @@ class TestDType(unittest.TestCase):
   @classmethod
   def setUpClass(cls):
     if cls.DTYPE is None: raise unittest.SkipTest("base class")
-    cls.DATA = rand_for_dtype(cls.DTYPE, 0x10, allow_subnormal=cls.DTYPE in supported_dtypes)
+    cls.DATA = rand_for_dtype(cls.DTYPE, 0x10, allow_subnormal=cls.DTYPE in supported_dtypes and cls.DTYPE not in dtypes.fp8s)
 
   def test_to_np(self):
-    _test_to_np(Tensor(self.DATA, dtype=self.DTYPE), _to_np_dtype(self.DTYPE), np.array(self.DATA, dtype=_to_np_dtype(self.DTYPE)))
+    a = Tensor(self.DATA, dtype=self.DTYPE)
+    self.assertEqual(a.dtype, self.DTYPE)
+    _test_to_np(a, _to_np_dtype(self.DTYPE), np.array(self.DATA, dtype=_to_np_dtype(self.DTYPE)))
 
   def test_casts_to(self):
     for dtype in get_available_cast_dtypes(self.DTYPE):
@@ -273,10 +278,11 @@ class TestBitCast(unittest.TestCase):
   @given(strat.sampled_from(dtype_ints + dtype_floats), strat.sampled_from(dtype_ints + dtype_floats))
   def test_shape_change_bitcast(self, dt1, dt2):
     data = rand_for_dtype(dt1, 32).reshape(2, 2, 8)
-    expected = torch.tensor(data.tolist(), dtype=_to_torch_storage_type(dt1)).view(_to_torch_dtype(dt2))
+    a = Tensor(data, dtype=dt1)
+    expected = _to_torch_storage(a).view(_to_torch_dtype(dt2))
     if dt2 in dtypes.fp8s:
       expected = torch.tensor([fp8_to_float(x, dt2) for x in expected.view(-1).tolist()]).view_as(expected)
-    _test_op(lambda: Tensor(data, dtype=dt1).bitcast(dt2), dt2, expected.tolist())
+    _test_op(lambda: a.bitcast(dt2), dt2, expected.tolist())
 
   def test_shape_change_bitcast_exceptions(self):
     with self.assertRaises(RuntimeError):
@@ -293,6 +299,11 @@ class TestBitCast(unittest.TestCase):
     b = a.bitcast(dtypes.float32)
     assert b.numpy()[0,0] == 1.
 
+  def test_bitcast_bf16_from_cast(self):
+    # a bfloat16 from a cast holds bfloat16 bits. 1.0 is 0x3f80 in bfloat16, which is 1.875 in half
+    a = Tensor([1.0], dtype=dtypes.float32).cast(dtypes.bfloat16)
+    assert a.bitcast(dtypes.half).numpy()[0] == 1.875
+
 class TestInt16DType(TestDType): DTYPE = dtypes.int16
 
 class TestUint16DType(TestDType):

test/backend/test_dtype_alu.py

@@ -291,6 +291,33 @@ class TestDTypeALU(unittest.TestCase):
   @Context(EMULATED_DTYPES="long")
   def test_emulated_int64(self, a, b, op): universal_test(a, b, dtypes.int64, op)
 
+  def _test_shl(self):
+    for dtype, values, distances in ((dtypes.int64, [-0x1234, 0x80000001, -1, 0x1234, 1], [0, 5, 31, 32, 62]),
+                                     (dtypes.uint64, [0x80000001, 0x80000001, 1, 0xFEDC, 1], [0, 5, 31, 32, 62]),
+                                     (dtypes.int8, [-3, 1, 7, -2, 1], [0, 1, 3, 5, 6]),
+                                     (dtypes.uint16, [3, 1, 0xFF, 7, 1], [0, 1, 7, 12, 15])):
+      with self.subTest(dtype=dtype):
+        result = Tensor(values, dtype=dtype) << Tensor(distances, dtype=dtype)
+        np.testing.assert_equal(result.numpy(), [x << d for x, d in zip(values, distances)])
+
+  def _test_shr(self):
+    for dtype, values, distances in ((dtypes.int64, [-(2**40), -1, -(2**50), -(2**40), 0x123456789ABCDEF], [0, 5, 31, 32, 63]),
+                                     (dtypes.uint64, [0xFEDCBA9876543210] * 5, [0, 5, 31, 32, 63]),
+                                     (dtypes.int8, [-128, -1, 64, -37, 1], [0, 1, 3, 5, 7]),
+                                     (dtypes.uint16, [0xFFFF] * 5, [0, 1, 8, 13, 15])):
+      with self.subTest(dtype=dtype):
+        result = Tensor(values, dtype=dtype) >> Tensor(distances, dtype=dtype)
+        np.testing.assert_equal(result.numpy(), [x >> d for x, d in zip(values, distances)])
+
+  def test_shl(self): self._test_shl()
+  def test_shr(self): self._test_shr()
+
+  @Context(EMULATED_DTYPES="long")
+  def test_emulated_shl(self): self._test_shl()
+
+  @Context(EMULATED_DTYPES="long")
+  def test_emulated_shr(self): self._test_shr()
+
   @given(ht.uint8, strat.sampled_from(integer_unary_operations))
   def test_uint8_unary(self, a, op): universal_test_unary(a, dtypes.uint8, op)
 

test/backend/test_graph.py

@@ -1,9 +1,9 @@
 import numpy as np
-import functools, unittest, ctypes
+import functools, unittest
 
 from tinygrad.device import Device, Buffer
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import Context, from_mv
+from tinygrad.helpers import Context
 from tinygrad.dtype import dtypes
 from tinygrad.engine.jit import MultiGraphRunner
 from tinygrad.engine.realize import run_linear, compile_linear
@@ -31,7 +31,7 @@ def make_buffer(device, size=BUF_SIZE, fill=False):
   buf = Buffer(device, size, dtypes.int).ensure_allocated()
   if fill:
     with Context(DEBUG=0):
-      buf.copyin(Tensor(np.random.randint(-10000, 10000, size=size, dtype=np.int32)).realize().uop.base.realized.as_memoryview())
+      buf.copy_from(Tensor(np.random.randint(-10000, 10000, size=size, dtype=np.int32)).realize().uop.base.realized)
   return buf
 
 def make_view(base, offset_elems, size_elems):
@@ -55,15 +55,12 @@ def run_schedule(calls:list[UOp]):
   run_linear(UOp(Ops.LINEAR, src=tuple(calls)))
 
 def zero_bufs(bufs):
-  for b in bufs:
-    mv = memoryview(bytearray(b.nbytes))
-    ctypes.memset(from_mv(mv), 0, len(mv))
-    b.copyin(mv)
+  for b in bufs: b.copy_from(Buffer("PYTHON", b.size, b.dtype, opaque=memoryview(bytearray(b.nbytes))))
 
 @unittest.skipUnless(Device[Device.DEFAULT].graph is not None, "graph support required")
 class TestGraph(unittest.TestCase):
   def skip_if_no_offset(self):
-    if not hasattr(Device[Device.DEFAULT].allocator, "_offset"): self.skipTest("device does not support _offset")
+    if Device.DEFAULT in {"WEBGPU", "CL"}: self.skipTest("device does not support _offset")
 
   def skip_if_not_multigraph(self):
     graph = g.func if isinstance(g:=(d:=Device[Device.DEFAULT]).graph, functools.partial) else g
@@ -213,8 +210,8 @@ class TestGraph(unittest.TestCase):
 
   def test_graph_offset_bufs(self):
     self.skip_if_not_multigraph()
+    self.skip_if_no_offset()
     d0 = Device.DEFAULT
-    if not hasattr(Device[d0].allocator, "_offset"): self.skipTest("device does not support _offset")
 
     b0 = make_buffer(d0, fill=True)
     b1 = make_view(b0, 0, b0.size)

test/backend/test_linearizer.py

@@ -59,6 +59,7 @@ class TestLinearizer(unittest.TestCase):
     self.assertEqual(len([x for x in uops if x.op is Ops.CAST]), 0)
 
   @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken on ptx")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_ranges, "test inspects ranges, which are rewritten to loops on this renderer")
   def test_late_bias_load(self):
     img = Tensor.empty(1, 3, 16, 16)
     w = Tensor.empty(16, 3, 3, 3)
@@ -238,6 +239,7 @@ class TestLinearizer(unittest.TestCase):
         helper_arg_acc_dtype(d.conv2d(w, dtype=acc_dtype), expected_dtype)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_ranges, "test inspects ranges, which are rewritten to loops on this renderer")
   def test_simple_unroll_no_between_phi_dependencies(self):
     x, y = Tensor.empty(64, 64), Tensor.empty(64, 64)
     r = (x@y).relu()
@@ -290,7 +292,7 @@ class TestLinearizer(unittest.TestCase):
   @unittest.skipIf(MOCKGPU and isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, CUDARenderer)), "PTX indexes differently. might be ok?")
   def test_where_fold(self):
     a = Tensor.ones(4, 4).contiguous().realize()
-    b = a.shrink(((1, 2), None)).pad(((1, 2), None))
+    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
@@ -299,6 +301,7 @@ class TestLinearizer(unittest.TestCase):
     program = to_program(replace_opts(linear.src[-1].src[0], []), renderer=Device[Device.DEFAULT].renderer)
     assert not any(u.op == Ops.WHERE for u in tuple(program.src[1].src)), "found where where where should be folded"
 
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_ranges, "test inspects ranges, which are rewritten to loops on this renderer")
   def test_phi_simplification(self):
     def helper(t, max_ops=0):
       ast = helper_linearizer_opt(t)
@@ -424,7 +427,7 @@ def copyout_outputs(outbufs:list[Buffer]) -> list[np.ndarray]:
   return [np.frombuffer(x.as_memoryview(), _to_np_dtype(x.dtype)) for x in outbufs]
 
 def reset_bufs(bufs:list[Buffer]):
-  for buf in bufs: buf.copyin(np.zeros((buf.size*buf.dtype.itemsize,), dtype=np.uint8).data)
+  for buf in bufs: buf.copy_from(Buffer("PYTHON", buf.size, buf.dtype, opaque=memoryview(bytearray(buf.nbytes))))
 
 def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[],
                                apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[]):

test/backend/test_linearizer_dumb.py

@@ -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.index, 512), 0, AxisType.GLOBAL)
-    c2 = UOp.range(UOp.const(dtypes.index, 784), 1, AxisType.GLOBAL)
-    c3 = UOp.range(UOp.const(dtypes.index, 10), 3, AxisType.GLOBAL)
+    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)
     c4 = UOp.param(1, dtypes.int, (512,))
     c5 = c4.index(c1.valid(UOp.const(dtypes.bool, True)))
-    c6 = UOp.range(UOp.const(dtypes.index, 6000), 1004, AxisType.REDUCE)
-    c7 = UOp.range(UOp.const(dtypes.index, 3750), 2006, AxisType.REDUCE)
-    c8 = UOp.range(UOp.const(dtypes.index, 16), 2007, AxisType.GROUP_REDUCE)
+    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)
     c9 = UOp.param(2, dtypes.uchar, (47040000,))
-    c10 = c9.index((((c3*UOp.const(dtypes.index, 4704000))+c2)+(c6*UOp.const(dtypes.index, 784))).valid(UOp.const(dtypes.bool, True)))
-    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.index, 6000))+c6)+((c7*UOp.const(dtypes.index, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.index, 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.index, 7840))+(c2*UOp.const(dtypes.index, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
+    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)
     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)
 

test/backend/test_llama_kernels.py

@@ -1,11 +1,14 @@
-import unittest
+import unittest, functools
 from tinygrad import Tensor, Device, dtypes, Context, GlobalCounters
 from tinygrad.helpers import getenv
 from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8
 from extra.llama_kernels.fused_ce import fused_ce_loss
 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.backend.test_asm_gemm import has_hipcc
 
 def run_fused_ce(bs:int, seqlen:int, vocab:int, label_smoothing:float=0.0) -> None:
   Tensor.manual_seed(0)
@@ -46,9 +49,10 @@ def run_quantize_fp8(shape:tuple[int, ...], delayed:bool=True) -> None:
   with Context(DEBUG=0): Tensor.realize(x, amax_state)
 
   if delayed:
-    fp8, inv_scale, new_amax, _ = quantize_fp8_delayed(x, amax_state, FP8_DTYPE)
+    amax_out = Tensor.zeros((), dtype=dtypes.float32, device=x.device).realize()
+    fp8, inv_scale = quantize_fp8_delayed(x, amax_state, amax_out, FP8_DTYPE)
     ref_fp8, ref_inv_scale, ref_new_amax = quantize_fp8(x, amax_state=amax_state)
-    Tensor.realize(fp8, inv_scale, new_amax)
+    Tensor.realize(fp8, inv_scale)
     Tensor.realize(ref_fp8, ref_inv_scale, ref_new_amax)
   else:
     fp8 = quantize_fp8_scalar(x, amax_state, FP8_DTYPE)
@@ -60,9 +64,10 @@ def run_quantize_fp8(shape:tuple[int, ...], delayed:bool=True) -> None:
     assert fp8.cast(dtypes.float).allclose(ref_fp8.cast(dtypes.float), atol=0, rtol=0).item(), "fp8 mismatch"
     if delayed:
       assert inv_scale.allclose(ref_inv_scale, atol=0, rtol=0).item(), "inv_scale mismatch"
-      assert new_amax.allclose(ref_new_amax, atol=0, rtol=0).item(), \
-        f"amax mismatch: got={new_amax.item()} ref={ref_new_amax.item()} diff={abs(new_amax.item()-ref_new_amax.item())}"
+      assert amax_out.allclose(ref_new_amax, atol=0, rtol=0).item(), \
+        f"amax mismatch: got={amax_out.item()} ref={ref_new_amax.item()} diff={abs(amax_out.item()-ref_new_amax.item())}"
 
+@unittest.skipUnless(Device.DEFAULT == "AMD", "requires atomic max")
 class TestQuantizeFP8(unittest.TestCase):
   def setUp(self):
     ren = Device[Device.DEFAULT].renderer
@@ -78,10 +83,11 @@ class TestQuantizeFP8(unittest.TestCase):
     x = Tensor.empty(2048*8, 1024, dtype=dtypes.bfloat16, device=devs).uop.multi(0)
     x = Tensor(x, device=devs)
     amax_state = Tensor.full((), 2.0, dtype=dtypes.float32, device=devs).contiguous()
-    fp8, _, new_amax, _ = quantize_fp8_delayed(x, amax_state, FP8_DTYPE)
-    Tensor.realize(fp8, new_amax)
+    amax_out = Tensor.zeros((), dtype=dtypes.float32, device=devs).realize()
+    fp8, _ = quantize_fp8_delayed(x, amax_state, amax_out, FP8_DTYPE)
+    Tensor.realize(fp8)
     assert fp8.uop.shape == x.uop.shape
-    assert new_amax.shape == ()
+    assert amax_out.shape == ()
 
 class TestLocalAmax(unittest.TestCase):
   def test_multi_tensor_local_shard_amax(self):
@@ -92,5 +98,68 @@ class TestLocalAmax(unittest.TestCase):
     self.assertEqual(GlobalCounters.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")
+class TestFusedQKVRoPE(unittest.TestCase):
+  SHAPE = (2, 8192, 32, 8, 128)
+
+  def rand_bf16(self, *shape:int) -> Tensor:
+    return (Tensor.randn(*shape) * 0.1).cast(dtypes.bfloat16).contiguous().realize()
+
+  def freqs_cis(self) -> Tensor:
+    _, N, _, _, D = self.SHAPE
+    return precompute_freqs_cis(D, N * 2).cast(dtypes.bfloat16).clone().realize()
+
+  def test_llama31_8b_forward(self):
+    Tensor.manual_seed(0)
+    B, N, H, H_KV, D = self.SHAPE
+    GROUP = H // H_KV
+    freqs_cis = self.freqs_cis()
+
+    x = self.rand_bf16(B, N, H_KV * (GROUP + 2) * D)
+    q, k, v = fused_qkv_rope(x, freqs_cis, H, H_KV, D)
+    Tensor.realize(q, k, v)
+    packed_ref = x.reshape(B, N, H_KV, GROUP + 2, D)
+    q_ref = packed_ref[:, :, :, :GROUP].reshape(B, N, H, D)
+    k_ref, v_ref = packed_ref[:, :, :, GROUP], packed_ref[:, :, :, GROUP+1]
+    q_ref, k_ref = apply_rotary_emb(q_ref, k_ref, freqs_cis[:, :N])
+    q_ref, k_ref, v_ref = q_ref.cast(dtypes.bfloat16), k_ref.cast(dtypes.bfloat16), v_ref.cast(dtypes.bfloat16)
+    Tensor.realize(q_ref, k_ref, v_ref)
+
+    with Context(DEBUG=0):
+      self.assertTrue(q.allclose(q_ref, atol=2e-2, rtol=0).item(), "Q forward mismatch")
+      self.assertTrue(k.allclose(k_ref, atol=2e-2, rtol=0).item(), "K forward mismatch")
+      self.assertTrue(v.allclose(v_ref, atol=0, rtol=0).item(), "V forward mismatch")
+
+  def test_llama31_8b_backward(self):
+    Tensor.manual_seed(1)
+    B, N, H, H_KV, D = self.SHAPE
+    PARTIALS = 2
+    GROUP = H // H_KV
+    freqs_cis = self.freqs_cis()
+    dq = self.rand_bf16(B, N, H, D)
+    dk_partial = self.rand_bf16(B * PARTIALS, N, H_KV, D)
+    dv_partial = self.rand_bf16(B * PARTIALS, N, H_KV, D)
+
+    # Invert Flash Attention's dQ layout transform to reproduce its native buffer.
+    dq_native = dq.transpose(1, 2).reshape(B, H, N//16, 4, 4, 4, 2, D//32, 2, 2) \
+      .permute(0, 1, 2, 5, 6, 8, 7, 3, 4, 9).reshape(B, H, N, D).contiguous().realize()
+    dx = Tensor.empty(B, N, H_KV * (GROUP + 2) * D, dtype=dtypes.bfloat16)
+    arch = Device[Device.DEFAULT].renderer.target.arch
+    fxn = functools.partial(custom_fused_qkv_rope_backward, device=Device.DEFAULT, arch=arch,
+                            B=B, N=N, H=H, H_KV=H_KV, D=D)
+    dx = Tensor.custom_kernel(dx, dq_native, dk_partial, dv_partial, freqs_cis, fxn=fxn)[0].realize()
+
+    def inverse_rope(x:Tensor) -> Tensor:
+      x = x.reshape(*x.shape[:-1], D//2, 2).float()
+      cs = freqs_cis[:, :N].float()
+      return Tensor.stack(x[..., 0] * cs[..., 0] + x[..., 1] * cs[..., 1],
+                          -x[..., 0] * cs[..., 1] + x[..., 1] * cs[..., 0], dim=-1).flatten(-2).cast(dtypes.bfloat16)
+
+    dq_ref = inverse_rope(dq).reshape(B, N, H_KV, GROUP, D)
+    dk_ref = inverse_rope(dk_partial.float().reshape(B, PARTIALS, N, H_KV, D).sum(1).cast(dtypes.bfloat16)).unsqueeze(3)
+    dv_ref = dv_partial.float().reshape(B, PARTIALS, N, H_KV, D).sum(1).cast(dtypes.bfloat16).unsqueeze(3)
+    ref = Tensor.cat(dq_ref, dk_ref, dv_ref, dim=3).reshape(*dx.shape).realize()
+    with Context(DEBUG=0): self.assertTrue(dx.allclose(ref, atol=2e-2, rtol=2e-2).item(), "backward mismatch")
+
 if __name__ == '__main__':
   unittest.main()

test/backend/test_multitensor.py

@@ -385,7 +385,7 @@ class TestMultiBufferView(unittest.TestCase):
     b_ref = view_fn(a_ref)
     b_multi = view_fn(a_multi).contiguous()
     linear, var_vals = b_multi.linear_with_vars()
-    if all(hasattr(Device[d].allocator, "_offset") for d in b_multi.device):
+    if all(not d.startswith(("WEBGPU", "CL")) for d in b_multi.device):
       compiled = [call for call in linear.src if call.src[0].op is Ops.SINK]
       self.assertEqual(len(compiled), 0, f"expected zero compiled kernels, got {len(compiled)}")
     run_linear(linear, var_vals)
@@ -417,7 +417,7 @@ class TestMultiBufferView(unittest.TestCase):
     a = Tensor.arange(8*12).reshape(8, 12).clone().shard(devices_4, axis=1).realize()
     out = a[5].contiguous()
     linear, var_vals = out.linear_with_vars()
-    if all(hasattr(Device[d].allocator, "_offset") for d in out.device):
+    if all(not d.startswith(("WEBGPU", "CL")) for d in out.device):
       compiled = [call for call in linear.src if call.src[0].op is Ops.SINK]
       self.assertEqual(len(compiled), 0)
     run_linear(linear, var_vals)
@@ -450,7 +450,7 @@ class TestMultiTransformer(unittest.TestCase):
       else: v.shard_(device, axis=None)
 
     last_tok = 0
-    for i in range(10):
+    for i in range(5):
       real_tok = real_model(Tensor([[last_tok]], device=Device.DEFAULT), i).item()
       shard_tok = shard_model(Tensor([[last_tok]], device=device), i).item()
 

test/backend/test_ops.py

@@ -5,7 +5,6 @@ import torch
 from tinygrad.helpers import getenv, DEBUG, DEV, IMAGE, Context
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.renderer.cstyle import QCOMCLRenderer
 from tinygrad.renderer.nir import NIRRenderer
 
 TINY_BACKEND = getenv("TINY_BACKEND")
@@ -244,7 +243,6 @@ class TestOps(unittest.TestCase):
 
     self.helper_test_exception([(8,)], lambda x: x.unfold(0, 9, 3), expected=RuntimeError)
     self.helper_test_exception([(8,)], lambda x: x.unfold(1, 8, 3), expected=IndexError)
-    self.helper_test_exception([(8,)], lambda x: x.unfold(0, 9, 3), expected=RuntimeError)
     self.helper_test_exception([(8,)], lambda x: x.unfold(0, 1, -1), expected=RuntimeError)
 
   def test_meshgrid(self):
@@ -285,6 +283,8 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: torch.arange(-128, 128, dtype=torch.int8), lambda: Tensor.arange(-128, 128, dtype=dtypes.int8), forward_only=True)
     helper_test_op([], lambda: torch.arange(127, -129, -1, dtype=torch.int8),
                    lambda: Tensor.arange(127, -129, -1, dtype=dtypes.int8), forward_only=True)
+    # an int range too large for default_int picks int64
+    self.assertEqual(Tensor.arange(2**31, 2**31+3).dtype, dtypes.int64)
     # overflow: tinygrad raises (torch silently wraps)
     with self.assertRaises(OverflowError): Tensor.arange(2**33, dtype=dtypes.int)
     with self.assertRaises(OverflowError): Tensor.arange(129, dtype=dtypes.int8)            # last=128 overflows
@@ -450,7 +450,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(45,35), (45,35), (45,35)], lambda x,y,z: x.lerp(y,z))
     helper_test_op(None, lambda x,y,z: x.lerp(y,z), vals=[[1.,2.,3.], [4.,5.,6.], 0.5])
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_tril(self):
     helper_test_op([(3,3)], lambda x: x.tril())
     helper_test_op([(3,3)], lambda x: x.tril(1))
@@ -468,7 +467,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(5,3,3)], lambda x: x.tril(1))
     helper_test_op(None, lambda x: x.tril(), vals=[[[True] * 3] * 3], forward_only=True)
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_triu(self):
     helper_test_op([(3,3)], lambda x: x.triu())
     helper_test_op([(3,3)], lambda x: x.triu(1))
@@ -771,6 +769,11 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: torch.tensor([2], dtype=torch.int) ** torch.tensor(-2, dtype=torch.int),
                        lambda: Tensor([2]) ** Tensor(-2), forward_only=True)
 
+  def test_pow_int_base_float_exponent(self):
+    for exponent in (0.5, 1.5, 2.0, -1.0, 0.0):
+      helper_test_op([], lambda: torch.tensor([1, 2, 3, 4], dtype=torch.int) ** exponent,
+                         lambda: Tensor([1, 2, 3, 4], dtype=dtypes.int32) ** exponent, forward_only=True)
+
   def test_sqrt(self):
     helper_test_op([(45,65)], lambda x: x.sqrt())
     helper_test_op(None, lambda x: x.sqrt(), vals=[[0.0]])
@@ -789,9 +792,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: tor^0x1337, lambda: ten^0x1337, forward_only=True)
     helper_test_op([], lambda: 0x1337^tor, lambda: 0x1337^ten, forward_only=True)
 
-    self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_xor(y), expected=RuntimeError)
-
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_and(self):
     data = [[1,-8,1],[32,1,6]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -807,9 +807,6 @@ class TestOps(unittest.TestCase):
 
     helper_test_op(None, lambda x: (1 < x) & (x < 2), forward_only=True, vals=[[1.2, 1.2, 1.2, 3.2]])
 
-    self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_and(y), expected=RuntimeError)
-
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_or(self):
     data = [[1,-8,1],[32,1,6]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -823,8 +820,6 @@ class TestOps(unittest.TestCase):
     ten0, ten1 = Tensor(data[0], dtype=dtypes.bool), Tensor(data[1], dtype=dtypes.bool)
     helper_test_op([], lambda: tor0|tor1, lambda: ten0|ten1, forward_only=True)
 
-    self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_or(y), expected=RuntimeError)
-
   def test_bitwise_not(self):
     data = [[1,-8,1],[32,1,6]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -838,8 +833,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: tor.bitwise_not(), lambda: ten.bitwise_not(), forward_only=True)
     helper_test_op([], lambda: ~tor, lambda: ~ten, forward_only=True)
 
-    self.helper_test_exception([(4)], lambda x: x.bitwise_not(), expected=RuntimeError)
-
   def test_lshift(self):
     data = [[0,1,2],[1<<8,1<<16,1<<31-1]]
     tor = torch.tensor(data, dtype=torch.int)
@@ -848,8 +841,13 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: tor << 0, lambda: (ten << 0).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor << 2, lambda: (ten << 2).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor << 31, lambda: (ten << 31).cast(dtypes.int32), forward_only=True)
+    helper_test_op([], lambda: tor << torch.tensor([0,2,4]).int(),
+                   lambda: (ten << Tensor([0,2,4], dtype=dtypes.uint32)).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor.__lshift__(2), lambda: ten.__lshift__(2).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor.bitwise_left_shift(2), lambda: ten.lshift(2).cast(dtypes.int32), forward_only=True)
+    self.helper_test_exception([], lambda: torch.tensor([1.0]) << 2, lambda: Tensor([1.0]) << 2, expected=RuntimeError)
+    self.helper_test_exception([], lambda: tor << torch.tensor([1.0]), lambda: ten << Tensor([1.0]), expected=RuntimeError)
+    self.helper_test_exception([], lambda: tor << 1.0, lambda: ten << 1.0, expected=RuntimeError)
 
   def test_rshift(self):
     data = [[0,1,2],[1<<8,1<<16,1<<31-1]]
@@ -859,8 +857,12 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: tor >> 0, lambda: (ten >> 0).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor >> 2, lambda: (ten >> 2).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor >> 31, lambda: (ten >> 31).cast(dtypes.int32), forward_only=True)
+    helper_test_op([], lambda: tor >> torch.tensor([0,2,4]).int(),
+                   lambda: (ten >> Tensor([0,2,4], dtype=dtypes.uint32)).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor.__rshift__(2), lambda: ten.__rshift__(2).cast(dtypes.int32), forward_only=True)
     helper_test_op([], lambda: tor.bitwise_right_shift(2), lambda: ten.rshift(2).cast(dtypes.int32), forward_only=True)
+    self.helper_test_exception([], lambda: torch.tensor([4.0]) >> 1, lambda: Tensor([4.0]) >> 1, expected=RuntimeError)
+    self.helper_test_exception([], lambda: tor >> torch.tensor([1.0]), lambda: ten >> Tensor([1.0]), expected=RuntimeError)
 
   def test_lshift_signed(self):
     data = [[-1, -3, 1, 7], [0, -2147483648, 2147483647, -1]]
@@ -870,6 +872,7 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: tor << 2, lambda: ten << 2, forward_only=True)
     helper_test_op([], lambda: tor << 8, lambda: ten << 8, forward_only=True)
     helper_test_op([], lambda: tor << 31, lambda: ten << 31, forward_only=True)
+    helper_test_op([], lambda: tor << torch.tensor([0,2,8,31]).int(), lambda: ten << Tensor([0,2,8,31], dtype=dtypes.int), forward_only=True)
 
   def test_rshift_signed(self):
     data = [[-1, -3, 1, 7], [0, -2147483648, 2147483647, -1]]
@@ -879,6 +882,7 @@ class TestOps(unittest.TestCase):
     helper_test_op([], lambda: tor >> 2, lambda: ten >> 2, forward_only=True)
     helper_test_op([], lambda: tor >> 8, lambda: ten >> 8, forward_only=True)
     helper_test_op([], lambda: tor >> 31, lambda: ten >> 31, forward_only=True)
+    helper_test_op([], lambda: tor >> torch.tensor([0,2,8,31]).int(), lambda: ten >> Tensor([0,2,8,31], dtype=dtypes.int), forward_only=True)
 
   def test_idiv_shift_rewrite_negative(self):
     a = Tensor(-5).div(2, rounding_mode="trunc").item()
@@ -1042,8 +1046,8 @@ class TestOps(unittest.TestCase):
     helper_test_op([(45,65)], torch.nn.functional.hardsigmoid, Tensor.hardsigmoid)
     helper_test_op([()], torch.nn.functional.hardsigmoid, Tensor.hardsigmoid)
   def test_hardsigmoid_extreme(self):
-    helper_test_op([(45,65)], torch.sigmoid, Tensor.sigmoid, low=300, high=400)
-    helper_test_op([(45,65)], torch.sigmoid, Tensor.sigmoid, low=-400, high=-300)
+    helper_test_op([(45,65)], torch.nn.functional.hardsigmoid, Tensor.hardsigmoid, low=300, high=400)
+    helper_test_op([(45,65)], torch.nn.functional.hardsigmoid, Tensor.hardsigmoid, low=-400, high=-300)
   def test_softplus(self):
     helper_test_op([(45,65)], torch.nn.functional.softplus, Tensor.softplus, grad_atol=1e-6)
     helper_test_op([(45,65)], lambda t: torch.nn.functional.softplus(t, beta=3), lambda t: Tensor.softplus(t, beta=3), grad_atol=1e-6)
@@ -1223,7 +1227,6 @@ class TestOps(unittest.TestCase):
     helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[False, True]])
     helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[True, False]])
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_argmin(self):
     # check if it returns the first index for multiple occurrences
     helper_test_op(None, lambda x: x.argmin().type(torch.int32), lambda x: x.argmin(), forward_only=True, vals=[[2, 2]])
@@ -1260,23 +1263,20 @@ class TestOps(unittest.TestCase):
                    lambda x: x.sort(descending=True)[1], forward_only=True, vals=[[0, 1] * 9])
 
   def test_argsort(self):
-    for dim in [-1, 0, 1]:
-      for descending in [True, False]:
-        helper_test_op([(8,8,6)], lambda x: torch.argsort(x, dim=dim, descending=descending, stable=True).type(torch.int32),
-                                  lambda x: x.argsort(dim, descending), forward_only=True)
+    helper_test_op([(8,8,6)], lambda x: torch.argsort(x, dim=1, descending=True, stable=True).type(torch.int32),
+                              lambda x: x.argsort(1, True), forward_only=True)
 
   def test_topk(self):
     helper_test_op([(8)], lambda x: x.topk(3).values, lambda x: x.topk(3)[0], forward_only=True)
     helper_test_op([(8)], lambda x: x.topk(3).indices.type(torch.int32), lambda x: x.topk(3)[1], forward_only=True)
-    for dim in [0, 1, -1]:
-      for largest in [True, False]:
-        for sorted_ in [True]: # TODO support False
-          helper_test_op([(5,5,4)],
-                          lambda x: x.topk(4, dim, largest, sorted_).values,
-                          lambda x: x.topk(4, dim, largest, sorted_)[0], forward_only=True)
-          helper_test_op([(5,5,4)],
-                          lambda x: x.topk(4, dim, largest, sorted_).indices.type(torch.int32),
-                          lambda x: x.topk(4, dim, largest, sorted_)[1], forward_only=True)
+    for dim, largest in [(0, True), (1, False)]:
+      for sorted_ in [True]: # TODO support False
+        helper_test_op([(5,5,4)],
+                        lambda x: x.topk(4, dim, largest, sorted_).values,
+                        lambda x: x.topk(4, dim, largest, sorted_)[0], forward_only=True)
+        helper_test_op([(5,5,4)],
+                        lambda x: x.topk(4, dim, largest, sorted_).indices.type(torch.int32),
+                        lambda x: x.topk(4, dim, largest, sorted_)[1], forward_only=True)
     # repeated values
     if not COMPILE_ONLY:
       value, indices = Tensor([1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0]).topk(3)
@@ -1529,7 +1529,6 @@ class TestOps(unittest.TestCase):
   def test_prod_dtype_arg(self):
     with self.assertRaises(AttributeError): Tensor([1.0, 2.0]).prod(dtype="")
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_min(self):
     helper_test_op([(3,3)], lambda x: x.min())
     helper_test_op([(45,3)], lambda x: x.min())
@@ -1569,7 +1568,6 @@ class TestOps(unittest.TestCase):
   def test_any_zero_axis(self):
     helper_test_op([(1,0,3,0,5)], lambda x: x.any(axis=(1,3)), forward_only=True)
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_all(self):
     helper_test_op([(3,4,5,6)], lambda x: x.all(), forward_only=True)
     helper_test_op(None, lambda x: x.all(), vals=[[True, True]], forward_only=True)
@@ -1901,9 +1899,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(3,3,3)], lambda x: x[-2:2])
     helper_test_op([(3,3,3)], lambda x: x[-2:-5])
 
-  def test_slice_empty(self):
-    helper_test_op([(10,10)], lambda x: x[1:1])
-
   def test_slice_zero_in_shape(self):
     helper_test_op([(10,10)], lambda x: x[1:1])  # x.shape = (0, 10)
     helper_test_op([(3,3,3)], lambda x: x[-2:-5])  # x.shape = (0, 3, 3)
@@ -2096,7 +2091,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(4,3,1,6)], lambda x: x.squeeze(1))
     helper_test_op([(4,3,6,6)], lambda x: x.squeeze(3))
     self.helper_test_exception([(4,3,6,6)], lambda x: x.squeeze(50), expected=IndexError)
-    self.helper_test_exception([(4,3,6,6)], lambda x: x.squeeze(50), expected=IndexError)
     helper_test_op([(4,3,6,1)], lambda x: x.squeeze(-1))
     helper_test_op([(4,3,6,6)], lambda x: x.squeeze())
     helper_test_op([(1,3,6,6)], lambda x: x.squeeze())
@@ -2369,9 +2363,10 @@ class TestOps(unittest.TestCase):
               lambda x,w: Tensor.conv2d(x,w,groups=groups), grad_rtol=1e-5)
   def test_conv2d(self): self._test_conv2d(bs=1, cin=3)
   @slow_test
+  @unittest.skip("redundant: bs/cout are loop dims, kernel×cin sweep covered by test_conv2d")
   def test_conv2d_bs_4_cin_3(self): self._test_conv2d(bs=4, cin=3, cout=2)
   def test_conv2d_bs_1_cin_1(self): self._test_conv2d(bs=1, cin=1)
-  @slow_test
+  @unittest.skip("redundant: cin=1 covered by test_conv2d_bs_1_cin_1")
   def test_conv2d_bs_4_cin_1(self): self._test_conv2d(bs=4, cin=1)
 
   def test_conv2d_errors(self):
@@ -2491,9 +2486,6 @@ class TestOps(unittest.TestCase):
             helper_test_op([(1,1,n,n), (1,1,k,k)],
               lambda x,w: torch.nn.functional.conv2d(torch.nn.functional.pad(x, p),w),
               lambda x,w: Tensor.conv2d(x,w,padding=p))
-            helper_test_op([(1,1,n,n), (1,1,k,k)],
-              lambda x,w: torch.nn.functional.conv2d(torch.nn.functional.pad(x, p),w),
-              lambda x,w: Tensor.conv2d(x,w,padding=p))
 
   def test_padded_conv2d_p21(self):
     bs,cin,H,W,padding = 4, 3, 3, 3, (2,1)
@@ -2542,7 +2534,7 @@ class TestOps(unittest.TestCase):
 
   @slow_test
   def test_max_pool2d(self):
-    for ksz in [(2,2), (3,3), 2, 3, (3,2), (5,5), (5,1)]:
+    for ksz in [2, (3,3), (3,2), (5,5), (5,1)]:
       with self.subTest(kernel_size=ksz):
         helper_test_op([(32,2,11,28)],
           lambda x: torch.nn.functional.max_pool2d(x, kernel_size=ksz),
@@ -2550,7 +2542,7 @@ class TestOps(unittest.TestCase):
 
   @slow_test
   def test_max_pool2d_padding(self):
-    for ksz in [(2,2), (3,3), 2, 3, (3,2)]:
+    for ksz in [(3,3), 2, (3,2)]:
       for p in [1, (1,0), (0,1)]:
         with self.subTest(kernel_size=ksz, padding=p):
           helper_test_op([(4,2,11,28)],
@@ -2613,7 +2605,7 @@ class TestOps(unittest.TestCase):
 
   def test_max_pool2d_ceil_mode(self):
     shape = (1,1,6,6)
-    for ksz in [(3,3), 3, (3,2), 4]:
+    for ksz in [(3,3), (3,2), 4]:
       with self.subTest(kernel_size=ksz):
         helper_test_op([shape],
           lambda x: torch.nn.functional.max_pool2d(x, kernel_size=ksz, padding=1, stride=3, ceil_mode=True),
@@ -2693,7 +2685,7 @@ class TestOps(unittest.TestCase):
   @slow_test
   def test_avg_pool2d(self):
     shape = (32,2,11,28)
-    for ksz in [(2,2), (3,3), (3,2), (5,5), (5,1)]:
+    for ksz in [2, (3,3), (3,2), (5,5), (5,1)]:
       with self.subTest(kernel_size=ksz):
         helper_test_op([shape],
           lambda x: torch.nn.functional.avg_pool2d(x, kernel_size=ksz),
@@ -2707,7 +2699,7 @@ class TestOps(unittest.TestCase):
   @slow_test
   def test_avg_pool2d_padding(self):
     shape = (32,2,11,28)
-    for ksz in [(2,2), (3,3), 2, 3, (3,2)]:
+    for ksz in [2, (3,3), (3,2)]:
       for p in [1, (1,0), (0,1)]:
         with self.subTest(kernel_size=ksz, padding=p):
           helper_test_op([shape],
@@ -2729,7 +2721,7 @@ class TestOps(unittest.TestCase):
   @slow_test
   def test_avg_pool2d_padding_not_counted(self):
     shape = (32,2,11,28)
-    for ksz in [(2,2), (3,3), 2, 3, (3,2)]:
+    for ksz in [(3,3), 2, (3,2)]:
       with self.subTest(kernel_size=ksz):
         helper_test_op([shape],
           lambda x: torch.nn.functional.avg_pool2d(x, kernel_size=ksz, padding=1, count_include_pad=False),
@@ -2737,7 +2729,7 @@ class TestOps(unittest.TestCase):
 
   def test_avg_pool2d_ceil_mode(self):
     shape = (1,1,6,6)
-    for ksz in [(3,3), 3, (3,2), 4]:
+    for ksz in [(3,3), (3,2), 4]:
       with self.subTest(kernel_size=ksz):
         helper_test_op([shape],
           lambda x: torch.nn.functional.avg_pool2d(x, kernel_size=ksz, padding=1, stride=3, ceil_mode=True),
@@ -2745,7 +2737,7 @@ class TestOps(unittest.TestCase):
 
   def test_avg_pool2d_ceil_mode_padding_not_counted(self):
     shape = (1,1,6,6)
-    for ksz in [(3,3), 3, (3,2), 4]:
+    for ksz in [(3,3), (3,2), 4]:
       with self.subTest(kernel_size=ksz):
         helper_test_op([shape],
           lambda x: torch.nn.functional.avg_pool2d(x, kernel_size=ksz, padding=1, stride=3, ceil_mode=True, count_include_pad=False),
@@ -2947,7 +2939,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[...,c,:,e], lambda x: x[...,k,:,p])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_dim_collapse_int(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     # dim collapse from int
@@ -2958,7 +2949,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[1,:,3:11:2,d,0:2], lambda x: x[1,:,3:11:2,o,0:2])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_dim_inject_none(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     # dim injection from None
@@ -2993,7 +2983,6 @@ class TestOps(unittest.TestCase):
                             lambda x: x[Tensor([[0,1,-1],[-1,-2,0]]), Tensor([2,1,-1])])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_list_indices(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[((0,),)])
@@ -3005,7 +2994,6 @@ class TestOps(unittest.TestCase):
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,(2,1,0),c,(-2,1,0),e], lambda x: x[i,(2,1,0),k,(-2,1,0),p])
 
   @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
   def test_slice_fancy_indexing_tuple_indices(self):
     a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
     helper_test_op([(2,5,6,5,3,4)], lambda x: x[(((0,),),)], lambda x: x[(((0,),),)])
@@ -3336,6 +3324,7 @@ class TestOps(unittest.TestCase):
   @unittest.skipIf((DEV.interface.startswith("MOCK") or Device.DEFAULT == "PYTHON"), "very slow on MOCKGPU because reduce does not fold")
   @unittest.skipIf(Device.DEFAULT == "WEBGPU", "webgpu runtime issue")
   @unittest.skipIf(Device.DEFAULT == "QCOM", "QCOM fails with: Resource deadlock avoided")
+  @unittest.skipIf(Device.DEFAULT == "CPU" and DEV.renderer == "LVP", "extremely slow with LVP")
   def test_masked_select(self):
     helper_test_op([(32, 10)], lambda x: x.masked_select(x>0.5), lambda x: x.masked_select(x>0.5), forward_only=True)
     helper_test_op([(32, 10)], lambda x: x.masked_select(torch.tensor(True)), lambda x: x.masked_select(Tensor(True)), forward_only=True)

test/backend/test_pickle.py

@@ -1,7 +1,7 @@
-import unittest, pickle, types
+import unittest, pickle, types, tracemalloc
 import numpy as np
-from tinygrad import Tensor, TinyJit, Variable, dtypes
-from tinygrad.helpers import GlobalCounters, ContextVar, Context
+from tinygrad import Tensor, Device, TinyJit, Variable, dtypes
+from tinygrad.helpers import GlobalCounters, ContextVar, Context, DEV
 from tinygrad.uop.ops import PatternMatcher, UPat, UOp
 
 class TestPickle(unittest.TestCase):
@@ -78,6 +78,22 @@ class TestPickle(unittest.TestCase):
     a2:UOp = pickle.loads(s)
     self.assertListEqual(a2.base.realized.as_memoryview().cast("I").tolist(), [0, 1, 2, 3])
 
+  @unittest.skipIf(DEV.interface.startswith("MOCK"), "mock device buffers live in host RAM, not VRAM")
+  def test_pickle_oob_ram(self):
+    N, M = 8, 10**6
+    ts = [Tensor.rand(M, dtype='float32').realize() for _ in range(N)]
+    tracemalloc.start()
+    st = pickle.dumps(ts, protocol=5, buffer_callback=lambda pb: pb.release())
+    self.assertLess(tracemalloc.get_traced_memory()[1], N*M*4)
+    tracemalloc.reset_peak()
+    def make_fake_buffers():
+      for _ in range(N):
+        Device[Device.DEFAULT].synchronize()
+        yield pickle.PickleBuffer(bytearray(M*4))
+    pickle.loads(st, buffers=make_fake_buffers())
+    self.assertLess(tracemalloc.get_traced_memory()[1], N*M*4)
+    tracemalloc.stop()
+
   def test_pickle_unrealized_tensor(self):
     t = Tensor.ones(10, 10)
     st = pickle.dumps(t)

test/backend/test_profiler.py

@@ -70,9 +70,9 @@ class TestProfiler(unittest.TestCase):
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
     with helper_collect_profile(TestProfiler.d0) as profile:
-      buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
+      buf1.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
 
-    kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]
+    kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith((TestProfiler.d0.device, "PYTHON"))]
     assert len(kernel_runs) == 1, "one kernel run is expected"
 
   def test_profile_multiops(self):
@@ -80,12 +80,12 @@ class TestProfiler(unittest.TestCase):
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
     with helper_collect_profile(TestProfiler.d0) as profile:
-      buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
+      buf1.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
       gs, ls = TestProfiler.prg.arg.launch_dims({})
       TestProfiler.runtime(buf1._buf, TestProfiler.a.uop.buffer._buf, global_size=gs, local_size=ls)
-      buf1.copyout(memoryview(bytearray(buf1.nbytes)))
+      buf1.as_memoryview()
 
-    evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]
+    evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith((TestProfiler.d0.device, "PYTHON"))]
 
     assert len(evs) == 3, "3 kernel runs are expected"
     # NOTE: order of events does not matter, the tool is responsible for sorting them
@@ -103,12 +103,12 @@ class TestProfiler(unittest.TestCase):
     buf2 = Buffer(f"{Device.DEFAULT}:1", 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
     with helper_collect_profile(TestProfiler.d0, d1) as profile:
-      buf1.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
-      buf2.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
+      buf1.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
+      buf2.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
 
     for dev in [TestProfiler.d0.device, d1.device]:
       evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and _dev_base(x.device) == dev]
-      assert len(evs) == 1, "one kernel runs are expected"
+      assert len(evs) == (0 if hasattr(TestProfiler.d0.allocator, '_as_buffer') else 1), "one kernel runs are expected"
 
   def test_profile_multidev_transfer(self):
     try: d1 = Device[f"{Device.DEFAULT}:1"]

test/backend/test_renderer_failures.py

@@ -1,6 +1,6 @@
 import unittest
 import numpy as np
-from tinygrad.device import Device
+from tinygrad.device import Device, Buffer
 from tinygrad.dtype import dtypes, ConstType
 from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
@@ -10,13 +10,13 @@ from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.wgsl import WGSLRenderer
 from tinygrad.runtime.ops_python import PythonRenderer
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, python_alu
-from tinygrad.tensor import Tensor, _to_np_dtype
+from tinygrad.tensor import Tensor
 
 def _test_uop_result(inputs:list[Tensor], sink:UOp, local_size=None):
   for x in inputs: x.realize()
   sz = 1 if local_size is None else prod(local_size)
   outs = [UOp.new_buffer(Device.DEFAULT, sz, u.src[1].dtype) for u in sink.src if u.op is Ops.STORE]
-  for u in outs: u.buffer.allocate().copyin(np.zeros(sz, dtype=_to_np_dtype(u.dtype)).data)
+  for u in outs: u.buffer.allocate().copy_from(Buffer("PYTHON", sz, u.dtype, opaque=memoryview(bytearray(u.buffer.nbytes))))
   run_linear(UOp(Ops.LINEAR, src=(sink.call(*outs, *(x.uop.base for x in inputs)),)))
   return [u.buffer.numpy() for u in outs]
 

test/backend/test_schedule.py

@@ -2,7 +2,7 @@
 # schedule confirms the right things are capable of fusing
 # NOTE: this has overlap with external_test_opt.py
 
-import unittest
+import unittest, time
 import numpy as np
 
 from tinygrad import nn, dtypes, Device, Tensor, Variable
@@ -197,6 +197,20 @@ class TestLimitBufs(unittest.TestCase):
         base = (idx >= i).where(a + b, base)
       assert all(x > 0 for x in base.tolist())
 
+  def test_limit_bufs_linear_scaling(self):
+    def sched_time(n):
+      with Context(TRACK_MATCH_STATS=0, DEBUG=0):
+        bufs = [Tensor.ones(16).contiguous().realize() for _ in range(4)]
+        root = bufs[0]
+        for i in range(n): root = root + bufs[i % 4]
+        with Context(MAX_KERNEL_BUFFERS=8, SCACHE=0):
+          st = time.perf_counter()
+          root.schedule_linear()
+          return time.perf_counter() - st
+    sched_time(400)
+    t1, t2 = min(sched_time(400) for _ in range(3)), min(sched_time(1600) for _ in range(3))
+    self.assertLess(t2/t1, 8, f"{t1*1e3:.1f}ms -> {t2*1e3:.1f}ms")
+
 class TestSwizzle(unittest.TestCase):
   def test_swizzle_simple(self):
     Tensor.manual_seed(0)

test/backend/test_subbuffer.py

@@ -4,11 +4,10 @@ from tinygrad.device import Buffer
 from tinygrad.helpers import Context, DEV
 from test.helpers import needs_second_gpu
 
-@unittest.skipUnless(hasattr(Device[Device.DEFAULT].allocator, "_offset"), "subbuffer not supported")
+@unittest.skipIf(Device.DEFAULT in {"WEBGPU", "CL"}, "subbuffer not supported")
 class TestSubBuffer(unittest.TestCase):
   def setUp(self):
-    self.buf = Buffer(Device.DEFAULT, 10, dtypes.uint8).ensure_allocated()
-    self.buf.copyin(memoryview(bytearray(range(10))))
+    self.buf = Buffer(Device.DEFAULT, 10, dtypes.uint8, initial_value=bytes(range(10)))
     self.buf_unalloc = Buffer(Device.DEFAULT, 10, dtypes.uint8)
 
   def test_subbuffer(self):
@@ -59,7 +58,7 @@ class TestSubBuffer(unittest.TestCase):
       _ = Buffer(Device.DEFAULT, 10, dtypes.uint8).ensure_allocated()
 
       self.buf.ensure_allocated()
-      self.buf.copyin(memoryview(bytearray(range(10, 20))))
+      self.buf.copy_from(Buffer("PYTHON", 10, dtypes.uint8, opaque=memoryview(bytearray(range(10, 20)))))
 
       vbuf.ensure_allocated()
 
@@ -109,15 +108,15 @@ class TestSubBuffer(unittest.TestCase):
   def test_subbuffer_copy_in_out(self):
     sub_buf = self.buf.view(3, dtypes.uint8, offset=3).ensure_allocated() # [3:6]
     data_out_sub = bytearray([0]*3)
-    sub_buf.copyout(memoryview(data_out_sub))
+    data_out_sub[:] = sub_buf.as_memoryview()
     assert data_out_sub == bytearray(range(3, 6))
-    sub_buf.copyin(memoryview(bytearray(range(3))))
+    sub_buf.copy_from(Buffer("PYTHON", 3, dtypes.uint8, opaque=memoryview(bytearray(range(3)))))
     assert sub_buf.as_memoryview().tolist() == list(range(3))
     assert self.buf.as_memoryview().tolist()[3:6] == list(range(3))
-    sub_buf.copyout(memoryview(data_out_sub))
+    data_out_sub[:] = sub_buf.as_memoryview()
     assert data_out_sub == bytearray(range(3))
     data_out_base = bytearray([0]*10)
-    self.buf.copyout(memoryview(data_out_base))
+    data_out_base[:] = self.buf.as_memoryview()
     assert data_out_base[0:3] == bytearray(range(0, 3))
     assert data_out_base[3:6] == data_out_sub
     assert data_out_base[6:10] == bytearray(range(6, 10))
@@ -129,27 +128,27 @@ class TestSubBuffer(unittest.TestCase):
     self.assertTrue(view2.is_allocated())
 
     data_in = bytearray([7, 8, 9])
-    view2.copyin(memoryview(data_in))
+    view2.copy_from(Buffer("PYTHON", 3, view2.dtype, opaque=memoryview(data_in)))
     data_out_v2 = bytearray([0]*3)
-    view2.copyout(memoryview(data_out_v2))
+    data_out_v2[:] = view2.as_memoryview()
     assert data_in == data_out_v2
 
     expected_base_data = memoryview(bytearray(range(10)))
     expected_base_data[4:7] = data_in
 
     data_out_base = bytearray([0]*10)
-    self.buf.copyout(memoryview(data_out_base))
+    data_out_base[:] = self.buf.as_memoryview()
     assert expected_base_data == data_out_base
 
   def test_subbuffer_alloc(self):
     sub_buf = self.buf.view(4, dtypes.int8, offset=3)
     sub_buf.allocate()
-    sub_buf.copyin(memoryview(bytearray(range(10, 14))))
+    sub_buf.copy_from(Buffer("PYTHON", 4, dtypes.int8, opaque=memoryview(bytearray(range(10, 14)))))
     assert self.buf.as_memoryview().tolist()[3:7] == sub_buf.as_memoryview().tolist()
 
     sub_buf = self.buf_unalloc.view(4, dtypes.int8, offset=3)
     sub_buf.allocate()
-    sub_buf.copyin(memoryview(bytearray(range(10, 14))))
+    sub_buf.copy_from(Buffer("PYTHON", 4, dtypes.int8, opaque=memoryview(bytearray(range(10, 14)))))
     assert self.buf_unalloc.as_memoryview().tolist()[3:7] == sub_buf.as_memoryview().tolist()
 
   def test_subbuffer_dealloc(self):

test/backend/test_uops.py

@@ -33,11 +33,9 @@ def _test_single_value(vals, op, dts):
   alu = uop(uops, op, output_dtype, loads)
   out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), alu))
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
-  buf2 = [Buffer(Device.DEFAULT, 1, dtype).allocate().copyin(np.array([a], dtype=_to_np_dtype(dtype)).data) for a,dtype in zip(vals, dts)]
+  buf2 = [Buffer(Device.DEFAULT, 1, dtype, initial_value=np.array([a], dtype=_to_np_dtype(dtype)).tobytes()) for a,dtype in zip(vals, dts)]
   run_uops([out], [buf]+buf2)
-  ret = np.empty(1, _to_np_dtype(output_dtype))
-  buf.copyout(ret.data)
-  return ret[0]
+  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
 
 def _test_single_value_const(vals, op, dts):
   uops = []
@@ -48,9 +46,7 @@ def _test_single_value_const(vals, op, dts):
   out = buf_store[UOp.const(dtypes.int32, 0)].store(alu)
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
   run_uops([out], [buf])
-  ret = np.empty(1, _to_np_dtype(output_dtype))
-  buf.copyout(ret.data)
-  return ret[0]
+  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
 
 def _test_uops_result(output_dtype, uops, res):
   # uops = []
@@ -59,9 +55,7 @@ def _test_uops_result(output_dtype, uops, res):
   out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), res))
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
   run_uops([out], [buf])
-  ret = np.empty(1, _to_np_dtype(output_dtype))
-  buf.copyout(ret.data)
-  return ret[0]
+  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
 
 class TestUOps(unittest.TestCase):
   def _equal(self, v1, v2):

test/backend/test_wait_loop.py

@@ -0,0 +1,130 @@
+import unittest, threading
+from tinygrad import Tensor, UOp
+from tinygrad.device import Device, Buffer, BufferSpec
+from tinygrad.dtype import AddrSpace, dtypes
+from tinygrad.engine.realize import run_linear
+from tinygrad.renderer.isa.x86 import X86Renderer
+from tinygrad.uop.ops import Ops, KernelInfo
+
+def wait_loop_kernel(C:UOp) -> UOp:
+  N = 10
+
+  # a RANGE with no src is a bound-less loop header: a jump target with no induction variable.
+  # the compare and conditional backedge are expanded by the renderers from the loop RANGE/END
+  l = UOp.loop(0)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+
+  # i = 0
+  i = i.after(i[0].store(0))
+
+  # i + 1, read loop-carried through after(l)
+  inc = i.after(l)[0].load() + 1
+
+  # i = inc; END(store, l, cond): conditional backedge, loop again while inc < N (do-while)
+  # NOTE: the cond uses the computed value, not a reload of the register
+  st = i[0].store(inc)
+  i = i.after(st.end(l, inc < N))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="wait_loop"))
+
+def nested_loop_kernel(C:UOp) -> UOp:
+  r = UOp.range(4, 0)
+  l = UOp.loop(1)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+  i = i.after(i[0].store(0))
+
+  inc = i.after(l, r)[0].load() + 1
+  st = i[0].store(inc)
+
+  lend = st.end(l, inc < (r.cast(dtypes.int)+1)*3)
+  i = i.after(lend.end(r))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="nested_loop", opts_to_apply=()))
+
+def wait_ext_kernel() -> UOp:
+  sig = UOp.param(0, dtypes.int, (1,), volatile=True)
+  l = UOp.loop(0)
+  v = sig.after(l)[0].load()
+  e = v.end(l, v < 1)
+  return e.sink(arg=KernelInfo(name="wait_ext"))
+
+def two_loops_kernel(C:UOp) -> UOp:
+  # two sequential loops on the same counter: ++ until 10, then ++ until 25
+  l1, l2 = UOp.loop(0), UOp.loop(1)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+  i = i.after(i[0].store(0))
+
+  inc1 = i.after(l1)[0].load() + 1
+  i = i.after(i[0].store(inc1).end(l1, inc1 < 10))
+
+  inc2 = i.after(l2)[0].load() + 1
+  i = i.after(i[0].store(inc2).end(l2, inc2 < 25))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="two_loops", opts_to_apply=()))
+
+def loop_in_loop_kernel(C:UOp) -> UOp:
+  # outer loop while i < 12, inner loop increments until i % 4 == 0 -> 12
+  l1, l2 = UOp.loop(0), UOp.loop(1)
+
+  i = UOp.placeholder((1,), dtypes.int, 0, addrspace=AddrSpace.REG)
+  i = i.after(i[0].store(0))
+
+  inc = i.after(l1, l2)[0].load() + 1
+  st = i[0].store(inc)
+
+  # the outer END closes the inner END, and its cond reloads the register after the inner loop (in scope at the outer level)
+  e2 = st.end(l2, inc % 4 != 0)
+  oc = i.after(e2)[0].load()
+  i = i.after(e2.end(l1, oc < 12))
+
+  return C[0].store(i[0].load()).sink(arg=KernelInfo(name="loop_in_loop", opts_to_apply=()))
+
+@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, X86Renderer), "loops are not supported in X86")
+class TestWaitLoop(unittest.TestCase):
+  def test_wait_loop(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=wait_loop_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 10)
+
+  def test_nested_loop_in_range(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=nested_loop_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 12)
+
+  def test_two_sequential_loops(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=two_loops_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 25)
+
+  def test_loop_in_loop(self):
+    c = Tensor.empty(1, dtype=dtypes.int)
+    c = Tensor.custom_kernel(c, fxn=loop_in_loop_kernel)[0]
+    c.realize()
+    self.assertEqual(c.item(), 12)
+
+@unittest.skipUnless(Device.DEFAULT in ("CPU", "AMD", "NV"), "need proper uncached=True handling")
+@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, X86Renderer), "loops are not supported in X86")
+class TestVolatileLoops(unittest.TestCase):
+  def test_async_wait_ext(self):
+    sig_buf = Buffer(Device.DEFAULT, 1, dtypes.int, options=BufferSpec(host=True, uncached=True, cpu_access=True), preallocate=True)
+    try: sig_view = sig_buf.as_memoryview(force_zero_copy=True).cast('i')
+    except (AssertionError, NotImplementedError): self.skipTest(f"{Device.DEFAULT} does not support host-visible buffers")
+    sig_view[0] = 0
+
+    def set_signal():
+      threading.Event().wait(0.3)
+      sig_view[0] = 1
+
+    sync = threading.Thread(target=set_signal, daemon=True)
+    sync.start()
+    run_linear(UOp(Ops.LINEAR, src=(wait_ext_kernel().call(UOp.from_buffer(sig_buf)),)), wait=True)
+    sync.join(timeout=3)
+
+
+if __name__ == "__main__": unittest.main()

test/device/test_amd_llvm.py

@@ -3,7 +3,7 @@ from tinygrad import Device
 from tinygrad.device import CompileError
 if Device.DEFAULT == "AMD":
   # NOTE: if you don't gate this, LVP fails on Mac
-  from tinygrad.runtime.support.compiler_amd import AMDLLVMCompiler
+  from tinygrad.runtime.support.compiler_llvm import AMDLLVMCompiler
 
 @unittest.skipUnless(Device.DEFAULT == "AMD", "Runs only on AMD")
 class TestAMDLLVM(unittest.TestCase):

test/device/test_hcq.py

@@ -31,8 +31,8 @@ class TestHCQ(unittest.TestCase):
 
   def setUp(self):
     TestHCQ.d0.synchronize()
-    TestHCQ.a.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
-    TestHCQ.b.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 0))))
+    TestHCQ.a.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
+    TestHCQ.b.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 0)))))
     TestHCQ.d0.synchronize() # wait for copyins to complete
 
   # Test signals
@@ -376,7 +376,7 @@ class TestHCQ(unittest.TestCase):
     SZ = 200_000_000
     b = Buffer(f"{Device.DEFAULT}:1", SZ, dtypes.uint8, options=BufferSpec(nolru=True)).allocate()
     a = Buffer(Device.DEFAULT, SZ, dtypes.uint8, options=BufferSpec(nolru=True)).allocate()
-    TestHCQ.d0.allocator.map(b._buf)
+    TestHCQ.d0.allocator._map(b._buf)
 
     sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
     TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
@@ -454,7 +454,7 @@ class TestHCQ(unittest.TestCase):
     buf1 = Buffer(Device.DEFAULT, 1, dtypes.int8, options=BufferSpec(nolru=True)).ensure_allocated()
     buf2 = Buffer(f"{Device.DEFAULT}:1", 1, dtypes.int8, options=BufferSpec(nolru=True)).ensure_allocated()
     buf3 = Buffer(Device.DEFAULT, 1, dtypes.int8, options=BufferSpec(host=True, nolru=True)).ensure_allocated()
-    TestHCQ.d0.allocator.map(buf2._buf)
+    TestHCQ.d0.allocator._map(buf2._buf)
 
     for i in range(256):
       ctypes.memset(buf3._buf.va_addr, i, 1)
@@ -569,7 +569,7 @@ class TestHCQ(unittest.TestCase):
 
       local_buf = Buffer(f"{Device.DEFAULT}:{devid}", sz, dtypes.uint8, options=BufferSpec(cpu_access=True)).ensure_allocated()
 
-      d.allocator.map(cpu_buffer._buf)
+      d.allocator._map(cpu_buffer._buf)
 
       d.hw_copy_queue_t().wait(d.timeline_signal, d.timeline_value - 1) \
                          .copy(local_buf._buf, cpu_buffer._buf, sz) \

test/device/test_ocl.py

@@ -34,7 +34,7 @@ class TestCLError(unittest.TestCase):
     data = list(range(65))
     unaligned = memoryview(bytearray(data))[1:]
     buffer = Buffer("CL", 64, dtypes.uint8).allocate()
-    buffer.copyin(unaligned)
+    buffer.copy_from(Buffer("PYTHON", 64, dtypes.uint8, opaque=unaligned))
     result = memoryview(bytearray(len(data) - 1))
-    buffer.copyout(result)
+    result[:] = buffer.as_memoryview()
     assert unaligned == result, "Unaligned data copied in must be equal to data copied out."

test/external/external_benchmark_op_conv.py

@@ -24,7 +24,7 @@ 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.index, Invalid))
+  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))
   c49 = UOp.param(2, dtypes.half, shape=(64, 49, 4))
   c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
@@ -50,7 +50,7 @@ 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.index, Invalid))
+  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))
   c49 = UOp.param(2, dtypes.half, shape=(128, 49, 4))
   c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))

test/external/external_test_hcq.py

@@ -47,8 +47,8 @@ class TestHCQ(unittest.TestCase):
 
   def setUp(self):
     TestHCQ.d0.synchronize()
-    TestHCQ.a.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 1))))
-    TestHCQ.b.uop.buffer.copyin(memoryview(bytearray(struct.pack("ff", 0, 0))))
+    TestHCQ.a.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 1)))))
+    TestHCQ.b.uop.buffer.copy_from(Buffer("PYTHON", 2, dtypes.float, opaque=memoryview(bytearray(struct.pack("ff", 0, 0)))))
     TestHCQ.d0.synchronize() # wait for copyins to complete
 
   def test_run_1000_times_one_submit(self):

test/external/external_test_usb_asm24.py

@@ -1,66 +1,31 @@
-import unittest, time
-from tinygrad.runtime.support.usb import ASM24Controller
-from tinygrad.helpers import Timing
+import unittest
+from tinygrad.helpers import Timing, getenv
 from tinygrad import Tensor, Device
 import numpy as np
 
-class TestASMController(unittest.TestCase):
-  @classmethod
-  def setUpClass(cls):
-    cls.ctrl = ASM24Controller()
-
-  def test_write_and_read(self):
-    base = 0xF000
-    data = b"hello!"
-    self.ctrl.write(base, data)
-    out = self.ctrl.read(base, len(data))
-    self.assertEqual(out, data)
-
-  def test_scsi_write_and_read_from_f000(self):
-    payload = bytes([0x5B]) * 4096
-    self.ctrl.scsi_write(payload, lba=0)
-    back = self.ctrl.read(0xF000, len(payload))
-    self.assertEqual(back, payload)
-
-  def test_scsi_write_speed_4k(self):
-    payload = bytes([0x5A]) * 4096
-    start = time.perf_counter()
-    self.ctrl.scsi_write(payload, lba=0)
-    dur_ms = (time.perf_counter() - start) * 1000
-    print(f"scsi_write 4K took {dur_ms:.3f} ms")
-
-  def test_read_speed_4k(self):
-    payload = bytes([0xA5]) * 4096
-    self.ctrl.write(0xF000, payload)
-    start = time.perf_counter()
-    out = self.ctrl.read(0xF000, 4096)
-    dur_ms = (time.perf_counter() - start) * 1000
-    print(f"read 4K took {dur_ms:.3f} ms")
-    self.assertEqual(out, payload)
-
 class TestDevCopySpeeds(unittest.TestCase):
   @classmethod
   def setUpClass(cls):
-    cls.sz = 512
+    cls.sz = getenv("SIZE", 2e6)
     cls.dev = Device["AMD"]
     if not cls.dev.is_usb(): raise unittest.SkipTest("only test this on USB devices")
 
   def testCopyCPUtoDefault(self):
     for _ in range(10):
-      t = Tensor.ones(self.sz, self.sz, device="CPU").contiguous().realize()
+      t = Tensor.ones(self.sz, device="CPU", dtype='uchar').contiguous().realize()
       with Timing(f"copyin of {t.nbytes()/1e6:.2f} MB:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns * 1e3:.2f} MB/s"): # noqa: F821
         t.to(Device.DEFAULT).realize()
         Device[Device.DEFAULT].synchronize()
       del t
 
   def testCopyDefaulttoCPU(self):
-    t = Tensor.ones(self.sz, self.sz).contiguous().realize()
+    t = Tensor.ones(self.sz, dtype='uchar').contiguous().realize()
     for _ in range(10):
       with Timing(f"copyout of {t.nbytes()/1e6:.2f} MB:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns * 1e3:.2f} MB/s"):
         t.to('CPU').realize()
 
   def testValidateCopies(self):
-    t = Tensor.randn(self.sz, self.sz, device="CPU").contiguous().realize()
+    t = Tensor.randn(self.sz, device="CPU", dtype='uchar').contiguous().realize()
     x = t.to(Device.DEFAULT).realize()
     Device[Device.DEFAULT].synchronize()
 
@@ -70,4 +35,4 @@ class TestDevCopySpeeds(unittest.TestCase):
     del x, y, t
 
 if __name__ == "__main__":
-  unittest.main()
+  unittest.main()

test/external/fuzz_graph.py

@@ -1,7 +1,7 @@
-import random, ctypes
+import random
 import numpy as np
 from tinygrad.device import Buffer, Device
-from tinygrad.helpers import Context, getenv, from_mv
+from tinygrad.helpers import Context, getenv
 from tinygrad.dtype import dtypes
 from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import BufferXfer, get_runner, ExecItem
@@ -29,7 +29,7 @@ def alloc_rawbuffer(device, fill=False):
   if fill:
     with Context(DEBUG=0):
       data = np.random.randint(-10000, 10000, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_memoryview())
+      rawbuf.copy_from(Tensor(data).realize().uop.base.realized)
   return rawbuf
 
 def gen_kernel_ji(device, deps):
@@ -84,9 +84,7 @@ def run_jit(jis, all_buffers, input_buffers, var_vals):
   with Context(DEBUG=0):
     for rawbuf in all_buffers:
       if rawbuf in input_buffers: continue
-      mv = memoryview(bytearray(rawbuf.nbytes))
-      ctypes.memset(from_mv(mv), 0, len(mv))
-      rawbuf.copyin(mv)
+      rawbuf.copy_from(Buffer("PYTHON", rawbuf.size, rawbuf.dtype, opaque=memoryview(bytearray(rawbuf.nbytes))))
 
   for ei in jis: ei.run(var_vals, jit=True)
 

test/external/fuzz_symbolic.py

@@ -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.index, random.randint(-10, 10))])
+  expr2 = random.choice([random_or_sub_expression_int(depth-1, expr1), UOp.const(dtypes.weakint, 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.index, {n1.as_long()}), UOp.const(dtypes.index, {n2.as_long()})," +\
-                f"UOp.const(dtypes.index, {n3.as_long()})\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" +\
             "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"

test/helpers.py

@@ -6,7 +6,7 @@ from tinygrad import Tensor, dtypes, Device
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.codegen import to_program
-from tinygrad.dtype import DType
+from tinygrad.dtype import DType, truncate
 from tinygrad.nn.state import get_parameters
 from tinygrad.helpers import T, Target, DEV
 from tinygrad.renderer import Renderer
@@ -38,6 +38,10 @@ def call_is_graph(call:UOp) -> bool:
   ast = call.src[0]
   return ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph"
 
+def call_is_hcq(call:UOp) -> bool:
+  ast = call.src[0]
+  return ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "hcq"
+
 def jit_cache_count(linear:UOp) -> int:
   n = 0
   for call in linear.src:
@@ -51,6 +55,7 @@ def assert_jit_cache_len(fxn, expected_len):
   if linear is None or not linear.src:
     assert expected_len == 0, expected_len
     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)
     inner = linear.src[0].src[0].src[0]  # LINEAR UOp inside CUSTOM_FUNCTION
@@ -58,6 +63,8 @@ def assert_jit_cache_len(fxn, expected_len):
   else:
     assert len(linear.src) == expected_len, f"expected {expected_len}, got {len(linear.src)}"
 
+def min_normal(dt:DType) -> float: return 2.0 ** (2 - (1 << (dtypes.finfo(dt)[0] - 1)))
+
 def rand_for_dtype(dt:DType, size:int, allow_subnormal=True):
   if dtypes.is_unsigned(dt):
     return np.random.randint(0, 100, size=size, dtype=_to_np_dtype(dt))
@@ -66,9 +73,8 @@ def rand_for_dtype(dt:DType, size:int, allow_subnormal=True):
   elif dt == dtypes.bool:
     return np.random.choice([True, False], size=size)
   ret = np.random.uniform(-10, 10, size=size).astype(_to_np_dtype(dt))
-  if not allow_subnormal:
-    min_normal = 2.0 ** (2 - (1 << (dtypes.finfo(dt)[0] - 1)))
-    ret = np.where(np.abs(ret) < min_normal, 0, ret)
+  if dt == dtypes.bfloat16 or dt in dtypes.fp8s: ret = np.array([truncate[dt](x) for x in ret], dtype=ret.dtype)
+  if not allow_subnormal: ret = np.where(np.abs(ret) < min_normal(dt), 0, ret)
   return ret
 
 def timeit(fxn:Callable[..., T], *args, **kwargs) -> tuple[T, float]:

test/mockgpu/amd/amdgpu.py

@@ -9,7 +9,7 @@ MOCKGPU_ARCH = "cdna4" if DEV.arch == "gfx950" else "rdna4" if DEV.arch.startswi
 assert (ma:=getenv("MOCKGPU_ARCH", "")) == "", "MOCKGPU_ARCH is deprecated, use DEV=" + \
   str(replace(DEV.value, arch={"cdna4":"gfx950", "rdna4":"gfx1201"}.get(ma, "gfx1100"))) # type: ignore
 GFX_TARGET_VERSION = {"rdna3": 110000, "rdna4": 120000, "cdna4": 90500}[MOCKGPU_ARCH]
-import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4
+import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4, tinygrad.runtime.autogen.am.sdma_6_0_0 as sdma
 
 SDMA_MAX_COPY_SIZE = 0x400000
 
@@ -275,6 +275,7 @@ class SDMAExecutor(AMDQueue):
       elif op == amd_gpu.SDMA_OP_POLL_REGMEM: cont = self._execute_poll_regmem()
       elif op == amd_gpu.SDMA_OP_GCR: self._execute_gcr()
       elif op == amd_gpu.SDMA_OP_COPY: self._execute_copy()
+      elif op == sdma.SDMA_OP_WRITE: self._execute_write()
       elif op == amd_gpu.SDMA_OP_TIMESTAMP: self._execute_timestamp()
       elif op == 32: self.rptr[0] += 4  # SDMA_OP_DUMMY_TRAP: pipeline flush, no interrupt
       else: raise RuntimeError(f"Unknown SDMA op {op}")
@@ -289,6 +290,12 @@ class SDMAExecutor(AMDQueue):
     struct = sdma_pkts.trap.from_address(self.base + self.rptr[0] % self.size)
     self.rptr[0] += ctypes.sizeof(struct)
 
+  def _execute_write(self):
+    packet = to_mv(self.base + self.rptr[0] % self.size, 16).cast('I')
+    addr, count = packet[1] | packet[2] << 32, packet[3] + 1
+    ctypes.memmove(self.gpu.translate_addr(addr), self.base + self.rptr[0] % self.size + 16, count * 4)
+    self.rptr[0] += (4 + count) * 4
+
   def _execute_poll_regmem(self):
     struct = sdma_pkts.poll_regmem.from_address(self.base + self.rptr[0] % self.size)
 

test/mockgpu/amd/emu.py

@@ -803,7 +803,7 @@ def _compile_sopp(inst: ir3.SOPP | ir4.SOPP, ctx: _Ctx) -> UOp:
     pcode = get_pcode(inst.op)
     pc_bytes = ctx.rpc()  # PC is already 64-bit byte address
     vcc, exec_val = ctx.rmask(_c(VCC_LO.offset)), ctx.rexec()
-    srcs = {'PC': pc_bytes.cast(dtypes.int64), 'SIMM16': simm16, 'SCC': ctx.rsgpr_dyn(_c(SCC.offset)), 'VCC': vcc,
+    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)}
     for dest, val in parse_pcode(pcode, srcs)[1]:
@@ -858,7 +858,7 @@ def _compile_sop(inst: ir3.SOP1|ir3.SOP2|ir3.SOPC|ir3.SOPK|ir4.SOP1|ir4.SOP2|ir4
     if isinstance(inst, ir4.SOPK): s0 = simm16
     elif isinstance(inst, irc.SOPK) and 'CMPK' not in op_name and 'SETREG' not in op_name: s0 = simm16_sext
     else: s0 = ctx.rsgpr_dyn(sdst_off)
-    srcs = {'S0': s0, 'S1': simm16_sext, 'SIMM16': simm16_sext, 'D0': ctx.rsgpr_dyn(sdst_off)}
+    srcs: dict[str, UOp|int] = {'S0': s0, 'S1': simm16_sext, 'SIMM16': simm16_sext, 'D0': ctx.rsgpr_dyn(sdst_off)}
     dst_off, dst_size = sdst_off, 1
     # S_GETREG_B32: extract bits from HW register. Handle as special case since HW_REGISTERS is not a normal variable.
     # HW register values are stored at SGPR[SGPR_COUNT-16 + hwRegId] by _init_wave.

test/mockgpu/amd/pcode.py

@@ -688,10 +688,10 @@ class Parser:
         return _extract_bits(base, hi, lo)
       # Dynamic bit slice: (base >> lo) & ((1 << (hi - lo + 1)) - 1)
       dt = dtypes.uint64 if base.dtype in (dtypes.uint64, dtypes.int64) else dtypes.uint32
-      hi, lo = first.cast(dt), second.cast(dt)
-      width = hi - lo + _const(dt, 1)
+      hi_u, lo_u = first.cast(dt), second.cast(dt)
+      width = hi_u - lo_u + _const(dt, 1)
       mask = (_const(dt, 1) << width) - _const(dt, 1)
-      return (base.cast(dt) >> lo) & mask
+      return (base.cast(dt) >> lo_u) & mask
     self.eat('RBRACKET')
     dt_suffix = None
     if self.try_eat('DOT'):
@@ -1123,11 +1123,11 @@ def parse_block(lines: list[str], start: int, env: dict[str, VarVal], funcs: dic
           val = parse_tokens(toks[j:], env, funcs)
           lo_dt, hi_dt = DTYPES.get(lo_type, dtypes.uint64), DTYPES.get(hi_type, dtypes.uint32)
           lo_bits = 64 if lo_dt in (dtypes.uint64, dtypes.int64) else 32
-          lo_val = val.cast(lo_dt) if val.dtype.itemsize * 8 <= lo_bits else (val & _const(val.dtype, (1 << lo_bits) - 1)).cast(lo_dt)
-          hi_val = (val >> _const(val.dtype, lo_bits)).cast(hi_dt)
-          block_assigns[lo_var] = env[lo_var] = lo_val
-          block_assigns[hi_var] = env[hi_var] = hi_val
-          if assigns is not None: assigns.extend([(f'{lo_var}.{lo_type}', lo_val), (f'{hi_var}.{hi_type}', hi_val)])
+          lo_u = val.cast(lo_dt) if val.dtype.itemsize * 8 <= lo_bits else (val & _const(val.dtype, (1 << lo_bits) - 1)).cast(lo_dt)
+          hi_u = (val >> _const(val.dtype, lo_bits)).cast(hi_dt)
+          block_assigns[lo_var] = env[lo_var] = lo_u
+          block_assigns[hi_var] = env[hi_var] = hi_u
+          if assigns is not None: assigns.extend([(f'{lo_var}.{lo_type}', lo_u), (f'{hi_var}.{hi_type}', hi_u)])
           i += 1
           continue
 

test/mockgpu/mockgpu.py

@@ -1,4 +1,4 @@
-import ctypes, time, os, builtins, fcntl
+import ctypes, time, os, builtins, fcntl, typing
 from tinygrad.helpers import DEV
 from tinygrad.runtime.support.hcq import FileIOInterface
 from tinygrad.runtime.autogen import libc
@@ -9,7 +9,7 @@ start = time.perf_counter()
 
 drivers = [cls() for t in DEV.value if (cls:={"MOCKPCI+AMD": AMDriver, "MOCKKFD+AMD": AMDDriver, "MOCK+AMD": AMDDriver, "MOCKUSB+AMD": AMUSBDriver,
                                               "MOCK+NV": NVDriver}.get(f"{t.interface}+{t.device}"))]
-tracked_fds = {}
+tracked_fds: dict[int, typing.Any] = {}
 
 original_memoryview = builtins.memoryview
 class TrackedMemoryView:

test/mockgpu/usb.py

@@ -6,38 +6,25 @@ class MockUSB:
   def __init__(self, mem):
     self.mem = mem
   def read(self, address, size): return bytes(self.mem[address:address+size])
-  def write(self, address, data, ignore_cache=False): self.mem[address:address+len(data)] = data
-  def pcie_mem_req(self, address, value=None, size=1):
-    if value is None: return int.from_bytes(self.mem[address:address+size], "little")
-    else: self.mem[address:address+size] = value.to_bytes(size, "little")
-  def pcie_mem_write(self, address, values, size):
-    for i, value in enumerate(values): self.pcie_mem_req(address + i * size, value, size)
+  def write(self, address, data): self.mem[address:address+len(data)] = data
+  def pcie_mem_read(self, address, nbytes): return bytes(self.mem[address:address+nbytes])
+  def pcie_mem_write(self, address, data): self.mem[address:address+len(data)] = data
 
 # *** ASM24 Controller Mock ***
 
 _mock_usb_state: MockASM24State|None = None
 
 class MockASM24State:
-  """Mock ASM24 controller: XRAM memory map, DMA windows, TLP engine, PCI config space.
+  """Mock custom ASM24 controller: XRAM, DMA windows, PCI config space, and GPU BARs.
 
   Memory map (64KB XRAM):
     0xA000-0xAFFF: DMA window -> sys 0x820000
     0xB000-0xB1FF: DMA window -> sys 0x800000
-    0xB200-0xB7FF: PCI MMIO (TLP engine)
+    0xB200-0xB7FF: controller PCI MMIO
     0xF000-0xFFFF: DMA window -> sys 0x200000 (512KB)
   """
   XRAM_SIZE = 0x10000
 
-  TLP_FMT_TYPE = 0xB210
-  TLP_BYTE_EN = 0xB217
-  TLP_ADDR_LO = 0xB218
-  TLP_ADDR_HI = 0xB21C
-  TLP_DATA = 0xB220
-  TLP_COMPL = 0xB22A
-  TLP_TRIGGER = 0xB254
-  TLP_LINK_STATUS = 0xB284
-  TLP_STATUS = 0xB296
-
   def __init__(self, gpu, driver, vram_size:int, doorbell_size:int, mmio_size:int):
     self.gpu, self.driver = gpu, driver
     self._xram = bytearray(self.XRAM_SIZE)
@@ -93,53 +80,7 @@ class MockASM24State:
       if ctrl_addr <= addr < ctrl_addr + dma_size:
         (ctypes.c_ubyte * 1).from_address(host_addr + (addr - ctrl_addr))[0] = value
         return
-    if addr == self.TLP_STATUS:
-      self._xram[addr] &= ~value & 0xFF
-      return
     self._xram[addr] = value
-    if addr == self.TLP_TRIGGER and value == 0x0F: self._process_tlp()
-
-  # --- TLP engine ---
-
-  def _process_tlp(self):
-    fmt_type, byte_en = self._xram[self.TLP_FMT_TYPE], self._xram[self.TLP_BYTE_EN]
-    addr_lo = int.from_bytes(self._xram[self.TLP_ADDR_LO:self.TLP_ADDR_LO+4], 'big')
-    addr_hi = int.from_bytes(self._xram[self.TLP_ADDR_HI:self.TLP_ADDR_HI+4], 'big')
-    address = addr_lo | (addr_hi << 32)
-
-    size, offset, tmp = 0, 0, byte_en
-    while tmp and not (tmp & 1):
-      offset += 1
-      tmp >>= 1
-    while tmp:
-      size += tmp & 1
-      tmp >>= 1
-
-    is_write, is_cfg = bool(fmt_type & 0x40), (fmt_type & 0xbe) == 0x04
-
-    if is_cfg:
-      bus, dev, fn, byte_addr = (address >> 24) & 0xFF, (address >> 19) & 0x1F, (address >> 16) & 0x7, address & 0xFFC
-      if is_write:
-        data = int.from_bytes(self._xram[self.TLP_DATA:self.TLP_DATA+4], 'big')
-        self._cfg_write(bus, dev, fn, byte_addr + offset, (data >> (8 * offset)) & ((1 << (8 * size)) - 1), size)
-      else:
-        self._xram[self.TLP_DATA:self.TLP_DATA+4] = int.from_bytes(self._get_cfg(bus, dev, fn)[byte_addr:byte_addr+4], 'little').to_bytes(4, 'big')
-      self._xram[self.TLP_COMPL:self.TLP_COMPL+2] = (4).to_bytes(2, 'big')
-      self._xram[self.TLP_LINK_STATUS] = 0x01 if not is_write else 0x00
-      self._xram[self.TLP_STATUS] = 0x02
-      return
-
-    if is_write:
-      data = int.from_bytes(self._xram[self.TLP_DATA:self.TLP_DATA+4], 'big')
-      self._pcie_dispatch(address + offset, (data >> (8 * offset)) & ((1 << (8 * size)) - 1), size)
-    else:
-      result = self._pcie_dispatch(address + offset, None, size)
-      if result is not None:
-        self._xram[self.TLP_DATA:self.TLP_DATA+4] = ((result << (8 * offset)) & 0xFFFFFFFF).to_bytes(4, 'big')
-
-    self._xram[self.TLP_COMPL:self.TLP_COMPL+2] = (size & 0xFFF).to_bytes(2, 'big')
-    self._xram[self.TLP_LINK_STATUS] = 0x01 if not is_write else 0x00
-    self._xram[self.TLP_STATUS] = 0x02
 
   def _cfg_write(self, bus:int, dev:int, fn:int, byte_addr:int, val:int, size:int):
     cfg = self._get_cfg(bus, dev, fn)
@@ -168,49 +109,104 @@ class MockASM24State:
     # Generic config write
     for i in range(size): cfg[byte_addr + i] = (val >> (8 * i)) & 0xFF
 
-  def _pcie_dispatch(self, address:int, value:int|None, size:int) -> int|None:
+  def _find_bar(self, address:int, size:int) -> tuple[int, int]:
     for reg_off, (bar_addr, bar_size) in self._bar_addrs.items():
-      if bar_addr <= address < bar_addr + bar_size:
-        offset = address - bar_addr
-        if reg_off == 0x10:  # BAR0 - VRAM
-          if value is None: return int.from_bytes(bytes(self.gpu.vram[offset:offset+size]), "little")
-          self.gpu.vram[offset:offset+size] = list(value.to_bytes(size, "little"))
-          return None
-        if reg_off == 0x18:  # BAR2 - Doorbell
-          if value is None: return int.from_bytes(bytes(self._doorbell[offset:offset+size]), "little")
-          for i, b in enumerate(value.to_bytes(size, "little")): self._doorbell[offset + i] = b
-          self.driver._emulate_execute()
-          return None
-        if reg_off == 0x24:  # BAR5 - MMIO
-          if value is None: return self.gpu.mmio[offset // 4]
-          self.gpu.mmio[offset // 4] = value
-          return None
-    raise ValueError(f"PCIe address {address:#x} not mapped to any BAR")
+      if bar_addr <= address and address + size <= bar_addr + bar_size: return reg_off, address - bar_addr
+    raise ValueError(f"PCIe range {address:#x}+{size:#x} not mapped to any BAR")
 
-  # --- CDB processing (called by MockUSB3.send_batch) ---
+  def _pcie_read(self, address:int, size:int) -> bytes:
+    reg_off, offset = self._find_bar(address, size)
+    if reg_off == 0x10: return bytes(self.gpu.vram[offset:offset+size])
+    if reg_off == 0x18: return bytes(self._doorbell[offset:offset+size])
+    if reg_off == 0x24: return bytes((self.gpu.mmio[(offset+i)//4] >> (8*((offset+i)&3))) & 0xFF for i in range(size))
+    raise RuntimeError(f"unsupported BAR register {reg_off:#x}")
 
-  def process_cdb(self, cdb:bytes, rlen:int, send_data:bytes|None) -> bytes|None:
-    op = cdb[0]
-    if op == 0xE5:  # write byte
-      self._xram_write_byte(((cdb[2] << 16) | (cdb[3] << 8) | cdb[4]) & 0xFFFF, cdb[1])
-      return None
-    if op == 0xE4:  # read
-      return self._xram_read(((cdb[2] << 16) | (cdb[3] << 8) | cdb[4]) & 0xFFFF, cdb[1])
-    if op == 0x8A and send_data is not None and 0xF000 in self._dma_regions:  # SCSI write
-      host_addr, dma_size = self._dma_regions[0xF000]
-      ctypes.memmove(host_addr, send_data, min(len(send_data), dma_size))
+  def _pcie_write(self, address:int, data:bytes):
+    reg_off, offset = self._find_bar(address, len(data))
+    if reg_off == 0x10: self.gpu.vram[offset:offset+len(data)] = list(data)
+    elif reg_off == 0x18:
+      self._doorbell[offset:offset+len(data)] = list(data)
+      self.driver._emulate_execute()
+    elif reg_off == 0x24:
+      updates: dict[int, int] = {}
+      for i, byte in enumerate(data):
+        idx, shift = (offset+i)//4, 8*((offset+i)&3)
+        updates[idx] = (updates.get(idx, self.gpu.mmio[idx]) & ~(0xFF << shift)) | (byte << shift)
+      for idx, val in updates.items(): self.gpu.mmio[idx] = val
+    else: raise RuntimeError(f"unsupported BAR register {reg_off:#x}")
+
+  def _pcie_dispatch(self, address:int, value:int|None, size:int) -> int|None:
+    if value is None: return int.from_bytes(self._pcie_read(address, size), 'little')
+    self._pcie_write(address, value.to_bytes(size, 'little'))
     return None
 
 class MockUSB3:
   @classmethod
   def list_devices(cls, vendor, dev): return [(0, "usb:mock")]
   def __init__(self, *args, **kwargs):
-    self.product, self.is_custom = "", False
-  def send_batch(self, cdbs:list[bytes], idata:list[int]|None=None, odata:list[bytes|None]|None=None) -> list[bytes|None]:
+    self.product = "custom mock"
+    self._bulk_read_op: tuple[str, int, int]|None = None
+    self._bulk_write_op: tuple[str, int, int]|None = None
+    self._f0_reply = bytes(8)
+
+  @property
+  def state(self) -> MockASM24State:
     assert _mock_usb_state is not None
-    idata, odata = idata or [0] * len(cdbs), odata or [None] * len(cdbs)
-    results: list[bytes|None] = []
-    for cdb, rlen, sdata in zip(cdbs, idata, odata):
-      result = _mock_usb_state.process_cdb(cdb, rlen, sdata)
-      results.append(result if rlen > 0 else None)
-    return results
+    return _mock_usb_state
+
+  def control_write(self, request:int, value:int=0, index:int=0, data:bytes=b'', timeout:int=1000):
+    if request == 0xF3:
+      self.state._xram[0xB450] = 0x78 if value else 0
+    elif request == 0xE5:
+      self.state._xram_write_byte(value, index)
+    elif request == 0xF2:
+      op = ("sram_read" if value & 0x8000 else "sram_write", 0xF000, (value & 0x7FFF) * 512)
+      if value & 0x8000: self._bulk_read_op = op
+      else: self._bulk_write_op = op
+    elif request == 0xF0:
+      address_lo, address_hi, payload = struct.unpack('<III', data)
+      address, fmt_type, byte_en = address_lo | (address_hi << 32), value & 0xFF, value >> 8
+      if index == 1: self._bulk_write_op = ("pcie_write", address, payload * 4)
+      elif index == 2: self._bulk_read_op = ("pcie_read", address, payload * 4)
+      else:
+        assert index == 0 and byte_en
+        offset = (byte_en & -byte_en).bit_length() - 1
+        size, is_write, is_cfg = byte_en.bit_count(), bool(fmt_type & 0x40), (fmt_type & 0xBE) == 0x04
+        if is_cfg:
+          bus, dev, fn, byte_addr = (address >> 24) & 0xFF, (address >> 19) & 0x1F, (address >> 16) & 0x7, address & 0xFFC
+          if is_write: self.state._cfg_write(bus, dev, fn, byte_addr + offset, (payload >> (8 * offset)) & ((1 << (8 * size))-1), size)
+          else: payload = int.from_bytes(self.state._get_cfg(bus, dev, fn)[byte_addr:byte_addr+4], 'little')
+        elif is_write:
+          self.state._pcie_dispatch(address + offset, (payload >> (8 * offset)) & ((1 << (8 * size))-1), size)
+        else: payload = (self.state._pcie_dispatch(address + offset, None, size) or 0) << (8 * offset)
+        self._f0_reply = struct.pack('<I', payload & 0xFFFFFFFF) + bytes(4)
+    else: raise ValueError(f"unsupported control OUT request 0x{request:02X}")
+
+  def control_read(self, request:int, length:int, value:int=0, index:int=0, timeout:int=1000) -> memoryview:
+    if request == 0xE4: data = self.state._xram_read(value, length)
+    elif request == 0xF0: data = self._f0_reply
+    else: raise ValueError(f"unsupported control IN request 0x{request:02X}")
+    return memoryview(data[:length])
+
+  def bulk_write(self, data:bytes, timeout:int=1000):
+    assert self._bulk_write_op is not None
+    op, address, size = self._bulk_write_op
+    assert len(data) == size
+    if op == "sram_write":
+      host_addr, region_size = self.state._dma_regions[address]
+      ctypes.memmove(host_addr, data, min(len(data), region_size))
+    elif op == "pcie_write": self.state._pcie_write(address, data)
+    else: raise RuntimeError(f"cannot bulk write for {op}")
+    self._bulk_write_op = None
+
+  def bulk_read(self, length:int, timeout:int=1000) -> memoryview:
+    assert self._bulk_read_op is not None
+    op, address, size = self._bulk_read_op
+    assert length == size
+    if op == "sram_read":
+      host_addr, region_size = self.state._dma_regions[address]
+      data = bytes((ctypes.c_ubyte * min(length, region_size)).from_address(host_addr))
+    elif op == "pcie_read": data = self.state._pcie_read(address, length)
+    else: raise RuntimeError(f"cannot bulk read for {op}")
+    self._bulk_read_op = None
+    return memoryview(data)

test/null/test_const_folding.py

@@ -1,6 +1,6 @@
 import unittest, itertools, math
 from tinygrad import Tensor, dtypes, Context
-from tinygrad.dtype import DType, ConstType
+from tinygrad.dtype import DType, ConstType, Invalid
 from tinygrad.uop.ops import Ops, UOp
 from test.helpers import full_rewrite
 import numpy as np
@@ -34,6 +34,24 @@ class TestUnaryOpsConstFolding(unittest.TestCase):
     x = x.clip(0, 1).realize()
     _check_ast_count(1, x.neg())
 
+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))
+
+  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))
+
+  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))
+
+  def test_invalid_poison(self):
+    self.assertIs(UOp.invalid().alu(Ops.CDIV, UOp.const(dtypes.weakint, 0)).simplify().arg, Invalid)
+
 class TestBinaryOpsConstFolding(unittest.TestCase):
   def test_add_literal_zero(self):
     _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + 0)

test/null/test_device.py

@@ -69,11 +69,13 @@ class TestDevice(unittest.TestCase):
   @unittest.skipIf(WIN, "skipping windows test") # TODO: subprocess causes memory violation?
   def test_env_overwrite_default_compiler(self):
     if Device.DEFAULT == "CPU":
-      from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangCompiler
+      from tinygrad.runtime.support.compiler_cpu import ClangCompiler
+      from tinygrad.runtime.support.compiler_llvm import CPULLVMCompiler
       try: _, _ = CPULLVMCompiler(), ClangCompiler()
       except Exception as e: self.skipTest(f"skipping compiler test: not all compilers: {e}")
 
-      imports = "from tinygrad import Device; from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangCompiler"
+      imports = ("from tinygrad import Device; from tinygrad.runtime.support.compiler_cpu import ClangCompiler; "
+                 "from tinygrad.runtime.support.compiler_llvm import CPULLVMCompiler")
       subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, CPULLVMCompiler)"'],
                         shell=True, check=True, env={**os.environ, "DEV": "CPU:LLVM"})
       subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, ClangCompiler)"'],
@@ -81,11 +83,13 @@ class TestDevice(unittest.TestCase):
       subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, ClangCompiler)"'],
                         shell=True, check=True, env={**os.environ, "DEV": "CPU:CLANG"})
     elif Device.DEFAULT == "AMD":
-      from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler
+      from tinygrad.runtime.support.compiler_amd import HIPCompiler
+      from tinygrad.runtime.support.compiler_llvm import AMDLLVMCompiler
       try: _, _ = HIPCompiler(Device[Device.DEFAULT].arch), AMDLLVMCompiler(Device[Device.DEFAULT].arch)
       except Exception as e: self.skipTest(f"skipping compiler test: not all compilers: {e}")
 
-      imports = "from tinygrad import Device; from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler"
+      imports = ("from tinygrad import Device; from tinygrad.runtime.support.compiler_amd import HIPCompiler; "
+                 "from tinygrad.runtime.support.compiler_amd import AMDLLVMCompiler")
       subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, AMDLLVMCompiler)"'],
                         shell=True, check=True, env={**os.environ, "DEV": "AMD:LLVM"})
       subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, HIPCompiler)"'],
@@ -96,7 +100,8 @@ class TestDevice(unittest.TestCase):
 
   @unittest.skipIf(WIN, "skipping windows test")
   def test_env_online(self):
-    from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangCompiler
+    from tinygrad.runtime.support.compiler_cpu import ClangCompiler
+    from tinygrad.runtime.support.compiler_llvm import CPULLVMCompiler
     try: _, _ = CPULLVMCompiler(), ClangCompiler()
     except Exception as e: self.skipTest(f"skipping compiler test: not all compilers: {e}")
 
@@ -111,7 +116,7 @@ class TestDevice(unittest.TestCase):
 
   @unittest.skipIf(Device.DEFAULT != "CPU", "only run on CPU")
   def test_compiler_autodetect_fallback(self):
-    from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler
+    from tinygrad.runtime.support.compiler_llvm import CPULLVMCompiler
 
     try: CPULLVMCompiler()
     except Exception as e: self.skipTest(f"skipping: LLVM not available: {e}")

test/null/test_dtype_spec.py

@@ -1,6 +1,6 @@
 import unittest, math, struct, operator
 from tinygrad import Tensor, Device
-from tinygrad.dtype import DTYPES_DICT, dtypes, truncate, float_to_fp16, float_to_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
+from tinygrad.dtype import DTYPES_DICT, dtypes, Invalid, truncate, float_to_fp16, float_to_bf16, _to_np_dtype, least_upper_dtype, least_upper_float
 
 from tinygrad.helpers import getenv
 from hypothesis import given, settings, strategies as strat
@@ -57,6 +57,7 @@ class TestHelpers(unittest.TestCase):
 
   def test_from_py(self):
     assert dtypes.from_py(True) == dtypes.bool
+    assert dtypes.from_py(Invalid) == dtypes.bool
     assert dtypes.from_py(2) == dtypes.default_int
     assert dtypes.from_py(3.0) == dtypes.default_float
     assert dtypes.from_py([]) == dtypes.default_float
@@ -223,30 +224,16 @@ class TestTypePromotion(unittest.TestCase):
     assert least_upper_dtype(dtypes.fp8e5m2, dtypes.uint64) == dtypes.fp8e5m2
 
   def test_weakint_promo(self):
-    # weakint with itself is weakint
-    assert least_upper_dtype(dtypes.weakint, dtypes.weakint) == dtypes.weakint
-    # weakint is above bool
-    assert least_upper_dtype(dtypes.weakint, dtypes.bool) == dtypes.weakint
-    # weakint defers to any concrete int type
-    assert least_upper_dtype(dtypes.weakint, dtypes.int8) == dtypes.int8
-    assert least_upper_dtype(dtypes.weakint, dtypes.uint8) == dtypes.uint8
-    assert least_upper_dtype(dtypes.weakint, dtypes.int16) == dtypes.int16
-    assert least_upper_dtype(dtypes.weakint, dtypes.int32) == dtypes.int32
-    assert least_upper_dtype(dtypes.weakint, dtypes.int64) == dtypes.int64
-    assert least_upper_dtype(dtypes.weakint, dtypes.uint64) == dtypes.uint64
-    # weakint defers to any float type
-    assert least_upper_dtype(dtypes.weakint, dtypes.float16) == dtypes.float16
-    assert least_upper_dtype(dtypes.weakint, dtypes.float32) == dtypes.float32
-    assert least_upper_dtype(dtypes.weakint, dtypes.float64) == dtypes.float64
+    with self.assertRaises(KeyError): least_upper_dtype(dtypes.weakint, dtypes.weakint)
+    with self.assertRaises(KeyError): least_upper_dtype(dtypes.weakint, dtypes.int8)
 
   def test_weakfloat_promo(self):
-    # weakfloat is a float, but like weakint it is not one of dtypes.floats
+    # weakfloat is a float, but is not one of dtypes.floats
     assert dtypes.is_float(dtypes.weakfloat) and dtypes.weakfloat not in dtypes.floats
     # weakfloat with itself is weakfloat
     assert least_upper_dtype(dtypes.weakfloat, dtypes.weakfloat) == dtypes.weakfloat
-    # weakfloat is above bool, weakint and any concrete int (they defer up to it)
+    # weakfloat is above bool and any concrete int (they defer up to it)
     assert least_upper_dtype(dtypes.weakfloat, dtypes.bool) == dtypes.weakfloat
-    assert least_upper_dtype(dtypes.weakfloat, dtypes.weakint) == dtypes.weakfloat
     assert least_upper_dtype(dtypes.weakfloat, dtypes.int32) == dtypes.weakfloat
     assert least_upper_dtype(dtypes.weakfloat, dtypes.uint64) == dtypes.weakfloat
     # weakfloat defers to any concrete float type
@@ -328,6 +315,19 @@ class TestAutoCastType(unittest.TestCase):
     assert (Tensor.ones(4, 4, dtype=dt) + 2).dtype == (dt if dtypes.is_float(dt) or dtypes.is_int(dt) else dtypes.default_int)
     assert (Tensor.ones(4, 4, dtype=dt) + True).dtype == dt
 
+  @given(strat.sampled_from(core_dtypes))
+  def test_pad_scalar(self, dt):
+    t = Tensor.ones(4, dtype=dt)
+    assert t.pad(((1, 1),), value=2.3).dtype == (dt if dtypes.is_float(dt) else dtypes.default_float)
+    assert t.pad(((1, 1),), value=2).dtype == (dt if dtypes.is_float(dt) or dtypes.is_int(dt) else dtypes.default_int)
+    assert t.pad(((1, 1),), value=True).dtype == dt
+
+  @given(strat.sampled_from(core_dtypes))
+  def test_sort(self, dt):
+    # sort pads with dtype.min/max, a scalar of its own dtype
+    assert Tensor([3, 1, 2], dtype=dt).sort()[0].dtype == dt
+    assert Tensor([3, 1, 2], dtype=dt).sort(descending=True)[0].dtype == dt
+
   @given(strat.sampled_from(dtype_floats))
   def test_int_div_int(self, default_float):
     dtypes.default_float = default_float
@@ -428,6 +428,9 @@ class TestAutoCastType(unittest.TestCase):
     self.check_where_alternate_input_other(3.1, True, dtypes.default_float)
     self.check_where_alternate_input_other(3, 2, dtypes.default_int)
     self.check_where_alternate_input_other(3, True, dtypes.default_int)
+
+  def test_where_non_bool_cond_raises(self):
+    with self.assertRaises(RuntimeError): Tensor([1, 0, 2]).where(1, 0)
     self.check_where_alternate_input_other(False, True, dtypes.bool)
 
   @given(strat.sampled_from(core_dtypes), strat.sampled_from(core_dtypes))

test/null/test_gpudims.py

@@ -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.index, 0)
+    flat = UOp.const(dtypes.weakint, 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})

test/null/test_graph_rewrite.py

@@ -107,21 +107,21 @@ class TestFoldingAndReduction(unittest.TestCase):
 class TestModuloAndDivisionFolding(unittest.TestCase):
   def test_full_graph_rewrite_modulo_folding_with_define_var(self):
     # index dtype because div-mod rules only work on index
-    x_var_uop = UOp.variable('x', 0, 100).cast(dtypes.index)
+    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)
 
   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.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)
 
   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.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)
@@ -140,7 +140,7 @@ class TestModuloAndDivisionFolding(unittest.TestCase):
   def test_full_graph_rewrite_modulo_large_divisor(self):
     # index dtype because div-mod rules only work on index
     x_var_uop = UOp.variable('x', 1, 5)
-    self.assertIs(apply_rewrite(x_var_uop.cast(dtypes.index) % 10).render(simplify=False), x_var_uop.render(simplify=False))
+    self.assertIs(apply_rewrite(x_var_uop.cast(dtypes.weakint) % 10).render(simplify=False), x_var_uop.render(simplify=False))
 
   def test_full_graph_rewrite_division_with_remainder(self):
     x_var_uop = UOp.variable('x', 7, 9)

test/null/test_hcq_iface.py

@@ -84,22 +84,22 @@ class TestUSBMMIOInterface(unittest.TestCase):
     self.mmio[2] = 0xFE
     self.assertEqual(full_view[2], 0xFE)
 
-  def test_pcimem_byte(self):
+  def test_pcimem_dword(self):
     usb2 = MockUSB(bytearray(self.size))
-    mmio_pci = USBMMIOInterface(usb2, 0, self.size, fmt='B', pcimem=True)
-    mmio_pci[3] = 0x11
-    self.assertEqual(mmio_pci[3], 0x11)
-    self.assertEqual(usb2.mem[3], 0x11)
+    mmio_pci = USBMMIOInterface(usb2, 0, self.size, fmt='I', pcimem=True)
+    mmio_pci[3] = 0x11223344
+    self.assertEqual(mmio_pci[3], 0x11223344)
+    self.assertEqual(usb2.mem[12:16], b'\x44\x33\x22\x11')
 
   def test_pcimem_slice(self):
     usb3 = MockUSB(bytearray(self.size))
     mmio_pci = USBMMIOInterface(usb3, 0, self.size, fmt='B', pcimem=True)
-    values = [2, 3, 4]
-    mmio_pci[4:7] = values
-    raw = mmio_pci[4:7]
+    values = [2, 3, 4, 5]
+    mmio_pci[4:8] = values
+    raw = mmio_pci[4:8]
     self.assertIsInstance(raw, bytes)
     self.assertEqual(list(raw), values)
-    self.assertEqual([mmio_pci[i] for i in range(4, 7)], values)
+    self.assertEqual(list(usb3.mem[4:8]), values)
 
 if __name__ == "__main__":
   unittest.main()

test/null/test_helpers.py

@@ -1,6 +1,7 @@
 import ctypes, gzip, unittest, timeit, pickle
 from tinygrad import Variable
-from tinygrad.helpers import Context, ContextVar, argfix, colored, word_wrap, is_numpy_ndarray, mv_address, count, all_same
+from tinygrad.helpers import Context, ContextVar, argfix, colored, word_wrap, mv_address, count, all_same
+from tinygrad.tensor import is_numpy_ndarray
 from tinygrad.helpers import merge_dicts, strip_parens, prod, round_up, fetch, fully_flatten, from_mv, to_mv, polyN, time_to_str, cdiv, cmod, getbits
 from tinygrad.helpers import ceildiv, ansistrip, get_shape
 from tinygrad.tensor import Tensor

test/null/test_linearizer_failures.py

@@ -8,12 +8,12 @@ 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.index, 2), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.index, 32), 2, AxisType.LOOP)
-    c3 = ((c1*UOp.const(dtypes.index, 32))+c2)
+    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)
     c4 = UOp.param(1, dtypes.float, (163840,))
-    c5 = UOp.range(UOp.const(dtypes.index, 2560), 0, AxisType.REDUCE)
-    c6 = c4.index(((((((c5//UOp.const(dtypes.index, 8))%UOp.const(dtypes.index, 8))*UOp.const(dtypes.index, 8))+(c5%UOp.const(dtypes.index, 8)))+(((c2*UOp.const(dtypes.index, 40))+(c5//UOp.const(dtypes.index, 64)))*UOp.const(dtypes.index, 64)))+(c1*UOp.const(dtypes.index, 81920))))
+    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))))
     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()

test/null/test_llm_server.py

@@ -1,4 +1,4 @@
-import unittest, threading, time
+import unittest, threading, time, json
 from unittest.mock import Mock
 
 class TestLLMServer(unittest.TestCase):
@@ -7,12 +7,9 @@ class TestLLMServer(unittest.TestCase):
   @classmethod
   def setUpClass(cls):
     cls.mock_tok = Mock()
-    cls.mock_tok.role = Mock(return_value=[100, 101])
     cls.mock_tok.encode = Mock(return_value=[200, 201, 202])
     cls.mock_tok.decode = Mock(return_value="Hello")
     cls.mock_tok.stream_decoder = Mock(return_value=lambda tid=None: "Hello" if tid is not None else "")
-    cls.mock_tok.end_turn = Mock(return_value=[998])
-    cls.mock_tok.prefix = Mock(return_value=[1])
     cls.mock_tok.preset = "llama3"
     cls.mock_tok.bos_id = 1
     cls.mock_tok.eos_id = 999
@@ -20,12 +17,14 @@ class TestLLMServer(unittest.TestCase):
     cls.mock_tok.is_end = Mock(side_effect=lambda tid: tid in (999,))
 
     cls.mock_model = Mock()
+    cls.mock_model.max_context = 4
     cls.mock_model.generate = Mock(side_effect=lambda ids, **kwargs: iter([300, 301, 999]))
     cls.mock_model.get_start_pos = Mock(return_value=0)
 
-    from tinygrad.llm.cli import LLMServer
+    from tinygrad.llm.cli import FallbackTemplate
+    from tinygrad.llm.serve import LLMServer
 
-    cls.server = LLMServer(('127.0.0.1', 0), cls.mock_model, "test-model", cls.mock_tok)
+    cls.server = LLMServer(('127.0.0.1', 0), cls.mock_model, "test-model", cls.mock_tok, FallbackTemplate(cls.mock_tok))
     cls.port = cls.server.server_address[1]
     cls.server_thread = threading.Thread(target=cls.server.serve_forever, daemon=True)
     cls.server_thread.start()
@@ -131,6 +130,16 @@ class TestLLMServer(unittest.TestCase):
     self.assertIsNotNone(resp.usage.prompt_tokens)
     self.assertIsNotNone(resp.usage.completion_tokens)
 
+  def test_context_length_error(self):
+    from openai import BadRequestError
+    self.mock_tok.encode.return_value = [200, 201, 202, 203]
+    try:
+      with self.assertRaises(BadRequestError) as err:
+        self.client.chat.completions.create(model="test-model", messages=[{"role":"user", "content":"too long"}])
+      self.assertEqual(err.exception.code, "context_length_exceeded")
+    finally:
+      self.mock_tok.encode.return_value = [200, 201, 202]
+
   def test_max_tokens_streaming(self):
     self.mock_model.generate = Mock(side_effect=lambda ids, **kwargs: iter([300, 301, 302, 303, 999]))
     stream = self.client.chat.completions.create(
@@ -149,50 +158,6 @@ class TestLLMServer(unittest.TestCase):
     self.assertEqual(resp.choices[0].finish_reason, "length")
     self.assertEqual(resp.usage.completion_tokens, 2)
 
-  def test_assistant_prefill(self):
-    """Last assistant message should be treated as prefill (not a completed turn)."""
-    self.mock_model.generate = Mock(side_effect=lambda ids, **kwargs: iter([300, 999]))
-    captured_ids = []
-    orig_generate = self.mock_model.generate.side_effect
-    def capture_generate(ids, **kwargs):
-      captured_ids.extend(ids)
-      return orig_generate(ids, **kwargs)
-    self.mock_model.generate = Mock(side_effect=capture_generate)
-
-    resp = self.client.chat.completions.create(
-      model="test", messages=[
-        {"role": "user", "content": "Hello"},
-        {"role": "assistant", "content": "Sure"}
-      ], stream=False
-    )
-    # prefill tokens should be in ids: role("assistant") + encode("Sure") but NO end_turn after it
-    # and NO extra role("assistant") appended
-    role_tokens = self.mock_tok.role.call_args_list
-    # last role() call should be for "assistant" (the prefill message), not an extra one
-    self.assertEqual(role_tokens[-1], unittest.mock.call("assistant"))
-    # end_turn should be called once less than role() — the prefill assistant msg doesn't get end_turn
-    # NOTE: this is flaky in random order
-    #self.assertEqual(self.mock_tok.end_turn.call_count, self.mock_tok.role.call_count - 1)
-    self.assertIsNotNone(resp.choices[0].message.content)
-
-  def test_assistant_prefill_not_last(self):
-    """Assistant message that's NOT last should be a normal completed turn."""
-    self.mock_model.generate = Mock(side_effect=lambda ids, **kwargs: iter([300, 999]))
-    self.mock_tok.role.reset_mock()
-    self.mock_tok.end_turn.reset_mock()
-    self.client.chat.completions.create(
-      model="test", messages=[
-        {"role": "user", "content": "Hello"},
-        {"role": "assistant", "content": "Sure"},
-        {"role": "user", "content": "Continue"}
-      ], stream=False
-    )
-    # all messages get end_turn, plus an extra role("assistant") at the end
-    # roles: user, assistant, user, assistant(generation prompt) = 4 role calls
-    # end_turns: user, assistant, user = 3 end_turn calls (one per message)
-    self.assertEqual(self.mock_tok.end_turn.call_count, 3)
-    self.assertEqual(self.mock_tok.role.call_count, 4)
-
   def test_models_endpoint(self):
     import requests as req
     resp = req.get(f"http://127.0.0.1:{self.port}/v1/models")
@@ -203,5 +168,103 @@ class TestLLMServer(unittest.TestCase):
     self.assertEqual(data["data"][0]["id"], "test-model")
     self.assertEqual(data["data"][0]["object"], "model")
 
+class TestLLMToolCalls(unittest.TestCase):
+  """Tool calling through the OpenAI-compatible HTTP API."""
+
+  @classmethod
+  def setUpClass(cls):
+    cls.mock_tok = Mock()
+    cls.mock_tok.encode = Mock(return_value=[200, 201, 202])
+    cls.mock_tok.decode = Mock(return_value="")
+    cls.mock_tok.preset = "qwen2"
+    cls.mock_tok.bos_id, cls.mock_tok.eos_id, cls.mock_tok.eot_id = None, 999, None
+    cls.mock_tok.is_end = Mock(return_value=False)
+
+    cls.mock_model = Mock()
+    cls.mock_model.max_context = 4
+    cls.mock_model.get_start_pos = Mock(return_value=0)
+
+    from tinygrad.llm.serve import LLMServer
+    import jinja2
+    # .items() matches tool-aware templates and ensures OpenAI JSON argument strings are normalized before rendering the next turn.
+    template = jinja2.Template("""{% for m in messages %}{{ m.content or '' }}{% for tc in m.tool_calls or [] %}
+      {% for key, value in tc.function.arguments.items() %}{{ key }}={{ value }}{% endfor %}{% endfor %}{% endfor %}""")
+    cls.server = LLMServer(('127.0.0.1', 0), cls.mock_model, "tool-model", cls.mock_tok, template)
+    cls.port = cls.server.server_address[1]
+    cls.server_thread = threading.Thread(target=cls.server.serve_forever, daemon=True)
+    cls.server_thread.start()
+    time.sleep(0.1)
+
+    from openai import OpenAI
+    cls.client = OpenAI(base_url=f"http://127.0.0.1:{cls.port}/v1", api_key="test")
+
+  @classmethod
+  def tearDownClass(cls):
+    cls.server.shutdown()
+    cls.server.server_close()
+
+  def set_output(self, text:str):
+    pieces = dict(enumerate(text, 1))
+    self.mock_tok.stream_decoder = Mock(return_value=lambda tid=None: pieces[tid] if tid is not None else "")
+    self.mock_model.generate = Mock(side_effect=lambda ids, **kwargs: iter(pieces))
+
+  @staticmethod
+  def tools():
+    return [{"type":"function", "function":{"name":"read", "description":"Read a file",
+      "parameters":{"type":"object", "properties":{"path":{"type":"string"}}, "required":["path"]}}}]
+
+  def test_streaming_tool_call(self):
+    self.set_output('before<tool_call>{"name":"read","arguments":{"path":"README.md"}}</tool_call>')
+    chunks = list(self.client.chat.completions.create(model="tool-model", messages=[{"role":"user", "content":"Read README.md"}],
+                                                     tools=self.tools(), stream=True))
+    self.assertEqual("".join(c.choices[0].delta.content or "" for c in chunks if c.choices), "before")
+    calls = [tc for c in chunks if c.choices for tc in c.choices[0].delta.tool_calls or []]
+    self.assertEqual(len(calls), 1)
+    self.assertEqual(calls[0].function.name, "read")
+    self.assertEqual(json.loads(calls[0].function.arguments), {"path":"README.md"})
+    self.assertEqual(chunks[-1].choices[0].finish_reason, "tool_calls")
+
+  def test_multiple_xml_tool_calls(self):
+    self.set_output("<tool_call><function=read><parameter=path>\"a\"</parameter></function></tool_call>"
+                    "<tool_call><function=read><parameter=path>\"b\"</parameter></function></tool_call>")
+    response = self.client.chat.completions.create(model="tool-model", messages=[{"role":"user", "content":"Read a and b"}],
+                                                   tools=self.tools())
+    self.assertEqual([json.loads(tc.function.arguments)["path"] for tc in response.choices[0].message.tool_calls], ["a", "b"])
+    self.assertEqual(response.choices[0].finish_reason, "tool_calls")
+
+  def test_multiline_tool_argument_preserves_trailing_newline(self):
+    self.set_output("<tool_call>\n<function=write>\n<parameter=content>\nfirst\nsecond\n\n</parameter>\n"
+                    "<parameter=filePath>\nout.txt\n</parameter>\n</function>\n</tool_call>")
+    response = self.client.chat.completions.create(model="tool-model", messages=[{"role":"user", "content":"Write out.txt"}], tools=self.tools())
+    args = json.loads(response.choices[0].message.tool_calls[0].function.arguments)
+    self.assertEqual(args, {"content":"first\nsecond\n", "filePath":"out.txt"})
+
+  def test_invalid_tool_call_becomes_content(self):
+    self.set_output("<tool_call>not a call</tool_call>")
+    response = self.client.chat.completions.create(model="tool-model", messages=[{"role":"user", "content":"Hello"}], tools=self.tools())
+    self.assertEqual(response.choices[0].message.content, "<tool_call>not a call</tool_call>")
+    self.assertIsNone(response.choices[0].message.tool_calls)
+    self.assertEqual(response.choices[0].finish_reason, "stop")
+
+  def test_tool_call_in_reasoning_is_not_executed(self):
+    self.set_output('<think>draft <tool_call>{"name":"wrong","arguments":{}}</tool_call></think>answer')
+    response = self.client.chat.completions.create(model="tool-model", messages=[{"role":"user", "content":"Hello"}], tools=self.tools())
+    self.assertEqual(response.choices[0].message.content, "answer")
+    self.assertIsNone(response.choices[0].message.tool_calls)
+    self.assertEqual(response.choices[0].finish_reason, "stop")
+
+  def test_tool_result_round_trip(self):
+    self.set_output('<tool_call>{"name":"read","arguments":{"path":"README.md"}}</tool_call>')
+    first = self.client.chat.completions.create(model="tool-model", messages=[{"role":"user", "content":"Read README.md"}], tools=self.tools())
+    call = first.choices[0].message.tool_calls[0]
+    self.set_output("done")
+    second = self.client.chat.completions.create(model="tool-model", messages=[
+      {"role":"user", "content":"Read README.md"},
+      {"role":"assistant", "content":None, "tool_calls":[call.model_dump()]},
+      {"role":"tool", "tool_call_id":call.id, "content":"file contents"},
+    ], tools=self.tools())
+    self.assertEqual(second.choices[0].message.content, "done")
+    self.assertEqual(second.choices[0].finish_reason, "stop")
+
 if __name__ == '__main__':
   unittest.main()

test/null/test_llm_tokenizer.py

@@ -1,5 +1,5 @@
-import unittest, base64, functools, sys
-from tinygrad.llm.cli import SimpleTokenizer
+import unittest, base64, functools, re, sys, time, unicodedata
+from tinygrad.llm.cli import SimpleTokenizer, FallbackTemplate
 from tinygrad.helpers import fetch
 
 @unittest.skipIf(sys.platform == 'win32', "fetch race condition on Windows")
@@ -46,6 +46,41 @@ class TestLLMTokenizer(unittest.TestCase):
   def test_llama_repeat(self): self._test_coding(self.llama_tok, "00000000000000000", [ 931, 931, 931, 931, 931, 410 ])
   def test_llama_pat(self): self._test_coding(self.llama_tok, "today\n  \n", [ 31213, 14211 ])
 
+  def test_split_regex_matches_naive_listing(self):
+    # the compacted codepoint ranges must match the same text as listing every codepoint
+    def naive(pre): return "".join(re.escape(chr(cp)) for cp in range(0x323b0) if unicodedata.category(chr(cp)).startswith(pre))
+    r_ws, r_p_N, r_p_L = r"\t\n\x0b\x0c\r\x85" + naive("Z"), naive("N"), naive("L")
+    naive_re = re.compile("(?i:'s|'t|'re|'ve|'m|'ll|'d)|" +
+      f"[^\\r\\n{r_p_N}{r_p_L}]?[{r_p_L}]+|[{r_p_N}]{{1,3}}| ?[^{r_ws}{r_p_N}{r_p_L}]+[\\r\\n]*|[{r_ws}]*[\\r\\n]+|[{r_ws}]+(?![^{r_ws}])|[{r_ws}]+")
+    sample = "hello world 한국어 中文 текст ١٢٣ 123 😊\n  \ttoday\n'équivalent ²³№ "
+    self.assertEqual(SimpleTokenizer({}, {})._split_to_word.findall(sample), naive_re.findall(sample))
+
+  def test_split_regex_speed(self):
+    # the naive listing compiles a 429KB pattern that takes 10+s to match a 225KB prompt; ranges keep it small and fast
+    tok = SimpleTokenizer({}, {})
+    self.assertLess(len(tok._split_to_word.pattern), 100_000)
+    text = "The quick brown fox jumps over the lazy dog. " * 5000
+    tok._split_to_word.findall(text)  # warmup
+    tms = []
+    for _ in range(5):
+      st = time.perf_counter()
+      words = tok._split_to_word.findall(text)
+      tms.append(time.perf_counter() - st)
+    self.assertLess(min(tms), 4)  # best-of-5 is robust to CI scheduling pauses; new code takes ~60ms
+    self.assertEqual(len(words), 50001)
+
+  def test_llama_continued_conversation(self):
+    self._test_coding(self.llama_tok, "hello <|eot_id|>world", [15339, 220, 128009, 14957])
+    self._test_coding(self.llama_tok, "hello <|eot_id|>world again", [15339, 220, 128009, 14957, 1578])
+    self._test_coding(self.llama_tok, "hello changed <|eot_id|>world again", [15339, 5614, 220, 128009, 14957, 1578])
+
+  def test_long_cached_prompt_matches_fresh_tokenization(self):
+    prefix = "system tools\n" * 700 + "<|eot_id|>"
+    first, changed = prefix + "run tower of hanoi", prefix + "run ls /"
+    expected = self.llama_tok.encode(changed)
+    self.llama_tok.encode(first)
+    self.assertEqual(self.llama_tok.encode(changed), expected)
+
   def test_tekken_from_gguf_kv(self):
     kv = {
       "tokenizer.ggml.tokens": ["<unk>", "<s>", "</s>", "[INST]", "[/INST]", "hello"],
@@ -54,10 +89,11 @@ class TestLLMTokenizer(unittest.TestCase):
       "tokenizer.ggml.eos_token_id": 2,
     }
     tok = SimpleTokenizer.from_gguf_kv(kv)
-    self.assertEqual(tok.role("user"), [3])
+    template = FallbackTemplate(tok)
+    self.assertEqual(template.role("user"), "[INST]")
     self.assertEqual(tok.encode("hello"), [5])
-    self.assertEqual(tok.end_turn(), [4])
-    self.assertEqual(tok.role("assistant"), [])
+    self.assertEqual(template.end_turn(), "[/INST]")
+    self.assertEqual(template.role("assistant"), "")
 
   def test_stream_decoder(self):
     """stream_decoder buffers incomplete UTF-8: token 25677 has 3/4 of emoji, token 138 completes it."""

test/null/test_pattern_matcher.py

@@ -51,8 +51,8 @@ class TestPatternMatcher(unittest.TestCase):
       ctx.append(True)
       assert len(x.src) == 0
       return x.replace(src=(UOp(Ops.NOOP),))
-    matcher = PatternMatcher([(UPat(Ops.CONST, src=(), name="x"), fxn)])
-    c1 = UOp.const(dtypes.float, 1.0)
+    matcher = PatternMatcher([(UPat(Ops.NOOP, src=(), name="x"), fxn)])
+    c1 = UOp(Ops.NOOP)
     # second rewrite shouldn't match anything
     ctx = []
     c1 = matcher.rewrite(c1, ctx)

test/null/test_schedule.py

@@ -1472,6 +1472,18 @@ class TestSchedule(unittest.TestCase):
     x.softmax().sum().backward()
     run_linear(*check_schedule(x.grad, 4))
 
+  def test_logsumexp_backward(self):
+    Tensor.manual_seed(0)
+    x = Tensor.randn(4, 12, 64, 64).realize()
+    x.logsumexp(-1).sum().backward()
+    run_linear(*check_schedule(x.grad, 3))
+
+  def test_logcumsumexp_backward(self):
+    Tensor.manual_seed(0)
+    x = Tensor.randn(4, 512).realize()
+    x.logcumsumexp(-1).sum().backward()
+    run_linear(*check_schedule(x.grad, 3))
+
   def test_scaled_dot_product_attention_fusion(self):
     x, y, z, m = (Tensor.empty(32, 8, 16, 16) for _ in range(4))
     out = Tensor.scaled_dot_product_attention(x, y, z, attn_mask=m)

test/null/test_simplify_valid_idx.py

@@ -1,6 +1,6 @@
 import unittest, itertools
 
-from tinygrad.codegen.late.coalese import indexing_simplify
+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.symbolic import simplify_valid, sym, pm_move_where_on_load
@@ -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.index, nmax),), arg=expr)
+def Special(expr, nmax): return UOp(Ops.SPECIAL, src=(UOp.const(dtypes.weakint, nmax),), arg=expr)
 def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
 def Range(n, nmax): return UOp.range(nmax, n)
 
@@ -409,7 +409,7 @@ class TestImageSimplification(unittest.TestCase):
     alu1 = ((idx2*1536)+(ridx4*768)+ridx3+(idx1*24)+(ridx5*3)+-771)//768
     valid = (((idx2+ridx4)<1)!=1)&(((idx1+ridx5)<1)!=1)
     load = get_load_image_uop((128, 768, 4), valid, (alu0, alu1))
-    self.check(load, None, "((((idx1*24)+r3)+(r5*3))+-3)", "(((idx2*2)+r4)+-1)")
+    self.check(load, None, "((((idx1*24)+(r5*3))+r3)+-3)", "(((idx2*2)+r4)+-1)")
 
   def test_simplify7(self):
     # DEBUG=2 ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1397 ALLOWED_GATED_READ_IMAGE=94 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916 # noqa: E501
@@ -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.index, 0)))
+    load = get_load_image_uop((1, 16384, 4), valid, (alu0, UOp.const(dtypes.weakint, 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.index, 0)))
+    load = get_load_image_uop((1, 8192, 4), valid, (alu0, UOp.const(dtypes.weakint, 0)))
     try:
       self.check(load, None, "(lidx2+gidx0*4+lidx0*1024+r0*1024+lidx1*256+-3264)", "0")
     except AssertionError:
@@ -488,18 +488,18 @@ 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.index, 0))
+    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(dtypes.weakint, 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")
 
 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
-    from tinygrad.codegen.late.coalese import indexing_simplify
+    from tinygrad.codegen.late.coalesce import indexing_simplify
     from tinygrad.uop.ops import graph_rewrite
     from tinygrad.uop.symbolic import sym
     buf = UOp.param(0, dtypes.int, (1,))
-    idx = UOp.const(dtypes.index, 0)
+    idx = UOp.const(dtypes.weakint, 0)
     true_gate = UOp.const(dtypes.bool, True)
     index_with_gate = UOp(Ops.INDEX, src=(buf, idx.valid(true_gate)))
     # apply the optimization
@@ -516,7 +516,7 @@ 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.index, 4), r)
+    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
     ranges = self.get_ranges(load.sink())
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 4)
@@ -524,8 +524,8 @@ class TestRangeShrink(unittest.TestCase):
   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.index, 4), r)
-    load2 = get_gated_load_uop(r < UOp.const(dtypes.index, 8), r)
+    load1 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
+    load2 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 8), r)
     ranges = self.get_ranges(UOp.sink(load1, load2))
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 8)
@@ -533,7 +533,7 @@ class TestRangeShrink(unittest.TestCase):
   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.index, 300), r)
+    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 300), r)
     ranges = self.get_ranges(load.sink())
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 204)
@@ -541,7 +541,7 @@ class TestRangeShrink(unittest.TestCase):
   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.index, 4), r)
+    load1 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 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)
@@ -550,7 +550,7 @@ class TestRangeShrink(unittest.TestCase):
   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.index, 4), r)
+    gated_load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 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)
@@ -559,7 +559,7 @@ class TestRangeShrink(unittest.TestCase):
   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.index, 1), r)
+    load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 1), r)
     ranges = self.get_ranges(load.sink())
     self.assertEqual(len(ranges), 0)
 
@@ -568,7 +568,7 @@ class TestRangeShrink(unittest.TestCase):
     from tinygrad.dtype import Invalid
     r = Range(0, 204)
     x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
-    ranges = self.get_ranges(UOp.param(0, dtypes.float, (204,)).index(r).store((r < 4).where(x, 0)).sink())
+    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)
 
@@ -577,7 +577,7 @@ class TestRangeShrink(unittest.TestCase):
     from tinygrad.dtype import Invalid
     r = Range(0, 204)
     x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
-    ranges = self.get_ranges(UOp.param(0, dtypes.float, (204,)).index(r).store((r < 4).where(0, x)).sink())
+    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)
 

test/null/test_tensor_uop_mixin.py

@@ -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.index, 1).stack(UOp.const(dtypes.index, 2)).shape, (2,))
+    self.assertEqual(UOp.const(dtypes.weakint, 1).stack(UOp.const(dtypes.weakint, 2)).shape, (2,))
 
 class TestTensorUOpConv2d(unittest.TestCase):
   def test_conv2d_basic(self):

test/null/test_uop_graph.py

@@ -2,7 +2,7 @@ import unittest, pytest
 from tinygrad import dtypes, Variable
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import DEBUG, Context
-from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, graph_rewrite, GroupOp, AxisType
+from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, graph_rewrite, GroupOp, AxisType, broadcast_axes
 from tinygrad.uop.symbolic import sym
 from test.helpers import to_uops_list
 
@@ -202,7 +202,7 @@ class TestUOpGraph(unittest.TestCase):
 
   def test_where_same_fold(self):
     v = UOp.variable('tmp', 0, 1)
-    c0 = UOp.const(dtypes.index, 0)
+    c0 = UOp.const(dtypes.weakint, 0)
     vc = v != c0
     c1 = UOp.const(dtypes.float, 1.0)
     out = vc.where(c1, c1)
@@ -309,66 +309,6 @@ class TestUOpGraph(unittest.TestCase):
         for uop, const in zip(uops, consts):
           self.assertEqual(uop, const)
 
-  @unittest.skip("no longer testable standalone")
-  def test_wmma_vectorize_fold(self):
-    for i in [2, 4, 8]:
-      vec = UOp(Ops.STACK, dtypes.half, tuple(UOp.const(dtypes.half, 0.0) for _ in range(i)))
-      var = UOp.variable("var", 0, 1, dtypes.half)
-      acc = UOp.variable('acc', 0, 1, dtypes.half)
-      wmma = UOp(Ops.WMMA, src=(vec, var, acc))
-      uops = to_uops_list([wmma])
-      self.assertEqual(uops[0], acc)
-      self.assertEqual(len(uops), 2)  # +1 for SINK
-
-    for i in [2, 4, 8]:
-      var = UOp.variable("var", 0, 1, dtypes.half)
-      vec = UOp(Ops.STACK, dtypes.half, tuple(UOp.const(dtypes.half, 0.0) for _ in range(i)))
-      acc = UOp.variable('acc', 0, 1, dtypes.half)
-      wmma = UOp(Ops.WMMA, src=(var, vec, acc))
-      uops = to_uops_list([wmma])
-      self.assertEqual(uops[0], acc)
-      self.assertEqual(len(uops), 2)  # +1 for SINK
-
-  @unittest.skip("wmma is wrong here, it needs an arg")
-  def test_wmma_vectorize_no_fold(self):
-    for i in [4, 8]:
-      vec = UOp(Ops.STACK, dtypes.half,
-                tuple(UOp.const(dtypes.half, 0.0) for _ in range(i//2)) +
-                tuple(UOp.variable(f'tmp{j}', 0, 1, dtypes.half) for j in range(i//2)))
-      var = UOp.variable(f'tmp{i}', 0, 1, dtypes.half)
-      acc = UOp.variable('acc', 0, 1, dtypes.half)
-      wmma = UOp(Ops.WMMA, src=(vec, var, acc))
-      uops = to_uops_list([wmma])
-      self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
-
-    for i in [4, 8]:
-      var = UOp.variable(f'tmp{i}', 0, 1, dtypes.half)
-      vec = UOp(Ops.STACK, dtypes.half,
-                tuple(UOp.const(dtypes.half, 0.0) for _ in range(i//2)) +
-                tuple(UOp.variable(f'tmp{j}', 0, 1, dtypes.half) for j in range(i//2)))
-      acc = UOp.variable('acc', 0, 1, dtypes.half)
-      wmma = UOp(Ops.WMMA, src=(var, vec, acc))
-      uops = to_uops_list([wmma])
-      self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
-
-    for i in [2, 4, 8]:
-      vec = UOp(Ops.STACK, dtypes.half,
-                tuple(UOp.const(dtypes.half, 1.0 if j == 0 else 0.0) for j in range(i)))
-      var = UOp.variable(f'tmp{i}', 0, 1, dtypes.half)
-      acc = UOp.variable('acc', 0, 1, dtypes.half)
-      wmma = UOp(Ops.WMMA, src=(vec, var, acc))
-      uops = to_uops_list([wmma])
-      self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
-
-    for i in [2, 4, 8]:
-      var = UOp.variable(f'tmp{i}', 0, 1, dtypes.half)
-      vec = UOp(Ops.STACK, dtypes.half,
-                tuple(UOp.const(dtypes.half, 1.0 if j == 0 else 0.0) for j in range(i)))
-      acc = UOp.variable('acc', 0, 1, dtypes.half)
-      wmma = UOp(Ops.WMMA, src=(var, vec, acc))
-      uops = to_uops_list([wmma])
-      self.assertEqual(uops[-2], wmma)  # -2 to skip SINK
-
   def test_cast_alu_fold(self):
     d0 = UOp.param(0, dtypes.bool, (1,))
     d1 = UOp.param(1, dtypes.int, (1,))
@@ -484,16 +424,16 @@ class TestUOpGraph(unittest.TestCase):
     # 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.index, 512), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.index, 250), 2, AxisType.LOOP)
+    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
+    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
     c3 = UOp.param(1, dtypes.int, (512,))
     c4 = c3.index(c1)
-    c5 = UOp.range(UOp.const(dtypes.index, 240), 0, AxisType.REDUCE)
-    c6 = ((c2*UOp.const(dtypes.index, 240))+c5)
+    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(dtypes.weakint, 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.index, 250))+c2)).store(c9).end(c1, c2)
+    c10 = c0.index(((c1*UOp.const(dtypes.weakint, 250))+c2)).store(c9).end(c1, c2)
     uops = to_uops_list([c10])
     for u in uops:
       self.assertNotEqual(u.dtype, dtypes.long)
@@ -501,19 +441,19 @@ 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.index, 512), 1, AxisType.LOOP)
-    c2 = UOp.range(UOp.const(dtypes.index, 250), 2, AxisType.LOOP)
+    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
+    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
     c3 = UOp.param(1, dtypes.int, (512,))
     c4 = c3.index(c1)  # c4 is a load
-    c5 = UOp.range(UOp.const(dtypes.index, 240), 0, AxisType.REDUCE)
-    c6 = ((c2*UOp.const(dtypes.index, 240))+c5)
+    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(dtypes.weakint, 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.index)
-    comparison = (loaded_idx + c5) < UOp.const(dtypes.index, 60000)
+    loaded_idx = c4.cast(dtypes.weakint)
+    comparison = (loaded_idx + c5) < UOp.const(dtypes.weakint, 60000)
     c9 = comparison.where(c8.cast(dtypes.uint).cast(dtypes.uchar), 0).reduce(c5, arg=Ops.ADD)
-    c10 = c0.index(((c1*UOp.const(dtypes.index, 250))+c2)).store(c9).end(c1, c2)
+    c10 = c0.index(((c1*UOp.const(dtypes.weakint, 250))+c2)).store(c9).end(c1, c2)
     uops = to_uops_list([c10])
     for u in uops:
       self.assertNotEqual(u.dtype, dtypes.long)
@@ -767,5 +707,16 @@ class TestUOpBroadcast(unittest.TestCase):
     c = a + b
     self.assertEqual(c.op, Ops.ADD)
 
+  def test_broadcast_axes(self):
+    t = Variable("t", 1, 10)
+    self.assertEqual(broadcast_axes((4, 8), (4, 8)), ())
+    self.assertEqual(broadcast_axes((8,), (4, 8)), (0,))
+    self.assertEqual(broadcast_axes((), (4, 8)), (0, 1))
+    self.assertEqual(broadcast_axes((3, 1), (4, 3, 8)), (0, 2))
+    self.assertEqual(broadcast_axes((1, 8), (1, 8)), ())
+    self.assertEqual(broadcast_axes((t, 8), (t, 8)), ())
+    self.assertEqual(broadcast_axes((1, 8), (t, 8)), (0,))
+    with self.assertRaises(RuntimeError): broadcast_axes((4, 8), (8,))
+
 if __name__ == '__main__':
   unittest.main(verbosity=2)

test/null/test_uop_symbolic.py

@@ -11,13 +11,13 @@ 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.index: s_eval = UOp.const(dtypes.index, s_eval)
+  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)
   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.index): return UOp.variable(name,min_val,max_val,dtype)
-def uconst(val): return UOp.const(dtypes.index, val)
+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 usum(ops): return functools.reduce(lambda x,y: x+y, ops)
 def uand(ops): return functools.reduce(lambda x,y: x*y, ops)
 
@@ -247,12 +247,12 @@ class TestSymbolic(unittest.TestCase):
     self.assertEqual((Variable("x", -10, 0)%Variable("y", 1, 10))._min_max, (0, 9))
 
   def test_range_div_its_symbolic_bound(self):
-    a = Variable("a", 1, 10, dtypes.index)
+    a = Variable("a", 1, 10, dtypes.weakint)
     ridx0 = UOp.range(a+2, 0)
     self.helper_test_variable(ridx0//(a+2), 0, 0, "0")
 
   def test_range_mod_its_symbolic_bound(self):
-    a = Variable("a", 1, 10, dtypes.index)
+    a = Variable("a", 1, 10, dtypes.weakint)
     ridx = UOp.range(a+2, 0)
     self.helper_test_variable(ridx%(a+2), 0, 11, "r0")
 
@@ -333,7 +333,7 @@ class TestSymbolic(unittest.TestCase):
   def test_mod_mod_wrong_sign(self):
     v1=Variable("v1", 0, 128)
     v3=Variable("v3", 0, 7)
-    self.helper_test_variable((((((v1%2)*2)+((v3+-1)%5))+-2)%5), 0, 4, "((v3+v1%2*2+-3)%5)")
+    self.helper_test_variable((((((v1%2)*2)+((v3+-1)%5))+-2)%5), 0, 4, "((v3+v1%2*2+2)%5)")
 
   def test_mod_mod_wrong_sign2(self):
     v2=Variable("v2", 0, 8)
@@ -365,7 +365,7 @@ class TestSymbolic(unittest.TestCase):
 
   def test_div_const_div_wrong_sign_divisor(self):
     a = Variable("a", 0, 124)
-    self.helper_test_variable(((a+10)//-2+10)//-4, -2, 14, "(((a+10)//-2+10)//-4)")
+    self.helper_test_variable(((a+10)//-2+10)//-4, -2, 14, "((a//-2+-3)//-4+-2)")
 
   def test_nested_div_negative_divisor(self):
     # (x//c1)//c2 -> x//(c1*c2) only when c2>0
@@ -437,7 +437,7 @@ class TestSymbolic(unittest.TestCase):
 
   def test_masked_shr_fold(self):
     x = UOp.variable('x', 0, 255, dtype=dtypes.uint32)
-    self.helper_test_variable((x & -4) >> 2, 0, 63, "(x>>2)", test_z3=False)
+    self.helper_test_variable((x & -4) >> 2, 0, 63, "(x>>2)")
 
   def test_bool_or_not_tautology(self):
     a = Variable("a", 0, 10)
@@ -450,8 +450,15 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(c & c.logical_not(), False, False, "False")
 
   def test_mod_factor_negative(self):
-    self.helper_test_variable(usum([uconst(-29), Variable("a", 0, 10), Variable("b", 0, 10)*28]) % 28, 0, 27, "((a+b*28+-29)%28)")
-    self.helper_test_variable(usum([uconst(-29), Variable("a", 0, 100), Variable("b", 0, 10)*28]) % 28, 0, 27, "((a+b*28+-29)%28)")
+    self.helper_test_variable(usum([uconst(-29), Variable("a", 0, 10), Variable("b", 0, 10)*28]) % 28, 0, 27, "((a+27)%28)")
+    self.helper_test_variable(usum([uconst(-29), Variable("a", 0, 100), Variable("b", 0, 10)*28]) % 28, 0, 27, "((a+27)%28)")
+
+  def test_mod_const_reduction_negative_offset(self):
+    # (x+c)%d -> (x+c%d)%d holds for any sign of x+c and d
+    x = Variable("x", 0, 100)
+    self.helper_test_variable((x-50)%3, 0, 2, "((x+1)%3)")
+    self.helper_test_variable((x-50)%-3, -2, 0, "((x+-2)%-3)")
+    self.helper_test_variable((x+7)%-13, -12, 0, "((x+-6)%-13)")
 
   def test_sum_combine_num(self):
     self.helper_test_variable(usum([uconst(29), Variable("a", 0, 10), uconst(-23)]), 6, 16, "(a+6)")
@@ -579,9 +586,9 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(x%12//4*4 + x%4 + x//12*12, 0, 23, "x")
 
   def test_div_neg_cancel(self):
-    self.helper_test_variable((-Variable("idx", 0, 100)+199)//-4 + 50, 0, 25, "((idx*-1+199)//-4+50)")
-    self.helper_test_variable((-Variable("idx", 0, 100)+200)//-4 + 50, 0, 25, "((idx*-1+200)//-4+50)")
-    self.helper_test_variable((-Variable("idx", 0, 100)+201)//-4 + 50, -1, 24, "((idx*-1+201)//-4+50)")
+    self.helper_test_variable((-Variable("idx", 0, 100)+199)//-4 + 50, 0, 25, "((idx*-1+-1)//-4)")
+    self.helper_test_variable((-Variable("idx", 0, 100)+200)//-4 + 50, 0, 25, "(idx*-1//-4)")
+    self.helper_test_variable((-Variable("idx", 0, 100)+201)//-4 + 50, -1, 24, "((idx*-1+-3)//-4+-1)")
     self.helper_test_variable((-Variable("idx", 0, 100))//2, -50, 0, "(idx*-1//2)")
     self.helper_test_variable(Variable("idx", 0, 100)//-2, -50, 0, "(idx//-2)")
 
@@ -658,20 +665,20 @@ class TestSymbolic(unittest.TestCase):
   def test_div_neg_all_range(self):
     gidx = Variable("gidx", 0, 124)
     lidx = Variable("lidx", 0, 7)
-    self.helper_test_variable((-gidx*8-lidx+999)//-4 + 250, 0, 250, "((gidx*-8+lidx*-1+999)//-4+250)")
-    self.helper_test_variable((-gidx*8-lidx+1000)//-4 + 250, 0, 249, "((gidx*-8+lidx*-1+1000)//-4+250)")
-    self.helper_test_variable((-gidx*8-lidx+1001)//-4 + 250, -1, 249, "((gidx*-8+lidx*-1+1001)//-4+250)")
-    self.helper_test_variable((-gidx*8-lidx+1002)//-4 + 250, -1, 249, "((gidx*-8+lidx*-1+1002)//-4+250)")
+    self.helper_test_variable((-gidx*8-lidx+999)//-4 + 250, 0, 250, "((lidx*-1+gidx*-8+-1)//-4)")
+    self.helper_test_variable((-gidx*8-lidx+1000)//-4 + 250, 0, 249, "((lidx*-1+gidx*-8)//-4)")
+    self.helper_test_variable((-gidx*8-lidx+1001)//-4 + 250, -1, 249, "((lidx*-1+gidx*-8+-3)//-4+-1)")
+    self.helper_test_variable((-gidx*8-lidx+1002)//-4 + 250, -1, 249, "((lidx*-1+gidx*-8+-2)//-4+-1)")
 
   def test_div_neg_then_neg(self):
     # taken from arange opts
     lidx0 = Variable("lidx0", 0, 7)
     lidx1 = Variable("lidx1", 0, 7)
     alu2 = -lidx0-lidx1
-    self.helper_test_variable((((alu2+14)//(-32))+4), 3, 4, "((lidx0*-1+lidx1*-1+14)//-32+4)")
-    self.helper_test_variable(-(((alu2+14)//(-32))+4), -4, -3, "((lidx0*-1+lidx1*-1+14)//-32*-1+-4)")
-    self.helper_test_variable((((alu2+134)//(-32))+4), -1, 0, "((lidx0*-1+lidx1*-1+134)//-32+4)")
-    self.helper_test_variable((((alu2+142)//(-32))+4), -1, 0, "((lidx0*-1+lidx1*-1+142)//-32+4)")
+    self.helper_test_variable((((alu2+14)//(-32))+4), 3, 4, "((lidx0*-1+lidx1*-1+-18)//-32+3)")
+    self.helper_test_variable(-(((alu2+14)//(-32))+4), -4, -3, "((lidx0*-1+lidx1*-1+-18)//-32*-1+-3)")
+    self.helper_test_variable((((alu2+134)//(-32))+4), -1, 0, "((lidx0*-1+lidx1*-1+-26)//-32+-1)")
+    self.helper_test_variable((((alu2+142)//(-32))+4), -1, 0, "((lidx0*-1+lidx1*-1+-18)//-32+-1)")
     self.helper_test_variable((((alu2+150)//(-32))+4), -1, -1, "-1")
     self.helper_test_variable((((alu2+158)//(-32))+4), -1, -1, "-1")
 
@@ -837,12 +844,12 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(x%(-3) + ((x//(-3))%5)*(-3), -14, 0, "x%-15")
 
   def test_div_mod_recombine_shifted_quotient(self):
-    # when vmin<0 blocks const reduction on the mod side, the quotient is stored const-shifted: (x-50)//3 -> (x+1)//3 - 17.
+    # const reduction stores mod/div const-shifted: (x-50)%3 -> (x+1)%3, (x-50)//3 -> (x+1)//3 - 17.
     # recombine only needs a quotient of some b congruent to base mod div, so the shift folds into the result
     x = Variable("x", 0, 100)
     y = Variable("y", 0, 99)
     self.helper_test_variable((x-50)%3 + ((x-50)//3)*3, -50, 50, "(x+-50)")                # shifted literal quotient
-    self.helper_test_variable((x-50)%3 + (((x-50)//3)%5)*3, 0, 14, "((x+-50)%15)")         # shift inside the partial's mod
+    self.helper_test_variable((x-50)%3 + (((x-50)//3)%5)*3, 0, 14, "((x+10)%15)")         # shift inside the partial's mod
     self.helper_test_variable(((y-50)//5)%4 + ((y-50)//20)*4, -10, 9, "(y//5+-10)")        # merged and shifted
 
   def test_div_mod_recombine_in_additive_sum(self):
@@ -872,12 +879,16 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((idx<4).where(idx//4, idx.const_like(-1)), -1, 6, "(idx<4).where((idx//4), -1)")
 
   def test_floordiv_lt(self):
-    # x//d<c <=> x<c*d for d>0
+    # x//d<c <=> x<c*d for d>0, and <=> c*d<x for d<0
     idx = Variable("idx", 0, 24)
     self.helper_test_variable((idx//4<3), 0, 1, "(idx<12)")
     self.helper_test_variable(((idx-20)//4<-3), 0, 1, "(idx<8)")
     self.helper_test_variable(((idx-10)//4<0), 0, 1, "(idx<10)")
-    self.helper_test_variable((idx//-4<-3), 0, 1, "((idx//-4)<-3)")
+    self.helper_test_variable((idx//-4<-3), 0, 1, "(12<idx)")
+    self.helper_test_variable((idx//-4<-5), 0, 1, "(20<idx)")
+    self.helper_test_variable((idx//-4<-6), 0, 0, "False")
+    self.helper_test_variable(((idx-10)//-4<0), 0, 1, "(8<(idx+-2))")
+    self.helper_test_variable(((idx-20)//-4<2), 0, 1, "(12<idx)")
 
   def test_nested_div_mod_negative_inner_divisor(self):
     # (x % (k*c)) // c -> (x // c) % k requires k>0; (x % (k*c)) % c -> x % c is unconditional for c>0
@@ -919,8 +930,8 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(cond.cast(dtypes.int).ne(2), 1, 1, "True")
     self.helper_test_variable(cond.cast(dtypes.int).ne(-1), 1, 1, "True")
     # CAST(bool -> index) folds too
-    self.helper_test_variable(cond.cast(dtypes.index).ne(0), 0, 1, "(a<2)")
-    self.helper_test_variable(cond.cast(dtypes.index).ne(1), 0, 1, "((a<2)!=True)")
+    self.helper_test_variable(cond.cast(dtypes.weakint).ne(0), 0, 1, "(a<2)")
+    self.helper_test_variable(cond.cast(dtypes.weakint).ne(1), 0, 1, "((a<2)!=True)")
 
   def test_where_removal(self):
     cond = Variable("a", 0, 3) < 2
@@ -956,6 +967,28 @@ class TestSymbolic(unittest.TestCase):
     # not combining  # TODO: can combine if one is identity element const
     self.helper_test_variable(aa+ab, 0, 6, "((x<2).where(a, b)+(x<2).where(a, 0))")
 
+  def test_where_combine_cross_zero(self):
+    cond = Variable("x", 0, 3) < 2
+    a = Variable("a", 0, 3)
+    b = Variable("b", 0, 3)
+    self.helper_test_variable(cond.where(a, a.ufix(0)) + cond.where(b.ufix(0), b), 0, 3, "(x<2).where(a, b)")
+    self.helper_test_variable(cond.where(a, a.ufix(0)) + cond.where(a.ufix(0), a), 0, 3, "a")
+
+  def test_where_or_dual(self):
+    m1 = Variable("x", 0, 3) < 2
+    m2 = Variable("y", 0, 3) < 2
+    a = Variable("a", 0, 3)
+    b = Variable("b", 0, 3)
+    self.helper_test_variable(m1.where(a, m2.where(a, b)), 0, 3, "((x<2)|(y<2)).where(a, b)")
+
+  def test_bool_ne_false(self):
+    cond = Variable("x", 0, 3) < 2
+    self.helper_test_variable(cond.ne(False), 0, 1, "(x<2)")
+
+  def test_bitcast_chain(self):
+    a = Variable("a", 0, 3)
+    self.assertIs(graph_rewrite(a.bitcast(dtypes.float32).bitcast(a.dtype), sym), a)
+
   def test_negation_in_where(self):
     cond = Variable("x", 0, 3) < 2
     a = Variable("a", 0, 3)
@@ -1001,7 +1034,6 @@ class TestSymbolic(unittest.TestCase):
     # (a if ((s<5)&(s<6)) else b) -> (a if (s<5) else b)
     self.helper_test_variable(expr, 0, 3, "(s<5).where(a, b)")
 
-  @unittest.expectedFailure
   def test_where_closure_folding(self):
     # cond.where(t, f) where f contains cond.where(a, b) should fold the inner where to b in false branch
     x = Variable("x", 0, 10)
@@ -1011,6 +1043,41 @@ class TestSymbolic(unittest.TestCase):
     # the inner where should be folded: true branch gets -x, false branch gets x
     self.helper_test_variable(outer, -20, 11, "(x<5).where((x*-2), (x+1))")
 
+  def test_where_closure_folding_deep(self):
+    x = Variable("x", 0, 10)
+    cond = x < 5
+    w1 = cond.where(-x, x)
+    w2 = cond.where(w1*2, w1+1)
+    self.helper_test_variable(cond.where(w2*3, w2+7), -60, 18, "(x<5).where((x*-6), (x+8))")
+
+  def test_where_closure_folding_different_cond(self):
+    # a nested where on a different condition is not folded
+    x = Variable("x", 0, 10)
+    a = Variable("a", 0, 3)
+    b = Variable("b", 0, 3)
+    expr = (x<5).where((x<7).where(a, b), (x<7).where(b, a))
+    self.helper_test_variable(expr, 0, 3, "(x<5).where((x<7).where(a, b), (x<7).where(b, a))")
+
+  def test_where_closure_folding_derived_cond(self):
+    # cond is a value inside the branch: (!cond).where(a, b) is b in the true branch
+    x = Variable("x", 0, 10)
+    a = Variable("a", 0, 3)
+    b = Variable("b", 0, 3)
+    c = Variable("c", 0, 3)
+    expr = (x<5).where((x<5).logical_not().where(a, b)*2, c)
+    self.helper_test_variable(expr, 0, 6, "(x<5).where((b*2), c)")
+
+  def test_where_closure_folding_valid(self):
+    # a valid gate on the same cond folds in the true branch, the live else value is kept
+    x = Variable("x", 0, 10)
+    a = Variable("a", 0, 3)
+    cond = x < 5
+    expr = cond.where(a.valid(cond), Variable("c", 0, 3))
+    self.assertIs(graph_rewrite(expr, sym), cond.where(a, Variable("c", 0, 3)))
+    # a same-cond valid gate in the false branch is Invalid there
+    expr = cond.where(Variable("t", 0, 3), a.valid(cond))
+    self.assertIs(graph_rewrite(expr, sym), cond.where(Variable("t", 0, 3), UOp.invalid()))
+
   def test_symbolic_div(self):
     # from symbolic arange
     a = Variable("a", 1, 10)
@@ -1021,7 +1088,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((numerator//denominator)<=0, 1, 1, "True")
 
   def test_symbolic_range_doesnt_collapse(self):
-    r0 = UOp.range((Variable("a", 1, 10)<5).cast(dtypes.index), 0)
+    r0 = UOp.range((Variable("a", 1, 10)<5).cast(dtypes.weakint), 0)
     self.helper_test_variable(r0, 0, 0, "r0")
 
   def test_const_reciprocal(self):
@@ -1043,8 +1110,8 @@ class TestSymbolic(unittest.TestCase):
   def test_nested_mod_negative_range(self):
     # (x%(k*c))%c = x%c for positive c
     x = Variable("x", 0, 1575)
-    self.helper_test_variable(((x + (-1064)) % 512) % 4, 0, 3, "((x+-1064)%4)")
-    self.helper_test_variable(((x + (-1064)) % 512) % 128, 0, 127, "((x+-1064)%128)")
+    self.helper_test_variable(((x + (-1064)) % 512) % 4, 0, 3, "(x%4)")
+    self.helper_test_variable(((x + (-1064)) % 512) % 128, 0, 127, "((x+88)%128)")
 
 class TestSymbolicNumeric(unittest.TestCase):
   def helper_test_numeric(self, f):
@@ -1254,23 +1321,21 @@ class TestSymbolicSymbolicOps(unittest.TestCase):
 """
 
 class TestInvalidIndex(unittest.TestCase):
+  def test_invalid_lift_keeps_live_else(self):
+    ridx = Variable("ridx", 0, 10)
+    cond = ridx < 5
+    expr = cond.where(cond.where(ridx, UOp.invalid()), ridx+100)
+    self.assertIs(expr.simplify(), cond.where(ridx, ridx+100))
+
   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(0, UOp.invalid()), "multiplying an index by 0 should preserve the invalid")
 
-  def test_invalid_comparison_drops_invalid(self):
-    # comparisons return a bool, and bools can't be invalid
-    ridx = Variable("ridx", 0, 10)
-    idx = (ridx<5).where(ridx, UOp.invalid())<3
-    self.assertIs(idx.simplify(), (ridx<3), "comparison of index should drop the invalid")
-    self.assertIs(idx.where(UOp.const(dtypes.int, 1), 0).simplify(), (ridx<3).where(UOp.const(dtypes.int, 1), 0),
-      "comparison of index should drop the invalid")
-
   def test_alu_moves_inside_invalid(self):
     ridx = Variable("ridx", 0, 10)
-    idx = (ridx<5).where(ridx, UOp.invalid())*10
-    self.assertIs(idx.simplify(), (ridx<5).where(ridx*10, UOp.invalid()), "multiplying an index by 0 should preserve the invalid")
+    self.assertIs((10*(ridx<5).where(ridx, UOp.invalid())).simplify(), (ridx<5).where(ridx*10, UOp.invalid()),
+      "Invalid should poison either binary operand position")
 
   def test_merge_invalid_conditions(self):
     ridx0 = Variable("ridx0", 0, 10)
@@ -1289,16 +1354,16 @@ 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.index, (1, 1, Invalid, Invalid))
-    c2 = UOp.const(dtypes.index, (1, Invalid, 1, 1))
-    self.assertIs((c1+c2).simplify(), UOp.const(dtypes.index, (2, Invalid, Invalid, Invalid)))
+    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)))
 
 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.index, 0))
+    index = buf.index(UOp.const(dtypes.weakint, 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

test/null/test_uop_vmin_vmax.py

@@ -127,6 +127,13 @@ class TestVminVmaxProperties(unittest.TestCase):
     self.assertEqual(x.vmin, 0)
     self.assertEqual(x.vmax, 10 >> 2)
 
+  def test_vmin_vmax_cast_unsigned(self):
+    # a fitting source keeps exact bounds: no wrap can occur
+    self.assertEqual(UOp.variable('x', 5, 10).cast(dtypes.uint8)._min_max, (5, 10))
+    # a possibly-negative or too-large source can wrap: conservative
+    self.assertEqual(UOp.variable('x', -1, 10).cast(dtypes.uint8)._min_max, (0, 255))
+    self.assertEqual(UOp.variable('x', 250, 260).cast(dtypes.uint8)._min_max, (0, 255))
+
   def test_vmin_vmax_xor_neg1(self):
     x = UOp.variable('x', 3, 7)
     uop = x ^ -1
@@ -160,7 +167,7 @@ class TestVminVmaxProperties(unittest.TestCase):
     self.assertNotEqual(i.vmin, i.vmax)
 
   def test_vmin_vmax_invalid_vconst(self):
-    x = UOp.const(dtypes.index, (0, 4, Invalid, Invalid))
+    x = UOp.const(dtypes.weakint, (0, 4, Invalid, Invalid))
     self.assertLess(x.vmin, 0)
     self.assertGreater(x.vmax, 4)
 
@@ -364,6 +371,10 @@ class TestConstFactor(unittest.TestCase):
     uop = (x * 3) * 5
     self.assertEqual(uop.const_factor(), 15)  # Constant multipliers are combined (3 * 5 = 15)
 
+  def test_const_factor_variable_multiple_of(self):
+    x = UOp.variable('x', 16, 32, multiple_of=4)
+    self.assertEqual(x.const_factor(), 4)
+
 class TestDivides(unittest.TestCase):
   def test_divides_constant_exact(self):
     # Divides a constant by an exact divisor
@@ -402,5 +413,15 @@ class TestDivides(unittest.TestCase):
     result = uop.divides(3)
     self.assertIsNone(result)  # Cannot divide by 3, since 4 is not divisible by 3
 
+  def test_divides_variable_multiple_of_exact(self):
+    x = UOp.variable('x', 16, 32, multiple_of=4)
+    result = x.divides(4)
+    self.assertIsNotNone(result)
+
+  def test_divides_variable_multiple_of_factor(self):
+    x = UOp.variable('x', 16, 32, multiple_of=4)
+    result = x.divides(2)
+    self.assertIsNotNone(result)
+
 if __name__ == '__main__':
   unittest.main()