Merge branch 'master' into move_gates_to_load_store

better order

.github/workflows/autogen.yml

@@ -33,12 +33,8 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: 'autogen'
-        opencl: 'true'
         amd: 'true'
-        cuda: 'true'
         llvm: 'true'
-        webgpu: 'true'
-        mesa: 'true'
         pydeps: 'pyyaml mako'
     - name: Install autogen support packages
       run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev
@@ -48,7 +44,7 @@ jobs:
         python3 -c "from tinygrad.runtime.autogen import opencl"
         python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv"
         python3 -c "from tinygrad.runtime.autogen import comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
-        python3 -c "from tinygrad.runtime.autogen.am import am, pm4_soc15, pm4_nv, sdma_4_0_0, sdma_5_0_0, sdma_6_0_0, smu_v13_0_0, smu_v13_0_6, smu_v13_0_12, smu_v14_0_2"
+        python3 -c "from tinygrad.runtime.autogen.am import *"
         python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
         python3 -c "from tinygrad.runtime.autogen import llvm"
         python3 -c "from tinygrad.runtime.autogen import webgpu"

.github/workflows/benchmark.yml

@@ -628,7 +628,7 @@ jobs:
     - name: IR3 openpilot compile3 0.11.0 driving_vision
       run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
     - name: openpilot compile3 0.11.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.11.0 dmonitoring
       run: BENCHMARK_LOG=openpilot_0_11_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
@@ -636,7 +636,7 @@ jobs:
     - name: openpilot compile3 0.10.1 driving_vision
       run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
     - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.10.1 dmonitoring
       run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: benchmark MobileNetV2 on DSP

.github/workflows/test.yml

@@ -333,7 +333,7 @@ jobs:
         deps: testing_unit
         python-version: '3.14'
     - name: Test SPEC=2
-      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" -k "not test_conv2d_ceildiv_edge_case" --splits 2 --group ${{ matrix.group }}
 
   fuzzing:
     name: Fuzzing
@@ -417,7 +417,7 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1486 ALLOWED_GATED_READ_IMAGE=17 FLOAT16=1 DEV=CL 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=1486 ALLOWED_GATED_READ_IMAGE=18 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Test openpilot CL compile fp16
         run: FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Test openpilot CL compile fp32 (test correctness)

docs/abstractions4.py

@@ -105,7 +105,7 @@ def example_3_custom_uop(a:Tensor, correct):
 def example_5_custom_assembly(a:Tensor, correct):
   # Kernel class copied from amd_asm_matmul
   class Kernel:
-    def __init__(self, arch='gfx1100'): self.instructions, self.labels, self.pos, self.arch = [], {}, 0, arch
+    def __init__(self): self.instructions, self.labels, self.pos = [], {}, 0
     def label(self, name): self.labels[name] = self.pos
     def emit(self, inst, target=None):
       self.instructions.append(inst)

docs/runtime.md

@@ -5,7 +5,7 @@ tinygrad supports various runtimes, enabling your code to scale across a wide ra
 | Runtime | Description | Compiler Options | Requirements |
 |---------|-------------|------------------|--------------|
 | [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`DEV=NV:PTX`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via [the `DEV` variable](env_vars.md#dev-variable). See [NV interfaces](#nv-interfaces) for details. |
-| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`DEV=AMD:LLVM`)<br>HIP/COMGR (`DEV=AMD:HIP`) | RDNA2 or newer GPUs.<br>You can select an interface via [the `DEV` variable](env_vars.md#dev-variable). See [AMD interfaces](#amd-interfaces) for details. |
+| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`DEV=AMD:LLVM`)<br>HIP/COMGR (`DEV=AMD:HIP`) | CDNA3, CDNA4, RDNA3 or RDNA4 GPUs.<br>You can select an interface via [the `DEV` variable](env_vars.md#dev-variable). See [AMD interfaces](#amd-interfaces) for details. |
 | [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
 | [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for `bfloat` support |
 | [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`DEV=CUDA:PTX`) | NVIDIA GPU with CUDA support |

docs/tensor/elementwise.md

@@ -68,6 +68,7 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.div
 ::: tinygrad.Tensor.idiv
 ::: tinygrad.Tensor.mod
+::: tinygrad.Tensor.fmod
 ::: tinygrad.Tensor.bitwise_xor
 ::: tinygrad.Tensor.bitwise_and
 ::: tinygrad.Tensor.bitwise_or

examples/mlperf/model_train.py

@@ -1357,6 +1357,7 @@ def train_llama3():
       MLLOGGER.event(key=mllog_constants.OPT_LR_WARMUP_STEPS, value=WARMUP_STEPS)
       MLLOGGER.event(key=mllog_constants.NUM_WARMUP_STEPS, value=WARMUP_STEPS)
       MLLOGGER.event(key=mllog_constants.OPT_LR_DECAY_STEPS, value=MAX_STEPS - WARMUP_STEPS)
+      MLLOGGER.event(key=mllog_constants.OPT_LR_DECAY_SCHEDULE, value="cosine with linear warmup")
       MLLOGGER.event(key=mllog_constants.OPT_GRADIENT_CLIP_NORM, value=1.0)
   else:
     MLLOGGER = None
@@ -1418,7 +1419,10 @@ def train_llama3():
 
   for p in optim.params:
     grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
-    p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device).contiguous()
+    if isinstance(p.device, tuple) and p.uop.axis is not None:
+      p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device[0]).shard_(p.device, axis=p.uop.axis).contiguous()
+    else:
+      p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device).contiguous()
   grads = [p.grad for p in optim.params]
 
   scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)
@@ -1433,6 +1437,7 @@ def train_llama3():
     load_state_dict(scheduler, safe_load(fn), realize=False)
 
   fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
+  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts] if hasattr(model, "_fp8_grad_amax") else []
   fp8_inv_scales = list(model._fp8_inv_scale.values())
 
   from tinygrad.nn.state import get_state_dict
@@ -1444,6 +1449,10 @@ def train_llama3():
       idx = next(j for j, p in enumerate(optim.params) if p is w)
       optim.master_params[idx].assign((optim.master_params[idx] * w._inv_scale.reshape(-1, *([1]*(w.ndim-1)))).contiguous())
 
+  # realize everything here
+  if optim.master_params: Tensor.realize(*optim.master_params)
+  Tensor.realize(*optim.params, *fp8_inv_scales, *fp8_amax, *fp8_grad_amax)
+
   @TinyJit
   def minibatch(tokens:Tensor):
     if is_dp: tokens = tokens.to(None).shard(device, 0)
@@ -1460,7 +1469,7 @@ def train_llama3():
       apply_grad(g, new_g.uop)
 
     loss_cpu = loss.flatten().float().to("CPU")
-    return loss_cpu.realize(*grads, *fp8_amax)
+    return loss_cpu.realize(*grads, *fp8_amax, *fp8_grad_amax)
 
   @TinyJit
   def optim_step():
@@ -1635,7 +1644,6 @@ def train_llama3():
         tqdm.write(f"target achieved after {sequences_seen} sequences")
         if MLLOGGER and RUNMLPERF:
           MLLOGGER.end(key=mllog_constants.EPOCH_STOP, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
-          MLLOGGER.event(key=mllog_constants.TRAIN_SAMPLES, value=sequences_seen)
           MLLOGGER.end(key=mllog_constants.RUN_STOP, metadata={mllog_constants.STATUS: mllog_constants.SUCCESS})
         if getenv("CKPT"):
           if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)

examples/mlperf/models/flat_llama.py

@@ -20,55 +20,73 @@ from extra.llama_kernels.rmsnorm import rmsnorm
 from extra.llama_kernels import FP8_MAX, local_abs_max
 
 ASM_GEMM = getenv("ASM_GEMM", 0)
+FUSED_INPUT_QUANTIZE = getenv("FUSED_INPUT_QUANTIZE", 0)
+FUSED_ADD_NORM_MUL_QUANTIZE = getenv("FUSED_ADD_NORM_MUL_QUANTIZE", 0)
+FUSED_SILU_W13 = getenv("FUSED_SILU_W13", 0)
 
 FP8_DTYPE = dtypes.fp8e4m3
 FP8_GRAD_DTYPE = dtypes.fp8e5m2
 
 def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
-  new_amax = (local_abs_max(x) if isinstance(x.device, tuple) else x.abs().max()).detach()
+  new_amax = (local_abs_max(x) if isinstance(x.device, tuple) else x.abs().max()).detach().cast(dtypes.float32)
   scale = FP8_MAX / ((amax_state if amax_state is not None else new_amax) + 1e-8)
   x_scaled = x * scale
   x_clamped = x_scaled + (x_scaled.detach().clamp(-FP8_MAX, FP8_MAX) - x_scaled.detach())  # STE
   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_scale:Tensor|None=None, x_new_amax:Tensor|None=None) -> tuple[Tensor,...]:
+           x_fp8:Tensor|None=None, x_scale:Tensor|None=None, x_new_amax:Tensor|None=None,
+           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
   if not fp8:
     if ASM_GEMM:
       from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
       if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T),)
     return (x @ w.T,)
   assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
-  if x_fp8 is None: x_fp8, x_scale, x_new_amax = quantize_fp8(x, amax_state=amax_x)
+  if x_fp8 is None:
+    if FUSED_INPUT_QUANTIZE and amax_x is not None:
+      from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed
+      x_fp8, x_scale, x_new_amax, _ = quantize_fp8_delayed(x, amax_x, FP8_DTYPE)
+    else:
+      x_fp8, x_scale, x_new_amax = quantize_fp8(x, amax_state=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): return asm_gemm(x_fp8, w.T, x_scale=x_scale, w_scale=w_inv_scale), x_new_amax, x_fp8, w
-  return (x_fp8.dot(w.T, dtype=dtypes.float) * x_scale * w_inv_scale).cast(dtypes.bfloat16), x_new_amax, x_fp8, w
+    if can_use_asm_gemm(x_fp8, w.T):
+      return asm_gemm(x_fp8, w.T, x_scale=x_scale, w_scale=w_inv_scale, grad_amax_state=grad_amax_state), x_new_amax, x_fp8, w
+  return x_fp8.dot(w.T, dtype=dtypes.float) * x_scale * w_inv_scale, x_new_amax, x_fp8, w
 
-def norm_mul_quantize_matmul(x:Tensor, norm:Tensor, amax_x, w_inv_scale, w:Tensor, eps:float):
-  FUSED_NORM_MUL_QUANTIZE = getenv("FUSED_NORM_MUL_QUANTIZE", 0)
-  normed, rrms = rmsnorm(x, eps)
-  if FUSED_NORM_MUL_QUANTIZE:
-    from extra.llama_kernels.fused_mul_quantize_fp8 import fused_mul_quantize_fp8
-    amax_s = amax_x if amax_x is not None else Tensor.full((), 1.0, dtype=dtypes.bfloat16, device=normed.device)
-    x_fp8, x_inv_scale, new_amax = fused_mul_quantize_fp8(normed, norm, amax_s, FP8_DTYPE)
+def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor, 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, x_inv_scale, 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, x_scale=x_inv_scale, x_new_amax=new_amax, grad_amax_state=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)
+  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):
+  if FUSED_ADD_NORM_MUL_QUANTIZE:
+    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
+    x_fp8, x_inv_scale, 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, x_scale=x_inv_scale, x_new_amax=new_amax)
-  else:
-    x = normed * norm
-    out, *ret = matmul(x, w, amax_x=amax_x, w_inv_scale=w_inv_scale)
-  return out, normed, rrms, ret
+    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)
+  return out, h, x_normed, rrms, ret
 
-def silu_w13_matmul(x_w13:Tensor, w2:Tensor, amax_x2, s_2):
-  FUSED_SILU_W13 = getenv("FUSED_SILU_W13", 0)
+def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
+                             amax_x2:Tensor,
+                             grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
   if FUSED_SILU_W13:
     from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
-    amax_s = amax_x2 if amax_x2 is not None else Tensor.full((), 1.0, dtype=dtypes.bfloat16, device=x_w13.device)
-    x2_fp8, x2_inv_scale, new_amax_x2 = fused_quantize_fp8_w13(x_w13, amax_s, FP8_DTYPE)
-    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, x_scale=x2_inv_scale, x_new_amax=new_amax_x2)
-  else:
-    hidden_dim = x_w13.shape[-1] // 2
-    x_w1, x_w3 = x_w13[..., :hidden_dim], x_w13[..., hidden_dim:]
-    out, *ret = matmul(x_w1.silu() * x_w3, w2, amax_x=amax_x2, w_inv_scale=s_2)
+    x2_fp8, x2_inv_scale, new_amax_x2 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13)
+    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, x_scale=x2_inv_scale, x_new_amax=new_amax_x2, grad_amax_state=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)
   return out, ret
 
 class FlatTransformer:
@@ -104,32 +122,33 @@ class FlatTransformer:
     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().requires_grad_(False)
 
-    def _amax(): return Tensor.full((), FP8_MAX).contiguous().requires_grad_(False)
+    def _amax(): return Tensor.full((), FP8_MAX, dtype=dtypes.float32).contiguous().requires_grad_(False)
     names = ["xqkv", "xo", "x13", "x2"]
     self._fp8_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
-    # per-weight inv_scale: single (n_layers,) float32 tensor per weight (kernel reads float* pointers)
+    grad_names = ["xqkv", "xo", "xw13", "xout"]
+    self._fp8_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
     w_names = ["wqkv", "wo", "w13", "w2"]
-    self._fp8_inv_scale = {}
-    for wname, inv_scales in zip(w_names, self._init_inv_scales):
-      self._fp8_inv_scale[wname] = inv_scales.float().contiguous().requires_grad_(False)
+    self._fp8_inv_scale = {wname: inv_scales.float().contiguous().requires_grad_(False)
+                           for wname, inv_scales in zip(w_names, self._init_inv_scales)}
     del self._init_inv_scales
 
   def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02):
-    if getenv("ZEROS", 0): w = Tensor.zeros(self.n_layers, out_features, in_features)
+    if getenv("ZEROS"): w = Tensor.zeros(self.n_layers, out_features, in_features)
     else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std)
-    # per-layer scaled fp8 cast: fill the fp8 range for best precision
     amax = w.abs().flatten(1).max(1).detach()
     scale = FP8_MAX / (amax + 1e-8)
-    self._init_inv_scales.append((amax + 1e-8) / FP8_MAX)  # save for inv_scale init
+    self._init_inv_scales.append((amax + 1e-8) / FP8_MAX)
     return (w * scale.reshape(-1, 1, 1)).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE)
 
   def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                amax_xqkv=None, amax_xo=None, s_qkv=None, s_o=None):
+                amax_xqkv:Tensor, amax_xo:Tensor, s_qkv:Tensor, s_o:Tensor,
+                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor):
     bsz, seqlen, _ = x.shape
     new_amaxs, saves = [], []
 
-    xqkv, normed, rrms, ret = norm_mul_quantize_matmul(x, attention_norm, amax_xqkv, s_qkv, wqkv, self.norm_eps)
-    saves.extend([normed, rrms])
+    xqkv, x_normed, rrms, ret = norm_quantize_matmul(x, attention_norm, wqkv, s_qkv, self.norm_eps,
+                                                     amax_x=amax_xqkv, grad_amax_state=grad_amax_xqkv)
+    saves.extend([x_normed, rrms])
     new_amaxs.extend(ret[:1])
     saves.extend(ret[1:] + [xqkv])
     xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
@@ -139,49 +158,52 @@ class FlatTransformer:
 
     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)
     if getenv("HK_FLASH_ATTENTION"):
       from extra.thunder.amd.fa import flash_attention
       attn, *save = flash_attention(xq, xk, xv, is_causal=True)
       saves.extend(save)
     else:
-      attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True)
-    attn = attn.transpose(1, 2).reshape(bsz, seqlen, -1)
+      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, *ret = matmul(attn, wo, amax_x=amax_xo, w_inv_scale=s_o)
+    out, *ret = matmul(attn, wo, amax_x=amax_xo, w_inv_scale=s_o, grad_amax_state=grad_amax_xo)
     new_amaxs.extend(ret[:1])
     saves.extend(ret[1:] + [out])
     return (out, *new_amaxs, *saves)
 
-  def feed_forward(self, x:Tensor, ffn_norm:Tensor, w13:Tensor, w2:Tensor,
-                   amax_x13=None, amax_x2=None, s_13=None, s_2=None):
+  def feed_forward(self, x:Tensor, residual:Tensor, ffn_norm:Tensor, w13:Tensor, w2:Tensor,
+                   amax_x13:Tensor, amax_x2:Tensor, s_13:Tensor, s_2:Tensor,
+                   grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
     new_amaxs, saves = [], []
 
-    x_w13, normed, rrms, ret = norm_mul_quantize_matmul(x, ffn_norm, amax_x13, s_13, w13, self.norm_eps)
-    saves.extend([normed, rrms])
+    x_w13, h, x_normed, rrms, ret = add_norm_quantize_matmul(x, residual, ffn_norm, w13, s_13, self.norm_eps,
+                                                             amax_x=amax_x13)
+    saves.extend([x_normed, rrms])
     new_amaxs.extend(ret[:1])
     saves.extend(ret[1:] + [x_w13])
 
-    out, ret = silu_w13_matmul(x_w13, w2, amax_x2, s_2)
+    out, ret = silu_w13_quantize_matmul(x_w13, w2, s_2, amax_x2=amax_x2, grad_amax_xw13=grad_amax_xw13, grad_amax_xout=grad_amax_xout)
     new_amaxs.extend(ret[:1])
     saves.extend(ret[1:] + [out])
-    return (out, *new_amaxs, *saves)
+    return (out, h, *new_amaxs, *saves)
 
   @function(precompile=True, precompile_backward=True)
   def run_layer(self, x:Tensor, freqs_cis:Tensor,
                 attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
                 ffn_norm:Tensor, w13:Tensor, w2:Tensor,
-                amax_xqkv=None, amax_xo=None,
-                amax_x13=None, amax_x2=None,
-                s_qkv=None, s_o=None, s_13=None, s_2=None):
+                amax_xqkv:Tensor, amax_xo:Tensor,
+                amax_x13:Tensor, amax_x2:Tensor,
+                s_qkv:Tensor, s_o:Tensor, s_13:Tensor, s_2:Tensor,
+                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor,
+                grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
     attn, *attn_ret = self.attention(x, freqs_cis, attention_norm, wqkv, wo,
-                                     amax_xqkv=amax_xqkv, amax_xo=amax_xo,
-                                     s_qkv=s_qkv, s_o=s_o)
+                                     amax_xqkv=amax_xqkv, amax_xo=amax_xo, s_qkv=s_qkv, s_o=s_o,
+                                     grad_amax_xqkv=grad_amax_xqkv, grad_amax_xo=grad_amax_xo)
     attn_amaxs, attn_saves = attn_ret[:2], attn_ret[2:]
-    h = x + attn
-    ffn, *ffn_ret = self.feed_forward(h, ffn_norm, w13, w2,
-                                      amax_x13=amax_x13, amax_x2=amax_x2,
-                                      s_13=s_13, s_2=s_2)
+    ffn, h, *ffn_ret = self.feed_forward(x, attn, ffn_norm, w13, w2,
+                                                   amax_x13=amax_x13, amax_x2=amax_x2, s_13=s_13, s_2=s_2,
+                                                   grad_amax_xw13=grad_amax_xw13, grad_amax_xout=grad_amax_xout)
     ffn_amaxs, ffn_saves = ffn_ret[:2], ffn_ret[2:]
     h = h + ffn
     return (h, *attn_amaxs, *ffn_amaxs, *attn_saves, *ffn_saves)
@@ -202,28 +224,31 @@ class FlatTransformer:
       self.tok_embeddings.weight.shard_(device, axis=0).realize()
       self.output.shard_(device, axis=1).realize()
       self.freqs_cis.shard_(device, axis=None).realize()
-      for name in self._fp8_amax:
-        for i in range(len(self._fp8_amax[name])):
-          self._fp8_amax[name][i] = self._fp8_amax[name][i].to(device).contiguous().requires_grad_(False)
+      for amax_dict in (self._fp8_amax, self._fp8_grad_amax):
+        for name in amax_dict:
+          for i in range(len(amax_dict[name])):
+            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().requires_grad_(False)
       for name in self._fp8_inv_scale:
         self._fp8_inv_scale[name] = self._fp8_inv_scale[name].to(device).contiguous().requires_grad_(False)
 
   def __call__(self, tokens:Tensor):
     h = self.tok_embeddings(tokens)
     freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    amaxs, inv_scales = self._fp8_amax, self._fp8_inv_scale
+    a, ga, s = self._fp8_amax, self._fp8_grad_amax, self._fp8_inv_scale
     for i in range(self.n_layers):
       h, *ret = self.run_layer(h, freqs_cis,
                                self.attention_norm[i], self.wqkv[i], self.wo[i],
                                self.ffn_norm[i], self.w13[i], self.w2[i],
-                               amax_xqkv=amaxs["xqkv"][i], amax_xo=amaxs["xo"][i],
-                               amax_x13=amaxs["x13"][i], amax_x2=amaxs["x2"][i],
-                               s_qkv=inv_scales["wqkv"][i], s_o=inv_scales["wo"][i],
-                               s_13=inv_scales["w13"][i], s_2=inv_scales["w2"][i])
+                               amax_xqkv=a["xqkv"][i], amax_xo=a["xo"][i],
+                               amax_x13=a["x13"][i], amax_x2=a["x2"][i],
+                               s_qkv=s["wqkv"][i], s_o=s["wo"][i],
+                               s_13=s["w13"][i], s_2=s["w2"][i],
+                               grad_amax_xqkv=ga["xqkv"][i], grad_amax_xo=ga["xo"][i],
+                               grad_amax_xw13=ga["xw13"][i], grad_amax_xout=ga["xout"][i])
       for name, new_val in zip(["xqkv", "xo", "x13", "x2"], ret[:5]):
-        amaxs[name][i].assign(new_val)
+        a[name][i].assign(new_val)
 
-    logits = matmul(self.norm(h).contiguous().contiguous_backward(), self.output[0], fp8=False)[0].contiguous_backward()
+    logits = matmul(self.norm(h), self.output[0], fp8=False)[0]
     return logits
 
 def _get_pads(uop:UOp) -> list[UOp]:
@@ -239,6 +264,11 @@ def apply_grad(grad_buf:Tensor, new_grad:UOp):
     return
   sorted_pads = sorted(pads, key=lambda p: p.marg[0][0] if p.op == Ops.PAD else 0)
   inners = [Tensor(p.src[0] if p.op == Ops.PAD else p, device=grad_buf.device).cast(grad_buf.dtype) for p in sorted_pads]
+  if getenv("FUSED_PAD_GRAD_ACCUM", 0):
+    from extra.llama_kernels.fused_pad_grad_accum import fused_pad_grad_accum, can_fused_pad_grad_accum
+    if can_fused_pad_grad_accum(grad_buf, inners):
+      grad_buf.uop = fused_pad_grad_accum(grad_buf, inners).uop
+      return
   grad_buf.assign(grad_buf + inners[0].cat(*inners[1:], dim=0))
 
 if __name__ == "__main__":

examples/mlperf/optim.py

@@ -81,7 +81,7 @@ class GradAccClipAdamW(Optimizer):
     if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16: return stochastic_round_bf16(new_w)
     if t.dtype in dtypes.fp8s:
       from examples.mlperf.models.flat_llama import FP8_MAX
-      amax = new_w.float().abs().flatten(1).max(1).detach()  # per-layer amax for (n_layers, out, in)
+      amax = new_w.float().abs().max(axis=tuple(range(1, new_w.ndim))).detach()  # per-layer amax for (n_layers, out, in)
       scale = FP8_MAX / (amax + 1e-8)
       fp8_w = (new_w * scale.reshape(-1, *([1]*(new_w.ndim-1)))).clamp(-FP8_MAX, FP8_MAX).cast(t.dtype)
       if hasattr(t, '_inv_scale'):

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama405b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -2,7 +2,6 @@
 
 export PYTHONPATH="."
 export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
 export DEVICE_IN_FUNCTION_BUG=1
@@ -10,14 +9,22 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
+export USE_ATOMICS=${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:-0}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export DP=${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/"
@@ -30,7 +37,7 @@ 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=1 BENCHMARK=10
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
   export LLAMA_LAYERS=2
 fi

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -15,7 +15,11 @@ export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
 export FP8=${FP8:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export FAST_CE=${FASE_CE:-1}
+export FAST_CE=${FAST_CE:-1}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}
 
 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}

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam_mp.sh

@@ -2,7 +2,6 @@
 
 export PYTHONPATH="."
 export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
 export DEVICE_IN_FUNCTION_BUG=1
@@ -10,9 +9,17 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
+export USE_ATOMICS=${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:-0}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
@@ -35,7 +42,7 @@ 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=1 BENCHMARK=10
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
   export LLAMA_LAYERS=2
 fi

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -15,7 +15,11 @@ export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
 export FP8=${FP8:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export FAST_CE=${FASE_CE:-1}
+export FAST_CE=${FAST_CE:-1}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}
 
 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}

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/run_and_time.sh

@@ -17,6 +17,10 @@ export MASTER_WEIGHTS=1
 export FP8=1
 export ALLREDUCE_CAST=1
 export FAST_CE=1
+export FUSED_INPUT_QUANTIZE=1
+export FUSED_ADD_NORM_MUL_QUANTIZE=1
+export FUSED_SILU_W13=1
+export FUSED_PAD_GRAD_ACCUM=1
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=8 MP=1 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=2

extra/amdpci/am_smi.py

@@ -64,7 +64,7 @@ def get_bar0_size(pcibus):
 
 class AMSMI(AMDev):
   def __init__(self, pcibus, vram_bar:MMIOInterface, doorbell_bar:MMIOInterface, mmio_bar:MMIOInterface):
-    self.pcibus = pcibus
+    self.pcibus, self.devfmt = pcibus, pcibus
     self.vram, self.doorbell64, self.mmio = vram_bar, doorbell_bar, mmio_bar
     self.pci_state = self.read_pci_state()
     if self.pci_state == "D0": self._init_from_d0()
@@ -91,6 +91,7 @@ class SMICtx:
     self.prev_lines_cnt = 0
     self.prev_terminal_width = 0
     self.prev_terminal_height = 0
+    self.prev_metrics = {}
 
     remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:", "Processing accelerators:"]
     lspci = subprocess.check_output(["lspci"]).decode("utf-8").splitlines()
@@ -235,6 +236,29 @@ class SMICtx:
       case (13,0,12): return self._smuq10_round(metrics.SocketPower), self._smuq10_round(metrics.SocketPowerLimit)
       case _: return metrics.SmuMetrics.AverageSocketPower, metrics.SmuMetrics.dGPU_W_MAX
 
+  def get_throttle_info(self, dev, metrics):
+    match dev.ip_ver[am.MP1_HWIP]:
+      case (13,0,6)|(13,0,12):
+        throttle_fields = [('ProchotResidencyAcc', 'Prochot'), ('PptResidencyAcc', 'PPT'),
+                           ('SocketThmResidencyAcc', 'Socket Thm'), ('VrThmResidencyAcc', 'VR Thm'), ('HbmThmResidencyAcc', 'HBM Thm')]
+        prev = self.prev_metrics.get(dev.pcibus)
+        active = []
+        if prev is not None:
+          acc_delta = metrics.AccumulationCounter - prev.AccumulationCounter
+          if acc_delta > 0:
+            for field, name in throttle_fields:
+              delta = getattr(metrics, field) - getattr(prev, field)
+              if delta > 0 and (pct := min(100, (delta * 100 + acc_delta // 2) // acc_delta)) > 0: active.append((name, pct))
+        return active
+      case _:
+        smu_mod = dev.smu.smu_mod
+        throttler_names = {getattr(smu_mod, a): a[len('THROTTLER_'):-len('_BIT')]
+                           for a in dir(smu_mod) if a.startswith('THROTTLER_') and a.endswith('_BIT')}
+        active = []
+        for i, pct in enumerate(metrics.SmuMetrics.ThrottlingPercentage):
+          if pct > 0: active.append((throttler_names.get(i, f"UNK_{i}"), int(pct)))
+        return active
+
   def get_mem_usage(self, dev):
     usage = 0
     pt_stack = [dev.mm.root_page_table]
@@ -281,6 +305,13 @@ class SMICtx:
                     + [f"MEM Activity {draw_bar(self.get_mem_activity(dev, metrics) / 100, activity_line_width)}"] \
                     + [f"MEM Usage    {draw_bar(mem_used / mem_total, activity_line_width, opt_text=mem_fmt)}"] \
 
+      throttle_info = self.get_throttle_info(dev, metrics)
+      if throttle_info:
+        throttle_text = colored(', '.join(f"{name} {pct}%" for name, pct in throttle_info), "red")
+      else:
+        throttle_text = colored("None", "green")
+      activity_line += [f"Throttle     {throttle_text}" + " " * (activity_line_width + 2)]
+
       temps_data, temps_data_compact = self.get_temps(dev, metrics), self.get_temps(dev, metrics, compact=True)
       temps_table = ["=== Temps (°C) ==="] + [f"{name:<16}: {color_temp(val)}" for name, val in temps_data.items()]
       temps_table_compact = ["Temps (°C):" + '/'.join([f"{color_temp(val)} {name}" for name, val in temps_data_compact.items()])]
@@ -324,6 +355,8 @@ class SMICtx:
 
       dev_content.append(device_line + activity_line + same_line([temps_table, power_table, frequency_table]))
 
+    self.prev_metrics = {dev.pcibus: m for dev, m in dev_metrics.items() if m is not None}
+
     raw_text = 'AM Monitor'.center(terminal_width) + "\n" + "=" * terminal_width + "\n\n"
     for i in range(0, len(dev_content), 2):
       if i + 1 < len(dev_content): raw_text += '\n'.join(same_line([dev_content[i], dev_content[i+1]], split=padding))

extra/amdpci/headers/amdgpu_irq.h

@@ -28,15 +28,7 @@
 // #include "soc15_ih_clientid.h"
 // #include "amdgpu_ih.h"
 
-#define int32_t int
-#define uint32_t unsigned int
-#define int8_t signed char
-#define uint8_t unsigned char
-#define uint16_t unsigned short
-#define int16_t short
-#define uint64_t unsigned long long
 #define bool _Bool
-#define u32 unsigned int
 
 #define AMDGPU_MAX_IRQ_SRC_ID		0x100
 #define AMDGPU_MAX_IRQ_CLIENT_ID	0x100

extra/amdpci/headers/amdgpu_smu.h

@@ -22,15 +22,7 @@
 #ifndef __AMDGPU_SMU_H__
 #define __AMDGPU_SMU_H__
 
-#define int32_t int
-#define uint32_t unsigned int
-#define int8_t signed char
-#define uint8_t unsigned char
-#define uint16_t unsigned short
-#define int16_t short
-#define uint64_t unsigned long long
 #define bool _Bool
-#define u32 unsigned int
 
 #define SMU_THERMAL_MINIMUM_ALERT_TEMP		0
 #define SMU_THERMAL_MAXIMUM_ALERT_TEMP		255

extra/amdpci/headers/amdgpu_ucode.h

@@ -24,15 +24,7 @@
 #define __AMDGPU_UCODE_H__
 
 // #include "amdgpu_socbb.h"
-#define int32_t int
-#define uint32_t unsigned int
-#define int8_t signed char
-#define uint8_t unsigned char
-#define uint16_t unsigned short
-#define int16_t short
-#define uint64_t unsigned long long
 #define bool _Bool
-#define u32 unsigned int
 
 struct common_firmware_header {
 	uint32_t size_bytes; /* size of the entire header+image(s) in bytes */

extra/gemm/amd_asm_matmul.py

@@ -167,7 +167,7 @@ PREFETCH_LOADS = [(V_LDS_A_DATA[4+2*i], V_LDS_A_DATA[4+2*i+1], V_GLOBAL_B_ADDR,
 # =============================================================================
 
 class Kernel:
-  def __init__(self, arch='gfx1100'): self.instructions, self.labels, self.pos, self.arch = [], {}, 0, arch
+  def __init__(self): self.instructions, self.labels, self.pos = [], {}, 0
   def label(self, name): self.labels[name] = self.pos
 
   def emit(self, inst, target=None):
@@ -196,10 +196,10 @@ class Kernel:
 # Kernel builder
 # =============================================================================
 
-def build_kernel(N, arch='gfx1100'):
+def build_kernel(N):
   assert N % 128 == 0, f"N must be a multiple of 128 (tile size), got {N}"
   assert N >= 256, f"N must be >= 256 (prefetch pipeline requires at least 2 K-blocks), got {N}"
-  k = Kernel(arch)
+  k = Kernel()
 
   # ===========================================================================
   # PROLOGUE: Load kernel arguments, compute tile coordinates and addresses
@@ -443,7 +443,7 @@ def test_matmul():
   dev = Device[Device.DEFAULT]
   print(f"Device arch: {dev.renderer.target.arch}")
 
-  insts = build_kernel(N, dev.renderer.target.arch)
+  insts = build_kernel(N)
 
   rng = np.random.default_rng(42)
   a = Tensor(rng.random((N, N), dtype=np.float32) - 0.5)

extra/gemm/cdna_asm_gemm.py

@@ -2628,20 +2628,24 @@ def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
 # ** FP8 GEMM custom kernel
 
 @functools.cache
-def custom_hk_fp8_gemm(C:UOp, A:UOp, B:UOp, X_s:UOp, W_s:UOp, dname:str) -> UOp:
-  # A is (batch, M, K), B is (N, K) transposed, X_s is x_scale, W_s is w_scale — kernel multiplies by both
+def custom_hk_fp8_gemm(C:UOp, A:UOp, B:UOp, *args:UOp, dname:str, scale_mode:int=3) -> UOp:
+  # scale_mode: 0=no scale, 1=x only, 2=w only, 3=both
+  n_scales = (1 if scale_mode & 1 else 0) + (1 if scale_mode & 2 else 0)
+  scales, extra = args[:n_scales], args[n_scales:]
   M, K = A.shape[0]*A.shape[1], A.shape[2]
   N, K2 = B.shape[(1 if B.ndim == 3 else 0):]
   assert K == K2, f"{A.shape} {B.shape}"
   block_size = 256
   threads = UOp.special(64 * 8, "lidx0")
   workgroups = UOp.special((M // block_size) * (N // block_size), "gidx0")
-  sink = UOp.sink(C.base, A.base, B.base, X_s.base, W_s.base, threads, workgroups,
+  sink_inputs = (C.base, A.base, B.base) + tuple(s.base for s in scales) + (threads, workgroups)
+  sink = UOp.sink(*sink_inputs,
                   arg=KernelInfo(f"hk_fp8_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K, mem=(M*K+N*K)*A.dtype.itemsize+M*N*C.dtype.itemsize)))
   kittens_path = pathlib.Path(__file__).parent.parent/"thunder"/"amd"
   src = (kittens_path/"gemm_fp8.cpp").read_text()
   lib = HIPCCCompiler("gfx950", [f"-I{(kittens_path/'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-ffast-math",
-                                 "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DGEMM_M={M}", f"-DGEMM_N={N}", f"-DGEMM_K={K}"]).compile_cached(src)
+                                 "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DGEMM_M={M}", f"-DGEMM_N={N}", f"-DGEMM_K={K}",
+                                 f"-DSCALE_MODE={scale_mode}"]).compile_cached(src)
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
                                UOp(Ops.BINARY, arg=lib)))
 
@@ -2698,19 +2702,31 @@ def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
 
 def custom_gemm_bw(gradient:UOp, kernel:UOp):
   inputs = kernel.src[1:]
-  # fp8 scaled gemm has 5 inputs (out, a, b, x_scale, w_scale), others have 3 (out, a, b)
-  if len(inputs) == 5:
-    out, a, b, s_x, s_w = inputs
+  if inputs[1].dtype == FP8_DTYPE:
+    grad_amax_state = inputs[5] if len(inputs) == 6 else None
+    out, a, b, s_x, s_w = inputs[:5]
     a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
     s_x_t, s_w_t = Tensor(s_x, device=a.device), Tensor(s_w, device=a.device)
     g_t = g_t[:a.shape[0]]
-    g_fp8, g_scale, _ = quantize_fp8(g_t)
+    from extra.llama_kernels.cast_amax import _grad_fp8_mailbox
+    from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed
+    gbase = gradient.base if hasattr(gradient, "base") else gradient
+    mailbox_entry = _grad_fp8_mailbox.pop(gbase, None) or _grad_fp8_mailbox.pop(gradient, None)
+    if mailbox_entry is not None:
+      g_fp8_u, inv_scale_u, _new_amax_u, store_effect = mailbox_entry
+      g_fp8 = Tensor(g_fp8_u, device=a.device)[:a.shape[0]]
+      g_scale = Tensor(inv_scale_u, device=a.device)
+    else:
+      assert grad_amax_state is not None, "fp8 matmul bwd needs either a mailbox entry or a grad_amax_state"
+      g_fp8, g_scale, _, store_effect = quantize_fp8_delayed(g_t, Tensor(grad_amax_state, device=a.device))
     # dgrad: uses g_scale * x_scale * w_scale
     grad_a = asm_gemm(g_fp8, b_t, x_scale=g_scale * s_x_t, w_scale=s_w_t)
     # wgrad: no w_scale
-    _one = Tensor(1.0, dtype=dtypes.float, device=a.device)
-    grad_b = asm_gemm(g_fp8.permute(2, 0, 1).reshape(g_t.shape[-1], -1), a_t.reshape(-1, a_t.shape[-1]), x_scale=g_scale * s_x_t, w_scale=_one)
-    return (None, grad_a.uop, grad_b.uop, None, None)
+    grad_b = asm_gemm(g_fp8.permute(2, 0, 1).reshape(g_t.shape[-1], -1), a_t.reshape(-1, a_t.shape[-1]), x_scale=g_scale * s_x_t)
+    # Attach the delayed-amax store effect (if any) to grad_a so realizing grads commits the amax update.
+    ret = (None, grad_a.uop.after(store_effect), grad_b.uop, None, None)
+    if len(inputs) == 6: ret = ret + (None,)
+    return ret
   else:
     out, a, b = inputs
     assert all_same([gradient.device, a.device, b.device, out.device])
@@ -2725,7 +2741,7 @@ def custom_gemm_bw(gradient:UOp, kernel:UOp):
 
 # ** main gemm function
 
-def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=None) -> Tensor:
+def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=None, grad_amax_state:Tensor|None=None) -> Tensor:
   assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
   counters["used"] += 1
   unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
@@ -2759,10 +2775,11 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
   if arch.startswith("gfx950") and getenv("USE_ASM", 1):
     # fp8 gemm computes [email protected], kernel multiplies output by x_scale * w_scale before bf16 store
     if a.dtype == FP8_DTYPE:
-      _one = lambda: Tensor(1.0, dtype=dtypes.float, device=a.device)
-      xs = x_scale if x_scale is not None else _one()
-      ws = w_scale if w_scale is not None else _one()
-      out = Tensor.custom_kernel(out, a, b.T, xs, ws, fxn=functools.partial(custom_hk_fp8_gemm, dname=dname), grad_fxn=custom_gemm_bw)[0]
+      scales = tuple(s for s in (x_scale, w_scale) if s is not None)
+      scale_mode = (1 if x_scale is not None else 0) | (2 if w_scale is not None else 0)
+      extra = [grad_amax_state] if grad_amax_state is not None else []
+      fxn = functools.partial(custom_hk_fp8_gemm, dname=dname, scale_mode=scale_mode)
+      out = Tensor.custom_kernel(out, a, b.T, *scales, *extra, fxn=fxn, grad_fxn=custom_gemm_bw)[0]
     else:
       out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname), grad_fxn=custom_gemm_bw)[0]
   else:

extra/gemm/triton_nv_matmul.py

@@ -4,7 +4,8 @@ import triton.language as tl
 from triton.compiler import AttrsDescriptor, ASTSource, compile as triton_compile
 import numpy as np
 from tinygrad import Tensor, dtypes, Device
-from tinygrad.engine.realize import CompiledRunner
+from tinygrad.engine.realize import get_runtime
+from tinygrad.codegen import to_program
 from tinygrad.uop.ops import Ops, UOp, KernelInfo, ProgramInfo
 from tinygrad.helpers import getenv
 np.set_printoptions(suppress=True)
@@ -92,13 +93,15 @@ if __name__ == "__main__":
   info = ProgramInfo(name="matmul_kernel",
                      global_size=(M//BLOCK_SIZE_M, N//BLOCK_SIZE_N, 1), local_size=(32*compiled.metadata.num_warps, 1, 1))
   sink = UOp.sink(arg=KernelInfo(name="matmul_kernel"))
-  prg_uop = UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR), UOp(Ops.SOURCE, arg=src)), arg=info)
-  runner = CompiledRunner(prg_uop, Device.DEFAULT)
+  prg_uop = to_program(UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR), UOp(Ops.SOURCE, arg=src)), arg=info),
+                       Device.default.renderer)
+  rt = get_runtime(Device.DEFAULT, prg_uop)
   all_bufs = [x.ensure_allocated() for x in bufs]
-  prg_bufs = [all_bufs[i] for i in runner.p.globals]
+  prg_bufs = [all_bufs[i] for i in info.globals]
+  gsize, lsize = info.launch_dims({})
   tflops = []
   for i in range(5):
-    tm = runner(prg_bufs, {}, wait=True)
+    tm = rt(*[b._buf for b in prg_bufs], global_size=gsize, local_size=lsize, vals=info.vals({}), wait=True)
     tflops.append((2*M*K*N/tm)*1e-12)
   print(f"TFLOPS: {max(tflops):.2f}")
 

extra/llama_kernels/__init__.py

@@ -29,7 +29,23 @@ def shard_shape(shape:tuple, axis:int, ndev:int) -> list:
   s[axis] //= ndev
   return s
 
+def dname_of(device) -> str:
+  if isinstance(device, tuple): return device[0].split(":")[0]
+  return device.split(":")[0] if isinstance(device, str) else device
+
+def alloc_like(shape, dtype, device, axis=None) -> Tensor:
+  if isinstance(device, tuple) and axis is not None:
+    return Tensor(Tensor.invalids(*shard_shape(shape, axis, len(device)), dtype=dtype, device=device).uop.multi(axis), device=device)
+  return Tensor.invalids(*shape, dtype=dtype, device=device)
+
+def alloc_local(shape, dtype, device, axis=None) -> Tensor:
+  if isinstance(device, tuple) and axis is not None:
+    return Tensor(Tensor.invalids(*shape, dtype=dtype, device=device).uop.multi(0), device=device)
+  return Tensor.invalids(*shape, dtype=dtype, device=device)
+
+def compile_hip(src:str, defines:list[str]):
+  return HIPCCCompiler("gfx950", ["-std=c++20", "-ffast-math", *defines]).compile_cached(src)
+
 def compile_cpp(cpp_dir:pathlib.Path, cpp_name:str, n_elems:int, hidden:int):
   src = (cpp_dir/cpp_name).read_text()
-  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={hidden}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
-  return src, HIPCCCompiler("gfx950", ["-std=c++20", "-ffast-math", *defines]).compile_cached(src)
+  return src, compile_hip(src, [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={hidden}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"])

extra/llama_kernels/cast_amax/__init__.py

@@ -3,26 +3,34 @@ 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, compile_cpp, shard_shape, scalar_amax
+from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, compile_cpp, alloc_like, alloc_local, scalar_amax, dname_of
+
+# module-level mailbox: grad_xw13 UOp -> (grad_xw13_fp8 UOp, inv_scale UOp, new_amax UOp, store_effect)
+# lets cdna_asm_gemm's bwd reuse the fp8 companion produced by the fused silu_mul bwd kernel
+# instead of doing a redundant bf16 -> fp8 quantize.
+_grad_fp8_mailbox:dict = {}
 
 @functools.cache
-def _custom_fused_bwd_w13(grad_xw13:UOp, xw13:UOp, grad_x2:UOp, amax_state:UOp, dname:str) -> UOp:
+def _custom_fused_bwd_w13(grad_xw13:UOp, grad_xw13_fp8:UOp, grad_amax_buf: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 * 5
-  sink = UOp.sink(grad_xw13.base, xw13.base, grad_x2.base, amax_state.base, threads, workgroups,
-                  arg=KernelInfo(f"fused_silu_mul_bwd_w13_{n_elems}", estimates=Estimates(ops=8*n_elems, mem=mem)))
+  mem = n_elems * 2 * 5 + n_elems * 2 + NUM_WG * 4 + 4
+  sink = UOp.sink(grad_xw13.base, grad_xw13_fp8.base, grad_amax_buf.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)
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                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, dname:str) -> UOp:
+def _custom_fused_cast_amax_w13(fp8_out:UOp, amax_buf:UOp, xw13:UOp, amax_state:UOp, 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 * 2
+  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,
                   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)
@@ -30,44 +38,37 @@ def _custom_fused_cast_amax_w13(fp8_out:UOp, amax_buf:UOp, xw13:UOp, amax_state:
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
 
 def _fused_quantize_bwd_w13(gradient:UOp, kernel:UOp):
-  # NOTE: inputs are (fp8_out, amax_buf, xw13, amax_state); grad for xw13 only
-  _, _, xw13, amax_state = kernel.src[1:]
+  _, _, xw13, amax_state, grad_amax_state = kernel.src[1:]
   device = xw13.device
-  if isinstance(device, tuple):
-    axis, ndev = xw13.axis, len(device)
-    assert axis in (0, 1), f"unsupported sharding axis={axis}"
-    grad_xw13 = Tensor(Tensor.invalids(*shard_shape(xw13.shape, axis, ndev), dtype=dtypes.bfloat16,
-                                        device=device).uop.multi(axis), device=device)
-    dname = device[0].split(":")[0]
-  else:
-    grad_xw13 = Tensor.invalids(*xw13.shape, dtype=dtypes.bfloat16, device=device)
-    dname = device.split(":")[0] if isinstance(device, str) else device
-  grad_x2_t = Tensor(gradient, device=device).cast(dtypes.bfloat16)
-  fxn = functools.partial(_custom_fused_bwd_w13, dname=dname)
-  grad_xw13, *_ = Tensor.custom_kernel(grad_xw13, Tensor(xw13, device=device), grad_x2_t,
-                                        Tensor(amax_state, device=device), fxn=fxn)
-  return (None, None, grad_xw13.uop, None)
+  axis = xw13.axis if isinstance(device, tuple) else None
+  grad_xw13     = alloc_like(xw13.shape, dtypes.bfloat16, device, axis)
+  grad_xw13_fp8 = alloc_like(xw13.shape, dtypes.fp8e4m3,  device, axis)
+  grad_amax_buf = alloc_local((NUM_WG,), dtypes.float32,  device, axis)
+  grad_amax_state_t = Tensor(grad_amax_state, device=device)
+  fxn = functools.partial(_custom_fused_bwd_w13, dname=dname_of(device))
+  grad_xw13, grad_xw13_fp8, grad_amax_buf, *_ = Tensor.custom_kernel(
+    grad_xw13, grad_xw13_fp8, grad_amax_buf,
+    Tensor(xw13, device=device), Tensor(gradient, device=device).cast(dtypes.bfloat16),
+    Tensor(amax_state, device=device), grad_amax_state_t, fxn=fxn)
+  inv_scale = (grad_amax_state_t.float() + 1e-8) / FP8_MAX
+  new_grad_amax = scalar_amax(grad_amax_buf)
+  store_effect = grad_amax_state_t.uop.store(new_grad_amax.uop)
+  # Stash fp8 companion + amax store for cdna_asm_gemm's bwd to attach to grad_a.
+  _grad_fp8_mailbox[grad_xw13.uop] = (grad_xw13_fp8.uop, inv_scale.uop, new_grad_amax.uop, store_effect)
+  return (None, None, grad_xw13.uop, None, None)
 
-def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype) -> tuple[Tensor, Tensor, Tensor]:
+def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_state:Tensor) -> tuple[Tensor, Tensor, Tensor]:
   # NOTE: silu(xw1)*xw3 -> fp8 + amax over fused xw13 layout. Returns (fp8, inv_scale, new_amax)
+  # 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
   assert H2 % 2 == 0, f"w13 last-axis must be even, got {H2}"
   HIDDEN = H2 // 2
-  if isinstance(xw13.device, tuple):
-    axis, ndev = xw13.uop.axis, len(xw13.device)
-    assert axis in (0, 1), f"unsupported sharding axis={axis}"
-    fp8_out = Tensor(Tensor.invalids(*shard_shape((MBS, SEQ, HIDDEN), axis, ndev), dtype=fp8_dtype,
-                                     device=xw13.device).uop.multi(axis), device=xw13.device)
-    amax_buf = Tensor(Tensor.invalids(NUM_WG, dtype=dtypes.bfloat16, device=xw13.device).uop.multi(0),
-                      device=xw13.device)
-    dname = xw13.device[0].split(":")[0]
-  else:
-    fp8_out = Tensor.invalids(MBS, SEQ, HIDDEN, dtype=fp8_dtype, device=xw13.device)
-    amax_buf = Tensor.invalids(NUM_WG, dtype=dtypes.bfloat16, device=xw13.device)
-    dname = xw13.device.split(":")[0] if isinstance(xw13.device, str) else xw13.device
-  fxn = functools.partial(_custom_fused_cast_amax_w13, dname=dname)
-  fp8_out, amax_buf, *_ = Tensor.custom_kernel(fp8_out, amax_buf, xw13, amax_state, fxn=fxn,
-                                                grad_fxn=_fused_quantize_bwd_w13)
+  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,
+                                                fxn=fxn, grad_fxn=_fused_quantize_bwd_w13)
   inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
   return fp8_out, inv_scale, scalar_amax(amax_buf)

extra/llama_kernels/cast_amax/cast_amax_bwd_w13.cpp

@@ -1,5 +1,6 @@
 #include <hip/hip_runtime.h>
 #include <hip/hip_bf16.h>
+#include <hip/hip_fp8.h>
 
 #ifndef N_ELEMS
 #define N_ELEMS 234881024
@@ -20,19 +21,32 @@ constexpr float FP8_MAX = 448.0f;
 static_assert(N_ELEMS % VEC == 0, "N_ELEMS must be divisible by VEC");
 static_assert(HIDDEN % VEC == 0, "HIDDEN must be divisible by VEC");
 
+// fused silu*mul backward, three outputs in a single HBM pass:
+//   1) bf16 grad_xw13      — consumed by downstream bf16 autograd chain
+//   2) fp8  grad_xw13_fp8  — delayed-scale quantize using grad_amax_state (mailbox to matmul bwd)
+//   3) fp32 grad_amax_buf  — per-WG partial |grad_xw13|, reduced into next step's grad_amax_state
+// 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_bfloat16*       __restrict__ grad_xw13_out,    // bf16, 2*N_ELEMS (interleaved layout)
-    const __hip_bfloat16* __restrict__ xw13,             // bf16, 2*N_ELEMS (interleaved)
-    const __hip_bfloat16* __restrict__ grad_x2,          // bf16, N_ELEMS
-    const __hip_bfloat16* __restrict__ amax_state)       // bf16 scalar
+    __hip_bfloat16*       __restrict__ grad_xw13_out,        // bf16, 2*N_ELEMS
+    __hip_fp8_storage_t*  __restrict__ grad_xw13_fp8_out,    // fp8,  2*N_ELEMS
+    float*                __restrict__ grad_amax_buf,        // fp32, NUM_WG per-WG partials
+    const __hip_bfloat16* __restrict__ xw13,                 // bf16, 2*N_ELEMS
+    const __hip_bfloat16* __restrict__ grad_x2,              // bf16, N_ELEMS
+    const float*          __restrict__ amax_state,           // fp32 scalar (fwd x2 amax)
+    const float*          __restrict__ grad_amax_state)      // fp32 scalar (delayed grad amax)
 {
+  __shared__ float sdata[THREADS_PER_WG];
+
   const int tid = threadIdx.x;
   const int wg  = blockIdx.x;
   const int gid = wg * THREADS_PER_WG + tid;
   const int stride_elems = NUM_WG * THREADS_PER_WG * VEC;
 
   const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+  const float g_scale = FP8_MAX / (static_cast<float>(*grad_amax_state) + 1e-8f);
+  float local_max = 0.0f;
 
   for (int base = gid * VEC; base < N_ELEMS; base += stride_elems) {
     const int outer = base / HIDDEN;
@@ -49,6 +63,7 @@ fused_silu_mul_bwd_w13(
     const __hip_bfloat16 *gv = reinterpret_cast<const __hip_bfloat16*>(&g_raw);
 
     __hip_bfloat16 out1[VEC], out3[VEC];
+    __hip_fp8_storage_t fp8_1[VEC], fp8_3[VEC];
     #pragma unroll
     for (int i = 0; i < VEC; i++) {
       const float f1 = static_cast<float>(x1[i]);
@@ -58,11 +73,26 @@ fused_silu_mul_bwd_w13(
       const float silu = f1 * sig;
       const float silu_prime = sig + silu * (1.0f - sig);
       const float gs = fg * scale;
-      out1[i] = static_cast<__hip_bfloat16>(gs * silu_prime * f3);
-      out3[i] = static_cast<__hip_bfloat16>(gs * silu);
+      const float g1 = gs * silu_prime * f3;
+      const float g3 = gs * silu;
+      out1[i] = static_cast<__hip_bfloat16>(g1);
+      out3[i] = static_cast<__hip_bfloat16>(g3);
+      local_max = fmaxf(local_max, fmaxf(fabsf(g1), fabsf(g3)));
+      fp8_1[i] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, g1 * g_scale)), __HIP_SATFINITE, __HIP_E4M3);
+      fp8_3[i] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, g3 * g_scale)), __HIP_SATFINITE, __HIP_E4M3);
     }
 
     *reinterpret_cast<float4*>(&grad_xw13_out[xw1_off]) = *reinterpret_cast<float4*>(out1);
     *reinterpret_cast<float4*>(&grad_xw13_out[xw3_off]) = *reinterpret_cast<float4*>(out3);
+    *reinterpret_cast<uint64_t*>(&grad_xw13_fp8_out[xw1_off]) = *reinterpret_cast<uint64_t*>(fp8_1);
+    *reinterpret_cast<uint64_t*>(&grad_xw13_fp8_out[xw3_off]) = *reinterpret_cast<uint64_t*>(fp8_3);
   }
+
+  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) grad_amax_buf[wg] = sdata[0];
 }

extra/llama_kernels/cast_amax/cast_amax_fwd_w13.cpp

@@ -24,9 +24,9 @@ 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
-    __hip_bfloat16*       __restrict__ amax_buf,        // bf16, NUM_WG (per-WG amaxes)
+    float*                __restrict__ amax_buf,        // fp32, NUM_WG (per-WG amaxes)
     const __hip_bfloat16* __restrict__ xw13,            // bf16, 2*N_ELEMS
-    const __hip_bfloat16* __restrict__ amax_state)      // bf16 scalar
+    const float*          __restrict__ amax_state)      // fp32 scalar
 {
   __shared__ float sdata[THREADS_PER_WG];
 
@@ -75,5 +75,5 @@ fused_silu_mul_cast_amax_w13(
     __syncthreads();
   }
 
-  if (tid == 0) amax_buf[wg] = static_cast<__hip_bfloat16>(sdata[0]);
+  if (tid == 0) amax_buf[wg] = sdata[0];
 }

extra/llama_kernels/fused_ce/__init__.py

@@ -42,8 +42,7 @@ def _fused_ce_loss_bwd(gradient:UOp, kernel:UOp, label_smoothing:float):
   # gradient is the upstream grad w.r.t. per-row loss (shape: (rows,) fp32)
   _, _, lse_u, logits_u, targets_u = kernel.src[1:]
   device = logits_u.device
-  rows_vocab = logits_u.shape  # (rows, VOCAB) after reshape
-  rows, VOCAB = rows_vocab
+  rows, VOCAB = logits_u.shape  # (rows, VOCAB) after reshape
   if isinstance(device, tuple):
     axis = logits_u.axis
     ndev = len(device)
@@ -54,9 +53,8 @@ def _fused_ce_loss_bwd(gradient:UOp, kernel:UOp, label_smoothing:float):
     d_logits = Tensor.invalids(rows, VOCAB, dtype=dtypes.bfloat16, device=device)
     dname = device.split(":")[0] if isinstance(device, str) else device
     rows_per_dev = rows
-  grad_t = Tensor(gradient, device=device).float().reshape(-1)  # (rows,) fp32
   # NOTE: .mean() backward gives same grad per row (1/N), so broadcast is safe; take scalar
-  scale = grad_t[0:1].contiguous()
+  scale = Tensor(gradient, device=device).float().reshape(-1)[0:1].contiguous()
   logits_t = Tensor(logits_u.after(kernel), device=device)
   lse_t = Tensor(lse_u.after(kernel), device=device)
   targets_t = Tensor(targets_u, device=device)

extra/llama_kernels/fused_mul_quantize_fp8/__init__.py

@@ -1,54 +0,0 @@
-from __future__ import annotations
-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, compile_cpp, shard_shape, scalar_amax
-
-@functools.cache
-def _custom_mul_quantize_fp8(fp8_out:UOp, amax_buf:UOp, x:UOp, weight:UOp, amax_state:UOp, dname:str) -> 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 + HIDDEN * 2 + n_elems + NUM_WG * 2
-  sink = UOp.sink(fp8_out.base, amax_buf.base, x.base, weight.base, amax_state.base, threads, workgroups,
-                  arg=KernelInfo(f"fused_mul_quantize_fp8_{n_elems}_h{HIDDEN}", estimates=Estimates(ops=3*n_elems, mem=mem)))
-  src, lib = compile_cpp(pathlib.Path(__file__).parent, "fused_mul_quantize_fp8.cpp", n_elems, HIDDEN)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
-                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
-
-def _fused_mul_quantize_fp8_bwd(gradient:UOp, kernel:UOp):
-  # NOTE: inputs are (fp8_out, amax_buf, x, weight, amax_state); grads for x and weight
-  _, _, x_u, weight_u, amax_state_u = kernel.src[1:]
-  device = x_u.device
-  grad_t = Tensor(gradient, device=device).cast(dtypes.bfloat16)
-  x_t, weight_t = Tensor(x_u, device=device), Tensor(weight_u, device=device)
-  scale = FP8_MAX / (Tensor(amax_state_u, device=device).float() + 1e-8)
-  grad_scaled = grad_t.float() * scale
-  # NOTE: grad_x stays bf16 to avoid CSE materializing a (MBS, SEQ, HIDDEN) fp32 intermediate
-  grad_x = (grad_scaled * weight_t.float()).cast(dtypes.bfloat16)
-  grad_weight = (grad_scaled * x_t.float()).sum(axis=(0, 1)).cast(dtypes.bfloat16)
-  return (None, None, grad_x.uop, grad_weight.uop, None)
-
-def fused_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, fp8_dtype) -> tuple[Tensor, Tensor, Tensor]:
-  # NOTE: (x * weight) -> fp8 + amax, delayed scaling. Returns (fp8, inv_scale, new_amax)
-  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}"
-  MBS, SEQ, HIDDEN = x.shape
-  if isinstance(x.device, tuple):
-    axis, ndev = x.uop.axis, len(x.device)
-    assert axis in (0, 1), f"unsupported sharding axis={axis}"
-    fp8_out = Tensor(Tensor.invalids(*shard_shape((MBS, SEQ, HIDDEN), axis, ndev), dtype=fp8_dtype,
-                                     device=x.device).uop.multi(axis), device=x.device)
-    amax_buf = Tensor(Tensor.invalids(NUM_WG, dtype=dtypes.bfloat16, device=x.device).uop.multi(0), device=x.device)
-    dname = x.device[0].split(":")[0]
-  else:
-    fp8_out = Tensor.invalids(MBS, SEQ, HIDDEN, dtype=fp8_dtype, device=x.device)
-    amax_buf = Tensor.invalids(NUM_WG, dtype=dtypes.bfloat16, device=x.device)
-    dname = x.device.split(":")[0] if isinstance(x.device, str) else x.device
-  fxn = functools.partial(_custom_mul_quantize_fp8, dname=dname)
-  fp8_out, amax_buf, *_ = Tensor.custom_kernel(fp8_out, amax_buf, x, weight, amax_state, fxn=fxn,
-                                                grad_fxn=_fused_mul_quantize_fp8_bwd)
-  new_amax = scalar_amax(amax_buf)
-  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
-  return fp8_out, inv_scale, new_amax

extra/llama_kernels/fused_pad_grad_accum/__init__.py

@@ -0,0 +1,55 @@
+from __future__ import annotations
+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 THREADS_PER_WG, dname_of, compile_hip
+
+ELEMS_PER_THREAD = 8  # vectorized 16-byte load (uint4 = 8 bf16)
+
+def _build_src(n_chunks:int) -> str:
+  template = (pathlib.Path(__file__).parent/"fused_pad_grad_accum.cpp").read_text()
+  params = "".join(f",\n    const __hip_bfloat16* __restrict__ chunk{i}" for i in range(n_chunks))
+  dispatch = "\n    ".join(f"case {i}: chunk_ptr = chunk{i}; break;" for i in range(n_chunks))
+  return (template.replace("__FUSED_PAD_GRAD_ACCUM_PARAMS", params)
+                  .replace("__FUSED_PAD_GRAD_ACCUM_DISPATCH", dispatch))
+
+@functools.cache
+def _custom_fused_pad_grad_accum(grad_buf:UOp, *chunk_uops, dname:str, n_chunks:int, chunk_size:int) -> UOp:
+  total = n_chunks * chunk_size
+  elems_per_block = THREADS_PER_WG * ELEMS_PER_THREAD
+  assert chunk_size % elems_per_block == 0, f"chunk_size {chunk_size} must be multiple of {elems_per_block}"
+  num_wg = total // elems_per_block
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(num_wg, "gidx0")
+  mem = total * 2 * 3
+  sink = UOp.sink(grad_buf.base, *(c.base for c in chunk_uops), threads, workgroups,
+                  arg=KernelInfo(f"fused_pad_grad_accum_n{n_chunks}_c{chunk_size}",
+                                 estimates=Estimates(ops=2*total, mem=mem)))
+  src = _build_src(n_chunks)
+  defines = [f"-DCHUNK_SIZE={chunk_size}", f"-DTHREADS_PER_WG={THREADS_PER_WG}", f"-DELEMS_PER_THREAD={ELEMS_PER_THREAD}"]
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def can_fused_pad_grad_accum(grad_buf:Tensor, chunks:list[Tensor]) -> bool:
+  if not chunks or grad_buf.dtype != dtypes.bfloat16: return False
+  if any(c.dtype != dtypes.bfloat16 for c in chunks): return False
+  chunk_shape = chunks[0].shape
+  if any(c.shape != chunk_shape for c in chunks): return False
+  chunk_size, total = 1, 1
+  for d in chunk_shape: chunk_size *= d
+  for d in grad_buf.shape: total *= d
+  return total == len(chunks) * chunk_size and chunk_size % (THREADS_PER_WG * ELEMS_PER_THREAD) == 0
+
+def fused_pad_grad_accum(grad_buf:Tensor, chunks:list[Tensor]) -> Tensor:
+  # NOTE: grad_buf += cat(*chunks, dim=0) in one HBM pass (in-place add). Returns new grad_buf Tensor.
+  # Requires uniform chunk shapes and chunk_size % (THREADS_PER_WG*ELEMS_PER_THREAD) == 0.
+  assert chunks and grad_buf.dtype == dtypes.bfloat16
+  for c in chunks: assert c.dtype == dtypes.bfloat16, f"chunk dtype must be bf16, got {c.dtype}"
+  chunk_size, total = 1, 1
+  for d in chunks[0].shape: chunk_size *= d
+  for d in grad_buf.shape: total *= d
+  assert total == len(chunks) * chunk_size, f"grad_buf size {total} != n_chunks {len(chunks)} * chunk_size {chunk_size}"
+  fxn = functools.partial(_custom_fused_pad_grad_accum, dname=dname_of(grad_buf.device),
+                          n_chunks=len(chunks), chunk_size=chunk_size)
+  out, *_ = Tensor.custom_kernel(grad_buf, *chunks, fxn=fxn)
+  return out

extra/llama_kernels/fused_pad_grad_accum/fused_pad_grad_accum.cpp

@@ -0,0 +1,63 @@
+// Fused custom kernel: grad_buf += cat(*chunks, dim=0) in one HBM pass.
+//
+// Template source — chunk parameter list and switch dispatch are filled by codegen
+// in cast_amax.py:_build_fused_pad_grad_accum_src to support arbitrary N.
+//
+// Defines required at compile time:
+//   CHUNK_SIZE       elements per chunk (must be multiple of THREADS_PER_WG * ELEMS_PER_THREAD)
+//   THREADS_PER_WG
+//   ELEMS_PER_THREAD (8 = one uint4 per thread = 16-byte vectorized load)
+//
+// Layout: one block-per-(slice-of-chunk) — blockIdx.x / BLOCKS_PER_CHUNK selects the chunk.
+// All threads in a block read the same chunk → switch is uniform → no warp divergence.
+
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+#ifndef ELEMS_PER_THREAD
+#define ELEMS_PER_THREAD 8
+#endif
+
+#define ELEMS_PER_BLOCK (THREADS_PER_WG * ELEMS_PER_THREAD)
+#define BLOCKS_PER_CHUNK (CHUNK_SIZE / ELEMS_PER_BLOCK)
+
+extern "C" __attribute__((global))
+__attribute__((amdgpu_flat_work_group_size(1, THREADS_PER_WG)))
+void fused_pad_grad_accum(
+    __hip_bfloat16* __restrict__ grad_buf
+    __FUSED_PAD_GRAD_ACCUM_PARAMS
+) {
+  const int bid = blockIdx.x;
+  const int chunk_idx = bid / BLOCKS_PER_CHUNK;
+  const int block_in_chunk = bid - chunk_idx * BLOCKS_PER_CHUNK;
+  const int tid = threadIdx.x;
+
+  const __hip_bfloat16* chunk_ptr;
+  switch (chunk_idx) {
+    __FUSED_PAD_GRAD_ACCUM_DISPATCH
+    default: chunk_ptr = (const __hip_bfloat16*)0; break;  // unreachable
+  }
+
+  // int64 for global_offset: at 32 chunks × 117M elements = 3.6B, int32 overflows → MEMVIOL.
+  const int local_offset = block_in_chunk * ELEMS_PER_BLOCK + tid * ELEMS_PER_THREAD;
+  const long long global_offset = (long long)chunk_idx * (long long)CHUNK_SIZE + (long long)local_offset;
+
+  // Vectorized 16-byte load (uint4 = 8 bf16). Requires CHUNK_SIZE % 8 == 0 and 16-byte alignment.
+  const uint4 chunk_v = *reinterpret_cast<const uint4*>(&chunk_ptr[local_offset]);
+  const uint4 grad_v  = *reinterpret_cast<const uint4*>(&grad_buf[global_offset]);
+  uint4 out_v;
+
+  const __hip_bfloat16* chunk_bf = reinterpret_cast<const __hip_bfloat16*>(&chunk_v);
+  const __hip_bfloat16* grad_bf  = reinterpret_cast<const __hip_bfloat16*>(&grad_v);
+  __hip_bfloat16*       out_bf   = reinterpret_cast<__hip_bfloat16*>(&out_v);
+
+  #pragma unroll
+  for (int i = 0; i < ELEMS_PER_THREAD; i++) {
+    out_bf[i] = (__hip_bfloat16)((float)grad_bf[i] + (float)chunk_bf[i]);
+  }
+
+  *reinterpret_cast<uint4*>(&grad_buf[global_offset]) = out_v;
+}

extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/__init__.py

@@ -0,0 +1,153 @@
+from __future__ import annotations
+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
+
+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,
+                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,
+                  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)))
+  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
+             f"-DEPS_LITERAL={eps_val}f"]
+  src = _src()
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               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,
+                    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,
+                  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)))
+  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
+             f"-DEPS_LITERAL={eps_val}f", f"-DHAS_RESIDUAL=1"]
+  src = _src()
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+@functools.cache
+def _custom_bwd(grad_x:UOp, grad_weight_partial:UOp,
+                grad_fp8:UOp, x_normed:UOp, rrms:UOp, weight:UOp, amax_state:UOp, dname:str) -> UOp:
+  MBS, SEQ, HIDDEN = x_normed.shape
+  n_elems = MBS * SEQ * HIDDEN
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 * 3 + NUM_WG * HIDDEN * 4 + MBS * SEQ * 4 + HIDDEN * 2 + 4
+  sink = UOp.sink(grad_x.base, grad_weight_partial.base,
+                  grad_fp8.base, x_normed.base, rrms.base, weight.base, amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"fused_rmsnorm_mul_quantize_fp8_bwd_{n_elems}_h{HIDDEN}",
+                                 estimates=Estimates(ops=8*n_elems, mem=mem)))
+  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  src = _src_bwd()
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel:UOp):
+  device = x_u.device
+  MBS, SEQ, HIDDEN = x_normed_u.shape
+  axis = x_normed_u.axis if isinstance(device, tuple) else None
+  grad_x = alloc_like((MBS, SEQ, HIDDEN), dtypes.bfloat16, device, axis)
+  grad_weight_partial = alloc_local((NUM_WG, HIDDEN), dtypes.float32, device, axis)
+  grad_h_from_fp8 = None
+  grad_weight_uop = None
+  if fp8_grad_u is not None:
+    fxn = functools.partial(_custom_bwd, dname=dname_of(device))
+    grad_x_t, grad_weight_partial_t, *_ = Tensor.custom_kernel(
+      grad_x, grad_weight_partial,
+      Tensor(fp8_grad_u, device=device).cast(dtypes.bfloat16),
+      Tensor(x_normed_u.after(kernel), device=device),
+      Tensor(rrms_u.after(kernel), device=device),
+      Tensor(weight_u, device=device),
+      Tensor(amax_state_u, device=device), fxn=fxn)
+    grad_h_from_fp8 = grad_x_t
+    grad_weight_uop = grad_weight_partial_t.sum(axis=0).cast(dtypes.bfloat16).uop
+  if h_grad_u is not None:
+    h_grad_t = Tensor(h_grad_u, device=device).cast(dtypes.bfloat16)
+    grad_total = (grad_h_from_fp8 + h_grad_t) if grad_h_from_fp8 is not None else h_grad_t
+  else:
+    grad_total = grad_h_from_fp8
+  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)
+  _, 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)
+
+def _fused_add_bwd(*args, **kwargs):
+  # Two invocation modes: 1 grad => positional; >1 grads => kwarg `call=`.
+  # Outputs: (fp8_out, h_out, x_normed_out, rrms_out, amax_buf). Both fp8 and h may be consumed
+  # downstream — TUPLE order in gradient.py preserves kernel-output slot order.
+  # Don't dispatch by dtype: matmul's bwd emits fp8 grad as bf16 (no explicit cast), so
+  # dtype-detection collapses both into h_grad and silently drops the rmsnorm-bwd path.
+  if 'call' in kwargs:
+    kernel, all_grads = kwargs['call'], list(args)
+  else:
+    gradient, kernel = args
+    all_grads = [gradient]
+  fp8_grad_u = h_grad_u = None
+  if len(all_grads) >= 2:
+    fp8_grad_u, h_grad_u = all_grads[0], all_grads[1]
+  elif len(all_grads) == 1:
+    g = all_grads[0]
+    if g.dtype == dtypes.bfloat16: h_grad_u = g
+    else: fp8_grad_u = g
+  _, _, x_normed_u, rrms_u, _, x_u, _, weight_u, amax_state_u = kernel.src[1:]
+  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, Tensor]:
+  # NOTE: rmsnorm(x) * weight -> fp8 + amax. Returns (fp8, inv_scale, new_amax, 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}"
+  MBS, SEQ, HIDDEN = x.shape
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  if isinstance(x.device, tuple): assert axis in (None, 0, 1), f"unsupported sharding axis={axis}"
+  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)
+  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
+  return fp8_out, inv_scale, scalar_amax(amax_buf), 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, Tensor]:
+  # NOTE: h = x + residual; y_normed = rmsnorm(h); fp8 = quantize(y_normed * weight).
+  # Returns (fp8, inv_scale, new_amax, 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
+  MBS, SEQ, HIDDEN = x.shape
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  if isinstance(x.device, tuple): assert axis in (None, 0, 1), f"unsupported sharding axis={axis}"
+  fp8_out      = alloc_like((MBS, SEQ, HIDDEN), fp8_dtype,       x.device, axis)
+  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,
+    fxn=fxn, grad_fxn=_fused_add_bwd)
+  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
+  return fp8_out, inv_scale, scalar_amax(amax_buf), h_out, x_normed_out, rrms_out

extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/fused_rmsnorm_mul_quantize_fp8.cpp

@@ -0,0 +1,155 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+#include <hip/hip_fp8.h>
+
+// Fuses the full pre-matmul preparation for a layer into a single HBM pass:
+//   y = rmsnorm(x) * weight            (reduce-mean-square + rsqrt + per-elem mul)
+//   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
+//
+// 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.
+
+#ifndef N_ELEMS
+#define N_ELEMS 67108864
+#endif
+#ifndef HIDDEN
+#define HIDDEN 4096
+#endif
+#ifndef NUM_WG
+#define NUM_WG 1024
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+#ifndef EPS_LITERAL
+#define EPS_LITERAL 1e-5f
+#endif
+#ifndef HAS_RESIDUAL
+#define HAS_RESIDUAL 0
+#endif
+
+constexpr int VEC = 8;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(N_ELEMS % HIDDEN == 0, "N_ELEMS must be a multiple of HIDDEN");
+static_assert(HIDDEN % (THREADS_PER_WG * VEC) == 0, "HIDDEN must be divisible by THREADS_PER_WG*VEC");
+
+constexpr int ROWS = N_ELEMS / HIDDEN;
+constexpr int ELEMS_PER_THREAD = HIDDEN / THREADS_PER_WG;  // each thread sees this many elems per row
+constexpr int VECS_PER_THREAD = ELEMS_PER_THREAD / VEC;    // number of 8-wide vec loads
+
+#if HAS_RESIDUAL
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fused_add_rmsnorm_mul_quantize_fp8(
+    __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, ROWS*HIDDEN
+    __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
+    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
+    const float*          __restrict__ amax_state)      // fp32 scalar
+{
+#else
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+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
+    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
+{
+#endif
+  __shared__ float sdata[THREADS_PER_WG];
+
+  const int tid = threadIdx.x;
+  const int wg  = blockIdx.x;
+
+  const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+  const float inv_hidden = 1.0f / static_cast<float>(HIDDEN);
+  float local_max = 0.0f;
+
+  // Grid-stride over rows. Each WG processes rows (wg, wg+NUM_WG, wg+2*NUM_WG, ...).
+  for (int row = wg; row < ROWS; row += NUM_WG) {
+    const int row_off = row * HIDDEN;
+
+    // Load row (+ residual if present) into registers.
+    float regs[ELEMS_PER_THREAD];
+    float sum_sq = 0.0f;
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      float4 raw = *reinterpret_cast<const float4*>(&x[row_off + h_base]);
+      const __hip_bfloat16 *xi = reinterpret_cast<const __hip_bfloat16*>(&raw);
+#if HAS_RESIDUAL
+      float4 res_raw = *reinterpret_cast<const float4*>(&residual[row_off + h_base]);
+      const __hip_bfloat16 *ri = reinterpret_cast<const __hip_bfloat16*>(&res_raw);
+      __hip_bfloat16 h_buf[VEC];
+#endif
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+#if HAS_RESIDUAL
+        const float f = static_cast<float>(xi[i]) + static_cast<float>(ri[i]);
+        h_buf[i] = static_cast<__hip_bfloat16>(f);
+#else
+        const float f = static_cast<float>(xi[i]);
+#endif
+        regs[v * VEC + i] = f;
+        sum_sq += f * f;
+      }
+#if HAS_RESIDUAL
+      *reinterpret_cast<float4*>(&h_out[row_off + h_base]) = *reinterpret_cast<float4*>(h_buf);
+#endif
+    }
+
+    // LDS tree-reduce sum_sq across the WG.
+    sdata[tid] = sum_sq;
+    __syncthreads();
+    for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
+      if (tid < s) sdata[tid] = sdata[tid] + sdata[tid + s];
+      __syncthreads();
+    }
+    const float mean_sq = sdata[0] * inv_hidden;
+    const float rrms = 1.0f / sqrtf(mean_sq + EPS_LITERAL);
+
+    if (tid == 0) rrms_out[row] = rrms;
+
+    // Normalize, multiply by weight, quantize. Also write x_normed (for rmsnorm bwd).
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      float4 w_raw = *reinterpret_cast<const float4*>(&weight[h_base]);
+      const __hip_bfloat16 *wi = reinterpret_cast<const __hip_bfloat16*>(&w_raw);
+
+      __hip_fp8_storage_t out[VEC];
+      __hip_bfloat16 xn[VEC];
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+        const float x_normed = regs[v * VEC + i] * rrms;
+        xn[i] = static_cast<__hip_bfloat16>(x_normed);
+        const float y = x_normed * static_cast<float>(wi[i]);
+        local_max = fmaxf(local_max, fabsf(y));
+        const float scaled = fmaxf(-FP8_MAX, fminf(FP8_MAX, y * scale));
+        out[i] = __hip_cvt_float_to_fp8(scaled, __HIP_SATFINITE, __HIP_E4M3);
+      }
+      *reinterpret_cast<uint64_t*>(&fp8_out[row_off + h_base]) = *reinterpret_cast<uint64_t*>(out);
+      *reinterpret_cast<float4*>(&x_normed_out[row_off + h_base]) = *reinterpret_cast<float4*>(xn);
+    }
+    __syncthreads();  // before next row's sum_sq reduce reuses sdata
+  }
+
+  // Final per-WG amax reduce.
+  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];
+}

extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/fused_rmsnorm_mul_quantize_fp8_bwd.cpp

@@ -0,0 +1,147 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+
+// Full backward for fused_rmsnorm_mul_quantize_fp8.cpp. One HBM pass per row produces:
+//   grad_x (bf16)              — gradient w.r.t. pre-rmsnorm x
+//   grad_weight_partial (fp32) — per-WG partial of the weight gradient, reduced later
+//
+// Input (all read):
+//   grad_fp8 (bf16)     — upstream grad w.r.t. fp8_out (bf16-typed gradient value)
+//   x_normed (bf16)     — saved from the fwd kernel, shape (ROWS, HIDDEN)
+//   rrms (fp32)         — saved rrms per row
+//   weight (bf16)       — per-HIDDEN rmsnorm weight
+//   amax_state (bf16)   — delayed amax used to compute the fp8 scale in fwd
+//
+// Chain: y = x_normed * weight; fp8 = sat(y * scale). Through STE: grad_y = grad_fp8 * scale.
+//        grad_x_normed = grad_y * weight.
+//        grad_weight   = sum_rows(grad_y * x_normed).
+//        grad_x        = rrms * (grad_x_normed - x_normed * mean(grad_x_normed * x_normed, last_dim)).
+
+#ifndef N_ELEMS
+#define N_ELEMS 67108864
+#endif
+#ifndef HIDDEN
+#define HIDDEN 4096
+#endif
+#ifndef NUM_WG
+#define NUM_WG 1024
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int VEC = 8;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(N_ELEMS % HIDDEN == 0, "N_ELEMS must be a multiple of HIDDEN");
+static_assert(HIDDEN % (THREADS_PER_WG * VEC) == 0, "HIDDEN must be divisible by THREADS_PER_WG*VEC");
+
+constexpr int ROWS = N_ELEMS / HIDDEN;
+constexpr int ELEMS_PER_THREAD = HIDDEN / THREADS_PER_WG;
+constexpr int VECS_PER_THREAD = ELEMS_PER_THREAD / VEC;
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fused_rmsnorm_mul_quantize_fp8_bwd(
+    __hip_bfloat16*       __restrict__ grad_x,              // out: bf16, ROWS*HIDDEN
+    float*                __restrict__ grad_weight_partial, // out: fp32, NUM_WG*HIDDEN
+    const __hip_bfloat16* __restrict__ grad_fp8,            // in:  bf16, ROWS*HIDDEN (grad of fp8_out)
+    const __hip_bfloat16* __restrict__ x_normed,            // in:  bf16, ROWS*HIDDEN
+    const float*          __restrict__ rrms,                // in:  fp32, ROWS
+    const __hip_bfloat16* __restrict__ weight,              // in:  bf16, HIDDEN
+    const float*          __restrict__ amax_state)          // in:  fp32 scalar
+{
+  __shared__ float sdata[THREADS_PER_WG];
+
+  const int tid = threadIdx.x;
+  const int wg  = blockIdx.x;
+
+  const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+  const float inv_hidden = 1.0f / static_cast<float>(HIDDEN);
+
+  // Per-thread accumulator for grad_weight (across all rows this WG touches).
+  float gw_accum[ELEMS_PER_THREAD];
+  #pragma unroll
+  for (int i = 0; i < ELEMS_PER_THREAD; i++) gw_accum[i] = 0.0f;
+
+  // Preload weight into registers (same across rows). Use ELEMS_PER_THREAD entries.
+  float w_regs[ELEMS_PER_THREAD];
+  #pragma unroll
+  for (int v = 0; v < VECS_PER_THREAD; v++) {
+    const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+    float4 w_raw = *reinterpret_cast<const float4*>(&weight[h_base]);
+    const __hip_bfloat16 *wi = reinterpret_cast<const __hip_bfloat16*>(&w_raw);
+    #pragma unroll
+    for (int i = 0; i < VEC; i++) w_regs[v * VEC + i] = static_cast<float>(wi[i]);
+  }
+
+  for (int row = wg; row < ROWS; row += NUM_WG) {
+    const int row_off = row * HIDDEN;
+    const float rrms_v = rrms[row];
+
+    // Load grad_fp8 and x_normed rows into registers, compute grad_y and grad_x_normed.
+    float g_y_regs[ELEMS_PER_THREAD];
+    float xn_regs[ELEMS_PER_THREAD];
+    float g_xn_regs[ELEMS_PER_THREAD];  // grad_x_normed
+    float local_dot = 0.0f;  // sum(grad_x_normed * x_normed) for mean
+
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      float4 g_raw  = *reinterpret_cast<const float4*>(&grad_fp8[row_off + h_base]);
+      float4 xn_raw = *reinterpret_cast<const float4*>(&x_normed[row_off + h_base]);
+      const __hip_bfloat16 *gi  = reinterpret_cast<const __hip_bfloat16*>(&g_raw);
+      const __hip_bfloat16 *xni = reinterpret_cast<const __hip_bfloat16*>(&xn_raw);
+
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+        const int idx = v * VEC + i;
+        const float g_y = static_cast<float>(gi[i]) * scale;
+        const float xn  = static_cast<float>(xni[i]);
+        g_y_regs[idx] = g_y;
+        xn_regs[idx]  = xn;
+        g_xn_regs[idx] = g_y * w_regs[idx];           // grad_x_normed = grad_y * weight
+        gw_accum[idx] += g_y * xn;                    // grad_weight contrib
+        local_dot += g_xn_regs[idx] * xn;             // for mean
+      }
+    }
+
+    // LDS reduce local_dot to sdata[0].
+    sdata[tid] = local_dot;
+    __syncthreads();
+    for (int s = THREADS_PER_WG / 2; s > 0; s >>= 1) {
+      if (tid < s) sdata[tid] = sdata[tid] + sdata[tid + s];
+      __syncthreads();
+    }
+    const float mean_term = sdata[0] * inv_hidden;
+
+    // Compute grad_x = rrms * (grad_x_normed - x_normed * mean_term) and write.
+    #pragma unroll
+    for (int v = 0; v < VECS_PER_THREAD; v++) {
+      const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+      __hip_bfloat16 out[VEC];
+      #pragma unroll
+      for (int i = 0; i < VEC; i++) {
+        const int idx = v * VEC + i;
+        const float dx = rrms_v * (g_xn_regs[idx] - xn_regs[idx] * mean_term);
+        out[i] = static_cast<__hip_bfloat16>(dx);
+      }
+      *reinterpret_cast<float4*>(&grad_x[row_off + h_base]) = *reinterpret_cast<float4*>(out);
+    }
+    __syncthreads();
+  }
+
+  // Write this WG's grad_weight partial to HBM (fp32, NUM_WG x HIDDEN layout).
+  const int gw_row_off = wg * HIDDEN;
+  #pragma unroll
+  for (int v = 0; v < VECS_PER_THREAD; v++) {
+    const int h_base = tid * VEC + v * THREADS_PER_WG * VEC;
+    // Write 8 fp32 values with two float4 stores.
+    float4 out_lo, out_hi;
+    out_lo.x = gw_accum[v * VEC + 0]; out_lo.y = gw_accum[v * VEC + 1];
+    out_lo.z = gw_accum[v * VEC + 2]; out_lo.w = gw_accum[v * VEC + 3];
+    out_hi.x = gw_accum[v * VEC + 4]; out_hi.y = gw_accum[v * VEC + 5];
+    out_hi.z = gw_accum[v * VEC + 6]; out_hi.w = gw_accum[v * VEC + 7];
+    *reinterpret_cast<float4*>(&grad_weight_partial[gw_row_off + h_base + 0]) = out_lo;
+    *reinterpret_cast<float4*>(&grad_weight_partial[gw_row_off + h_base + 4]) = out_hi;
+  }
+}

extra/llama_kernels/quantize_fp8_delayed/__init__.py

@@ -0,0 +1,67 @@
+from __future__ import annotations
+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
+
+@functools.cache
+def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_state:UOp, dname:str) -> UOp:
+  n_elems = 1
+  for d in x.shape: n_elems *= d
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 + n_elems + 4 + NUM_WG * 4
+  sink = UOp.sink(fp8_out.base, amax_partial.base, x.base, amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"quantize_fp8_with_amax_{n_elems}", estimates=Estimates(ops=3*n_elems, mem=mem)))
+  src = (pathlib.Path(__file__).parent/"quantize_fp8_with_amax.cpp").read_text()
+  defines = [f"-DN_ELEMS={n_elems}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+@functools.cache
+def _custom_quantize_fp8_scalar(fp8_out:UOp, x:UOp, amax_state:UOp, dname:str) -> UOp:
+  n_elems = 1
+  for d in x.shape: n_elems *= d
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
+  mem = n_elems * 2 + n_elems
+  sink = UOp.sink(fp8_out.base, x.base, amax_state.base, threads, workgroups,
+                  arg=KernelInfo(f"quantize_fp8_scalar_{n_elems}", estimates=Estimates(ops=2*n_elems, mem=mem)))
+  src = (pathlib.Path(__file__).parent/"quantize_fp8_scalar.cpp").read_text()
+  defines = [f"-DN_ELEMS={n_elems}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def _quantize_fp8_delayed_bwd(gradient:UOp, kernel:UOp):
+  # NOTE: STE-equivalent backward — grad_x = grad_fp8 * scale, scale = FP8_MAX / amax_state.
+  # `gradient` is bf16 grad w.r.t. fp8 output (asm_gemm bwd already applied x_scale).
+  _, _, x, amax_state = kernel.src[1:]
+  device = x.device
+  scale = FP8_MAX / (Tensor(amax_state, device=device).float() + 1e-8)
+  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.
+  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)
+  amax_partial = alloc_local((NUM_WG,), dtypes.float32, x.device, axis)
+  fxn = functools.partial(_custom_quantize_fp8_with_amax, dname=dname_of(x.device))
+  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)
+  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
+
+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.
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  fp8_out = alloc_like(x.shape, fp8_dtype, x.device, axis)
+  fxn = functools.partial(_custom_quantize_fp8_scalar, dname=dname_of(x.device))
+  fp8_out, *_ = Tensor.custom_kernel(fp8_out, x, amax_state, fxn=fxn)
+  return fp8_out

extra/llama_kernels/quantize_fp8_delayed/quantize_fp8_scalar.cpp

@@ -0,0 +1,48 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+#include <hip/hip_fp8.h>
+
+// Pure one-pass bf16 -> fp8 quantize with delayed scalar scale. No amax computation.
+
+#ifndef N_ELEMS
+#define N_ELEMS 67108864
+#endif
+#ifndef NUM_WG
+#define NUM_WG 1024
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int VEC = 8;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(N_ELEMS % VEC == 0, "N_ELEMS must be divisible by VEC");
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+quantize_fp8_scalar(
+    __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, N_ELEMS
+    const __hip_bfloat16* __restrict__ x,               // bf16, N_ELEMS
+    const float*          __restrict__ amax_state)      // fp32 scalar (delayed)
+{
+  const int tid = threadIdx.x;
+  const int wg  = blockIdx.x;
+  const int gid = wg * THREADS_PER_WG + tid;
+  const int stride_elems = NUM_WG * THREADS_PER_WG * VEC;
+
+  const float scale = FP8_MAX / (static_cast<float>(*amax_state) + 1e-8f);
+
+  for (int base = gid * VEC; base < N_ELEMS; base += stride_elems) {
+    float4 x_raw = *reinterpret_cast<const float4*>(&x[base]);
+    const __hip_bfloat16 *xi = reinterpret_cast<const __hip_bfloat16*>(&x_raw);
+
+    __hip_fp8_storage_t out[VEC];
+    #pragma unroll
+    for (int i = 0; i < VEC; i++) {
+      const float v = static_cast<float>(xi[i]);
+      const float scaled = fmaxf(-FP8_MAX, fminf(FP8_MAX, v * scale));
+      out[i] = __hip_cvt_float_to_fp8(scaled, __HIP_SATFINITE, __HIP_E4M3);
+    }
+    *reinterpret_cast<uint64_t*>(&fp8_out[base]) = *reinterpret_cast<uint64_t*>(out);
+  }
+}

extra/llama_kernels/quantize_fp8_delayed/quantize_fp8_with_amax.cpp

@@ -2,12 +2,13 @@
 #include <hip/hip_bf16.h>
 #include <hip/hip_fp8.h>
 
+// One-pass bf16 -> fp8 quantize using a scalar delayed amax state,
+// AND simultaneously computes per-WG |x| max partials for the next step's amax state.
+// Saves one full HBM pass over the grad tensor vs. doing quantize + separate abs().max().
+
 #ifndef N_ELEMS
 #define N_ELEMS 67108864
 #endif
-#ifndef HIDDEN
-#define HIDDEN 4096
-#endif
 #ifndef NUM_WG
 #define NUM_WG 1024
 #endif
@@ -19,15 +20,13 @@ constexpr int VEC = 8;
 constexpr float FP8_MAX = 448.0f;
 
 static_assert(N_ELEMS % VEC == 0, "N_ELEMS must be divisible by VEC");
-static_assert(HIDDEN % VEC == 0, "HIDDEN must be divisible by VEC");
 
 extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
-fused_mul_quantize_fp8(
-    __hip_fp8_storage_t*  __restrict__ fp8_out,         // fp8, N_ELEMS
-    __hip_bfloat16*       __restrict__ amax_buf,        // bf16, NUM_WG
-    const __hip_bfloat16* __restrict__ x,               // bf16, N_ELEMS
-    const __hip_bfloat16* __restrict__ weight,          // bf16, HIDDEN (per-hidden scale)
-    const __hip_bfloat16* __restrict__ amax_state)      // bf16 scalar
+quantize_fp8_with_amax(
+    __hip_fp8_storage_t*  __restrict__ fp8_out,       // out: fp8, N_ELEMS
+    float*                __restrict__ amax_partial,  // out: fp32, NUM_WG per-WG partials
+    const __hip_bfloat16* __restrict__ x,             // in:  bf16, N_ELEMS
+    const float*          __restrict__ amax_state)    // in:  fp32 scalar (delayed)
 {
   __shared__ float sdata[THREADS_PER_WG];
 
@@ -40,32 +39,25 @@ fused_mul_quantize_fp8(
   float local_max = 0.0f;
 
   for (int base = gid * VEC; base < N_ELEMS; base += stride_elems) {
-    const int h = base % HIDDEN;   // 0..HIDDEN-VEC, 8-aligned (since base is 8-aligned and HIDDEN divides VEC)
     float4 x_raw = *reinterpret_cast<const float4*>(&x[base]);
-    float4 w_raw = *reinterpret_cast<const float4*>(&weight[h]);
-
     const __hip_bfloat16 *xi = reinterpret_cast<const __hip_bfloat16*>(&x_raw);
-    const __hip_bfloat16 *wi = reinterpret_cast<const __hip_bfloat16*>(&w_raw);
 
     __hip_fp8_storage_t out[VEC];
     #pragma unroll
     for (int i = 0; i < VEC; i++) {
-      const float val = static_cast<float>(xi[i]) * static_cast<float>(wi[i]);
-      local_max = fmaxf(local_max, fabsf(val));
-      const float scaled = fmaxf(-FP8_MAX, fminf(FP8_MAX, val * scale));
+      const float v = static_cast<float>(xi[i]);
+      local_max = fmaxf(local_max, fabsf(v));
+      const float scaled = fmaxf(-FP8_MAX, fminf(FP8_MAX, v * scale));
       out[i] = __hip_cvt_float_to_fp8(scaled, __HIP_SATFINITE, __HIP_E4M3);
     }
-
     *reinterpret_cast<uint64_t*>(&fp8_out[base]) = *reinterpret_cast<uint64_t*>(out);
   }
 
-  // LDS tree-reduce per-WG amax
   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] = static_cast<__hip_bfloat16>(sdata[0]);
+  if (tid == 0) amax_partial[wg] = sdata[0];
 }

extra/sqtt/roc.py

@@ -1,10 +1,13 @@
 #!/usr/bin/env python3
-import ctypes, pathlib, argparse, pickle, dataclasses, threading
+import ctypes, pathlib, argparse, pickle, dataclasses, threading, itertools
+from decimal import Decimal
 from typing import Generator
 from tinygrad.helpers import temp, unwrap, DEBUG
 from tinygrad.runtime.ops_amd import ProfileSQTTEvent
 from tinygrad.runtime.autogen import rocprof
 from tinygrad.renderer.amd.dsl import Inst
+from tinygrad.helpers import ProfileEvent, ProfileRangeEvent, ProfilePointEvent
+from tinygrad.device import ProfileProgramEvent
 from test.amd.disasm import disasm
 
 @dataclasses.dataclass(frozen=True)
@@ -126,6 +129,71 @@ def decode(sqtt_evs:list[ProfileSQTTEvent], disasms:dict[str, dict[int, Inst]])
     raise exc
   return ROCParseCtx
 
+def unpack_occ(viz_data, i:int, j:int, key:tuple[str, int], data:list, p:ProfileProgramEvent, target:str) -> dict:
+  from tinygrad.viz.serve import amd_decode, create_step, row_tuple
+  steps = viz_data.ctxs[i]["steps"]
+  if len(steps[j+1:]) > 0: return {"steps":[{k:v for k,v in s.items() if k != "data"} for s in steps[j+1:]]}
+  base = unwrap(p.base)
+  disasm:dict[int, Inst] = {addr+base:inst for addr,inst in amd_decode(unwrap(p.lib), target).items()}
+  rctx = decode(data, {p.tag:disasm})
+  cu_events:dict[str, list[ProfileEvent]] = {}
+  # ** inst traces
+  wave_insts:dict[str, dict[str, dict]] = {}
+  inst_units:dict[str, itertools.count] = {}
+  for w in rctx.inst_execs.get(key, []):
+    if (u:=w.wave_loc) not in inst_units: inst_units[u] = itertools.count(0)
+    n = next(inst_units[u])
+    if (events:=cu_events.get(w.cu_loc)) is None: cu_events[w.cu_loc] = events = []
+    events.append(ProfileRangeEvent(f"SIMD:{w.simd}", loc:=f"INST WAVE:{w.wave_id} N:{n}", Decimal(w.begin_time), Decimal(w.end_time)))
+    wave_insts.setdefault(w.cu_loc, {})[f"{u} N:{n}"] = {"wave":w, "disasm":disasm, "prg":p, "run_number":n, "loc":loc}
+  # ** occ traces (only WAVESTART/WAVEEND)
+  units:dict[str, itertools.count] = {}
+  wave_start:dict[str, int] = {}
+  for occ in rctx.occ_events.get(key, []):
+    if (u:=occ.wave_loc) not in units: units[u] = itertools.count(0)
+    if u in inst_units: continue
+    if occ.start: wave_start[u] = occ.time
+    else:
+      if (events:=cu_events.get(occ.cu_loc)) is None: cu_events[occ.cu_loc] = events = []
+      events.append(ProfileRangeEvent(f"SIMD:{occ.simd}", f"OCC WAVE:{occ.wave_id} N:{next(units[u])}", Decimal(wave_start.pop(u)),Decimal(occ.time)))
+  # ** split graph by CU
+  for cu in sorted(cu_events, key=row_tuple):
+    steps.append(create_step(f"{cu} {len(cu_events[cu])}", ("/cu-sqtt", i, len(steps)), depth=1,
+                             data=[ProfilePointEvent(unit, "start", unit, ts=Decimal(0)) for unit in units]+cu_events[cu]))
+    for k in sorted(wave_insts.get(cu, []), key=row_tuple):
+      wd = wave_insts[cu][k]
+      steps.append(create_step(k.replace(cu, ""), ("/amd-sqtt-insts", i, len(steps)), loc=wd["loc"], depth=2,
+                               data={"fxn":unpack_insts, "args":(wd,)}))
+  return {"steps":[{k:v for k,v in s.items() if k != "data"} for s in steps[j+1:]]}
+
+def unpack_insts(viz_data, i:int, j:int, data:dict) -> dict:
+  columns = ["PC", "Instruction", "Hits", "Cycles", "Stall", "Type"]
+  inst_columns = ["N", "Clk", "Idle", "Dur", "Stall"]
+  # Idle:     The total time gap between the completion of previous instruction and the beginning of the current instruction.
+  #           The idle time can be caused by:
+  #             * Arbiter loss
+  #             * Source or destination register dependency
+  #             * Instruction cache miss
+  # Stall:    The total number of cycles the hardware pipe couldn't issue an instruction.
+  # Duration: Total latency in cycles, defined as "Stall time + Issue time" for gfx9 or "Stall time + Execute time" for gfx10+.
+  prev_instr = (w:=data["wave"]).begin_time
+  pc_to_inst = data["disasm"]
+  start_pc = None
+  rows:dict[int, dict] = {}
+  for pc, inst in pc_to_inst.items():
+    if start_pc is None: start_pc = pc
+    rows[pc] = {"pc":pc-start_pc, "inst":str(inst), "hit_count":0, "dur":0, "stall":0, "type":"", "hits":{"cols":inst_columns, "rows":[]}}
+  for e in w.unpack_insts():
+    if not (inst:=rows[e.pc]).get("type"): inst["type"] = str(e.typ).split("_")[-1]
+    inst["hit_count"] += 1
+    inst["dur"] += e.dur
+    inst["stall"] += e.stall
+    inst["hits"]["rows"].append((inst["hit_count"]-1, e.time, max(0, e.time-prev_instr), e.dur, e.stall))
+    prev_instr = max(prev_instr, e.time + e.dur)
+  summary = [{"label":"Total Cycles", "value":w.end_time-w.begin_time}, {"label":"SE", "value":w.se}, {"label":"CU", "value":w.cu},
+             {"label":"SIMD", "value":w.simd}, {"label":"Wave ID", "value":w.wave_id}, {"label":"Run number", "value":data["run_number"]}]
+  return {"rows":[tuple(v.values()) for v in rows.values()], "cols":columns, "metadata":[summary], "ref":viz_data.ref_map.get(data["prg"].name)}
+
 def print_data(data:dict) -> None:
   from tabulate import tabulate
   # plaintext

extra/thunder/amd/fa.py

@@ -55,8 +55,6 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
   assert attn_mask is None, "attn_mask not supported"
   assert is_causal, "only causal attention supported"
 
-  xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-
   B, N, H, D = xq.shape
   H_KV = xk.shape[2]
   assert D == 128, "only D=128 supported"
@@ -81,7 +79,7 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
 
   attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D), grad_fxn=grad)[:2]
 
-  return attn.transpose(1, 2), attn, l_vec
+  return attn, attn, l_vec
 
 @functools.cache
 def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):

extra/thunder/amd/gemm_fp8.cpp

@@ -93,7 +93,20 @@ constexpr int NUM_WARPS = 8;
 
 using G = kittens::group<NUM_WARPS>;
 
-__global__ __launch_bounds__(512, 2) void hk_fp8_gemm(bf16 *C_ptr, fp8e4m3 *A_ptr, fp8e4m3 *B_ptr, float *x_scale_ptr, float *w_scale_ptr) {
+// scale_mode: 0=no scale, 1=x only, 2=w only, 3=both
+#ifndef SCALE_MODE
+#define SCALE_MODE 3
+#endif
+
+__global__ __launch_bounds__(512, 2) void hk_fp8_gemm(bf16 *C_ptr, fp8e4m3 *A_ptr, fp8e4m3 *B_ptr
+#if SCALE_MODE == 1
+    , float *x_scale_ptr
+#elif SCALE_MODE == 2
+    , float *w_scale_ptr
+#elif SCALE_MODE == 3
+    , float *x_scale_ptr, float *w_scale_ptr
+#endif
+) {
     constexpr int M = GEMM_M, N = GEMM_N, K = GEMM_K;
 
     kittens::gl<fp8e4m3, 1, 1, M, K> A{A_ptr, nullptr, nullptr, nullptr, nullptr};
@@ -333,11 +346,25 @@ __global__ __launch_bounds__(512, 2) void hk_fp8_gemm(bf16 *C_ptr, fp8e4m3 *A_pt
     }
 
     // apply x_scale * w_scale before bf16 store to prevent overflow
+#if SCALE_MODE == 1
+    float scale = *x_scale_ptr;
+    mul(cA, cA, scale);
+    mul(cB, cB, scale);
+    mul(cC, cC, scale);
+    mul(cD, cD, scale);
+#elif SCALE_MODE == 2
+    float scale = *w_scale_ptr;
+    mul(cA, cA, scale);
+    mul(cB, cB, scale);
+    mul(cC, cC, scale);
+    mul(cD, cD, scale);
+#elif SCALE_MODE == 3
     float scale = *x_scale_ptr * *w_scale_ptr;
     mul(cA, cA, scale);
     mul(cB, cB, scale);
     mul(cC, cC, scale);
     mul(cD, cD, scale);
+#endif
 
     store(C, cA, {0, 0, block_row * WARPS_ROW * 2 + warp_m, block_col * WARPS_COL * 2 + warp_n});
     store(C, cB, {0, 0, block_row * WARPS_ROW * 2 + warp_m, block_col * WARPS_COL * 2 + WARPS_COL + warp_n});

pyproject.toml

@@ -251,8 +251,7 @@ select = [
   "F541",
   "F841",
 ]
-"tinygrad/runtime/autogen/**/*.py" = ["E501", "F401", "E722", "E731", "F821", "A006", "A002", "F811"]
-"tinygrad/runtime/autogen/amd/**/*.py" = ["E501"]
+"tinygrad/runtime/autogen/**/*.py" = ["E501", "F401", "E731", "F821", "A006", "A002", "F811", "F822"]
 "test/amd/**/*.py" = ["F403", "F405"]
 
 [tool.ruff.format]

test/amd/test_custom_kernel.py

@@ -3,7 +3,7 @@ import functools
 import numpy as np
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.engine.realize import run_linear, estimate_uop
+from tinygrad.engine.realize import run_linear, estimate_uop, compile_linear
 from tinygrad.renderer import Estimates
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import getenv
@@ -99,13 +99,13 @@ def custom_lds_sync(A:UOp, arch:str) -> UOp:
   sink = UOp.sink(A.base, lds, threads, wg, arg=KernelInfo("custom_lds_sync"))
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
 
-def custom_handwritten(A:UOp, arch:str) -> UOp:
+def custom_handwritten(A:UOp) -> UOp:
   A = A.flatten()
   threads = UOp.special(128, "lidx0")
   wg = UOp.special(1, "gidx0")
   lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=512, addrspace=AddrSpace.LOCAL), (), 'lds')  # 128 * 4 bytes
   pipes = {getenv("PIPE", "")} if getenv("PIPE", "") else {"SALU", "VALU", "TRANSCENDENTAL", "WMMA"}
-  k = Kernel(arch)
+  k = Kernel()
   # wrap in loop to filter out icache misses
   LOOP_N, UNROLL_N = 8, 5
   k.emit(r4.s_mov_b32(s[1], LOOP_N))
@@ -145,10 +145,10 @@ def custom_handwritten(A:UOp, arch:str) -> UOp:
   sink = UOp.sink(A.base, threads, wg, lds, arg=KernelInfo("custom_handwritten"))
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
 
-def custom_data_deps(A:UOp, arch:str) -> UOp:
+def custom_data_deps(A:UOp) -> UOp:
   A = A.flatten()
   threads = UOp.special(A.numel(), "lidx0")
-  k = Kernel(arch)
+  k = Kernel()
   k.emit(s_load_b64(s[0:1], s[0:1], soffset=NULL))
   k.emit(s_waitcnt_lgkmcnt(sdst=NULL, simm16=0))
   k.emit(v_lshlrev_b32_e32(v[0], 2, v[0]))
@@ -169,7 +169,7 @@ class TestCustomKernel(unittest.TestCase):
     if self.arch != "rdna3": self.skipTest("only rdna3")
     a = Tensor.full((16, 16), 1.).contiguous().realize()
     a = Tensor.custom_kernel(a, fxn=custom_add_one)[0]
-    linear = a.schedule_linear()
+    linear = compile_linear(a.schedule_linear())
     est = estimate_uop(linear.src[-1])
     self.assertEqual(est.ops, a.numel())
     self.assertEqual(est.mem, a.nbytes()*2)
@@ -198,13 +198,13 @@ class TestCustomKernel(unittest.TestCase):
   def test_handwritten(self):
     if self.arch != "rdna4": self.skipTest("only tested on rdna4")
     a = Tensor.empty(1024, dtype=dtypes.int32).contiguous().realize()
-    a = Tensor.custom_kernel(a, fxn=functools.partial(custom_handwritten, arch=self.arch))[0]
+    a = Tensor.custom_kernel(a, fxn=custom_handwritten)[0]
     a.realize()
 
   def test_data_deps(self):
     if self.arch != "rdna3": self.skipTest("only tested on rdna3")
     a = Tensor(np.full(32, 5.0, dtype=np.float32)).realize()
-    a = Tensor.custom_kernel(a, fxn=functools.partial(custom_data_deps, arch=self.arch))[0]
+    a = Tensor.custom_kernel(a, fxn=custom_data_deps)[0]
     a.realize()
     self.assertTrue((a.numpy() == 6.0).all())
 

test/amd/test_mockgpu_invalid.py

@@ -8,17 +8,14 @@ class TestMockGPUInvalidInstruction(unittest.TestCase):
     test_code = '''
 import struct
 from tinygrad import Device, Tensor
-from tinygrad.engine.realize import get_runner
+from tinygrad.engine.realize import compile_linear
 from tinygrad.runtime.ops_amd import AMDProgram
 
 dev = Device["AMD"]
 a = Tensor([1.0]).realize()
 b = a + 1
-si = b.schedule_linear().src[-1]
-runner = get_runner(dev.device, si.src[0])
-
-prg = runner._prg
-lib = bytearray(prg.lib)
+linear = compile_linear(b.schedule_linear())
+lib = bytearray(linear.src[-1].src[0].src[4].arg)
 
 # Find s_endpgm (0xBFB00000) and replace with V_MOVRELD_B32 (op=66) which has no pcode
 # VOP1 encoding: bits[31:25]=0x7E, op=bits[16:9], so op=66 -> 66<<9 = 0x8400

test/amd/test_sqttmap.py

@@ -1,5 +1,5 @@
 # test to compare every packet with the rocprof decoder
-import unittest, pickle, functools
+import unittest, pickle, functools, json
 from typing import Iterator
 from pathlib import Path
 from tinygrad.helpers import DEBUG, getenv, temp, ansistrip, Context
@@ -156,7 +156,7 @@ class TestSQTTMapRDNA4(TestSQTTMapBase):
     row_counts:dict[str, int] = {}
     for e in sqtt_timeline(events[1].blob, lib, target):
       if type(e).__name__ != "ProfileRangeEvent": continue
-      info = e.name.ret or ""
+      info = json.loads(e.name.ret) if e.name.ret else {}
       if e.device.startswith("WAVE"):
         idx = row_counts.get(e.device, 0)
         dispatch_st[f"{e.device}-{idx}"] = int(e.st)

test/backend/test_arange.py

@@ -2,14 +2,14 @@ import unittest
 import numpy as np
 from tinygrad import Tensor, GlobalCounters, dtypes, nn, Device, Variable
 from tinygrad.helpers import Context, getenv, DEV
-from tinygrad.engine.realize import run_linear, estimate_uop
+from tinygrad.engine.realize import run_linear, estimate_uop, compile_linear
 from tinygrad.renderer.ptx import PTXRenderer
 from test.helpers import needs_second_gpu
 
 class TestArange(unittest.TestCase):
   def _get_flops(self, tensor, desired):
     GlobalCounters.reset()
-    linear = tensor.schedule_linear()
+    linear = compile_linear(tensor.schedule_linear())
     self.assertEqual(len(linear.src), 1)
     run_linear(linear)
     np.testing.assert_equal(tensor.numpy(), desired)
@@ -36,7 +36,7 @@ class TestArange(unittest.TestCase):
   def test_tri_complexity(self):
     with Context(NOOPT=1):
       t = Tensor.ones(256, 256).contiguous().realize()
-      linear = t.triu().schedule_linear()
+      linear = compile_linear(t.triu().schedule_linear())
       self.assertLessEqual(estimate_uop(linear.src[-1]).ops, 4 * 256 * 256)
 
 DSET, DDIM = 2048, 32
@@ -229,7 +229,7 @@ class TestIndexing(unittest.TestCase):
     xq = xq.reshape(bs, seqlen, n_heads, head_dim)
     xq_rope, _ = apply_rotary_emb(xq, xq, freqs_cis)
     xq_rope.sum().backward()
-    linear = wq.grad.schedule_linear()
+    linear = compile_linear(wq.grad.schedule_linear())
     assert len(linear.src) == 1, f"expected one kernel for backward, got: {len(linear.src)}"
     bwd_ops = estimate_uop(linear.src[0]).ops
     # bfloat16 on non CDNA4 has ~10x ops overhead because of the software emulation

test/backend/test_custom_kernel.py

@@ -1,5 +1,5 @@
 import unittest
-from tinygrad import Tensor, UOp
+from tinygrad import Tensor, UOp, GlobalCounters
 from tinygrad.dtype import AddrSpace, dtypes
 from tinygrad.uop.ops import KernelInfo, AxisType
 
@@ -308,6 +308,36 @@ class TestCustomKernel(unittest.TestCase):
     expected = (3+2)*2+2
     assert all(x == expected for x in result), f"expected all {expected}, got {result}"
 
+  def test_custom_kernel_sched(self, use_custom=False):
+    x = Tensor.arange(32).reshape(8, 4).realize()
+    y = Tensor.empty_like(x)
+    y = Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
+    if use_custom:
+      z = Tensor.empty_like(x)
+      z = Tensor.custom_kernel(y, y.T.T, fxn=custom_add_one_kernel)[0]
+    else: z = y.T.T+1
+    GlobalCounters.reset()
+    z.realize()
+    self.assertEqual(GlobalCounters.kernel_count, 2)
+    self.assertEqual(z.tolist(), x.add(2).tolist())
+
+  @unittest.expectedFailure
+  def test_custom_kernel_sched_copy(self): self.test_custom_kernel_sched(use_custom=True)
+
+  @unittest.expectedFailure
+  def test_sliced_buffer_function(self):
+    x = Tensor.arange(32).reshape(8, 4).realize()
+    from tinygrad import function
+    @function(precompile=True)
+    def run(x:Tensor) -> Tensor:
+      y = Tensor.invalids(*x.shape, dtype=x.dtype)
+      return Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
+    GlobalCounters.reset()
+    y = run(x[0]).realize()
+    # it's copying the input and the output
+    self.assertEqual(GlobalCounters.kernel_count, 1)
+    self.assertEqual(y.tolist(), [1, 2, 3, 4])
+
 class TestUOpReduce(unittest.TestCase):
   def test_uop_sum(self):
     a = Tensor([1.0, 2, 3, 4, 5])

test/backend/test_jit.py

@@ -4,6 +4,7 @@ import numpy as np
 
 from hypothesis import given, settings, strategies as strat
 from test.helpers import assert_jit_cache_len, call_is_graph, not_support_multi_device, needs_second_gpu
+from tinygrad import Variable
 from tinygrad.tensor import Tensor
 from tinygrad.engine.jit import TinyJit, JitError, graph_class
 from tinygrad.device import Device
@@ -39,6 +40,19 @@ class TestJit(unittest.TestCase):
     def add(a, b): return (a+b).realize()
     _simple_test(add)
 
+  @unittest.skipUnless(Device.DEFAULT == "CPU", "core_id is a CPU runtimevar")
+  def test_hcq_core_id_runtimevar_merge(self):
+    N = 262144
+    @TinyJit
+    def f(x, st):
+      y = (x + 1).contiguous().realize()
+      z = x.shrink(((st, st + N),)).contiguous().realize()
+      return y, z
+    x = Tensor.arange(2*N).contiguous().realize()
+    for _ in range(3): y, z = f(x, Variable("a", 0, N).bind(0))
+    self.assertEqual(y.shape, (2*N,))
+    self.assertEqual(z.shape, (N,))
+
   def test_jitbeam_triggers_beam(self):
     from unittest.mock import patch
     from tinygrad.helpers import getenv as _getenv

test/backend/test_linearizer.py

@@ -2,10 +2,10 @@ import numpy as np
 import unittest
 
 from tinygrad.codegen.opt import Opt, OptOps
-from tinygrad.uop.ops import UOp, Ops, GroupOp, AxisType
+from tinygrad.uop.ops import UOp, Ops, GroupOp, AxisType, buffers
 from tinygrad.device import Device, Buffer, is_dtype_supported
 from tinygrad.tensor import Tensor, _to_np_dtype
-from tinygrad.engine.realize import run_linear, CompiledRunner
+from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
 from tinygrad.helpers import Context, flatten, dedup, TC_SELECT, TC_OPT, DEV
 from tinygrad.dtype import DType, dtypes, PtrDType, AddrSpace
@@ -14,7 +14,7 @@ from tinygrad.renderer.cstyle import CUDARenderer
 from test.helpers import replace_opts
 MOCKGPU = DEV.interface.startswith("MOCK")
 
-from tinygrad.uop.ops import print_uops # noqa: F401 # pylint: disable=unused-import
+from tinygrad.uop.render import print_uops # noqa: F401 # pylint: disable=unused-import
 
 class TestLinearizer(unittest.TestCase):
   def test_arg_dedup(self):
@@ -274,7 +274,7 @@ class TestLinearizer(unittest.TestCase):
     sched = [si for si in t.schedule_linear().src if si.src[0].op is Ops.SINK]
     # sum_collapse is a full collapse now
     assert len(sched) == 1
-    assert not any(u.op is Ops.REDUCE_AXIS for u in sched[0].src[0].toposort()), "found reduce in sum collapse"
+    assert not any(u.op is Ops.REDUCE and len(u.arg[1]) > 0 for u in sched[0].src[0].toposort()), "found reduce in sum collapse"
     #lin = Kernel(sched[0].ast)
     #assert not any(u.op is Ops.RANGE for u in lin.linearize().uops), "found loop in sum collapse"
 
@@ -424,30 +424,28 @@ def reset_bufs(bufs:list[Buffer]):
 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=[]):
   outbufs = real_bufs[:len(realized_ast.src)]
-  device = real_bufs[0].device
   wanna_output = [np.array(x).flatten() for x in wanna_output]
+  buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in real_bufs]
+  for u,b in zip(buf_uops, real_bufs): buffers[u] = b
 
-  def get_prg(opts):
+  def run_prg(opts):
     ast = realized_ast if opts is None else replace_opts(realized_ast, list(opts))
-    return CompiledRunner(to_program(ast, renderer=Device[Device.DEFAULT].renderer), device)
+    run_linear(UOp(Ops.LINEAR, src=(ast.call(*buf_uops),)))
 
   def check_opt(opts):
-    prg = get_prg(opts=opts)
     reset_bufs(outbufs)
-    prg.exec(real_bufs)
+    run_prg(opts)
     for x,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(x, want, atol=atol, rtol=rtol)
 
   # Get baseline if it is not provided, which is not optimized at all.
-  prg = get_prg(opts=())
-  prg.exec(real_bufs)
+  run_prg(opts=())
   if len(wanna_output) == 0: wanna_output = copyout_outputs(outbufs)
   else:
     for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)
 
   # Check correctness of handcoded optimiztions.
-  prg = get_prg(opts=None)
   reset_bufs(outbufs)
-  prg.exec(real_bufs)
+  run_prg(opts=None)
   for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)
   for x in opts: # Check custom transformations if any.
     check_opt(([Opt(OptOps.TC, 0, (TC_SELECT.value, TC_OPT.value, 1))] if apply_tc else [])+x)

test/backend/test_ops.py

@@ -636,6 +636,17 @@ class TestOps(unittest.TestCase):
         helper_test_op(None, lambda x: 100%x, forward_only=True, vals=[va])
         helper_test_op(None, lambda x: 100.5%x, forward_only=True, vals=[va])
 
+  def test_fmod(self):
+    a = [-4, 7, 5, 4, -7, 8, -9]
+    b = [2, -3, 8, -2, 3, 5, -5]
+    for float_a in [True, False]:
+      for float_b in [True, False]:
+        va = [float(ai) for ai in a] if float_a else a
+        vb = [float(bi) for bi in b] if float_b else b
+        helper_test_op(None, lambda x,y: x.fmod(y), forward_only=True, vals=[va, vb])
+        helper_test_op(None, lambda x: x.fmod(2), forward_only=True, vals=[va])
+        helper_test_op(None, lambda x: x.fmod(3.5), forward_only=True, vals=[va])
+
   def test_mul_naninf(self):
     helper_test_op([(45,65)], lambda x: x*math.inf)
     helper_test_op([(45,65)], lambda x: x*-math.inf)

test/backend/test_profiler.py

@@ -3,7 +3,8 @@ from tinygrad import Device, Tensor, dtypes, TinyJit
 from tinygrad.helpers import CI, DEV, Context, ProfileRangeEvent, cpu_profile, cpu_events, ProfilePointEvent, dedup
 from tinygrad.device import Buffer, BufferSpec, Compiled, ProfileDeviceEvent, ProfileGraphEvent
 from tinygrad.runtime.support.hcq import HCQCompiled
-from tinygrad.engine.realize import get_runner
+from tinygrad.engine.realize import get_runtime
+from tinygrad.codegen import to_program
 
 MOCKGPU = DEV.interface.startswith("MOCK")
 def _dev_base(d):
@@ -46,13 +47,15 @@ class TestProfiler(unittest.TestCase):
     TestProfiler.b = self.a + 1
     si = self.b.schedule_linear().src[-1]
 
-    TestProfiler.runner = get_runner(TestProfiler.d0.device, si.src[0])
+    TestProfiler.prg = to_program(si.src[0], TestProfiler.d0.renderer)
+    TestProfiler.runtime = get_runtime(TestProfiler.d0.device, TestProfiler.prg)
     TestProfiler.b.uop.buffer.allocate()
 
-  def test_profile_kernel_run(self):
-    runner_name = TestProfiler.runner._prg.name
+  def test_profile_kernel_run(self, wait=False):
+    runner_name = TestProfiler.runtime.name
     with helper_collect_profile(TestProfiler.d0) as profile:
-      TestProfiler.runner([TestProfiler.b.uop.buffer, TestProfiler.a.uop.buffer], var_vals={})
+      gs, ls = TestProfiler.prg.arg.launch_dims({})
+      TestProfiler.runtime(TestProfiler.b.uop.buffer._buf, TestProfiler.a.uop.buffer._buf, global_size=gs, local_size=ls, wait=wait)
 
     profile, _ = helper_profile_filter_device(profile, TestProfiler.d0.device)
     kernel_runs = [x for x in profile if isinstance(x, ProfileRangeEvent)]
@@ -60,6 +63,9 @@ class TestProfiler(unittest.TestCase):
     assert kernel_runs[0].name == runner_name, "kernel name is not correct"
     assert _dev_base(kernel_runs[0].device) == kernel_runs[0].device, "kernel should not be on a sub-device"
 
+  def test_profile_kernel_run_wait(self):
+    self.test_profile_kernel_run(wait=True)
+
   def test_profile_copyin(self):
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
@@ -70,12 +76,13 @@ class TestProfiler(unittest.TestCase):
     assert len(kernel_runs) == 1, "one kernel run is expected"
 
   def test_profile_multiops(self):
-    runner_name = TestProfiler.runner._prg.name
+    runner_name = TestProfiler.runtime.name
     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))))
-      TestProfiler.runner([buf1, TestProfiler.a.uop.buffer], var_vals={})
+      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)))
 
     evs = [x for x in profile if isinstance(x, ProfileRangeEvent) and x.device.startswith(TestProfiler.d0.device)]

test/backend/test_randomness.py

@@ -307,17 +307,26 @@ class TestRandomness(unittest.TestCase):
     with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0, high=3.5)
     with self.assertRaises(TypeError): Tensor.randint((3, 4), low=1, high=3, dtype="float")
     with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0, high=3, dtype=dtypes.float32)
+    # check low < high
+    with self.assertRaises(ValueError): Tensor.randint((3, 4), low=10, high=5)
+    with self.assertRaises(ValueError): Tensor.randint((3, 4), low=10, high=10)
+    np.testing.assert_array_equal(Tensor.randint(16, low=5, high=6).numpy(), 5)
 
   def test_normal(self):
     self.assertTrue(normal_test(Tensor.normal))
     self.assertTrue(equal_distribution(Tensor.normal, lambda x: torch.nn.init.normal_(torch.empty(x), mean=0, std=1),
                                                       lambda x: np.random.normal(loc=0, scale=1, size=x)))
+    # check std >= 0
+    with self.assertRaises(ValueError): Tensor.normal((3, 4), mean=0, std=-1)
 
   def test_uniform(self):
     self.assertFalse(normal_test(Tensor.uniform))
     self.assertTrue(equal_distribution(Tensor.uniform, lambda x: torch.nn.init.uniform_(torch.empty(x)), lambda x: np.random.uniform(size=x)))
     self.assertTrue(equal_distribution(partial(Tensor.uniform, low=-100, high=100, dtype=dtypes.int32),
                                        numpy_func=lambda x: np.random.randint(low=-100, high=100, size=x)))
+    # check low < high
+    with self.assertRaises(ValueError): Tensor.uniform((3, 4), low=5.0, high=3.0)
+    with self.assertRaises(ValueError): Tensor.uniform((3, 4), low=1.0, high=1.0)
 
   def test_scaled_uniform(self):
     self.assertFalse(normal_test(Tensor.scaled_uniform))
@@ -352,7 +361,7 @@ class TestRandomness(unittest.TestCase):
     _check_with_torch(w=[0.231, 0., 1., 0.5], num_samples=300, replacement=True)
     _check_with_torch(w=[[0.2, 0.8]], num_samples=300, replacement=True)  # 2D but only 1 row
     _check_with_torch(w=[[0.453, 0., 1., 0.81], [0.1, 0.8, 0., 0.1]], num_samples=300, replacement=True)
-    # no-replacement isn't supported, unless taking only one sample
+    # no-replacement
     w = [0.1, 0.9]
     self.assertRaises(AssertionError, lambda: Tensor(w).multinomial(100, replacement=False))
 
@@ -363,6 +372,23 @@ class TestRandomness(unittest.TestCase):
     torch_samples = [torch.tensor(w).multinomial(1, replacement=False).item() for _ in range(1000)]
     self.assertTrue(equal_distribution(lambda *_: Tensor(tiny_samples), lambda _: torch.tensor(torch_samples)))
 
+    w = list(range(32))
+    s1 = Tensor(w).multinomial(5, replacement=False).numpy()
+    self.assertEqual(len(set(s1.tolist())), 5)
+    s2 = Tensor(w).multinomial(5, replacement=False).numpy()
+    self.assertFalse(np.array_equal(s1, s2))
+    full = Tensor(w).multinomial(len(w), replacement=False).numpy()
+    self.assertEqual(sorted(full.tolist()), w)
+
+    w = [0.1, 0.2, 0.3, 0.4]
+    @TinyJit
+    def sample_three(): return Tensor(w).multinomial(3, replacement=False).realize()
+
+    tiny_draws = np.array([sample_three().numpy() for _ in range(1000)])
+    torch_draws = np.array([torch.tensor(w).multinomial(3, replacement=False).numpy() for _ in range(1000)])
+    for pos in range(3):
+      self.assertTrue(equal_distribution(lambda *_: Tensor(tiny_draws[:, pos]), lambda _: torch.tensor(torch_draws[:, pos])))
+
   @unittest.skip("this test is flaky")
   def test_multinomial_counterexample(self):
     tiny_res = Tensor([0.3, 0.6, 0.1]).multinomial(4000, replacement=True)

test/backend/test_renderer_failures.py

@@ -1,9 +1,8 @@
 import unittest
 import numpy as np
-from dataclasses import replace
-from tinygrad.device import Buffer, Device, is_dtype_supported
+from tinygrad.device import Device, is_dtype_supported
 from tinygrad.dtype import dtypes, ConstType
-from tinygrad.engine.realize import CompiledRunner
+from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
 from tinygrad.helpers import prod
 from tinygrad.renderer.cstyle import CStyleLanguage
@@ -13,17 +12,13 @@ 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
 
-def _test_uop_result(inputs:list[Tensor], prg:UOp, local_size=None):
+def _test_uop_result(inputs:list[Tensor], sink:UOp, local_size=None):
   for x in inputs: x.realize()
-  uops = prg.src[2].src
-  outbufs = [Buffer(Device.DEFAULT, sz:=(1 if local_size is None else prod(local_size)), (dtype:=u.src[1].dtype), \
-      initial_value=np.zeros(sz, dtype=_to_np_dtype(dtype)).data) for u in uops if u.op is Ops.STORE]
-  inbufs = [x.uop.base.buffer for x in inputs]
-  info = prg.arg
-  if local_size is not None: info = replace(info, local_size=tuple(local_size))
-  ei = CompiledRunner(prg.replace(arg=info), Device.DEFAULT)
-  ei.exec(outbufs+inbufs)
-  return [np.frombuffer(x.as_memoryview(), _to_np_dtype(x.dtype)) for x in outbufs]
+  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)
+  run_linear(UOp(Ops.LINEAR, src=(sink.call(*outs, *(x.uop.base for x in inputs)),)))
+  return [u.buffer.numpy() for u in outs]
 
 def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
   dtype = alu_src_uops[0].dtype
@@ -33,9 +28,7 @@ def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
   ld = b.index(idx)
   alu = ld.alu(alu_op, *alu_src_uops)
   store = UOp.store(a.index(idx), alu)
-  sink = UOp(Ops.SINK, dtypes.void, (store,), arg=KernelInfo())
-  prg = to_program(sink, Device[Device.DEFAULT].renderer)
-  return _test_uop_result([Tensor([input_val])], prg)[0]
+  return _test_uop_result([Tensor([input_val])], UOp(Ops.SINK, dtypes.void, (store,), arg=KernelInfo()))[0]
 
 class TestRendererFailures(unittest.TestCase):
   @unittest.skipIf(not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, PythonRenderer)), "test is for ptx or python renderer")
@@ -44,8 +37,7 @@ class TestRendererFailures(unittest.TestCase):
     gate_alu = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
     gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0.valid(gate_alu)), UOp.const(dtypes.int, 1)))
     sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,), arg=KernelInfo())
-    prg = to_program(sink, Device[Device.DEFAULT].renderer)
-    ret = _test_uop_result([], prg, local_size=[4, 1, 1])[0]
+    ret = _test_uop_result([], sink, local_size=[4, 1, 1])[0]
     np.testing.assert_equal(ret, [0, 1, 1, 1])
 
   @unittest.skipIf(not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, PythonRenderer)), "test is for ptx or python renderer")
@@ -55,8 +47,7 @@ class TestRendererFailures(unittest.TestCase):
     gate_alu_1 = (lidx1:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 2),), 'lidx1')).ne(0)
     gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(dtypes.int, 1)))
     sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,), arg=KernelInfo())
-    prg = to_program(sink, Device[Device.DEFAULT].renderer)
-    ret = _test_uop_result([], prg, local_size=[4, 2, 1])[0]
+    ret = _test_uop_result([], sink, local_size=[4, 2, 1])[0]
     np.testing.assert_equal(ret, [0, 0, 0, 0, 0, 1, 1, 1])
 
 @unittest.skipIf(not isinstance(Device[Device.DEFAULT].renderer, CStyleLanguage), "uops are for cstyle")
@@ -102,8 +93,7 @@ class TestPTXFailures(unittest.TestCase):
     if_uop = UOp(Ops.IF, dtypes.void, (gate_alu,))
     gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0, if_uop), val))
     sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,), arg=KernelInfo())
-    prg = to_program(sink, Device[Device.DEFAULT].renderer)
-    ret = _test_uop_result([], prg, local_size=[4, 1, 1])[0]
+    ret = _test_uop_result([], sink, local_size=[4, 1, 1])[0]
     np.testing.assert_equal(ret, [0, 1, 1, 1])
 
   @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")

test/backend/test_uops.py

@@ -5,18 +5,19 @@ from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.helpers import CI, Context
 from tinygrad.dtype import dtypes, DType, AddrSpace, ConstFloat  # noqa: F401
 from tinygrad.device import Buffer, Device
-from tinygrad.uop.ops import Ops, UOp, KernelInfo, AxisType
+from tinygrad.uop.ops import Ops, UOp, KernelInfo, AxisType, buffers
 from tinygrad.renderer.cstyle import CStyleLanguage
-from tinygrad.engine.realize import CompiledRunner, run_linear
+from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
 from tinygrad.device import is_dtype_supported
 from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.renderer.ptx import PTXRenderer
 from test.helpers import to_uops_list
 
-def _uops_to_prg(uops_list):
-  prg = to_program(UOp.sink(*uops_list, arg=KernelInfo()), Device[Device.DEFAULT].renderer)
-  return CompiledRunner(prg, Device.DEFAULT)
+def run_uops(uops_list:list[UOp], bufs:list[Buffer]):
+  buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in bufs]
+  for u,b in zip(buf_uops, bufs): buffers[u] = b
+  run_linear(UOp(Ops.LINEAR, src=(UOp.sink(*uops_list, arg=KernelInfo()).call(*buf_uops),)))
 
 def uop(uops:list[UOp], op:Ops, dtype:Optional[DType], src:tuple[UOp, ...], arg:Any=None) -> UOp:
   if op is Ops.CONST: uops.append(UOp.const(dtype, arg))
@@ -33,8 +34,7 @@ def _test_single_value(vals, op, dts):
   out = uop(uops, Ops.STORE, dtypes.void, (buf_store.index(uop(uops, Ops.CONST, dtypes.int32, (), 0), ptr=True), 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)]
-  prg = _uops_to_prg([out])
-  prg.exec([buf]+buf2)
+  run_uops([out], [buf]+buf2)
   ret = np.empty(1, _to_np_dtype(output_dtype))
   buf.copyout(ret.data)
   return ret[0]
@@ -47,8 +47,7 @@ def _test_single_value_const(vals, op, dts):
   alu = uop(uops, op, output_dtype, loads)
   out = buf_store[UOp.const(dtypes.int32, 0)].store(alu)
   buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
-  prg = _uops_to_prg([out])
-  prg.exec([buf])
+  run_uops([out], [buf])
   ret = np.empty(1, _to_np_dtype(output_dtype))
   buf.copyout(ret.data)
   return ret[0]
@@ -59,8 +58,7 @@ def _test_uops_result(output_dtype, uops, res):
   # res = output_fn(uops)
   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()
-  prg = _uops_to_prg([out])
-  prg.exec([buf])
+  run_uops([out], [buf])
   ret = np.empty(1, _to_np_dtype(output_dtype))
   buf.copyout(ret.data)
   return ret[0]

test/device/test_hcq.py

@@ -6,7 +6,7 @@ from tinygrad.device import Buffer, BufferSpec
 from tinygrad.runtime.support.hcq import HCQCompiled, HCQBuffer
 from tinygrad.runtime.autogen import libc
 from tinygrad.runtime.support.system import PCIIfaceBase
-from tinygrad.engine.realize import get_runner, CompiledRunner
+from tinygrad.engine.realize import get_runtime
 from tinygrad.codegen import to_program
 from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad import Variable
@@ -22,11 +22,12 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.b = self.a + 1
     si = self.b.schedule_linear().src[-1]
 
-    TestHCQ.runner = get_runner(TestHCQ.d0.device, si.src[0])
+    TestHCQ.prg = to_program(si.src[0], TestHCQ.d0.renderer)
+    TestHCQ.runtime = get_runtime(TestHCQ.d0.device, TestHCQ.prg)
     TestHCQ.b.uop.buffer.allocate()
 
-    TestHCQ.kernargs_ba_ptr = TestHCQ.runner._prg.fill_kernargs([TestHCQ.b.uop.buffer._buf, TestHCQ.a.uop.buffer._buf])
-    TestHCQ.kernargs_ab_ptr = TestHCQ.runner._prg.fill_kernargs([TestHCQ.a.uop.buffer._buf, TestHCQ.b.uop.buffer._buf])
+    TestHCQ.kernargs_ba_ptr = TestHCQ.runtime.fill_kernargs([TestHCQ.b.uop.buffer._buf, TestHCQ.a.uop.buffer._buf])
+    TestHCQ.kernargs_ab_ptr = TestHCQ.runtime.fill_kernargs([TestHCQ.a.uop.buffer._buf, TestHCQ.b.uop.buffer._buf])
 
   def setUp(self):
     TestHCQ.d0.synchronize()
@@ -114,7 +115,7 @@ class TestHCQ(unittest.TestCase):
 
   # Test exec
   def test_exec_one_kernel(self):
-    TestHCQ.d0.hw_compute_queue_t().exec(TestHCQ.runner._prg, TestHCQ.kernargs_ba_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size) \
+    TestHCQ.d0.hw_compute_queue_t().exec(TestHCQ.runtime, TestHCQ.kernargs_ba_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size) \
                                    .signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
 
     TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
@@ -128,8 +129,8 @@ class TestHCQ(unittest.TestCase):
 
     q = TestHCQ.d0.hw_compute_queue_t()
     q.wait(TestHCQ.d0.timeline_signal, virt_val - 1) \
-     .exec(TestHCQ.runner._prg, TestHCQ.kernargs_ba_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size) \
-     .exec(TestHCQ.runner._prg, TestHCQ.kernargs_ab_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size) \
+     .exec(TestHCQ.runtime, TestHCQ.kernargs_ba_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size) \
+     .exec(TestHCQ.runtime, TestHCQ.kernargs_ab_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size) \
      .signal(TestHCQ.d0.timeline_signal, virt_val)
 
     for _ in range(100):
@@ -141,11 +142,11 @@ class TestHCQ(unittest.TestCase):
 
   @unittest.skipIf(Device.DEFAULT in {"CPU"}, "No globals/locals on LLVM/CPU")
   def test_exec_update(self):
-    sint_global = (Variable("sint_global", 0, 0xffffffff, dtypes.uint32),) + tuple(TestHCQ.runner.p.global_size[1:])
-    sint_local = (Variable("sint_local", 0, 0xffffffff, dtypes.uint32),) + tuple(TestHCQ.runner.p.local_size[1:])
+    sint_global = (Variable("sint_global", 0, 0xffffffff, dtypes.uint32),) + tuple(TestHCQ.prg.arg.global_size[1:])
+    sint_local = (Variable("sint_local", 0, 0xffffffff, dtypes.uint32),) + tuple(TestHCQ.prg.arg.local_size[1:])
 
     q = TestHCQ.d0.hw_compute_queue_t()
-    q.exec(TestHCQ.runner._prg, TestHCQ.kernargs_ba_ptr, sint_global, sint_local) \
+    q.exec(TestHCQ.runtime, TestHCQ.kernargs_ba_ptr, sint_global, sint_local) \
      .signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
 
     q.submit(TestHCQ.d0, {sint_global[0].expr: 1, sint_local[0].expr: 1})
@@ -166,17 +167,17 @@ class TestHCQ(unittest.TestCase):
     b = a + 1
     si = b.schedule_linear().src[-1]
 
-    runner = CompiledRunner(to_program(replace_opts(si.src[0], [Opt(op=OptOps.LOCAL, axis=0, arg=3) for _ in range(3)]), TestHCQ.d0.renderer),
-                            Device.DEFAULT)
+    prg = to_program(replace_opts(si.src[0], [Opt(op=OptOps.LOCAL, axis=0, arg=3) for _ in range(3)]), TestHCQ.d0.renderer)
+    runtime = get_runtime(Device.DEFAULT, prg)
 
     zb = Buffer(Device.DEFAULT, 3 * 3 * 3, dtypes.int, options=BufferSpec(cpu_access=True, nolru=True)).ensure_allocated()
     zt = Buffer(Device.DEFAULT, 3 * 3 * 3, dtypes.int, options=BufferSpec(cpu_access=True, nolru=True)).ensure_allocated()
     ctypes.memset(zb._buf.va_addr, 0, zb.nbytes)
-    kernargs = runner._prg.fill_kernargs([zt._buf, zb._buf])
+    kernargs = runtime.fill_kernargs([zt._buf, zb._buf])
 
     q = TestHCQ.d0.hw_compute_queue_t()
     q.memory_barrier() \
-     .exec(runner._prg, kernargs, (1,1,1), virt_local) \
+     .exec(runtime, kernargs, (1,1,1), virt_local) \
      .signal(TestHCQ.d0.timeline_signal, virt_val)
 
     for x in range(1, 4):
@@ -330,7 +331,7 @@ class TestHCQ(unittest.TestCase):
   def test_speed_exec_time(self):
     sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
     TestHCQ.d0.hw_compute_queue_t().timestamp(sig_st) \
-                                   .exec(TestHCQ.runner._prg, TestHCQ.kernargs_ba_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size) \
+                                   .exec(TestHCQ.runtime, TestHCQ.kernargs_ba_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size) \
                                    .timestamp(sig_en) \
                                    .signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
 
@@ -470,12 +471,13 @@ class TestHCQ(unittest.TestCase):
   def test_memory_barrier(self):
     a = Tensor([0, 1], device=Device.DEFAULT, dtype=dtypes.int8).realize()
     b = a + 1
-    runner = get_runner(TestHCQ.d0.device, b.schedule_linear().src[-1].src[0])
+    prg = to_program(b.schedule_linear().src[-1].src[0], TestHCQ.d0.renderer)
+    runtime = get_runtime(TestHCQ.d0.device, prg)
 
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.int8, options=BufferSpec(nolru=True)).ensure_allocated()
     buf2 = Buffer(Device.DEFAULT, 2, dtypes.int8, options=BufferSpec(cpu_access=True, nolru=True)).ensure_allocated()
 
-    kernargs_ptr = runner._prg.fill_kernargs([buf1._buf, buf2._buf])
+    kernargs_ptr = runtime.fill_kernargs([buf1._buf, buf2._buf])
 
     for i in range(255):
       ctypes.memset(buf2._buf.va_addr, i, 2)
@@ -483,7 +485,7 @@ class TestHCQ(unittest.TestCase):
       # Need memory_barrier after direct write to vram
       TestHCQ.d0.hw_compute_queue_t().wait(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value - 1) \
                                      .memory_barrier() \
-                                     .exec(runner._prg, kernargs_ptr, runner.p.global_size, runner.p.local_size) \
+                                     .exec(runtime, kernargs_ptr, prg.arg.global_size, prg.arg.local_size) \
                                      .signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
       TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1

test/device/test_metal.py

@@ -50,19 +50,6 @@ kernel void r_5(device int* data0, const device int* data1, uint3 gid [[threadgr
       compiled = compiled[:40] # corrupt the compiled program
       MetalProgram(device, "r_5", compiled)
 
-  def test_wait_skips_in_flight(self):
-    device = MetalDevice("metal")
-    compiled = MetalCompiler().compile("""
-#include <metal_stdlib>
-kernel void noop(uint3 gid [[threadgroup_position_in_grid]], uint3 lid [[thread_position_in_threadgroup]]) {}
-""")
-    prg = MetalProgram(device, "noop", compiled)
-    self.assertIsInstance(prg(wait=True), float)
-    self.assertEqual(device.mtl_buffers_in_flight, [])
-    self.assertIsNone(prg(wait=False))
-    self.assertEqual(len(device.mtl_buffers_in_flight), 1)
-    device.synchronize()
-
   def test_free(self):
     size = 2**16
     device = Device['METAL']

test/external/external_benchmark_op_conv.py

@@ -3,7 +3,7 @@ from dataclasses import replace
 from tinygrad import dtypes, Device
 from tinygrad.uop.ops import UOp, AxisType, Ops, KernelInfo
 from tinygrad.codegen.opt import Opt, OptOps # pylint: disable=unused-import
-from tinygrad.engine.realize import CompiledRunner
+from tinygrad.engine.realize import get_runtime
 from tinygrad.codegen import to_program
 from tinygrad.helpers import dedup, getenv
 from tinygrad.device import Buffer
@@ -90,12 +90,13 @@ renderer = Device.default.renderer
 allocator = Device.default.allocator
 
 ps = to_program(ast, renderer)
-cr = CompiledRunner(ps, Device.DEFAULT)
+rt = get_runtime(Device.DEFAULT, ps)
 
 gs = sorted(dedup([u for u in ast.toposort() if u.op is Ops.PARAM]), key=lambda u: u.arg)
 # print(len(gs))
 # print([g.dtype for g in gs])
 bufs = [Buffer(ps.arg.device, g.size, g.dtype if isinstance(g.dtype, ImageDType) else g.dtype._base).ensure_allocated() for g in gs]
 
-t = cr(bufs, wait=True)
+gsize, lsize = ps.arg.launch_dims({})
+t = rt(*[b._buf for b in bufs], global_size=gsize, local_size=lsize, vals=ps.arg.vals({}), wait=True)
 print(f"{t*1e6:.2f} us")

test/external/external_fuzz_beam_timeout_recovery.py

@@ -3,7 +3,7 @@
 Stress test for beam timeout + device recovery on AM devices.
 
 Usage:
-  DEV=AMD python test/external/external_test_beam_timeout_recovery.py
+  DEV=AMD python test/external/external_fuzz_beam_timeout_recovery.py
 """
 from tinygrad import Tensor, Device
 from tinygrad.helpers import Context

test/external/external_test_hcq.py

@@ -2,7 +2,8 @@ import unittest, ctypes, struct, time, array
 from tinygrad import Device, Tensor, dtypes
 from tinygrad.helpers import to_mv, CI
 from tinygrad.device import Buffer, BufferSpec
-from tinygrad.engine.realize import get_runner
+from tinygrad.engine.realize import get_runtime
+from tinygrad.codegen import to_program
 
 def _time_queue(q, d):
   st = time.perf_counter()
@@ -21,13 +22,14 @@ class TestHCQ(unittest.TestCase):
     TestHCQ.a = Tensor([0.,1.], device=Device.DEFAULT).realize()
     TestHCQ.b = self.a + 1
     linear = self.b.schedule_linear()
-    TestHCQ.runner = get_runner(TestHCQ.d0.device, linear.src[-1].src[0])
+    TestHCQ.prg = to_program(linear.src[-1].src[0], TestHCQ.d0.renderer)
+    TestHCQ.runtime = get_runtime(TestHCQ.d0.device, TestHCQ.prg)
     TestHCQ.b.uop.buffer.allocate()
     # wow that's a lot of abstraction layers
     TestHCQ.addr = struct.pack("QQ", TestHCQ.b.uop.buffer._buf, TestHCQ.a.uop.buffer._buf)
     TestHCQ.addr2 = struct.pack("QQ", TestHCQ.a.uop.buffer._buf, TestHCQ.b.uop.buffer._buf)
-    TestHCQ.kernargs_off = TestHCQ.runner._prg.kernargs_offset
-    TestHCQ.kernargs_size = TestHCQ.runner._prg.kernargs_alloc_size
+    TestHCQ.kernargs_off = TestHCQ.runtime.kernargs_offset
+    TestHCQ.kernargs_size = TestHCQ.runtime.kernargs_alloc_size
     ctypes.memmove(TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_off, TestHCQ.addr, len(TestHCQ.addr))
     ctypes.memmove(TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size+TestHCQ.kernargs_off, TestHCQ.addr2, len(TestHCQ.addr2))
 
@@ -38,8 +40,8 @@ class TestHCQ(unittest.TestCase):
     elif Device.DEFAULT == "NV":
       from tinygrad.runtime.ops_nv import HWQueue, HWQueue
       # nv need to copy constbuffer there as well
-      to_mv(TestHCQ.d0.kernargs_ptr, 0x160).cast('I')[:] = array.array('I', TestHCQ.runner._prg.constbuffer_0)
-      to_mv(TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, 0x160).cast('I')[:] = array.array('I', TestHCQ.runner._prg.constbuffer_0)
+      to_mv(TestHCQ.d0.kernargs_ptr, 0x160).cast('I')[:] = array.array('I', TestHCQ.runtime.constbuffer_0)
+      to_mv(TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, 0x160).cast('I')[:] = array.array('I', TestHCQ.runtime.constbuffer_0)
       TestHCQ.compute_queue = HWQueue
       TestHCQ.copy_queue = HWQueue
 
@@ -53,11 +55,11 @@ class TestHCQ(unittest.TestCase):
     temp_signal, temp_value = TestHCQ.d0._alloc_signal(value=0), 0
     q = TestHCQ.compute_queue()
     for _ in range(1000):
-      q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+      q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
       q.signal(temp_signal, temp_value + 1).wait(temp_signal, temp_value + 1)
       temp_value += 1
 
-      q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+      q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
       q.signal(temp_signal, temp_value + 1).wait(temp_signal, temp_value + 1)
       temp_value += 1
 
@@ -71,10 +73,10 @@ class TestHCQ(unittest.TestCase):
   def test_run_1000_times(self):
     temp_signal = TestHCQ.d0._alloc_signal(value=0)
     q = TestHCQ.compute_queue()
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(temp_signal, 2).wait(temp_signal, 2)
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.runner.p.global_size,
-           TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.prg.arg.global_size,
+           TestHCQ.prg.arg.local_size)
     for _ in range(1000):
       TestHCQ.d0._set_signal(temp_signal, 1)
       q.submit(TestHCQ.d0)
@@ -87,11 +89,11 @@ class TestHCQ(unittest.TestCase):
   def test_run_to_3(self):
     temp_signal = TestHCQ.d0._alloc_signal(value=0)
     q = TestHCQ.compute_queue()
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(temp_signal, 1).wait(temp_signal, 1)
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(temp_signal, 2).wait(temp_signal, 2)
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
@@ -101,7 +103,7 @@ class TestHCQ(unittest.TestCase):
   def test_update_exec(self):
     q = TestHCQ.compute_queue()
     exec_cmd_idx = len(q)
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.update_exec(exec_cmd_idx, (1,1,1), (1,1,1))
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
@@ -115,10 +117,10 @@ class TestHCQ(unittest.TestCase):
   def test_bind_run(self):
     temp_signal = TestHCQ.d0._alloc_signal(value=0)
     q = TestHCQ.compute_queue()
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(temp_signal, 2).wait(temp_signal, 2)
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.runner.p.global_size,
-           TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr+TestHCQ.kernargs_size, TestHCQ.prg.arg.global_size,
+           TestHCQ.prg.arg.local_size)
     q.bind(TestHCQ.d0)
     for _ in range(1000):
       TestHCQ.d0._set_signal(temp_signal, 1)
@@ -133,7 +135,7 @@ class TestHCQ(unittest.TestCase):
   def test_update_exec_binded(self):
     q = TestHCQ.compute_queue()
     exec_ptr = q.ptr()
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     q.bind(TestHCQ.d0)
 
@@ -170,7 +172,7 @@ class TestHCQ(unittest.TestCase):
 
   def test_run_normal(self):
     q = TestHCQ.compute_queue()
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     TestHCQ.d0.timeline_value += 1
@@ -201,7 +203,7 @@ class TestHCQ(unittest.TestCase):
 
   def test_run_signal(self):
     q = TestHCQ.compute_queue()
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
     q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
     q.submit(TestHCQ.d0)
     TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
@@ -278,7 +280,7 @@ class TestHCQ(unittest.TestCase):
   def test_interleave_compute_and_copy(self):
     q = TestHCQ.compute_queue()
     qc = TestHCQ.copy_queue()
-    q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)  # b = [1, 2]
+    q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)  # b = [1, 2]
     q.signal(sig:=TestHCQ.d0._alloc_signal(value=0), value=1)
     qc.wait(sig, value=1)
     qc.copy(TestHCQ.a.uop.buffer._buf, TestHCQ.b.uop.buffer._buf, 8)
@@ -315,7 +317,7 @@ class TestHCQ(unittest.TestCase):
     for _ in range(40):
       q = TestHCQ.compute_queue()
       q.wait(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value - 1)
-      q.exec(TestHCQ.runner._prg, TestHCQ.d0.kernargs_ptr, TestHCQ.runner.p.global_size, TestHCQ.runner.p.local_size)
+      q.exec(TestHCQ.runtime, TestHCQ.d0.kernargs_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size)
       q.signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
       TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)
       TestHCQ.d0.timeline_value += 1

test/external/external_test_speed_llama.py

@@ -5,7 +5,7 @@ from tinygrad.tensor import Tensor
 from tinygrad import Device
 from tinygrad.nn.state import get_state_dict
 from tinygrad.device import Allocator, Compiled
-from tinygrad.engine.realize import method_cache
+from tinygrad.codegen import to_program_cache
 from tinygrad.helpers import Profiling
 
 class FakeProgram:
@@ -31,8 +31,8 @@ class TestLLaMASpeed(unittest.TestCase):
     for v in get_state_dict(model).values(): v.assign(Tensor.empty(*v.shape, dtype=v.dtype))
     print("assigned empty tensors, doing warmup")
 
-    def run_llama(st, empty_method_cache=True):
-      if empty_method_cache: method_cache.clear()
+    def run_llama(st, empty_cache=True):
+      if empty_cache: to_program_cache.clear()
       tms = [time.perf_counter()]
       for i in range(5):
         model(Tensor([[1,2,3,4]]), i).realize()

test/external/external_uop_gc.py

@@ -1,7 +1,6 @@
 import gc
 from tinygrad import Tensor, UOp, Device, nn
 from tinygrad.schedule import schedule_cache
-from tinygrad.engine.realize import method_cache
 from tinygrad.codegen import to_program, to_program_cache
 from tinygrad.schedule.indexing import apply_movement_op, _apply_reshape
 from tinygrad.uop.divandmod import fold_divmod_general
@@ -71,7 +70,6 @@ if __name__ == "__main__":
 
     # these caches will keep uops alive
     schedule_cache.clear()
-    method_cache.clear()
     to_program_cache.clear()
     apply_movement_op.cache_clear()
     _apply_reshape.cache_clear()

test/external/fuzz_symbolic.py

@@ -7,7 +7,7 @@ import z3
 from tinygrad import Variable, dtypes
 from tinygrad.uop.ops import UOp
 from tinygrad.uop.validate import uops_to_z3
-from tinygrad.helpers import DEBUG, Context
+from tinygrad.helpers import DEBUG
 
 seed = int(sys.argv[1]) if len(sys.argv) > 1 else random.randint(0, 100)
 print(f"Seed: {seed}", flush=True)
@@ -56,8 +56,7 @@ if __name__ == "__main__":
     v = [u1,u2,u3]
     expr = random_int_expr(6)
 
-    with Context(CORRECT_DIVMOD_FOLDING=1):
-      simplified_expr = expr.simplify()
+    simplified_expr = expr.simplify()
 
     solver = z3.Solver(ctx=z3.Context())
     solver.set(timeout=5000)  # some expressions take very long verify, but its very unlikely they actually return sat
@@ -74,10 +73,9 @@ if __name__ == "__main__":
       m = solver.model()
       n1, n2, n3 = m[v1], m[v2], m[v3]
       u1_val, u2_val, u3_val = u1.const_like(n1.as_long()), u2.const_like(n2.as_long()), u3.const_like(n3.as_long())
-      with Context(CORRECT_DIVMOD_FOLDING=1):
-        num = expr.simplify().substitute({u1:u1_val, u2:u2_val, u3:u3_val}).ssimplify()
-        rn = expr.substitute({u1:u1_val, u2:u2_val, u3:u3_val}).ssimplify()
-        if num==rn: print("z3 found a mismatch but the expressions are equal!!")
+      num = expr.simplify().substitute({u1:u1_val, u2:u2_val, u3:u3_val}).ssimplify()
+      rn = expr.substitute({u1:u1_val, u2:u2_val, u3:u3_val}).ssimplify()
+      if num==rn: print("z3 found a mismatch but the expressions are equal!!")
       assert False, f"mismatched {expr.render()} at v1={m[v1]}; v2={m[v2]}; v3={m[v3]} = {num} != {rn}\n" +\
             "Reproduce with:\n" +\
             f"v1=Variable(\"{u1.arg[0]}\", {u1.arg[1]}, {u1.arg[2]})\n" +\

test/external/fuzz_symbolic_symbolic_div.py

@@ -2,7 +2,7 @@ import random, sys
 import z3
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.uop.validate import uops_to_z3
-from tinygrad.helpers import DEBUG, Context, colored
+from tinygrad.helpers import DEBUG, colored
 
 seed = int(sys.argv[1]) if len(sys.argv) > 1 else random.randint(0, 100)
 print(f"Seed: {seed}", flush=True)
@@ -36,8 +36,7 @@ if __name__ == "__main__":
     variable_names += [f"r{i}" for i in range(num_ranges)]
     expr = get_random_expr(ranges, factors)
 
-    with Context(CORRECT_DIVMOD_FOLDING=1):
-      simplified_expr = expr.simplify()
+    simplified_expr = expr.simplify()
 
     if DEBUG>=1:
       print(expr.render(simplify=False), "  -->  ", simplified_expr.render(simplify=False))

test/helpers.py

@@ -17,11 +17,14 @@ from tinygrad.codegen.late.linearizer import linearize
 slow = unittest.skipUnless(os.getenv("RUN_SLOW"), "slow test, set RUN_SLOW=1 to run")
 from tinygrad.runtime.ops_python import PythonProgram, PythonRenderer, PythonCompiler
 
-def get_uops(sink:UOp, ren:Renderer|None=None) -> list[UOp]:
-  """Extract linearized UOps from a sink. Test helper that only does linearization (no render)."""
+def full_rewrite(sink:UOp, ren:Renderer|None=None) -> UOp:
   if ren is None: ren = Renderer(Target())
   if sink.arg is None: sink = sink.replace(arg=KernelInfo())
-  full_sink = full_rewrite_to_sink(sink, ren, optimize=sink.tag is None)
+  return full_rewrite_to_sink(sink, ren, optimize=sink.tag is None)
+
+def get_uops(sink:UOp, ren:Renderer|None=None) -> list[UOp]:
+  """Extract linearized UOps from a sink. Test helper that only does linearization (no render)."""
+  full_sink = full_rewrite(sink, ren)
   return line_rewrite(linearize(full_sink), pm_linearize_cleanups)
 
 def replace_opts(ast:UOp, opts:list) -> UOp: return ast.replace(arg=replace(ast.arg, opts_to_apply=tuple(opts)))

test/mockgpu/am/amgpu.py

@@ -133,8 +133,7 @@ class MockPSP(MockIPBlock):
 
 class MockSMU(MockIPBlock):
   def __init__(self, gpu, mmio):
-    try: regs = import_asic_regs('mp', (11, 0), cls=functools.partial(AMDReg, bases={0: IP_BASES[am.MP1_HWIP]}))
-    except Exception: regs = {}
+    regs = import_asic_regs('mp', (11, 0, 0), cls=functools.partial(AMDReg, bases={0: IP_BASES[am.MP1_HWIP]}))
     super().__init__(gpu, mmio, regs)
     self._msg_pending = False
     def r(n): return self.reg(f"mmMP1_SMN_C2PMSG_{n}")

test/mockgpu/amd/emu.py

@@ -53,10 +53,11 @@ class _MXCSRContext:
 
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
 from tinygrad.dtype import dtypes, AddrSpace
-from tinygrad.device import Buffer, BufferSpec
+from tinygrad.device import Buffer, BufferSpec, Device
 from tinygrad.runtime.autogen import hsa
 from tinygrad.helpers import Context, DEBUG, PROFILE, colored
-from tinygrad.engine.realize import get_runner
+from tinygrad.engine.realize import get_runtime
+from tinygrad.codegen import to_program
 
 from tinygrad.renderer.amd import decode_inst
 from tinygrad.runtime.autogen.amd.rdna3.str_pcode import PCODE as PCODE_RDNA3
@@ -374,7 +375,7 @@ def _mem_store(mem: UOp, addr: UOp, val: UOp, active: UOp, addr_bits: int = 32,
   """Conditional memory store with sub-word support. Returns list of store UOps."""
   adt = dtypes.uint64 if addr_bits == 64 else dtypes.uint32
   word_addr = addr >> UOp.const(adt, 2)
-  idx = mem.index(word_addr.cast(dtypes.int), active)
+  idx = mem.index(word_addr.cast(dtypes.int).valid(active))
   if data_bits == 32: return [idx.store(active.where(_to_u32(val), idx))]
   # Sub-word store: read-modify-write with mask
   byte_pos = addr.cast(dtypes.uint32) & _c(3)
@@ -387,7 +388,7 @@ def _mem_store(mem: UOp, addr: UOp, val: UOp, active: UOp, addr_bits: int = 32,
   is_cross = byte_pos.eq(_c(3))
   cross_word0 = (idx & _c(0x00FFFFFF)) | ((val_u32 & _c(0xFF)) << _c(24))
   store0 = idx.store(active.where(is_cross.where(cross_word0, new_word), idx))
-  next_idx = mem.index((word_addr + UOp.const(adt, 1)).cast(dtypes.int), active & is_cross)
+  next_idx = mem.index((word_addr + UOp.const(adt, 1)).cast(dtypes.int).valid(active & is_cross))
   cross_word1 = (next_idx & _c(0xFFFFFF00)) | ((val_u32 >> _c(8)) & _c(0xFF))
   return [store0, next_idx.store((active & is_cross).where(cross_word1, next_idx))]
 
@@ -397,7 +398,7 @@ def _mem_store_bytes(mem: UOp, addr: UOp, val: UOp, active: UOp, data_bits: int
   val_u32 = val.cast(dtypes.uint32) if val.dtype != dtypes.uint32 else val
   for i in range(data_bits // 8):
     byte_val = (val_u32 >> UOp.const(dtypes.uint32, i * 8)) & UOp.const(dtypes.uint32, 0xFF)
-    stores.append(mem.index((addr + UOp.const(dtypes.uint64, i)).cast(dtypes.int), active).store(byte_val.cast(dtypes.uint8)))
+    stores.append(mem.index((addr + UOp.const(dtypes.uint64, i)).cast(dtypes.int).valid(active)).store(byte_val.cast(dtypes.uint8)))
   return stores
 
 def _collect_data_slices(assigns: list[tuple[str, UOp]], data_prefix: str, pcode_vars: dict | None = None, op_name: str = "") -> dict[int, UOp]:
@@ -515,14 +516,14 @@ class _Ctx:
   # Dynamic register access (takes UOp index instead of int)
   def rsgpr_dyn(self, reg: UOp, valid: UOp | None = None) -> UOp:
     """Read SGPR with dynamic register index."""
-    if valid is not None: return self.sgpr.index(reg.cast(dtypes.int), valid, ptr=True).load()
+    if valid is not None: return self.sgpr.index(reg.cast(dtypes.int).valid(valid), ptr=True).load()
     return self.sgpr.index(reg.cast(dtypes.int), ptr=True).load()
 
   def wsgpr_dyn(self, reg: UOp, val: UOp) -> UOp:
     """Write SGPR with dynamic register index. On RDNA, index 124 = NULL (writes discarded). On CDNA, index 124 = M0 (read/write)."""
     # RDNA: NULL (124) discards writes. CDNA: M0 (124) is writable.
     valid = None if self.wave_size == 64 else reg.ne(_c(124))
-    return self.sgpr.index(reg.cast(dtypes.int), valid).store(val.cast(dtypes.uint32))
+    return self.sgpr.index(reg.cast(dtypes.int).valid(valid) if valid is not None else reg.cast(dtypes.int)).store(val.cast(dtypes.uint32))
 
   def wmask(self, reg: UOp, val: UOp) -> list[UOp]:
     """Write a lane mask (VCC/EXEC). Splits into lo/hi for wave64."""
@@ -539,24 +540,24 @@ class _Ctx:
   def rvgpr_dyn(self, reg: UOp, lane: UOp, valid: UOp | None = None) -> UOp:
     """Read VGPR with dynamic register index."""
     idx = reg.cast(dtypes.int) * _c(self.wave_size, dtypes.int) + lane.cast(dtypes.int)
-    return self.vgpr.index(idx, valid, ptr=True).load() if valid is not None else self.vgpr.index(idx, ptr=True).load()
+    return self.vgpr.index(idx.valid(valid), ptr=True).load() if valid is not None else self.vgpr.index(idx, ptr=True).load()
 
   def wvgpr_dyn(self, reg: UOp, lane: UOp, val: UOp, exec_mask: UOp, after: UOp | None = None) -> UOp:
     """Write VGPR with dynamic register index."""
     buf = self.vgpr.after(after) if after is not None else self.vgpr
     offset = reg.cast(dtypes.int) * _c(self.wave_size, dtypes.int) + lane.cast(dtypes.int)
-    return buf.index(offset, _lane_active(exec_mask, lane)).store(val.cast(dtypes.uint32))
+    return buf.index(offset.valid(_lane_active(exec_mask, lane))).store(val.cast(dtypes.uint32))
 
   def raccvgpr_dyn(self, reg: UOp, lane: UOp, valid: UOp | None = None) -> UOp:
     """Read ACCVGPR with dynamic register index (CDNA only)."""
     idx = reg.cast(dtypes.int) * _c(self.wave_size, dtypes.int) + lane.cast(dtypes.int)
-    return self.accvgpr.index(idx, valid, ptr=True).load() if valid is not None else self.accvgpr.index(idx, ptr=True).load()
+    return self.accvgpr.index(idx.valid(valid), ptr=True).load() if valid is not None else self.accvgpr.index(idx, ptr=True).load()
 
   def waccvgpr_dyn(self, reg: UOp, lane: UOp, val: UOp, exec_mask: UOp, after: UOp | None = None) -> UOp:
     """Write ACCVGPR with dynamic register index (CDNA only)."""
     buf = self.accvgpr.after(after) if after is not None else self.accvgpr
     offset = reg.cast(dtypes.int) * _c(self.wave_size, dtypes.int) + lane.cast(dtypes.int)
-    return buf.index(offset, _lane_active(exec_mask, lane)).store(val.cast(dtypes.uint32))
+    return buf.index(offset.valid(_lane_active(exec_mask, lane))).store(val.cast(dtypes.uint32))
 
   def rsrc_dyn(self, off: UOp, lane: UOp | None, bits: int = 32, literal: UOp | None = None, is_f64: bool = False, do_cast: bool = True) -> UOp:
     """Read source operand with dynamic offset. Handles SGPR/inline constants (<256), VGPR (>=256).
@@ -712,7 +713,7 @@ class _Ctx:
         old = self.vgpr.index(val[0].cast(dtypes.int), ptr=True).load()
         new_val = _set_bits(old, _val_to_bits(val[1]), width, lo_bit).cast(dtypes.uint32)
         active = _lane_active(exec_mask, lane)
-        raw_stores.append(('vgpr_direct', self.vgpr.index(val[0].cast(dtypes.int), active).store(new_val)))
+        raw_stores.append(('vgpr_direct', self.vgpr.index(val[0].cast(dtypes.int).valid(active)).store(new_val)))
         continue
       if 'D0' in dest and '[laneId]' in dest:
         old_vcc = self.rmask(_c(VCC_LO.offset))
@@ -1846,7 +1847,7 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
     if data_bits < 32:
       # Sub-dword LDS write: read-modify-write within the uint32 slot
       word_addr = (addr >> addr_shift).cast(dtypes.int)
-      idx = mem.index(word_addr, active)
+      idx = mem.index(word_addr.valid(active))
       byte_pos = addr.cast(dtypes.uint32) & _c(3)
       byte_shift = byte_pos * _c(8)
       size_mask = _c(0xFF if data_bits == 8 else 0xFFFF)
@@ -2004,17 +2005,18 @@ def _compile_mubuf(inst: irc.MUBUF, ctx: _Ctx) -> UOp:
       word_addr = (addr + UOp.const(dtypes.uint64, i * 4)) >> UOp.const(dtypes.uint64, 2)
       val = in_bounds.where(mem.index(word_addr.cast(dtypes.int64), ptr=True).load(), _c(0))
       lds_idx = ((lds_addr + _c(i * 4)) >> _c(2)).cast(dtypes.int)
-      stores.append(ctx.lds.index(lds_idx, active).store(active.where(val, ctx.lds.index(lds_idx, active))))
+      lds_slot = ctx.lds.index(lds_idx.valid(active))
+      stores.append(lds_slot.store(active.where(val, lds_slot)))
   elif is_store:
     for i in range(n_dwords):
       word_addr = (addr + UOp.const(dtypes.uint64, i * 4)) >> UOp.const(dtypes.uint64, 2)
-      idx = mem.index(word_addr.cast(dtypes.int64), in_bounds)
+      idx = mem.index(word_addr.cast(dtypes.int64).valid(in_bounds))
       val = (ctx.raccvgpr_dyn if use_acc else ctx.rvgpr_dyn)(vdata + _c(i), lane)
       stores.append(idx.store(in_bounds.where(_to_u32(val), idx)))
   else:
     for i in range(n_dwords):
       word_addr = (addr + UOp.const(dtypes.uint64, i * 4)) >> UOp.const(dtypes.uint64, 2)
-      val = in_bounds.where(mem.index(word_addr.cast(dtypes.int64), in_bounds, ptr=True).load(), _c(0))
+      val = in_bounds.where(mem.index(word_addr.cast(dtypes.int64).valid(in_bounds), ptr=True).load(), _c(0))
       stores.append((ctx.waccvgpr_dyn if use_acc else ctx.wvgpr_dyn)(vdata + _c(i), lane, val, exec_mask))
   return UOp.sink(UOp.group(*stores).end(lane), *ctx.inc_pc())
 
@@ -2045,18 +2047,18 @@ _INST_HANDLERS: dict[type, Callable[..., UOp]] = {
 # PROGRAM DECODE AND COMPILATION
 # ═══════════════════════════════════════════════════════════════════════════════
 
-_canonical_runner_cache: list[tuple[type, int, int, int, object]] = []  # [(inst_type, base, mask, size, runner), ...]
+_canonical_runner_cache: list[tuple[type, int, int, int, tuple[UOp, object]]] = []  # [(inst_type, base, mask, size, (prg, runtime)), ...]
 
 @functools.cache
 def _get_runner(inst_bytes: bytes, arch: str = "rdna3"):
-  """Build and compile instruction to CompiledRunner. Cached by instruction bytes, with canonical dedup."""
+  """Build and compile instruction to (prg, runtime). Cached by instruction bytes, with canonical dedup."""
   inst = decode_inst(inst_bytes, arch)
   inst_size = inst.size()
   inst_int = int.from_bytes(inst_bytes[:inst_size], 'little')
 
   # Check if instruction matches any cached canonical pattern (must also match instruction type to avoid variant conflicts)
-  for inst_type, base, mask, size, runner in _canonical_runner_cache:
-    if type(inst) is inst_type and inst_size == size and (inst_int & mask) == base: return runner
+  for inst_type, base, mask, size, entry in _canonical_runner_cache:
+    if type(inst) is inst_type and inst_size == size and (inst_int & mask) == base: return entry
 
   # Look up handler by type, falling back to base classes for _LIT variants
   handler = _INST_HANDLERS.get(type(inst))
@@ -2075,9 +2077,10 @@ def _get_runner(inst_bytes: bytes, arch: str = "rdna3"):
 
   # NOTE: renderer output is not reproducible because of _MXCSRContext. PROFILE=0 prevents emulator instruction runners from polluting profiling.
   with Context(NOOPT=1, CHECK_OOB=0, TUPLE_ORDER=0, EMULATED_DTYPES="", CAPTURE_PROCESS_REPLAY=0, PROFILE=0):
-    runner = get_runner('CPU', sink)
-  _canonical_runner_cache.append((type(inst), base, mask, size, runner))
-  return runner
+    prg = to_program(sink, Device['CPU'].renderer)
+    runtime = get_runtime('CPU', prg)
+  _canonical_runner_cache.append((type(inst), base, mask, size, (prg, runtime)))
+  return prg, runtime
 
 _BARRIER_OPS = {ir3.SOPPOp.S_BARRIER, irc.SOPPOp.S_BARRIER}
 if hasattr(ir4.SOPPOp, 'S_BARRIER_WAIT'): _BARRIER_OPS.add(ir4.SOPPOp.S_BARRIER_WAIT)
@@ -2208,10 +2211,10 @@ def run_asm(lib: int, lib_sz: int, gx: int, gy: int, gz: int, lx: int, ly: int,
   def _ensure_compiled(pc: int) -> tuple[Callable, list[int], bool, Inst]:
     if pc not in program:
       prev_len = len(_canonical_runner_cache)
-      runner, inst = _decode_at(pc, arch)
+      (prg, runtime), inst = _decode_at(pc, arch)
       is_barrier = (isinstance(inst, (ir3.SOPP, ir4.SOPP, irc.SOPP)) and inst.op in _BARRIER_OPS) or \
                    (isinstance(inst, (ir4.SOP1,)) and inst.op in _BARRIER_SOP1_OPS)
-      program[pc] = (runner._prg.fxn, runner.p.globals, is_barrier, inst)
+      program[pc] = (runtime.fxn, prg.arg.globals, is_barrier, inst)
       if DEBUG >= 3:
         msg = f"[emu] PC={pc - lib}: {inst!r}"
         print(colored(msg, 'green') if len(_canonical_runner_cache) > prev_len else msg)

test/mockgpu/amd/pcode.py

@@ -828,28 +828,28 @@ class Parser:
     assert mem is not None, "memory load requires _vmem or _lds"
     adt = dtypes.uint64 if addr.dtype == dtypes.uint64 else dtypes.uint32
     active = self.vars.get('_active')
-    gate = (active,) if active is not None else ()
+    def mindex(idx:UOp, ptr=False): return mem.index(idx.valid(active) if active is not None else idx, ptr=ptr)
     byte_mem = mem.dtype.base == dtypes.uint8
     if byte_mem:
       idx = addr.cast(dtypes.int)
       if dt in (dtypes.uint64, dtypes.int64, dtypes.float64):
         val = _u32(0).cast(dtypes.uint64)
-        for i in range(8): val = val | (mem.index(idx + _const(dtypes.int, i), *gate, ptr=True).load().cast(dtypes.uint64) << _u64(i * 8))
+        for i in range(8): val = val | (mindex(idx + _const(dtypes.int, i), ptr=True).load().cast(dtypes.uint64) << _u64(i * 8))
       elif dt in (dtypes.uint8, dtypes.int8):
-        val = mem.index(idx, *gate, ptr=True).load().cast(dt)
+        val = mindex(idx, ptr=True).load().cast(dt)
       elif dt in (dtypes.uint16, dtypes.int16, dtypes.short):
-        lo = mem.index(idx, *gate, ptr=True).load().cast(dtypes.uint32)
-        hi = mem.index(idx + _const(dtypes.int, 1), *gate, ptr=True).load().cast(dtypes.uint32)
+        lo = mindex(idx, ptr=True).load().cast(dtypes.uint32)
+        hi = mindex(idx + _const(dtypes.int, 1), ptr=True).load().cast(dtypes.uint32)
         val = (lo | (hi << _u32(8))).cast(dt)
       else:
         val = _u32(0)
-        for i in range(4): val = val | (mem.index(idx + _const(dtypes.int, i), *gate, ptr=True).load().cast(dtypes.uint32) << _u32(i * 8))
+        for i in range(4): val = val | (mindex(idx + _const(dtypes.int, i), ptr=True).load().cast(dtypes.uint32) << _u32(i * 8))
     else:
       idx = (addr >> _const(addr.dtype, 2)).cast(dtypes.int)
-      val = mem.index(idx, *gate)
+      val = mindex(idx)
       if dt in (dtypes.uint64, dtypes.int64, dtypes.float64):
         idx2 = ((addr + _const(adt, 4)) >> _const(adt, 2)).cast(dtypes.int)
-        val = val.cast(dtypes.uint64) | (mem.index(idx2, *gate).cast(dtypes.uint64) << _u64(32))
+        val = val.cast(dtypes.uint64) | (mindex(idx2).cast(dtypes.uint64) << _u64(32))
       elif dt in (dtypes.uint8, dtypes.int8): val = (val >> ((addr & _const(adt, 3)).cast(dtypes.uint32) * _u32(8))) & _u32(0xFF)
       elif dt in (dtypes.uint16, dtypes.int16):
         val = (val >> (((addr >> _const(adt, 1)) & _const(adt, 1)).cast(dtypes.uint32) * _u32(16))) & _u32(0xFFFF)
@@ -862,7 +862,7 @@ class Parser:
         idx_native = (addr >> _const(adt, 2)).cast(dtypes.int64)
         idx_hi_native = ((addr + _const(adt, 4)) >> _const(adt, 2)).cast(dtypes.int64)
         safe_idx_hi = is_unaligned.where(idx_hi_native, idx_native)
-        hi = mem.index(safe_idx_hi, *gate)
+        hi = mindex(safe_idx_hi)
         combined = val.cast(dtypes.uint64) | (hi.cast(dtypes.uint64) << UOp.const(dtypes.uint64, 32))
         val = is_unaligned.where((combined >> (byte_off.cast(dtypes.uint64) * UOp.const(dtypes.uint64, 8))).cast(dtypes.uint32), val)
     return _cast_to(val, dt)

test/null/test_const_folding.py

@@ -2,7 +2,7 @@ import unittest, itertools, math
 from tinygrad import Tensor, dtypes, Context
 from tinygrad.dtype import DType, ConstType
 from tinygrad.uop.ops import Ops, UOp
-from tinygrad.codegen import full_rewrite_to_sink
+from test.helpers import full_rewrite
 import numpy as np
 
 def _check_ast_count(desired_count:int, t:Tensor):
@@ -103,7 +103,7 @@ class TestBitcastConstFolding(unittest.TestCase):
     def t(cases: dict[DType, ConstType]):
       for (from_dt, from_v), (to_dt, to_v) in itertools.product(cases.items(), cases.items()):
         if not math.isnan(from_v):
-          r = full_rewrite_to_sink(UOp.const(from_dt, from_v).bitcast(to_dt).sink()).src[0]
+          r = full_rewrite(UOp.const(from_dt, from_v).bitcast(to_dt).sink()).src[0]
           self.assertEqual(r.op, Ops.CONST, msg:=f"{from_dt} -> {to_dt} ({from_v} -> {to_v})")
           self.assertEqual(r.dtype, to_dt, msg)
           np.testing.assert_equal(r.arg, to_v, msg)
@@ -127,7 +127,7 @@ class TestBitcastConstFolding(unittest.TestCase):
 
   def test_vec_bitcast(self):
     with Context(SPEC=0):
-      r = full_rewrite_to_sink(UOp.const(dtypes.int32.vec(3), (-1, -2**31, 75)).bitcast(dtypes.uint32.vec(3)).sink()).src[0]
+      r = full_rewrite(UOp.const(dtypes.int32.vec(3), (-1, -2**31, 75)).bitcast(dtypes.uint32.vec(3)).sink()).src[0]
     self.assertEqual(r.op, Ops.STACK)
     self.assertEqual(r.dtype, dtypes.uint32.vec(3))
     self.assertEqual(tuple(x.arg for x in r.src), (2**32-1, 2**31, 75))

test/null/test_graph_rewrite.py

@@ -2,13 +2,13 @@ import unittest, math
 from tinygrad import dtypes
 from tinygrad.helpers import all_same, Context
 from tinygrad.uop.ops import GroupOp, UOp, Ops, exec_alu, PatternMatcher, TrackedPatternMatcher, UPat
-from tinygrad.codegen import full_rewrite_to_sink
+from test.helpers import full_rewrite
 from hypothesis import given, strategies as strat
 
 # Helper function to apply the graph rewrite
 @Context(SPEC=0)
 def apply_rewrite(expr):
-  return full_rewrite_to_sink(expr.sink()).src[0]
+  return full_rewrite(expr.sink()).src[0]
 
 def evaluate_uop(uop, variables):
   if uop.op == Ops.CONST:
@@ -151,7 +151,7 @@ class TestModuloAndDivisionFolding(unittest.TestCase):
 
 class TestEdgeCasesAndSpecialOperations(unittest.TestCase):
   def test_full_graph_rewrite_transcendental_edge_cases(self):
-    optimized_sink = full_rewrite_to_sink(UOp.const(dtypes.float32, -1.0).log2().sink(UOp.const(dtypes.float32, 0.0).reciprocal()))
+    optimized_sink = full_rewrite(UOp.const(dtypes.float32, -1.0).log2().sink(UOp.const(dtypes.float32, 0.0).reciprocal()))
     optimized_log2_neg, optimized_recip_zero = optimized_sink.src
     self.assertTrue(math.isnan(optimized_log2_neg.arg), f"Expected NaN for log2(-1.0), got {optimized_log2_neg.arg}")
     self.assertTrue(math.isinf(optimized_recip_zero.arg) and optimized_recip_zero.arg > 0,
@@ -160,14 +160,14 @@ class TestEdgeCasesAndSpecialOperations(unittest.TestCase):
   @unittest.skip("broken")
   def test_full_graph_rewrite_modulo_negative_dividend(self):
     x_var_uop = UOp.variable('x', -5, -1)
-    optimized_sink = full_rewrite_to_sink((x_var_uop % 3).sink())
+    optimized_sink = full_rewrite((x_var_uop % 3).sink())
     for x_value in range(-5, 0):
       self.assertEqual(x_value % 3, evaluate_uop(optimized_sink.src[0], {'x': x_value}))
 
   @unittest.skip("broken")
   def test_full_graph_rewrite_division_negative_divisor(self):
     x_var_uop = UOp.variable('x', 1, 5)
-    optimized_sink = full_rewrite_to_sink((x_var_uop // -2).sink())
+    optimized_sink = full_rewrite((x_var_uop // -2).sink())
     for x_value in range(1, 6):
       self.assertEqual(x_value // -2, evaluate_uop(optimized_sink.src[0], {'x': x_value}))
 

test/null/test_helpers.py

@@ -209,6 +209,11 @@ class TestFetch(unittest.TestCase):
           headers={"Range": "bytes=0-100"}).read_bytes()
     assert len(x) == 101, f"{len(x) != 101}"
 
+  def test_fetch_sha(self):
+    self.assertRaises(Exception, fetch, "https://ftp.gnu.org/gnu/gzip/gzip-1.13.tar.gz", allow_caching=False, sha256="a")
+    fetch("https://ftp.gnu.org/gnu/gzip/gzip-1.13.tar.gz", allow_caching=False,
+          sha256="20fc818aeebae87cdbf209d35141ad9d3cf312b35a5e6be61bfcfbf9eddd212a")
+
 class TestFullyFlatten(unittest.TestCase):
   def test_fully_flatten(self):
     self.assertEqual(fully_flatten([[1, 3], [1, 2]]), [1, 3, 1, 2])

test/null/test_schedule.py

@@ -380,7 +380,7 @@ class TestSchedule(unittest.TestCase):
     r1 = (x - r0).sum(axis=0).div(2)
     out = r0 + r1
     linear, _ = check_schedule(out, 2)
-    reduceops = [x for si in linear.src for x in si.src[0].toposort() if x.op in {Ops.REDUCE_AXIS, Ops.REDUCE}]
+    reduceops = [x for si in linear.src for x in si.src[0].toposort() if x.op is Ops.REDUCE]
     assert len(reduceops) == 2
 
   def test_cache_reduce_multiple_children(self):
@@ -391,7 +391,7 @@ class TestSchedule(unittest.TestCase):
     out0 = r0 + y
     out1 = r1 + y
     linear, _ = check_schedule([out0, out1], 3)
-    reduceops = [x for si in linear.src for x in si.src[0].toposort() if x.op in {Ops.REDUCE_AXIS, Ops.REDUCE}]
+    reduceops = [x for si in linear.src for x in si.src[0].toposort() if x.op is Ops.REDUCE]
     self.assertEqual(len(reduceops), 2) # why is RANGEIFY different?
 
   def test_dedup_assign(self):

test/null/test_simplify_valid_idx.py

@@ -1,12 +1,18 @@
 import unittest, itertools
 
-from tinygrad.codegen import full_rewrite_to_sink
+from tinygrad.codegen.late.devectorizer import load_store_indexing
 from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import UOp, Ops
-from tinygrad.uop.symbolic import simplify_valid
+from tinygrad.uop.ops import UOp, Ops, graph_rewrite
+from tinygrad.uop.symbolic import simplify_valid, sym, pm_move_where_on_load
 from tinygrad.helpers import Context
+from test.helpers import full_rewrite
 from test.null.test_uop_symbolic import check_uop_against_string
 
+# symbolic-only idx + valid simplification (no late lowering of FLOORDIV/FLOORMOD)
+def simplify_valid_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move_where_on_load, name="simplify_valid_idx")
+# image-aware idx + valid simplification: adds the codegen-layer matcher that drops provably in-bounds gates
+def simplify_image_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move_where_on_load+load_store_indexing, name="simplify_image_idx")
+
 def get_gated_load_uop(valid:UOp, idx:UOp):
   return UOp(Ops.LOAD, dtypes.float, (
     UOp(Ops.PARAM, dtypes.float.ptr(), arg=0).index(idx.valid(valid), ptr=True),
@@ -47,11 +53,10 @@ class TestHelpers(unittest.TestCase):
 
 class TestValidIdxSimplification(unittest.TestCase):
   def check(self, load, sidx, svalid, extra=()):
-    with Context(NOOPT=1, SPEC=0):
-      load = full_rewrite_to_sink(UOp.sink(load, *extra)).src[0]
-    idx, valid = load.src[0].src[1], load.src[0].src[2]
-    check_uop_against_string(self, idx, sidx)
-    check_uop_against_string(self, valid, svalid)
+    load = simplify_valid_idx(UOp.sink(load, *extra)).src[0]
+    off = load.src[0].src[1]
+    check_uop_against_string(self, off.get_idx(), sidx)
+    check_uop_against_string(self, off.get_valid(), svalid)
 
   def test_cumsum(self):
     gidx0 = Special("gidx0", 5)
@@ -157,7 +162,7 @@ class TestValidIdxSimplification(unittest.TestCase):
     valid = (ridx2<1)&(ridx1<6)
     load = get_gated_load_uop(valid, idx)
     # prevent ridx1 and ridx2 from being shrunk
-    red = UOp(Ops.REDUCE, dtypes.float, (load, ridx1, ridx2), Ops.ADD)
+    red = load.reduce(ridx1, ridx2, arg=Ops.ADD)
     self.check(load,
       "(r0*1568)",
       "((r2<1)&(r1<6))",
@@ -216,18 +221,18 @@ class TestValidIdxSimplification(unittest.TestCase):
 
 class TestImageSimplification(unittest.TestCase):
   def check(self, load, svalid, sidx0, sidx1):
-    with Context(NOOPT=1, SPEC=0):
-      load = full_rewrite_to_sink(load.sink()).src[0]
-    idx = load.src[0].src[1]
+    load = simplify_image_idx(load.sink()).src[0]
+    off = load.src[0].src[1]
+    idx = off.get_idx()
     self.assertEqual(idx.op, Ops.STACK)
     self.assertEqual(len(idx.src), 2)
     idx0, idx1 = idx.src[0], idx.src[1]
     check_uop_against_string(self, idx0, sidx0)
     check_uop_against_string(self, idx1, sidx1)
     if svalid is not None:
-      check_uop_against_string(self, load.src[0].src[2], svalid)
+      check_uop_against_string(self, off.get_valid(), svalid)
     else:
-      self.assertEqual(len(load.src[0].src), 2, "svalid is None but load still has a valid")
+      self.assertEqual(off.get_valid(), UOp.const(dtypes.bool, True), "svalid is None but valid is not True")
 
   def test_idx_gt_c(self):
     # (idx1 < c+1).ne(True) ? (..., idx1-1+c) : 0 can drop the valid
@@ -287,7 +292,7 @@ class TestImageSimplification(unittest.TestCase):
     # empty -> invalid
     load = get_load_image_uop(shape, (gidx0<8) & (gidx0<8).ne(True), idx)
     with Context(NOOPT=1, SPEC=0):
-      load = full_rewrite_to_sink(load.sink()).src[0]
+      load = full_rewrite(load.sink()).src[0]
     self.assertEqual(load.op, Ops.STACK)
     self.assertEqual(load.dtype.count, 4)
 
@@ -447,12 +452,12 @@ class TestImageSimplification(unittest.TestCase):
     load = get_load_image_uop((32, 1024, 4), valid, (alu0, alu1))
     self.check(load, None, "(lidx1*128+gidx0//2+144)", "(lidx0*2+r0+-3)")
 
-    # TODO: this is the same idx as above, but simplifying idx too early makes it hard to drop the valid
+    # same idx, written without the inline simplification of the inner div/mod
     alu0 = ((gidx0*2+lidx1*512+(lidx0*8192+r0*4096)+-11711)//4%1024)
     alu1 = (lidx0*2+r0+-3)
     valid = ((lidx1<7)&((((lidx0*2+r0)<3)!=1)&((lidx0*2+r0)<35)))
     load = get_load_image_uop((32, 1024, 4), valid, (alu0, alu1))
-    self.check(load, "(lidx1<7)", "((gidx0*2+lidx1*512+(lidx0*8192+r0*4096)+-11711)//4%1024)", "(lidx0*2+r0+-3)")
+    self.check(load, None, "(lidx1*128+gidx0//2+144)", "(lidx0*2+r0+-3)")
 
   def test_simplify8(self):
     # from openpilot compile3, kernel r_4_16_8_16_4_4_3_3n1
@@ -508,7 +513,7 @@ class TestUnfoldableImage(unittest.TestCase):
     with Context(SPEC=0):
       lidx = Special("lidx", 2)
       load = UOp(Ops.LOAD, dtypes.float, (UOp(Ops.PARAM, dtypes.imagef((10, 10, 4)), arg=0).index(lidx, ptr=True), UOp.const(dtypes.float, 0)))
-      res = full_rewrite_to_sink(load.sink()).src[0]
+      res = full_rewrite(load.sink()).src[0]
       self.assertEqual(res.src[0].src[0].dtype, dtypes.float.ptr(400))
 
 class TestDropTrueGate(unittest.TestCase):
@@ -528,7 +533,7 @@ class TestDropTrueGate(unittest.TestCase):
 class TestRangeShrink(unittest.TestCase):
   def get_ranges(self, sink):
     with Context(NOOPT=1, SPEC=0):
-      result = full_rewrite_to_sink(sink)
+      result = full_rewrite(sink)
     return [u for u in result.toposort() if u.op is Ops.RANGE]
 
   def test_range_shrink_single_guard(self):
@@ -569,7 +574,7 @@ class TestRangeShrink(unittest.TestCase):
     # range used in both a gated load AND directly in the reduce expression -> no shrink
     r = Range(0, 204)
     gated_load = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
-    red = UOp(Ops.REDUCE, dtypes.float, (r.cast(dtypes.float) + gated_load, r), Ops.ADD)
+    red = (r.cast(dtypes.float) + gated_load).reduce(r, arg=Ops.ADD)
     ranges = self.get_ranges(red.sink())
     self.assertEqual(len(ranges), 1)
     self.assertEqual(ranges[0].src[0].arg, 204)

test/null/test_symbolic_failures.py

@@ -1,16 +1,8 @@
 import unittest
 from tinygrad import Variable
-from tinygrad.helpers import Context
 
 
 class TestFuzzFailure(unittest.TestCase):
-  def setUp(self):
-    self.context = Context(CORRECT_DIVMOD_FOLDING=1)
-    self.context.__enter__()
-
-  def tearDown(self):
-    self.context.__exit__(None, None, None)
-
   def test_fuzz_failure1(self):
     v1=Variable('v1', 0, 8)
     v2=Variable('v2', 0, 2)

test/null/test_tensor_uop_mixin.py

@@ -137,6 +137,22 @@ class TestTensorUOpAllclose(unittest.TestCase):
     a, b = _t(4).float(), _t(4).float()
     self.assertIs(_strip_unique(a.allclose(b).uop), _strip_unique(a.uop.allclose(b.uop)))
 
+class TestTensorUOpBitcast(unittest.TestCase):
+  def test_bitcast_same_dtype(self): _check(self, _t(4).float(), lambda x: x.bitcast(dtypes.float32))
+
+class TestTensorUOpRand(unittest.TestCase):
+  def test_random_bits(self):
+    k = UOp.empty((2,), dtype=dtypes.uint32)
+    c = UOp.zeros(2, dtype=dtypes.uint32)
+    for num in (1, 4, 7, 1024):
+      self.assertIs(_strip_unique(Tensor.random_bits(Tensor(k), Tensor(c), num).uop),
+                    _strip_unique(UOp.random_bits(k, c, num)))
+  def test_bits_to_rand_float32(self):
+    bits_uop = UOp.empty((8,), dtype=dtypes.uint32)
+    for shape in ((8,), (2, 4), (5,)):
+      self.assertIs(_strip_unique(Tensor._bits_to_rand(Tensor(bits_uop), shape, dtypes.float32).uop),
+                    _strip_unique(UOp._bits_to_rand(bits_uop, shape, dtypes.float32)))
+
 class TestTensorUOpGather(unittest.TestCase):
   def _check(self, t, dim, idx):
     self.assertIs(_strip_unique(t.gather(dim, idx).uop), _strip_unique(t.uop.gather(dim, idx.uop)))

test/null/test_uop_graph.py

@@ -424,7 +424,7 @@ class TestUOpGraph(unittest.TestCase):
     ld = d0.index(ridx0.valid(ridx0<50))
     w = (ridx0<50).where(ld, 5)
     # prevent ridx0 from being shrunk
-    red = UOp(Ops.REDUCE, dtypes.long, (ridx0.cast(dtypes.long), ridx0), Ops.ADD)
+    red = ridx0.cast(dtypes.long).reduce(ridx0, arg=Ops.ADD)
     uops = to_uops_list([w, red])
     for u in uops:
       assert u.op is not Ops.WHERE
@@ -447,7 +447,7 @@ class TestUOpGraph(unittest.TestCase):
     ld = d0.index(gate_idx).cast(dtypes.float)
     w = (ridx0<50).where(ld, 5.0)
     # prevent ridx0 from being shrunk
-    red = UOp(Ops.REDUCE, dtypes.long, (ridx0.cast(dtypes.long), ridx0), Ops.ADD)
+    red = ridx0.cast(dtypes.long).reduce(ridx0, arg=Ops.ADD)
     uops = to_uops_list([w, red])
     for u in uops:
       assert u.op is not Ops.WHERE
@@ -459,7 +459,7 @@ class TestUOpGraph(unittest.TestCase):
     ld = d0.index(ridx0.valid(ridx0<50))
     w = ((ridx0<50) & (ridx0>30)).where(ld, UOp.const(dtypes.float, 0)).cast(dtypes.half)
     # prevent ridx0 from being shrunk
-    red = UOp(Ops.REDUCE, dtypes.long, (ridx0.cast(dtypes.long), ridx0), Ops.ADD)
+    red = ridx0.cast(dtypes.long).reduce(ridx0, arg=Ops.ADD)
     uops = to_uops_list([w, red])
     for u in uops:
       assert u.op is not Ops.WHERE
@@ -470,7 +470,7 @@ class TestUOpGraph(unittest.TestCase):
     ld = d0.index(ridx0.valid(ridx0<50))
     w = ((ridx0<50) & (ridx0>30)).where(UOp.const(dtypes.float, 0), ld).cast(dtypes.half)
     # prevent ridx0 from being shrunk
-    red = UOp(Ops.REDUCE, dtypes.long, (ridx0.cast(dtypes.long), ridx0), Ops.ADD)
+    red = ridx0.cast(dtypes.long).reduce(ridx0, arg=Ops.ADD)
     uops = to_uops_list([w, red])
     for u in uops:
       assert u.op is not Ops.WHERE
@@ -679,7 +679,7 @@ class TestExpander(unittest.TestCase):
   @unittest.skip("no longer supported")
   def test_reduce_known_axis(self):
     e1 = UOp(Ops.UNROLL, dtypes.int, tuple(UOp.const(dtypes.int, x) for x in range(4)), ((1,4),))
-    sink = UOp(Ops.REDUCE, dtypes.int, (3*e1,e1), Ops.ADD)
+    sink = (3*e1).reduce(e1, arg=Ops.ADD)
     sink = expander_rewrite(sink)
     assert sink.op is Ops.CONST
     self.assertEqual(sink.arg, 3*(0+1+2+3))
@@ -687,7 +687,7 @@ class TestExpander(unittest.TestCase):
   @unittest.skip("no longer supported")
   def test_reduce_const(self):
     e1 = UOp(Ops.UNROLL, dtypes.int, tuple(UOp.const(dtypes.int, x) for x in range(4)), ((1,4),))
-    sink = UOp(Ops.REDUCE, dtypes.int, (UOp.const(dtypes.int, 3), e1), Ops.ADD)
+    sink = UOp.const(dtypes.int, 3).reduce(e1, arg=Ops.ADD)
     sink = expander_rewrite(sink)
     assert sink.op is Ops.CONST
     self.assertEqual(sink.arg, 3*4)
@@ -728,7 +728,7 @@ class TestExpander(unittest.TestCase):
   def test_reduce_different_axis(self):
     e1 = UOp(Ops.UNROLL, dtypes.int, tuple(UOp.const(dtypes.int, x) for x in range(4)), ((1,4),))
     e2 = UOp(Ops.UNROLL, dtypes.int, tuple(UOp.const(dtypes.int, x) for x in range(4)), ((2,4),))
-    sink = UOp(Ops.REDUCE, dtypes.int, (e1,e2), Ops.ADD)
+    sink = e1.reduce(e2, arg=Ops.ADD)
     sink = expander_rewrite(sink)
     print(sink)
 

test/null/test_uop_symbolic.py

@@ -3,7 +3,6 @@ import unittest, pickle, functools, math
 import z3
 
 from tinygrad.dtype import dtypes, ConstType, DType, Invalid
-from tinygrad.helpers import Context
 from test.helpers import get_uops
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad.uop.symbolic import sym, commutative, pm_simplify_valid, pm_move_where_on_load
@@ -181,8 +180,8 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(Variable("a", 0, 8)*1, 0, 8, "a")
 
   def test_mul_neg_1(self):
-    self.helper_test_variable((Variable("a", 0, 2)*-1)//3, 0, 0, "0")
-    self.helper_test_variable((Variable("a", 2, 7)*-1)//3, -2, 0, "((a//3)*-1)")
+    self.helper_test_variable((Variable("a", 0, 2)*-1)//3, -1, 0, "((a*-1)//3)")
+    self.helper_test_variable((Variable("a", 2, 7)*-1)//3, -3, -1, "((a*-1)//3)")
 
   def test_mul_2(self):
     self.helper_test_variable(Variable("a", 0, 8)*2, 0, 16, "(a*2)")
@@ -203,8 +202,8 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(Variable("a", 0, 7) // 20, 0, 0, "0")
 
   def test_div_neg_min_max(self):
-    self.helper_test_variable(Variable("a", 1, 7) // -2, -3, 0, "((a//2)*-1)")
-    self.helper_test_variable(Variable("a", 0, 6) // -2, -3, 0, "((a//2)*-1)")
+    self.helper_test_variable(Variable("a", 1, 7) // -2, -4, -1, "(a//-2)")
+    self.helper_test_variable(Variable("a", 0, 6) // -2, -3, 0, "(a//-2)")
 
   def test_div_mod_zero(self):
     with self.assertRaises(ZeroDivisionError):
@@ -238,14 +237,14 @@ class TestSymbolic(unittest.TestCase):
 
   def test_mod_min_max(self):
     self.helper_test_variable(Variable("x", 0, 10)%Variable("y", 1, 10), 0, 9, "(x%y)")
-    self.helper_test_variable(Variable("x", -10, 0)%Variable("y", 1, 10), -9, 0, "(((x*-1)%y)*-1)")
-    self.helper_test_variable(Variable("x", 0, 10)%Variable("y", -10, -1), 0, 9, "(x%(y*-1))")
-    self.helper_test_variable(Variable("x", -10, 0)%Variable("y", -10, -1), -9, 0, "(((x*-1)%(y*-1))*-1)")
-    self.helper_test_variable(Variable("x", -10, 10)%Variable("y", -10, -1), -9, 9, "(x%(y*-1))")
+    self.helper_test_variable(Variable("x", -10, 0)%Variable("y", 1, 10), 0, 9, "(x%y)")
+    self.helper_test_variable(Variable("x", 0, 10)%Variable("y", -10, -1), -9, 0, "(x%y)")
+    self.helper_test_variable(Variable("x", -10, 0)%Variable("y", -10, -1), -9, 0, "(x%y)")
+    self.helper_test_variable(Variable("x", -10, 10)%Variable("y", -10, -1), -9, 0, "(x%y)")
 
-    # test _min_max directly without the rewrite taking out the sign
+    # test _min_max directly: floor mod with positive divisor is in [0, c-1]; with negative divisor in [c+1, 0]
     self.assertEqual((Variable("x", -10, 0)%Variable("y", -10, -1))._min_max, (-9, 0))
-    self.assertEqual((Variable("x", -10, 0)%Variable("y", 1, 10))._min_max, (-9, 0))
+    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.weakint)
@@ -262,12 +261,12 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(Variable("a", 0, 6) // 2, 0, 3, "(a//2)")
 
     self.helper_test_variable(Variable("x", 0, 10)//Variable("y", 1, 10), 0, 10, "(x//y)")
-    self.helper_test_variable(Variable("x", -10, 0)//Variable("y", 1, 10), -10, 0, "(((x*-1)//y)*-1)")
-    self.helper_test_variable(Variable("x", 0, 10)//Variable("y", -10, -1), -10, 0, "((x//(y*-1))*-1)")
-    self.helper_test_variable(Variable("x", -10, 0)//Variable("y", -10, -1), 0, 10, "((x*-1)//(y*-1))")
+    self.helper_test_variable(Variable("x", -10, 0)//Variable("y", 1, 10), -10, 0, "(x//y)")
+    self.helper_test_variable(Variable("x", 0, 10)//Variable("y", -10, -1), -10, 0, "(x//y)")
+    self.helper_test_variable(Variable("x", -10, 0)//Variable("y", -10, -1), 0, 10, "(x//y)")
 
     self.helper_test_variable(Variable("x", -10, 10)//Variable("y", 1, 10), -10, 10, "(x//y)")
-    self.helper_test_variable(Variable("x", -10, 10)//Variable("y", -10, -1), -10, 10, "((x//(y*-1))*-1)")
+    self.helper_test_variable(Variable("x", -10, 10)//Variable("y", -10, -1), -10, 10, "(x//y)")
 
   def test_mod_factor(self):
     self.helper_test_variable(usum([Variable("a", 0, 7)*100, Variable("b", 0, 3)*50]) % 100, 0, 50, "((b%2)*50)")
@@ -334,12 +333,12 @@ 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), -3, 4, "(v1%2*2+(v3+-1)%5+-2)")
+    self.helper_test_variable((((((v1%2)*2)+((v3+-1)%5))+-2)%5), 0, 4, "((v3+v1%2*2+-3)%5)")
 
   def test_mod_mod_wrong_sign2(self):
     v2=Variable("v2", 0, 8)
     v3=Variable("v3", 0, 4)
-    self.helper_test_variable((((((v3+3)%7)+(v2+-2))%7)%7), -2, 6, "(((v2+((v3+3)%7))+-2)%7)")
+    self.helper_test_variable((((((v3+3)%7)+(v2+-2))%7)%7), 0, 6, "((v2+v3+1)%7)")
 
   def test_mul_mul(self):
     self.helper_test_variable((Variable("a", 0, 5)*10)*9, 0, 5*10*9, "(a*90)")
@@ -357,21 +356,28 @@ class TestSymbolic(unittest.TestCase):
   def test_div_const_div(self):
     a = Variable("a", 0, 124)
     self.helper_test_variable((a//2+1)//2, 0, 31, "((a+2)//4)")
-    self.helper_test_variable(((-a)//2-1)//2, -31, 0, "(((a+2)//4)*-1)")
-    self.helper_test_variable(((-a)//2+10)//2, -26, 5, "((((a//2)*-1)+10)//2)")
+    self.helper_test_variable(((-a)//2-1)//2, -32, -1, "((a*-1+2)//4+-1)")
+    self.helper_test_variable(((-a)//2+10)//2, -26, 5, "(a*-1//4+5)")
 
   def test_div_const_div_wrong_sign(self):
     a = Variable("a", 0, 124)
-    self.helper_test_variable(((a-10)//2+10)//2, 2, 33, "((((a+-10)//2)+10)//2)")
+    self.helper_test_variable(((a-10)//2+10)//2, 2, 33, "((a+2)//4+2)")
 
   def test_div_const_div_wrong_sign_divisor(self):
     a = Variable("a", 0, 124)
-    self.helper_test_variable(((a+10)//-2+10)//-4, -1, 14, "(((((a//2)*-1)+5)//4)*-1)")
+    self.helper_test_variable(((a+10)//-2+10)//-4, -2, 14, "(((a+10)//-2+10)//-4)")
+
+  def test_nested_div_negative_divisor(self):
+    # (x//c1)//c2 -> x//(c1*c2) only when c2>0
+    a = Variable("a", 0, 124)
+    self.helper_test_variable((a//-2)//-3, 0, 20, "((a//-2)//-3)")
+    self.helper_test_variable((a//2)//-3, -21, 0, "((a//2)//-3)")
+    self.helper_test_variable((a//-2)//3, -21, 0, "(a//-6)")
 
   def test_neg_mod(self):
     a = Variable("a", 0, 124)
-    self.helper_test_variable((-a)%4, -3, 0, "((a%4)*-1)")
-    self.helper_test_variable(a%-4, 0, 3, "(a%4)")
+    self.helper_test_variable((-a)%4, 0, 3, "(a*-1%4)")
+    self.helper_test_variable(a%-4, -3, 0, "(a%-4)")
 
   def test_distribute_mul(self):
     self.helper_test_variable(usum([Variable("a", 0, 3), Variable("b", 0, 5)])*3, 0, 24, "((a*3)+(b*3))")
@@ -387,11 +393,11 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(Variable("a", 0, 6)%100, 0, 6, "a")
 
   def test_big_mod(self):
-    self.helper_test_variable(Variable("a", -20, 20)%10, -9, 9, "(a%10)")
-    self.helper_test_variable(Variable("a", -20, 0)%10, -9, 0, "(((a*-1)%10)*-1)")
-    self.helper_test_variable(Variable("a", -20, 1)%10, -9, 1, "(a%10)")
+    self.helper_test_variable(Variable("a", -20, 20)%10, 0, 9, "(a%10)")
+    self.helper_test_variable(Variable("a", -20, 0)%10, 0, 9, "(a%10)")
+    self.helper_test_variable(Variable("a", -20, 1)%10, 0, 9, "(a%10)")
     self.helper_test_variable(Variable("a", 0, 20)%10, 0, 9, "(a%10)")
-    self.helper_test_variable(Variable("a", -1, 20)%10, -1, 9, "(a%10)")
+    self.helper_test_variable(Variable("a", -1, 20)%10, 0, 9, "(a%10)")
 
   def test_ge_remove(self):
     self.helper_test_variable(Variable("a", 0, 6) >= 25, 0, 0, "False")
@@ -439,8 +445,8 @@ 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, -27, 27, "(((a+(b*28))+-29)%28)")
-    self.helper_test_variable(usum([uconst(-29), Variable("a", 0, 100), Variable("b", 0, 10)*28]) % 28, -27, 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+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)")
 
   def test_sum_combine_num(self):
     self.helper_test_variable(usum([uconst(29), Variable("a", 0, 10), uconst(-23)]), 6, 16, "(a+6)")
@@ -448,22 +454,12 @@ class TestSymbolic(unittest.TestCase):
   def test_sum_num_hoisted_and_factors_cancel_out(self):
     self.helper_test_variable(usum([Variable("a", 0, 1) * -4 + 1, Variable("a", 0, 1) * 4]), 1, 1, "1")
 
-  @unittest.expectedFailure  # only correct for floordiv, not truncdiv
   def test_div_cancel(self):
     self.helper_test_variable(usum([uconst(-40), Variable("a", 0, 10)*2, Variable("b", 0, 10)*40])//40, -1, 9, "(b+-1)")
 
-  def test_div_cancel_correct(self):
-    with Context(CORRECT_DIVMOD_FOLDING=1):
-      self.helper_test_variable(usum([uconst(-40), Variable("a", 0, 10)*2, Variable("b", 0, 10)*40])//40, -1, 9, "(((a+(b*20))+-20)//20)")
-
-  @unittest.expectedFailure  # only correct for floordiv, not truncdiv
   def test_mod_cancel(self):
     self.helper_test_variable(usum([uconst(-40), Variable("a", 0, 10)*2, Variable("b", 0, 10)*40]) % 40, 0, 20, "(a*2)")
 
-  def test_mod_cancel_correct(self):
-    with Context(CORRECT_DIVMOD_FOLDING=1):
-      self.helper_test_variable(usum([uconst(-40), Variable("a", 0, 10)*2, Variable("b", 0, 10)*40]) % 40, -38, 38, "((((a+(b*20))+-20)%20)*2)")
-
   def test_mul_div(self):
     self.helper_test_variable((Variable("a", 0, 10)*4)//4, 0, 10, "a")
 
@@ -475,22 +471,22 @@ class TestSymbolic(unittest.TestCase):
     lidx1 = UOp.variable("lidx1", 0, 1)
     ridx1005 = UOp.variable("ridx1005", 0, 2)
     ridx1006 = UOp.variable("ridx1006", 0, 2)
-    self.helper_test_variable((lidx1+((gidx1*18)+(ridx1005*18)+(lidx0*162))+(gidx0*2)+(ridx1006*2)+-40)//18, -2, 20,
-      "(((((lidx1+(((gidx1*18)+(ridx1005*18))+(lidx0*162)))+(gidx0*2))+(ridx1006*2))+-40)//18)")
+    self.helper_test_variable((lidx1+((gidx1*18)+(ridx1005*18)+(lidx0*162))+(gidx0*2)+(ridx1006*2)+-40)//18, -3, 20,
+      "(gidx1+ridx1005+lidx0*9+(gidx0+ridx1006+7)//9+-3)")
 
   def test_add_div(self):
     # careful about the lower bounds and upper bounds
-    self.helper_test_variable((Variable("a", 0, 5)-2)//4, 0, 0, "0")
-    self.helper_test_variable((Variable("a", 0, 5)-1)//4, 0, 1, "((a+-1)//4)")
+    self.helper_test_variable((Variable("a", 0, 5)-2)//4, -1, 0, "((a+2)//4+-1)")
+    self.helper_test_variable((Variable("a", 0, 5)-1)//4, -1, 1, "((a+3)//4+-1)")
     self.helper_test_variable((Variable("a", 0, 5))//4, 0, 1, "(a//4)")
     self.helper_test_variable((Variable("a", 0, 5)+1)//4, 0, 1, "((a+1)//4)")
     self.helper_test_variable((Variable("a", 0, 5)+2)//4, 0, 1, "((a+2)//4)")
     self.helper_test_variable((Variable("a", 0, 5)+3)//4, 0, 2, "((a+3)//4)")
-    self.helper_test_variable((Variable("a", 0, 5)+4)//4, 1, 2, "((a//4)+1)")
-    self.helper_test_variable((Variable("a", 0, 5)+5)//4, 1, 2, "(((a+1)//4)+1)")
+    self.helper_test_variable((Variable("a", 0, 5)+4)//4, 1, 2, "(a//4+1)")
+    self.helper_test_variable((Variable("a", 0, 5)+5)//4, 1, 2, "((a+1)//4+1)")
 
   def test_div_neg_rem(self):
-    self.helper_test_variable((-Variable("a", 0, 255)+256)//2, 0, 128, "((((a+1)//2)*-1)+128)")
+    self.helper_test_variable((-Variable("a", 0, 255)+256)//2, 0, 128, "(a*-1//2+128)")
 
   def test_mul_div_factor_mul(self):
     self.helper_test_variable((Variable("a", 0, 10)*8)//4, 0, 20, "(a*2)")
@@ -502,7 +498,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((Variable("a", 0, 10)*4)//8, 0, 5, "(a//2)")
 
   def test_mul_div_factor_div_neg(self):
-    self.helper_test_variable((Variable("a", 0, 10)*-4+4)//8, -4, 0, "(((a*-1)+1)//2)")
+    self.helper_test_variable((Variable("a", 0, 10)*-4+4)//8, -5, 0, "((a*-1+1)//2)")
 
   def test_div_symbolic_const_gcd(self):
     a = Variable("a", -10, 10)
@@ -520,8 +516,8 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((d1*a*d2*b*d1)//(d1*d2),  -1000, 1000, "(a*(b*d1))", test_z3=False)
     self.helper_test_variable((d1*a + b*d1)//(d1), -20, 20, "(a+b)", test_z3=False)
     self.helper_test_variable((d1*a + b*d1 + c*d1)//(d1), -30, 30, "(c+(a+b))", test_z3=False)
-    self.helper_test_variable((3*a*d1 + 9*b*d1)//(3*d1*d2), -40, 40, "(((a+(b*3))//(d2*-1))*-1)", test_z3=False)
-    self.helper_test_variable((3*a*d1 + 9*b*d1+3)//(3*d1*d2), -401, 399, "(((((a*d1)+((b*d1)*3))+1)//((d1*d2)*-1))*-1)", test_z3=False)
+    self.helper_test_variable((3*a*d1 + 9*b*d1)//(3*d1*d2), -40, 40, "((a+b*3)//d2)", test_z3=False)
+    self.helper_test_variable((3*a*d1 + 9*b*d1+3)//(3*d1*d2), -401, 399, "((a*d1+b*d1*3+1)//(d1*d2))", test_z3=False)
 
   def test_symbolic_factor_remainder_div(self):
     a = Variable("a", 0, 10)
@@ -532,7 +528,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((d*a*20+b*d*5+10)//(5*d), 0, 52, "((b+(a*4))+(2//d))")
 
   def test_mod_gcd_factor_neg(self):
-    self.helper_test_variable((Variable("a", 0, 10)*-4+4)%8, -4, 4, "((((a*-1)+1)%2)*4)")
+    self.helper_test_variable((Variable("a", 0, 10)*-4+4)%8, 0, 4, "((a*-1+1)%2*4)")
 
   def test_mod_gcd_fold_neg(self):
     self.helper_test_variable((Variable("a", 0, 10)*-8+20)%4, 0, 0, "0")
@@ -540,22 +536,32 @@ class TestSymbolic(unittest.TestCase):
   def test_sum_div_partial_remove(self):
     self.helper_test_variable(usum([Variable("idx0", 0, 127)*4, Variable("idx2", 0, 3)])//4, 0, 127, "idx0")
 
-  def test_cdiv_const_evaluation(self):
-    self.helper_test_variable((Variable("a", 0, 2)-12)//8, -1, -1, "-1")
-    self.helper_test_variable((-Variable("a", 0, 2))//7, 0, 0, "0")
+  def test_floordiv_const_evaluation(self):
+    self.helper_test_variable((Variable("a", 0, 2)-12)//8, -2, -2, "-2")
+    self.helper_test_variable((-Variable("a", 0, 2))//7, -1, 0, "(a*-1//7)")
 
-  def test_cmod_const_evaluation(self):
-    self.helper_test_variable((Variable("a", 1, 1)*-3)%8, -3, -3, "-3")
-    self.helper_test_variable((-Variable("a", 10, 10))%7, -3, -3, "-3")
+  def test_floormod_const_evaluation(self):
+    self.helper_test_variable((Variable("a", 1, 1)*-3)%8, 5, 5, "5")
+    self.helper_test_variable((-Variable("a", 10, 10))%7, 4, 4, "4")
 
   def test_div_numerator_negative(self):
-    with Context(CORRECT_DIVMOD_FOLDING=1):
-      self.helper_test_variable((Variable("idx", 0, 9)*-10)//11, -8, 0, "(((idx*10)//11)*-1)")
+    self.helper_test_variable((Variable("idx", 0, 9)*-10)//11, -9, 0, "(idx*-1)")
 
   def test_nest_div_negative_factor(self):
     ridx0=Variable("ridx0", 0, 9)
     ridx1=Variable("ridx1", 0, 6)
-    self.helper_test_variable(((((ridx0*-7)+ridx1)+63)//35), 0, 1, "(((ridx0//5)*-1)+1)")
+    self.helper_test_variable(((((ridx0*-7)+ridx1)+63)//35), 0, 1, "((ridx0*-1+4)//5+1)")
+
+  def test_floordiv_factor_nest_negative_numerator(self):
+    # x//c = (x//f)//(c//f) for f|c, any sign of x
+    a = Variable("a", -10, 10)
+    b = Variable("b", 0, 3)
+    self.helper_test_variable((a*4 + b)//12, -4, 3, "(a//3)")
+
+  def test_floordiv_gcd_with_remainder_negative_numerator(self):
+    # factor gcd from numerator, even when x crosses zero, as long as the shifted numerator stays nonneg
+    a = Variable("a", -1, 5)
+    self.helper_test_variable((a*2 + 7)//8, 0, 2, "((a+3)//4)")
 
   def test_div_into_mod(self):
     self.helper_test_variable((Variable("idx", 0, 16)*4)%8//4, 0, 1, "(idx%2)")
@@ -568,11 +574,11 @@ 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, 1, 26, "((idx//4)+1)")
-    self.helper_test_variable((-Variable("idx", 0, 100)+200)//-4 + 50, 0, 25, "((idx+3)//4)")
-    self.helper_test_variable((-Variable("idx", 0, 100)+201)//-4 + 50, 0, 25, "((idx+2)//4)")
-    self.helper_test_variable((-Variable("idx", 0, 100))//2, -50, 0, "((idx//2)*-1)")
-    self.helper_test_variable(Variable("idx", 0, 100)//-2, -50, 0, "((idx//2)*-1)")
+    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))//2, -50, 0, "(idx*-1//2)")
+    self.helper_test_variable(Variable("idx", 0, 100)//-2, -50, 0, "(idx//-2)")
 
   def test_sum_div_big_const(self):
     gidx0 = Variable("gidx0", 0, 24)
@@ -647,22 +653,22 @@ 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, 1, 250, "(((gidx*2)+(lidx//4))+1)")
-    self.helper_test_variable((-gidx*8-lidx+1000)//-4 + 250, 0, 250, "((gidx*2)+((lidx+3)//4))")
-    self.helper_test_variable((-gidx*8-lidx+1001)//-4 + 250, 0, 250, "((gidx*2)+((lidx+2)//4))")
-    self.helper_test_variable((-gidx*8-lidx+1002)//-4 + 250, 0, 250, "((gidx*2)+((lidx+1)//4))")
+    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)")
 
   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), 4, 4, "4")
-    self.helper_test_variable(-(((alu2+14)//(-32))+4), -4, -4, "-4")
-    self.helper_test_variable((((alu2+134)//(-32))+4), 0, 1, "(((lidx0+lidx1)+25)//32)")
-    self.helper_test_variable((((alu2+142)//(-32))+4), 0, 0, "0")
-    self.helper_test_variable((((alu2+150)//(-32))+4), 0, 0, "0")
-    self.helper_test_variable((((alu2+158)//(-32))+4), 0, 0, "0")
+    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+150)//(-32))+4), -1, -1, "-1")
+    self.helper_test_variable((((alu2+158)//(-32))+4), -1, -1, "-1")
 
   def test_div_mod_recombine(self):
     gidx = Variable("gidx", 0, 124)
@@ -696,7 +702,7 @@ class TestSymbolic(unittest.TestCase):
     # negative variable range
     xn = Variable("x", -1000, 1000)
     self.helper_test_variable(xn//3%224*3 + xn%3 + xn//672*672, -1000, 1000, "x")
-    self.helper_test_variable(xn//3%7*3 + xn//21*21, -999, 999, "(x//3*3)")
+    self.helper_test_variable(xn//3%7*3 + xn//21*21, -1002, 999, "(x//3*3)")
     # should NOT simplify: a*c1 != b (3*224 != 600)
     self.helper_test_variable(gidx//3%224*3 + gidx//600*600, 0, 150669, "(gidx//600*600+gidx//3%224*3)")
     # should NOT simplify: c1*c2 != c3 (224*3 != 700)
@@ -709,7 +715,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((30 * b + 1) % 18 + ((30 * b + 1) // 18) * 18, 1, 3001, "((b*30)+1)")
 
   def test_div_partial_quotient(self):
-    # IDIV should extract partial quotients when const_factor > divisor, matching what MOD already does
+    # FLOORDIV should extract partial quotients when const_factor > divisor, matching what FLOORMOD already does
     # (f*x+c)//d -> (f%d*x+c)//d + (f//d)*x  when f >= d
     b = Variable("b", 0, 100)
     self.helper_test_variable((31*b+1)//18, 0, 172, "(((b*13)+1)//18+b)")
@@ -730,8 +736,7 @@ class TestSymbolic(unittest.TestCase):
   def test_div_by_factor_tie_break(self):
     a = Variable("a", 0, 1)
     b = Variable("b", 0, 1)
-    with Context(CORRECT_DIVMOD_FOLDING=1):
-      self.helper_test_variable((a*2+b*3+2)//6, 0, 1, "((a+b+1)//3)")
+    self.helper_test_variable((a*2+b*3+2)//6, 0, 1, "((a+b+1)//3)")
 
   def test_div_mod_recombine_large_coeff(self):
     # recombine must work even when coeff > divisor: both mod and div reduce the coeff the same way
@@ -741,7 +746,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((25*a+3)%10 + ((25*a+3)//10)*10, 3, 253, "((a*25)+3)")
 
   def test_mod_nest_by_factor(self):
-    # (a*f+b) % (f*k) = (a%k)*f + b when 0<=b<f — mirrors nest_div_by_factor for MOD
+    # (a*f+b) % (f*k) = (a%k)*f + b when 0<=b<f — mirrors nest_div_by_factor for FLOORMOD
     gidx0 = Variable("gidx0", 0, 15)
     lidx0 = Variable("lidx0", 0, 3)
     # f=4, k=2, c=8: (gidx0*4+lidx0)%8 = (gidx0%2)*4 + lidx0
@@ -755,7 +760,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((a*3+b)%9, 0, 8, "(b+a%3*3)")
 
   def test_mod_nest_by_factor_with_const(self):
-    # nest_by_factor MOD with non-zero constant offset: (a*f+b+const) % (f*k) = (a%k)*f + b + const when 0<=b+const<f
+    # nest_by_factor FLOORMOD with non-zero constant offset: (a*f+b+const) % (f*k) = (a%k)*f + b + const when 0<=b+const<f
     a = Variable("a", 0, 7)
     b = Variable("b", 0, 1)
     # f=4, k=2, const=2: (a*4+b+2)%8 = (a%2)*4 + b + 2
@@ -767,7 +772,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable((a*3+b+1)%6, 1, 5, "(b+a%2*3+1)")
 
   def test_div_nest_by_factor_with_const(self):
-    # nest_by_factor IDIV: (160*a + 5*b + 4*c + K) // 60 should pick div=5 (clean) over div=4 (dirty)
+    # nest_by_factor FLOORDIV: (160*a + 5*b + 4*c + K) // 60 should pick div=5 (clean) over div=4 (dirty)
     a = Variable("a", 0, 2)
     b = Variable("b", 0, 31)
     c = Variable("c", 0, 1)
@@ -827,12 +832,26 @@ class TestSymbolic(unittest.TestCase):
     # TODO: simplify the true branch
     self.helper_test_variable((idx<4).where(idx//4, idx.const_like(-1)), -1, 6, "(idx<4).where((idx//4), -1)")
 
-  def test_idiv_lt(self):
+  def test_floordiv_lt(self):
+    # x//d<c <=> x<c*d 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<5)")
-    self.helper_test_variable(((idx-10)//4<0), 0, 1, "(idx<7)")
-    self.helper_test_variable((idx//-4<-3), 0, 1, "(((idx//4)*-1)<-3)")
+    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)")
+
+  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
+    a = Variable("a", 0, 100)
+    self.helper_test_variable((a % -8) // 2, -4, 0, "(a%-8//2)")
+    self.helper_test_variable((a % -8) % 2, 0, 1, "(a%2)")
+
+  def test_floordiv_lt_negative_c(self):
+    # x//d<c with negative c also reduces to x<c*d for d>0
+    idx = Variable("idx", -20, 20)
+    self.helper_test_variable((idx//4 < 0), 0, 1, "(idx<0)")
+    self.helper_test_variable((idx//4 < -1), 0, 1, "(idx<-4)")
+    self.helper_test_variable((idx//4 < -2), 0, 1, "(idx<-8)")
 
   def test_simplex_lt(self):
     a = Variable("a", 0, 3)
@@ -981,10 +1000,10 @@ class TestSymbolic(unittest.TestCase):
     self.assertIn((a.cast(dtypes.long)*b.cast(dtypes.long)).render(), "(long)((a*b))")
 
   def test_nested_mod_negative_range(self):
-    # (x%(k*c))%c = x%c holds for cmod regardless of signs since sign(x%(k*c)) = sign(x)
+    # (x%(k*c))%c = x%c for positive c
     x = Variable("x", 0, 1575)
-    self.helper_test_variable(((x + (-1064)) % 512) % 4, -3, 3, "((x+-1064)%4)")
-    self.helper_test_variable(((x + (-1064)) % 512) % 128, -127, 127, "((x+-1064)%128)")
+    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)")
 
 class TestSymbolicNumeric(unittest.TestCase):
   def helper_test_numeric(self, f):
@@ -1062,12 +1081,13 @@ class TestSymInfer(unittest.TestCase):
     assert sym_infer(a+b+c, var_vals) == 9
     assert sym_infer(a*b, var_vals) == 6
     assert sym_infer(a*b+c, var_vals) == 10
-  def test_sym_infer_cdiv_cmod(self):
+  def test_sym_infer_floordiv_floormod(self):
     a = Variable("a", -1000, 1)
     b = Variable("b", -1000, 1)
     var_vals = {a.expr: 1, b.expr: -1000}
-    assert sym_infer(a%b, var_vals) == 1
-    assert sym_infer(a//b, var_vals) == 0
+    # floor: 1 % -1000 = -999, 1 // -1000 = -1
+    assert sym_infer(a%b, var_vals) == -999
+    assert sym_infer(a//b, var_vals) == -1
   def test_sym_infer_with_bitcast(self):
     a = Variable("a", 1, 10, dtypes.int)
     expr = ((a.bitcast(dtypes.uint) << UOp.const(dtypes.uint, 1)).bitcast(dtypes.int) + 2)
@@ -1286,7 +1306,8 @@ class TestGatedUopGivenValid(unittest.TestCase):
 
     idx:UOp = (r0 < 3).where((r0 + uconst(-1)) // uconst(3), UOp.invalid())
     idx = graph_rewrite(idx, pm_simplify_valid)
-    self.assertEqual(idx, (r0 < 3).where(uconst(0), UOp.invalid()))
+    # (r0-1)//3 = (r0+2)//3 - 1 (constant offset split)
+    self.assertEqual(idx, (r0 < 3).where((r0 + uconst(2)) // uconst(3) + uconst(-1), UOp.invalid()))
 
   def test_invalid_gate_simplifies_vectorize(self):
     r0 = Variable("r0", 0, 2)
@@ -1295,8 +1316,8 @@ class TestGatedUopGivenValid(unittest.TestCase):
     idx1 = r0 % uconst(3)
     idx:UOp = (r0 < 3).where(UOp(Ops.STACK, dtypes.weakint.vec(2), (idx0, idx1)), UOp.invalid())
     idx = graph_rewrite(idx, pm_simplify_valid)
-    # NOTE: independent simplification: (r0-1)//3 -> 0, r0%3 -> r0 when r0 in [0,2]
-    expected_vec = UOp(Ops.STACK, dtypes.weakint.vec(2), (uconst(0), r0))
+    # independent simplification: (r0-1)//3 -> (r0+2)//3 - 1, and r0%3 -> r0 when r0 in [0,2]
+    expected_vec = UOp(Ops.STACK, dtypes.weakint.vec(2), ((r0 + uconst(2)) // uconst(3) + uconst(-1), r0))
     self.assertEqual(idx, (r0 < 3).where(expected_vec, UOp.invalid()))
 
 class TestRangeSplitting(unittest.TestCase):
@@ -1335,8 +1356,8 @@ class TestBounds(unittest.TestCase):
     alu0 = gidx0 * -1
     assert alu0.vmin == -2559 and alu0.vmax == 0
     assert (alu0+2559).vmin == 0 and (alu0+2559).vmax == 2559
-    assert ((alu0+2559)//-4).vmin == -639 and ((alu0+2559)//-4).vmax == 0
-    assert (((alu0+2559)//-4)*(-1)).vmin == 0 and (((alu0+2559)//-4)*(-1)).vmax == 639
+    assert ((alu0+2559)//-4).vmin == -640 and ((alu0+2559)//-4).vmax == 0
+    assert (((alu0+2559)//-4)*(-1)).vmin == 0 and (((alu0+2559)//-4)*(-1)).vmax == 640
 
 class TestFuzzFailure(unittest.TestCase):
   def test_fuzz_failure1(self):

test/null/test_uop_vmin_vmax.py

@@ -173,17 +173,15 @@ class TestVminVmaxDivMod(unittest.TestCase):
     self.assertEqual(uop.vmax, 10)
 
   def test_vmin_vmax_division_negative(self):
-    # vmin and vmax for division of a variable by a negative constant
-    # always positive
+    # floor division of a variable by a negative constant
     x = UOp.variable('x', 10, 20)
     uop = x // -2
     self.assertEqual(uop.vmin, -10)
     self.assertEqual(uop.vmax, -5)
     uop = x // -3
-    self.assertEqual(uop.vmin, -6)
-    self.assertEqual(uop.vmax, -3)
+    self.assertEqual(uop.vmin, -7)
+    self.assertEqual(uop.vmax, -4)
 
-    # always negative
     x = UOp.variable('x', -20, -10)
     uop = x // -2
     self.assertEqual(uop.vmin, 5)
@@ -192,38 +190,61 @@ class TestVminVmaxDivMod(unittest.TestCase):
     self.assertEqual(uop.vmin, 3)
     self.assertEqual(uop.vmax, 6)
 
+  def test_vmin_vmax_floordiv_floormod(self):
+    x = UOp.variable('x', -7, 7)
+    floordiv = x.alu(Ops.FLOORDIV, x.const_like(3))
+    self.assertEqual(floordiv.vmin, -3)
+    self.assertEqual(floordiv.vmax, 2)
+    floormod = x.alu(Ops.FLOORMOD, x.const_like(3))
+    self.assertEqual(floormod.vmin, 0)
+    self.assertEqual(floormod.vmax, 2)
+    # negative const divisor: floormod range is [c+1, 0]
+    floormod_neg = x.alu(Ops.FLOORMOD, x.const_like(-3))
+    self.assertEqual(floormod_neg.vmin, -2)
+    self.assertEqual(floormod_neg.vmax, 0)
+
     # cross 0
     x = UOp.variable('x', -10, 10)
     uop = x // -2
     self.assertEqual(uop.vmin, -5)
     self.assertEqual(uop.vmax, 5)
     uop = x // -3
-    self.assertEqual(uop.vmin, -3)
+    self.assertEqual(uop.vmin, -4)
     self.assertEqual(uop.vmax, 3)
 
+  def test_vmin_vmax_floordiv_floormod_empty_range(self):
+    # empty numerator range (vmin > vmax, e.g. RANGE with end=0) short-circuits to (0, 0)
+    rng = UOp.range(0, 0)
+    self.assertEqual(rng.vmin, 0)
+    self.assertEqual(rng.vmax, -1)
+    self.assertEqual((rng // 4).vmin, 0)
+    self.assertEqual((rng // 4).vmax, 0)
+    self.assertEqual((rng % 4).vmin, 0)
+    self.assertEqual((rng % 4).vmax, 0)
+
   def test_vmin_vmax_div_symbolic(self):
     x = UOp.variable('x', 1, 10)
     y = UOp.variable('y', 3, 5)
     self.assertEqual((x//y).vmin, 0)
     self.assertEqual((x//y).vmax, 3)
-    self.assertEqual(((-x)//y).vmin, -3)
-    self.assertEqual(((-x)//y).vmax, 0)
-    self.assertEqual((x//(-y)).vmin, -3)
-    self.assertEqual((x//(-y)).vmax, 0)
+    self.assertEqual(((-x)//y).vmin, -4)
+    self.assertEqual(((-x)//y).vmax, -1)
+    self.assertEqual((x//(-y)).vmin, -4)
+    self.assertEqual((x//(-y)).vmax, -1)
     self.assertEqual(((-x)//(-y)).vmin, 0)
     self.assertEqual(((-x)//(-y)).vmax, 3)
 
     self.assertEqual((100//y).vmin, 20)
     self.assertEqual((100//y).vmax, 33)
-    self.assertEqual(((-100)//y).vmin, -33)
+    self.assertEqual(((-100)//y).vmin, -34)
     self.assertEqual(((-100)//y).vmax, -20)
-    self.assertEqual((100//(-y)).vmin, -33)
+    self.assertEqual((100//(-y)).vmin, -34)
     self.assertEqual((100//(-y)).vmax, -20)
     self.assertEqual(((-100)//(-y)).vmin, 20)
     self.assertEqual(((-100)//(-y)).vmax, 33)
 
   def test_vmin_vmax_mod_positive(self):
-    # vmin and vmax for modulo of a variable by a positive constant
+    # floor mod with positive divisor: result in [0, c-1] regardless of dividend sign
     positive = UOp.variable('positive', 10, 20)
     uop = positive % 3
     self.assertEqual(uop.vmin, 0)
@@ -231,20 +252,20 @@ class TestVminVmaxDivMod(unittest.TestCase):
 
     negative = UOp.variable('negative', -20, -10)
     uop = negative % 3
-    self.assertEqual(uop.vmin, -2)
-    self.assertEqual(uop.vmax, 0)
+    self.assertEqual(uop.vmin, 0)
+    self.assertEqual(uop.vmax, 2)
 
     mixed = UOp.variable('mixed', -20, 20)
     uop = mixed % 3
-    self.assertEqual(uop.vmin, -2)
+    self.assertEqual(uop.vmin, 0)
     self.assertEqual(uop.vmax, 2)
 
   def test_vmin_vmax_mod_negative(self):
-    # vmin and vmax for modulo of a variable by a negative constant
+    # floor mod with negative divisor: result in [c+1, 0] regardless of dividend sign
     positive = UOp.variable('positive', 10, 20)
     uop = positive % -3
-    self.assertEqual(uop.vmin, 0)
-    self.assertEqual(uop.vmax, 2)
+    self.assertEqual(uop.vmin, -2)
+    self.assertEqual(uop.vmax, 0)
 
     negative = UOp.variable('negative', -20, -10)
     uop = negative % -3
@@ -254,7 +275,7 @@ class TestVminVmaxDivMod(unittest.TestCase):
     mixed = UOp.variable('mixed', -20, 20)
     uop = mixed % -3
     self.assertEqual(uop.vmin, -2)
-    self.assertEqual(uop.vmax, 2)
+    self.assertEqual(uop.vmax, 0)
 
 class TestVminVmaxVConst(unittest.TestCase):
   def test_vmin_vmax_vconst_single_element(self):
@@ -297,7 +318,7 @@ class TestVminVmaxVConst(unittest.TestCase):
     # vmin and vmax for a vector constant of bool values
     d1 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 1)
     idx = UOp.const(dtypes.int, 0)
-    val = UOp(Ops.LOAD, dtypes.int.vec(2), (d1.index(idx),))
+    val = UOp(Ops.LOAD, dtypes.int.vec(2), (d1.index(idx).cast(dtypes.int.vec(2).ptr()),))
     uop = (val // 32).gep(0)
     self.assertEqual(uop.vmin, -67108864)
     self.assertEqual(uop.vmax, 67108863)

test/null/test_uops.py

@@ -46,6 +46,20 @@ class TestExecALU(unittest.TestCase):
     self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (7, -3)), -2)
     self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (-50, 6)), -8)
 
+  def test_floordiv(self):
+    self.assertEqual(exec_alu(Ops.FLOORDIV, dtypes.int8, (8, 2)), 4)
+    self.assertEqual(exec_alu(Ops.FLOORDIV, dtypes.int8, (7, 3)), 2)
+    self.assertEqual(exec_alu(Ops.FLOORDIV, dtypes.int8, (7, -3)), -3)
+    self.assertEqual(exec_alu(Ops.FLOORDIV, dtypes.int8, (-7, 3)), -3)
+    self.assertEqual(exec_alu(Ops.FLOORDIV, dtypes.int8, (-50, 6)), -9)
+
+  def test_floormod(self):
+    self.assertEqual(exec_alu(Ops.FLOORMOD, dtypes.int8, (8, 2)), 0)
+    self.assertEqual(exec_alu(Ops.FLOORMOD, dtypes.int8, (7, 3)), 1)
+    self.assertEqual(exec_alu(Ops.FLOORMOD, dtypes.int8, (7, -3)), -2)
+    self.assertEqual(exec_alu(Ops.FLOORMOD, dtypes.int8, (-7, 3)), 2)
+    self.assertEqual(exec_alu(Ops.FLOORMOD, dtypes.int8, (-50, 6)), 4)
+
     np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIPROCAL, dtypes.float32, (3.0,)))), 2+(1.0/3.0))
     np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIPROCAL, dtypes.float32, (-3.0,)))), -2-(1.0/3.0))
 
@@ -163,6 +177,30 @@ class TestFastIdiv(unittest.TestCase):
       self.assertIn(Ops.SHR, ops, f"For dtype={dt} divison by power of two did not simplify to shift")
       self.assertNotIn(Ops.IDIV, ops, f"For dtype={dt} divison by power of two did not simplify to shift")
 
+  def test_floormod_power_of_two(self):
+    # FLOORMOD by a power of two lowers to AND (correct floor mod for any sign in two's complement)
+    for dt in (dtypes.int32, dtypes.uint32):
+      g = UOp(Ops.PARAM, dt.ptr(), (), 0)
+      c = UOp.const(dt, 8)
+      a = UOp(Ops.FLOORMOD, dt, (g.index(c), c))
+      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
+      ops = [x.op for x in uops]
+      self.assertIn(Ops.AND, ops, f"For dtype={dt} FLOORMOD by pow2 did not simplify to AND")
+      self.assertNotIn(Ops.MOD, ops, f"For dtype={dt} FLOORMOD by pow2 left a MOD")
+      self.assertNotIn(Ops.FLOORMOD, ops, f"For dtype={dt} FLOORMOD survived past late rewrite")
+
+  def test_floordiv_power_of_two_uint(self):
+    # uint FLOORDIV by a power of two lowers to a shift, leaving no IDIV/FLOORDIV in the kernel
+    for dt in (dtypes.uint32, dtypes.uint64):
+      g = UOp(Ops.PARAM, dt.ptr(), (), 0)
+      c = UOp.const(dt, 2)
+      a = UOp(Ops.FLOORDIV, dt, (g.index(c), c))
+      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
+      ops = [x.op for x in uops]
+      self.assertIn(Ops.SHR, ops, f"For dtype={dt} FLOORDIV by power of two did not simplify to shift")
+      self.assertNotIn(Ops.IDIV, ops, f"For dtype={dt} FLOORDIV by power of two did not simplify to shift")
+      self.assertNotIn(Ops.FLOORDIV, ops, f"For dtype={dt} FLOORDIV survived past late rewrite")
+
   @unittest.skipIf(Device.DEFAULT == "WEBGPU", "WEBGPU doesn't support long")
   def test_fast_idiv_and_mod(self):
     g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)

test/null/test_viz.py

@@ -1,4 +1,4 @@
-import unittest, decimal, sys, json, contextlib, tempfile, pickle, io, itertools
+import unittest, decimal, sys, json, contextlib, tempfile, pickle, io
 from pathlib import Path
 from dataclasses import dataclass
 from typing import Generator
@@ -11,7 +11,7 @@ from tinygrad.helpers import cpu_profile, ProfilePointEvent, unwrap
 from tinygrad.device import Buffer
 
 from tinygrad.uop.ops import tracked_keys, tracked_ctxs, uop_fields, active_rewrites, active_group, _name_cnt, RewriteTrace
-from tinygrad.viz.serve import load_rewrites, get_full_rewrite, uop_to_json, VizData
+from tinygrad.viz.serve import load_rewrites, get_full_rewrite, uop_to_json, VizData, get_render
 from tinygrad.codegen import to_program_cache
 from tinygrad.codegen import to_program
 
@@ -320,8 +320,7 @@ class TestVizGC(unittest.TestCase):
 
 # VIZ integrates with other parts of tinygrad
 
-from tinygrad import Tensor, Device
-from tinygrad.engine.realize import get_runner
+from tinygrad import Tensor, Device, TinyJit, Variable
 
 class TestVizIntegration(unittest.TestCase):
   # codegen supports rendering of code blocks
@@ -409,6 +408,18 @@ class TestVizIntegration(unittest.TestCase):
     lst = viz.list_items()
     assert len(lst) == 1
 
+  def test_jit(self):
+    with save_viz():
+      @TinyJit
+      def f(a, b, c): return (a+b).contiguous().mul(3), c.add(1).contiguous().assign(a.to(c.device)), b.assign(c.to(b.device))
+      a, b, c = Tensor.empty(16, device="NULL"), Tensor.empty(16, device="NULL"), Tensor.empty(16, device="NULL:1")
+      for _ in range(3): Tensor.realize(*f(a, b, c))
+    out = load_profile(cpu_events)
+    self.assertEqual(["NULL", "NULL Graph", "NULL:SDMA:0", "NULL:1", "NULL:1:SDMA:0"], [k for k in out["layout"] if k.startswith("NULL")])
+    self.assertEqual(len(out["layout"]["NULL"]["events"]), 2*3)
+    self.assertEqual(len(out["layout"]["NULL:SDMA:0"]["events"]), 3)
+    self.assertEqual(len(out["layout"]["NULL Graph"]["events"]), 2)
+
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEvent, ProfileGraphEntry
 from tinygrad.viz.serve import get_profile
 from tinygrad.viz.cli import decode_profile
@@ -422,9 +433,9 @@ class TestVizProfiler(unittest.TestCase):
       a.to("NULL:1").realize()
     range_events = [e for e in cpu_events if isinstance(e, ProfileRangeEvent)]
     compute_events = [e for e in range_events if e.device == "NULL"]
-    copy_events = [e for e in range_events if e.device.endswith(":COPY")]
+    copy_events = [e for e in range_events if e.device.endswith(":SDMA:0")]
     self.assertGreater(len(compute_events), 0, "expected compute events on base device")
-    self.assertGreater(len(copy_events), 0, "transfer must produce events with ':COPY' device suffix")
+    self.assertGreater(len(copy_events), 0, "transfer must produce events with ':SDMA' device suffix")
 
   def test_node(self):
     prof = [ProfileRangeEvent(device='NV', name='E_2', st=decimal.Decimal(1000), en=decimal.Decimal(1010)),
@@ -465,8 +476,7 @@ class TestVizProfiler(unittest.TestCase):
             ProfileDeviceEvent(device='NV:SDMA:0', tdiff=decimal.Decimal(-1000))]
     j = load_profile(prof)
     event = j['layout']['NV:SDMA:0']['events'][0]
-    gbs = sz/(dur*1e-6)*1e-9
-    self.assertEqual(event['fmt'], f"{gbs:.0f} GB/s\n{sz/1e6:.0f} MB")
+    self.assertEqual(event['fmt'], {"B/s": sz/(dur*1e-6), "B": sz})
 
   def test_graph(self):
     prof = [ProfileDeviceEvent(device='NV', tdiff=decimal.Decimal(-1000)),
@@ -507,8 +517,7 @@ class TestVizProfiler(unittest.TestCase):
 
     j = load_profile(prof)
     sdma_events = j['layout']['NV:1:SDMA:0']['events']
-    gbs = sz/(dur*1e-6)*1e-9
-    self.assertEqual(sdma_events[0]["fmt"], f"{gbs:.0f} GB/s\n{sz/1e6:.0f} MB")
+    self.assertEqual(sdma_events[0]["fmt"], {"B/s": sz/(dur*1e-6), "B": sz})
 
   def test_block_ordering(self):
     prof = [ProfileDeviceEvent(device='NV', tdiff=decimal.Decimal(-1000)),
@@ -572,7 +581,7 @@ class TestVizProfiler(unittest.TestCase):
     profile_ret = load_profile(cpu_events)
     e = profile_ret["layout"]["CUSTOM"]["events"][0]
     self.assertEqual(e["name"], "test_fxn")
-    runtime_trace = json.loads(e["fmt"].replace("TB:", ""))
+    runtime_trace = e["fmt"]["tb"]
     assert any(fxn.__code__.co_filename == f and fxn.__code__.co_firstlineno+1 == l for f,l,*_ in runtime_trace), str(runtime_trace)
 
   # can pack up to 1hr 11 min of trace events
@@ -707,7 +716,6 @@ class TestVizMemoryLayout(unittest.TestCase):
     self.assertEqual(len(programs), len(set(users)), n)
 
 from tinygrad.uop.ops import KernelInfo
-from tinygrad.viz.serve import amdgpu_cfg
 from tinygrad.renderer.amd.dsl import s
 from tinygrad.runtime.autogen.amd.rdna3.ins import (s_add_u32, s_branch, s_cbranch_execz, s_cbranch_scc0, s_cbranch_scc1, s_cmp_eq_i32,
                                                     s_cmp_eq_u64, s_code_end, s_endpgm, s_mov_b32, s_nop)
@@ -723,13 +731,16 @@ class TestCfg(unittest.TestCase):
       gidx = UOp.special(1, "gidx0")
       sink = UOp.sink(out.base, lidx, gidx, arg=KernelInfo(name=name))
       return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="NULL"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
-    with Context(DEV=f"NULL::{self.arch}"):
-      out = Tensor.custom_kernel(Tensor.empty(1), fxn=fxn)[0]
-      runner = get_runner(out.device, out.schedule_linear().src[-1].src[0])
-      return amdgpu_cfg(runner.prg.src[4].arg, self.arch)
+    with save_viz() as viz:
+      with Context(DEV=f"NULL::{self.arch}"):
+        out = Tensor.custom_kernel(Tensor.empty(1), fxn=fxn)[0]
+        _ = to_program(out.schedule_linear().src[-1].src[0], Device[out.device].renderer)
+    codegen_rewrites = next(s for s in viz.list_items() if s["name"] == name)
+    disasm = next(s for s in codegen_rewrites["steps"] if s["name"] == "View Disassembly")
+    return get_render(viz.data, disasm["query"])
 
   def test_simple(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_branch(), target="bb1")
     k.label("bb1")
@@ -739,7 +750,7 @@ class TestCfg(unittest.TestCase):
     self.assertEqual(len(cfg["blocks"]), 2)
 
   def test_diamond(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_mov_b32(s[0], 0))
     k.emit(s_mov_b32(s[1], 0))
@@ -773,7 +784,7 @@ class TestCfg(unittest.TestCase):
       assert st.startswith("s_code_end") and st.endswith("x)"), st
 
   def test_loop(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_mov_b32(s[1], 4))
     k.label("loop")
@@ -785,7 +796,7 @@ class TestCfg(unittest.TestCase):
     self.get_cfg("simple_loop", k)
 
   def test_loop_branch(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_mov_b32(s[1], 4))
     k.label("loop")
@@ -803,7 +814,7 @@ class TestCfg(unittest.TestCase):
     self.get_cfg("loop_if", k)
 
   def test_loop_break(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_mov_b32(s[1], 8))
     k.label("loop")
@@ -818,7 +829,7 @@ class TestCfg(unittest.TestCase):
     self.get_cfg("loop_break", k)
 
   def test_switch(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_cmp_eq_i32(s[0], 0))
     k.emit(s_cbranch_scc1(), target="case0")
@@ -840,7 +851,7 @@ class TestCfg(unittest.TestCase):
     self.get_cfg("switch_case", k)
 
   def test_ping_pong(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_cmp_eq_i32(s[0], 0))
     k.emit(s_cbranch_scc1(), target="ping")
@@ -859,7 +870,7 @@ class TestCfg(unittest.TestCase):
 
   def test_colored_blocks(self):
     N = 10
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_branch(), target="init0")
     for i in range(N):
@@ -879,7 +890,7 @@ class TestCfg(unittest.TestCase):
     self.get_cfg("test_colored_blocks", k)
 
   def test_jump_back_to_end(self):
-    k = Kernel(arch=self.arch)
+    k = Kernel()
     k.label("entry")
     k.emit(s_mov_b32(s[1], 2))
     k.emit(s_cbranch_execz(), target="loop")
@@ -900,46 +911,62 @@ def run_cli(*cli_args) -> str:
     main(args)
   return buf.getvalue().strip()
 
+def call_cli(fxn, *cli_args, debug=2) -> str:
+  with save_viz() as viz:
+    fxn()
+  with tempfile.TemporaryDirectory() as tmpdir:
+    (r:=Path(tmpdir)/"rewrites.pkl").write_bytes(pickle.dumps(viz.data.trace))
+    (p:=Path(tmpdir)/"profile.pkl").write_bytes(pickle.dumps(cpu_events))
+    with Context(DEBUG=debug, NO_COLOR=1):
+      stdout = run_cli("--rewrites-path", str(r), "--profile-path", str(p), *cli_args)
+  return stdout
+
 class TestCLI(unittest.TestCase):
-  def test_simple(self):
-    a = Tensor.empty(1, device="NULL")+2.0
-    empty_counter = itertools.count(0)
-    def custom_empty_prg(B:UOp, A:UOp) -> UOp:
-      sink = UOp(Ops.SINK, arg=KernelInfo(name=f"custom_empty_n{next(empty_counter)}"))
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=a.device), UOp(Ops.LINEAR, src=(sink,))))
-    def custom_empty_src(B:UOp, A:UOp) -> UOp:
-      sink = UOp(Ops.SINK, arg=KernelInfo(name=f"custom_empty_n{next(empty_counter)}"))
-      src = "void custom_empty_src() { 0; }"
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=a.device), UOp(Ops.LINEAR, src=(sink,)), UOp(Ops.SOURCE, arg=src)))
-    b = Tensor.custom_kernel(Tensor.empty_like(a), a, fxn=custom_empty_prg)[0]
-    c = Tensor.custom_kernel(Tensor.empty_like(a), a, fxn=custom_empty_prg)[0]
-    d = Tensor.custom_kernel(Tensor.empty_like(a), a, fxn=custom_empty_src)[0]
-    with save_viz() as viz:
-      b.realize()
+  def test_reconstruct_debug(self):
+    def fxn():
+      Tensor.empty(1, device="NULL").add(2.0).realize()
       profile_marker("marker @ 1")
-      c.realize()
-      d.realize()
-    # save trace to disk for CLI to consume it
-    with tempfile.TemporaryDirectory() as tmpdir:
-      (r:=Path(tmpdir)/"rewrites.pkl").write_bytes(pickle.dumps(viz.data.trace))
-      (p:=Path(tmpdir)/"profile.pkl").write_bytes(pickle.dumps(cpu_events))
-      # reconstruct DEBUG=4 output and see all markers.
-      with Context(DEBUG=4):
-        kernels = run_cli("--rewrites-path", str(r), "--profile-path", str(p), "-s", "NULL")
-      self.assertIn("void custom_empty_n0", kernels)
-      self.assertIn("marker @ 1", kernels)
-      self.assertIn("void custom_empty_n1", kernels)
-      self.assertIn("void custom_empty_src", kernels)
-      self.assertIn("E", kernels)
-      self.assertIn("UOp.const", kernels)
-      # get the top slowest functions across all devices
-      with Context(DEBUG=2):
-        times = run_cli("--rewrites-path", str(r), "--profile-path", str(p), "-s", "ALL", "--top", "-1")
-      self.assertIn("TINY", times)
-      self.assertIn("NULL", times)
-      with Context(DEBUG=3):
-        json_lines = run_cli("--rewrites-path", str(r), "--profile-path", str(p), "-s", "ALL", "--json")
-      for line in json_lines.split("\n"): _ = json.loads(line)
+      Tensor.empty(1, device="NULL").add(3.0).realize()
+    out = call_cli(fxn, "-s", "NULL", debug=4)
+    self.assertIn("void E", out)
+    self.assertIn("marker @ 1", out)
+
+  def test_aggregate(self):
+    N, CNT = 1024, 5
+    def fxn():
+      for _ in range(CNT):
+        (Tensor.empty(N, N, device="NULL")@Tensor.empty(N, N, device="NULL")).realize()
+      for _ in range(CNT):
+        (Tensor.empty(N, N, device="NULL").assign(Tensor.empty(N, N, device="NULL"))).realize()
+    kernels = [json.loads(line) for line in call_cli(fxn, "-s", "NULL", "-t", "--json").splitlines()]
+    self.assertEqual(len(kernels), 2)
+    gemm_summary = [s for s in kernels if s["name"].startswith("r_")][0]
+    copy_summary = [s for s in kernels if s["name"].startswith("E_")][0]
+    self.assertEqual(gemm_summary["count"], CNT)
+    self.assertEqual(copy_summary["count"], CNT)
+
+  def test_flops(self):
+    test_n = [(8, 16), (16, 32), (32, 64)]
+    def fxn():
+      @TinyJit
+      def f(a, b): return (a@a.T), (b@b.T)
+      a = Tensor.empty(64, 64, device="NULL")
+      b = Tensor.empty(64, 64, device="NULL")
+      for i_val, j_val in test_n:
+        i = Variable("i", 1, 64).bind(i_val)
+        j = Variable("j", 1, 64).bind(j_val)
+        Tensor.realize(*f(a[:i], b[:j]))
+    out = [json.loads(line) for line in call_cli(fxn, "-s", "NULL", "--json").splitlines()]
+    self.assertEqual(len(out), 3*2)
+    # flops increases as N gets larger
+    gflops = [row["fmt"]["FLOPS"] for row in out]
+    self.assertGreater(gflops[4], gflops[2])
+    self.assertGreater(gflops[5], gflops[3])
+    # aggregate flops
+    out = [json.loads(line) for line in call_cli(fxn, "-s", "NULL", "-t", "--json").splitlines()]
+    self.assertEqual(len(out), 2)
+    agg_gflops = [row["fmt"]["FLOPS"] for row in out]
+    assert all(min(gflops) < v < max(gflops) for v in agg_gflops), f"{agg_gflops}"
 
 if __name__ == "__main__":
   unittest.main()

test/opt/test_tensor_cores.py

@@ -3,12 +3,12 @@ import unittest
 
 from tinygrad import Device, Tensor, dtypes
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.uop.ops import Ops
+from tinygrad.uop.ops import Ops, UOp, buffers
 from tinygrad.dtype import DType
-from tinygrad.device import is_dtype_supported
+from tinygrad.device import Buffer, is_dtype_supported
 from tinygrad.helpers import DEV, Context
 from test.helpers import slow, replace_opts
-from tinygrad.engine.realize import CompiledRunner
+from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
 from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 from tinygrad.codegen.opt.tc import amd_cdna_1616128
@@ -20,6 +20,11 @@ from test.backend.test_linearizer import helper_realized_ast, helper_linearizer_
 
 AMX = "AMX" in DEV.arch
 
+def run_program(prg:UOp, bufs:list[Buffer]):
+  buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in bufs]
+  for u,b in zip(buf_uops, bufs): buffers[u] = b
+  run_linear(UOp(Ops.LINEAR, src=(prg.call(*buf_uops),)))
+
 def helper_tc_ensure_uops_and_opts_count(N: int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0,
                                          ensure_triggered:bool=True):
   a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
@@ -47,11 +52,11 @@ def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axi
   if dtype_in == dtypes.bfloat16: r = r.float()
   realized_ast, bufs = helper_realized_ast(r)
   opts = [Opt(op=OptOps.TC, axis=axis, arg=(tc_select, tc_opt, use_tensor_cores))]
-  pu = to_program(replace_opts(realized_ast, opts), Device[Device.DEFAULT].renderer)
+  ast = replace_opts(realized_ast, opts)
+  pu = to_program(ast, Device[Device.DEFAULT].renderer)
   if use_tensor_cores == 1: assert len([uop for uop in pu.src[2].src if uop.op is Ops.WMMA]) > 0, "wmma not triggered"
   assert len([x for x in pu.src[0].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
-  prg = CompiledRunner(pu, Device.DEFAULT)
-  prg.exec(bufs)
+  run_program(ast, bufs)
   if dtype_in == dtypes.half: tc_atol, tc_rtol = 1e-2, 1e-3
   elif dtype_in == dtypes.bfloat16: tc_atol, tc_rtol = (1e-1, 2e-2) if dtype_out == dtypes.bfloat16 else (1e-2, 1e-2)
   else: tc_atol, tc_rtol = 5e-3, 1e-4
@@ -145,15 +150,15 @@ class TestTensorCores(unittest.TestCase):
         c = a.conv2d(b, padding=1, dtype=tc.dtype_out)
         realized_ast, real_bufs = helper_realized_ast(c)
 
-        program = to_program(replace_opts(realized_ast, [Opt(OptOps.TC, axis, (-1, 2, 1))]), Device[Device.DEFAULT].renderer)
+        ast = replace_opts(realized_ast, [Opt(OptOps.TC, axis, (-1, 2, 1))])
+        program = to_program(ast, Device[Device.DEFAULT].renderer)
         assert len([uop for uop in tuple(program.src[2].src) if uop.op is Ops.WMMA]) > 0, "tensor core not triggered"
         assert len([x for x in program.src[0].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
 
-        prg = CompiledRunner(program, Device.DEFAULT)
         # TODO: support this even if numpy doesn't
         if _to_np_dtype(real_bufs[0].dtype) is None: continue
         real_bufs[0].copyin(np.zeros((real_bufs[0].size, ), dtype=_to_np_dtype(real_bufs[0].dtype)).data) # Zero to check that all values are filled
-        prg.exec(real_bufs)
+        run_program(ast, real_bufs)
         result = np.frombuffer(real_bufs[0].as_memoryview(), _to_np_dtype(real_bufs[0].dtype))
 
         # ensure the results for each choice of axis matches

test/unit/test_assign.py

@@ -903,8 +903,8 @@ class TestAssignToUnrealizedView(unittest.TestCase):
 
   def test_reduce(self):
     a = Tensor([[1,2],[3,4]]).contiguous().realize()
-    r = a.sum(axis=0)  # unrealized REDUCE_AXIS
-    self.assertIs(r.uop.base.op, Ops.REDUCE_AXIS)
+    r = a.sum(axis=0)  # unrealized REDUCE
+    self.assertIs(r.uop.base.op, Ops.REDUCE)
     r[:1].assign(Tensor([99]).realize())
     try:
       self.assertEqual(r.tolist(), [99,6])

test/unit/test_function.py

@@ -449,8 +449,8 @@ class TestFunctionTuple(unittest.TestCase):
       store_d = D[i].store(A[i] * 3.0)
       return UOp.group(store_c, store_d).end(i).sink(arg=KernelInfo(name="my_kernel"))
 
-    def my_grad(d_combined:UOp, call:UOp):
-      return (None, None, Tensor(d_combined).uop)
+    def my_grad(d_c:UOp, d_d:UOp, call:UOp):
+      return (None, None, (Tensor(d_c) + Tensor(d_d)).uop)
 
     @function(precompile=True, precompile_backward=True)
     def f(a:Tensor):

test/unit/test_gradient.py

@@ -2,6 +2,7 @@ import unittest
 import numpy as np
 from tinygrad import Tensor
 from tinygrad.dtype import dtypes
+from tinygrad.uop.ops import UOp, KernelInfo
 
 class TestTensorGradient(unittest.TestCase):
   def test_example(self):
@@ -76,6 +77,71 @@ class TestTensorGradient(unittest.TestCase):
     x = Tensor.randn(4, 4)
     np.testing.assert_allclose(x.pad(((1,0),(0,0))).gradient(x, gradient=g2)[0].numpy(), np.zeros((4, 4)))
 
+class TestMultiOutputGradient(unittest.TestCase):
+  @staticmethod
+  def addmul_kernel(C:UOp, D:UOp, A:UOp, B:UOp) -> UOp:
+    C, D, A, B = C.flatten(), D.flatten(), A.flatten(), B.flatten()
+    i = UOp.range(C.numel(), 0)
+    store_c = C[i].store(A[i] + B[i])
+    store_d = D[i].store(A[i] * B[i])
+    return UOp.group(store_c, store_d).end(i).sink(arg=KernelInfo(name="addmul")).simplify()
+  @staticmethod
+  def backward_addmul(grad_c, grad_d, call):
+    _c, _d, a, b = call.src[1:]
+    grad_a = (Tensor(grad_c) + Tensor(grad_d) * Tensor(b)).uop
+    grad_b = (Tensor(grad_c) + Tensor(grad_d) * Tensor(a)).uop
+    return (None, None, grad_a, grad_b)
+
+  def test_custom_kernel_multi_output_backward(self):
+    a_np, b_np = np.random.randn(4, 4).astype(np.float32), np.random.randn(4, 4).astype(np.float32)
+    a_ref, b_ref = Tensor(a_np, requires_grad=True), Tensor(b_np, requires_grad=True)
+    ((a_ref + b_ref).sum() + (a_ref * b_ref).sum()).backward()
+
+    a, b = Tensor(a_np, requires_grad=True), Tensor(b_np, requires_grad=True)
+    Tensor.realize(a, b)
+    c, d, _, _ = Tensor.custom_kernel(Tensor.empty(4, 4), Tensor.empty(4, 4), a, b, fxn=self.addmul_kernel, grad_fxn=self.backward_addmul)
+    (c.sum() + d.sum()).backward()
+    np.testing.assert_allclose(a.grad.numpy(), a_ref.grad.numpy(), rtol=1e-5)
+    np.testing.assert_allclose(b.grad.numpy(), b_ref.grad.numpy(), rtol=1e-5)
+
+  def test_custom_kernel_multi_output_backward_interacting(self):
+    a_np, b_np = np.random.randn(4, 4).astype(np.float32), np.random.randn(4, 4).astype(np.float32)
+    a_ref, b_ref = Tensor(a_np, requires_grad=True), Tensor(b_np, requires_grad=True)
+    ((a_ref + b_ref) * (a_ref * b_ref)).sum().backward()
+
+    a, b = Tensor(a_np, requires_grad=True), Tensor(b_np, requires_grad=True)
+    Tensor.realize(a, b)
+    c, d, _, _ = Tensor.custom_kernel(Tensor.empty(4, 4), Tensor.empty(4, 4), a, b, fxn=self.addmul_kernel, grad_fxn=self.backward_addmul)
+    (c * d).sum().backward()
+    np.testing.assert_allclose(a.grad.numpy(), a_ref.grad.numpy(), rtol=1e-5)
+    np.testing.assert_allclose(b.grad.numpy(), b_ref.grad.numpy(), rtol=1e-5)
+
+  def test_custom_kernel_three_output_backward(self):
+    def addmulsub_kernel(C:UOp, D:UOp, E:UOp, A:UOp, B:UOp) -> UOp:
+      C, D, E, A, B = C.flatten(), D.flatten(), E.flatten(), A.flatten(), B.flatten()
+      i = UOp.range(C.numel(), 0)
+      store_c = C[i].store(A[i] + B[i])
+      store_d = D[i].store(A[i] * B[i])
+      store_e = E[i].store(A[i] - B[i])
+      return UOp.group(store_c, store_d, store_e).end(i).sink(arg=KernelInfo(name="addmulsub")).simplify()
+    def backward_addmulsub(grad_c, grad_d, grad_e, call):
+      _c, _d, _e, a, b = call.src[1:]
+      grad_a = (Tensor(grad_c) + Tensor(grad_d) * Tensor(b) + Tensor(grad_e)).uop
+      grad_b = (Tensor(grad_c) + Tensor(grad_d) * Tensor(a) - Tensor(grad_e)).uop
+      return (None, None, None, grad_a, grad_b)
+
+    a_np, b_np = np.random.randn(4, 4).astype(np.float32), np.random.randn(4, 4).astype(np.float32)
+    a_ref, b_ref = Tensor(a_np, requires_grad=True), Tensor(b_np, requires_grad=True)
+    ((a_ref + b_ref).sum() + (a_ref * b_ref).sum() + (a_ref - b_ref).sum()).backward()
+
+    a, b = Tensor(a_np, requires_grad=True), Tensor(b_np, requires_grad=True)
+    Tensor.realize(a, b)
+    c, d, e, _, _ = Tensor.custom_kernel(Tensor.empty(4, 4), Tensor.empty(4, 4), Tensor.empty(4, 4), a, b,
+                                          fxn=addmulsub_kernel, grad_fxn=backward_addmulsub)
+    (c.sum() + d.sum() + e.sum()).backward()
+    np.testing.assert_allclose(a.grad.numpy(), a_ref.grad.numpy(), rtol=1e-5)
+    np.testing.assert_allclose(b.grad.numpy(), b_ref.grad.numpy(), rtol=1e-5)
+
 class TestViewGradient(unittest.TestCase):
   def test_expand(self):
     x = Tensor.randn(5,2)

test/unit/test_hcq_graph.py

@@ -11,7 +11,7 @@ from test.mockgpu.usb import MockUSB
 @unittest.skipUnless(issubclass(type(Device[Device.DEFAULT]), HCQCompiled), "HCQ device required to run")
 class TestHCQUnit(unittest.TestCase):
   @unittest.skipIf(Device.DEFAULT == "CPU", "requires non-CPU HCQ device")
-  def test_supports_exec_item(self):
+  def test_supports_uop(self):
     d0, cpu_dev = Device[Device.DEFAULT], Device["CPU"]
 
     @TinyJit
@@ -20,23 +20,23 @@ class TestHCQUnit(unittest.TestCase):
     inp, inp_cpu = Tensor.randn(10, 10, device=Device.DEFAULT).realize(), Tensor.randn(10, 10, device="CPU").realize()
     for _ in range(5): f(inp, inp_cpu)
 
-    # construct minimal CALL UOps for supports_exec_item (graphs only see PROGRAMs after compile_linear)
+    # construct minimal CALL UOps for supports_uop (graphs only see PROGRAMs after compile_linear)
     gpu_call = UOp(Ops.PROGRAM).call(UOp.new_buffer(Device.DEFAULT, 1, dtypes.float))
     cpu_call = UOp(Ops.PROGRAM).call(UOp.new_buffer("CPU", 1, dtypes.float))
     gpu_devs = [d0]
 
     # local MMIO: GPU works alone and with CPU in batch (cpu_support=True)
-    assert HCQGraph.supports_exec_item(gpu_devs, gpu_call) is True
-    assert HCQGraph.supports_exec_item(gpu_devs, cpu_call) is True
-    assert HCQGraph.supports_exec_item(gpu_devs + [cpu_dev], gpu_call) is True
+    assert HCQGraph.supports_uop(gpu_devs, gpu_call) is True
+    assert HCQGraph.supports_uop(gpu_devs, cpu_call) is True
+    assert HCQGraph.supports_uop(gpu_devs + [cpu_dev], gpu_call) is True
 
     # USB MMIO: GPU-only still works, but CPU batching must be rejected (cpu_support=False)
     orig_view = d0.timeline_signal.base_buf.view
     try:
       d0.timeline_signal.base_buf.view = USBMMIOInterface(MockUSB(bytearray(256)), 0, 16, fmt='B')
-      assert HCQGraph.supports_exec_item(gpu_devs, gpu_call) is True
-      assert HCQGraph.supports_exec_item(gpu_devs, cpu_call) is False
-      assert HCQGraph.supports_exec_item(gpu_devs + [cpu_dev], gpu_call) is False
+      assert HCQGraph.supports_uop(gpu_devs, gpu_call) is True
+      assert HCQGraph.supports_uop(gpu_devs, cpu_call) is False
+      assert HCQGraph.supports_uop(gpu_devs + [cpu_dev], gpu_call) is False
     finally:
       d0.timeline_signal.base_buf.view = orig_view
 

test/unit/test_metal_graph.py

@@ -27,18 +27,18 @@ class TestMetalGraph(unittest.TestCase):
     c.src = (MagicMock(op=Ops.PROGRAM),) + tuple(bufs)
     return c
 
-  def test_supports_exec_item_normal_offset(self):
-    assert self.MetalGraph.supports_exec_item([self.dev], self.call(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF))) is True
+  def test_supports_uop_normal_offset(self):
+    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF))) is True
 
-  def test_supports_exec_item_overflow_offset(self):
-    assert self.MetalGraph.supports_exec_item([self.dev], self.call(self.metal_buf(0), self.metal_buf(0x100000000))) is False
+  def test_supports_uop_overflow_offset(self):
+    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(0x100000000))) is False
 
-  def test_supports_exec_item_nonmetal_buf(self):
+  def test_supports_uop_nonmetal_buf(self):
     # non-BUFFER_VIEW ops should not be checked for offset
     buf = MagicMock()
     buf.op = Ops.BUFFER
     buf.device = Device.DEFAULT
-    self.MetalGraph.supports_exec_item([self.dev], self.call(buf))
+    self.MetalGraph.supports_uop([self.dev], self.call(buf))
 
 if __name__ == "__main__":
   unittest.main()

tinygrad/codegen/__init__.py

@@ -2,8 +2,9 @@ from typing import cast
 from dataclasses import replace
 import itertools
 from tinygrad.helpers import DISABLE_FAST_IDIV, DEVECTORIZE, TRANSCENDENTAL, SPEC, DEBUG, VIZ, IMAGE, NOOPT, EMULATED_DTYPES, NOLOCALS, USE_TC
-from tinygrad.helpers import ALLOW_TF32, TracingKey, Context, Target, panic
-from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype, Ops, UPat, track_rewrites, KernelInfo, ProgramInfo, pyrender
+from tinygrad.helpers import ALLOW_TF32, TracingKey, Context, panic
+from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype, Ops, UPat, track_rewrites, KernelInfo, ProgramInfo
+from tinygrad.uop.render import pyrender
 from tinygrad.uop.spec import type_verify, program_spec, kernel_spec
 from tinygrad.renderer import Renderer, Estimates
 from tinygrad.dtype import dtypes
@@ -15,20 +16,19 @@ from tinygrad.uop.decompositions import get_late_rewrite_patterns, get_transcend
 from tinygrad.codegen.late.expander import expander, pm_pre_expander, pm_group_for_reduce
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
   ReduceContext, correct_load_store, pm_render, pm_add_loads, pm_make_images
+from tinygrad.codegen.late.gater import pm_move_gates_from_index
 from tinygrad.codegen.opt.postrange import apply_opts
 from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse
 from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen, pm_mops, pm_syntactic_sugar, pm_store_ranges
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
 
-def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True, beam:int=0) -> UOp:
-  if ren is None: ren = Renderer(Target())
-
-  if VIZ: graph_rewrite(sink, PatternMatcher([]), name="View Base AST")
-  if DEBUG >= 5: print(pyrender(sink))
-  if SPEC: type_verify(sink, kernel_spec)
+def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
+  if VIZ: graph_rewrite(ast, PatternMatcher([]), name="View Base AST")
+  if DEBUG >= 5: print(pyrender(ast))
+  if SPEC: type_verify(ast, kernel_spec)
 
   # preprocess
-  sink = graph_rewrite(sink, pm_mops+pm_syntactic_sugar+pm_store_ranges, ctx=itertools.count(1000), name="early movement ops", bottom_up=True)
+  sink = graph_rewrite(ast, pm_mops+pm_syntactic_sugar+pm_store_ranges, ctx=itertools.count(1000), name="early movement ops", bottom_up=True)
 
   # first we optimize
   if optimize:
@@ -45,7 +45,7 @@ def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True, b
     sink = graph_rewrite(sink, pm_flatten_range+pm_simplify_ranges, ctx={}, name="simplify ranges")
 
     # do postrange optimization, BEAM or hand_coded_optimizations
-    sink = apply_opts(sink, ren, beam=beam)
+    sink = apply_opts(sink, ren, beam=ast.arg.beam)
 
   # ** expander (expand_rewrite) **
   sink = graph_rewrite(sink, sym+pm_move_where_on_load, name="postopt symbolic")
@@ -77,8 +77,13 @@ def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True, b
   else: pm_devectorize = sym+load_store_folding+correct_load_store+load_store_indexing
   if DEVECTORIZE >= 0: sink = graph_rewrite(sink, pm_devectorize, ctx=ren, name="devectorize")
 
-  # lower the index dtype to a concrete int
+  # lower the index dtype to a concrete int. this needs to happen while gates are still present
   sink = graph_rewrite(sink, pm_lower_index_dtype+load_store_indexing+gep_pushing, name="lower all index dtypes")
+
+  # move the gates from index onto the loads and stores
+  sink = graph_rewrite(sink, pm_move_gates_from_index, name="move gates from index")
+
+  # a final symbolic
   sink = graph_rewrite(sink, symbolic, name="post index symbolic")
 
   # optional pre matcher
@@ -107,8 +112,8 @@ def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True, b
 pm_linearize_cleanups = PatternMatcher([
   # if statements are not allowed in the graph
   (UPat((Ops.IF, Ops.ENDIF)), lambda: panic(RuntimeError, "if not allowed in graph")),
-  # gated INDEX becomes IF-STORE-ENDIF. this is the only use of IF-ENDIF
-  (UPat(Ops.STORE, name="u", src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat(name="gate", dtype=dtypes.bool))).or_casted(), UPat())),
+  # gated STORE becomes IF-STORE-ENDIF. this is the only use of IF-ENDIF
+  (UPat(Ops.STORE, name="u", src=(UPat(Ops.INDEX).or_casted(), UPat(), UPat(name="gate", dtype=dtypes.bool))),
    lambda u, gate: (u, [mif:=UOp(Ops.IF, src=(gate, u.src[0])), u, UOp(Ops.ENDIF, src=(mif,))]))
 ])
 
@@ -170,7 +175,7 @@ def do_to_program(ast:UOp, renderer:Renderer) -> UOp:
   if ast.op is Ops.PROGRAM: prg = ast
   elif ast.op is Ops.SINK:
     assert isinstance(ast.arg, KernelInfo), "requires KernelInfo on arg to to_program"
-    full_sink = full_rewrite_to_sink(ast, renderer, optimize=ast.tag is None, beam=ast.arg.beam)
+    full_sink = full_rewrite_to_sink(ast, renderer, optimize=ast.tag is None)
     prg = UOp(Ops.PROGRAM, src=(full_sink, UOp(Ops.DEVICE, arg=renderer.target.device)), arg=ProgramInfo.from_sink(full_sink))
   else: raise RuntimeError(f"can't call to_program on {ast.op}")
   if not isinstance(prg.arg, ProgramInfo): prg = prg.replace(arg=ProgramInfo.from_sink(prg.src[0]))
@@ -180,8 +185,6 @@ def do_to_program(ast:UOp, renderer:Renderer) -> UOp:
 
 to_program_cache: dict[tuple, UOp] = {}
 def to_program(ast:UOp, renderer:Renderer) -> UOp:
-  if ast.op is Ops.PROGRAM and len(ast.src) >= 5 and ast.src[4].op is Ops.BINARY:
-    return ast if isinstance(ast.arg, ProgramInfo) else ast.replace(arg=ProgramInfo.from_sink(ast.src[0]))
   config = (NOOPT, DEVECTORIZE, EMULATED_DTYPES, NOLOCALS, USE_TC, IMAGE, DISABLE_FAST_IDIV, TRANSCENDENTAL, ALLOW_TF32)
   key = (ast.key, type(renderer), renderer.target, *[x.value for x in config])
   if (prg:=to_program_cache.get(key)) is None: to_program_cache[key] = prg = do_to_program(ast, renderer)

tinygrad/codegen/late/devectorizer.py

@@ -37,6 +37,7 @@ def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
   return drop_stmt
 
 def simplify_valid_load(buf:UOp, start_idx:UOp, valid:UOp) -> UOp|None:
+  start_idx = start_idx.simplify()  # if you don't do this, uop_given_valid may simplify things and this might inf loop
   idx = uop_given_valid(valid, start_idx)
   if not isinstance(buf.dtype, ImageDType): return None if idx is start_idx else buf.index(idx.valid(valid), ptr=True)
 
@@ -116,22 +117,22 @@ def fold_expanded_index(midx:UOp):
   post_cat = UOp(Ops.PTRCAT, buf.ptrdtype.base.ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace).vec(global_offset), tuple(ret))
   return post_cat.gep(tuple(cast(list[int], idxs)))
 
-def cat_after_store(cat:UOp, data:UOp, sto:UOp):
+def cat_after_store(cat:UOp, data:UOp):
   # TODO: this is written in many places
   offset = 0
   ret: list[UOp] = []
   for s in cat.src:
-    ret.append(s.store(data.gep(tuple(range(offset, offset+s.dtype.count))), *sto.src[2:]))
+    ret.append(s.store(data.gep(tuple(range(offset, offset+s.dtype.count)))))
     offset += s.dtype.count
   return UOp.group(*ret)
 
-def gep_on_store(gep:UOp, st:UOp, sto:UOp):
+def gep_on_store(gep:UOp, st:UOp):
   # NOTE: we need to invert the gep here, but it may be an expanding gep
   # fake argsort. TODO: handle duplicates
   a = {}
   for i,x in enumerate(gep.arg): a[x] = i
   new_arg = tuple(x[1] for x in sorted(a.items()))
-  return gep.src[0].store(st.gep(new_arg), *sto.src[2:])
+  return gep.src[0].store(st.gep(new_arg))
 
 load_store_folding = PatternMatcher([
   (UPat(Ops.INDEX, src=(UPat(Ops.STACK, src=UPat(GroupOp.Defines).or_after(name="buf")), UPat.var("vec"))), expand_index),
@@ -140,12 +141,12 @@ load_store_folding = PatternMatcher([
   (UPat(Ops.LOAD, src=(UPat(Ops.GEP, name="gep"),), name="ld", allow_any_len=True),
    lambda gep, ld: ld.replace(dtype=ld.dtype.scalar().vec(gep.dtype.count), src=(gep.src[0],)+ld.src[1:]).gep(gep.arg)),
   # GEP on data of STORE
-  (UPat(Ops.STORE, src=(UPat(Ops.GEP, name="gep"), UPat.var("st")), name="sto"), gep_on_store),
+  (UPat(Ops.STORE, src=(UPat(Ops.GEP, name="gep"), UPat.var("st"))), gep_on_store),
   # put PTRCAT after LOAD
   (UPat(Ops.LOAD, src=(UPat(Ops.PTRCAT, name="cat"),), name="ld", allow_any_len=True),
    lambda cat,ld: UOp(Ops.VCAT, cat.dtype.base.vec(cat.dtype.vcount), tuple(ld.replace(dtype=x.dtype.base, src=(x,)+ld.src[1:]) for x in cat.src))),
   # put PTRCAT after STORE
-  (UPat(Ops.STORE, src=(UPat(Ops.PTRCAT, name="cat"), UPat(name="data")), name="sto"), cat_after_store),
+  (UPat(Ops.STORE, src=(UPat(Ops.PTRCAT, name="cat"), UPat(name="data"))), cat_after_store),
 ])
 
 # *** correct load/store ***
@@ -187,7 +188,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
       if global_offset+fold_length > sz: continue
       lidx = buf.index((offset + global_offset).valid(mask), ptr=True)
       if fold_length > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(fold_length).ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace))
-      if ls.op is Ops.STORE: ret.append(ls.replace(src=(lidx,ls.src[1].gep(tuple(range(global_offset, global_offset+fold_length))))+ls.src[2:]))
+      if ls.op is Ops.STORE: ret.append(ls.replace(src=(lidx,ls.src[1].gep(tuple(range(global_offset, global_offset+fold_length))))))
       else: ret.append(ls.replace(src=(lidx,)+ls.src[1:], dtype=ls.dtype.scalar().vec(fold_length)))
       global_offset += fold_length
       break
@@ -197,8 +198,9 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
   return UOp(Ops.VCAT, ls.dtype, tuple(ret)) if ls.op is Ops.LOAD else UOp.group(*ret)
 
 def get_image_idx(idx:UOp, width:int):
-  oidx = UOp(Ops.STACK, dtypes.weakint.vec(2), (((x:=idx.src[1].get_idx()) // 4) % width, (x // (4*width))))
-  return idx.replace(src=(idx.src[0], oidx.valid(idx.src[1].get_valid())))
+  x, valid = idx.src[1].get_idx(), idx.src[1].get_valid()
+  idx_x, idx_y = (x // 4) % width, x // (4*width)
+  return idx.replace(src=(idx.src[0], UOp.vectorize(idx_x, idx_y).valid(valid)))
 
 def image_fixup(ls:UOp):
   # normal image load or store, with the CAST from expand_index
@@ -280,18 +282,13 @@ pm_render = PatternMatcher([
   (UPat(Ops.GEP, name='gep'), lambda gep: UOp(Ops.STACK, gep.dtype, tuple(gep.src[0].gep(x) for x in gep.arg)) if len(gep.arg) > 1 else None),
   (UPat(Ops.GEP, name='gep'), lambda gep: gep.src[0] if gep.src[0].dtype.vcount == 1 and gep.arg == (0,) else None),
   (UPat(Ops.STACK, src=(UPat(name='x'),)), lambda x: x),
-  # give any loads that are masked an alt value
-  (UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat())).or_casted(),), allow_any_len=True, name="x"),
-    lambda x: x.replace(src=(x.src[0], x.const_like(0))+x.src[1:])
-      if len(x.src) == 1 or x.src[1].op in (Ops.CUSTOM, Ops.STORE, Ops.BARRIER) else None),
   # Where after gated load becomes alt value
   # NOTE: if a is CAST and a.src[0].dtype == l.dtype, use a.src[0] to avoid roundtrip cast (e.g. uint->float->uint)
-  (UPat.var("c").where(UPat(Ops.LOAD, src=(UPat().index(UPat(), UPat.var("c")).or_casted(),), allow_any_len=True, name="l").or_casted(),
-    UPat.var("a")), lambda c,l,a: l.replace(src=(l.src[0], a.src[0] if a.op is Ops.CAST and a.src[0].dtype == l.dtype else a.cast(l.dtype))+
-                                            l.src[2:]).cast(a.dtype)),
-  (UPat.var("c").where(UPat.var("a"), UPat(Ops.LOAD, src=(UPat().index(UPat(), UPat.var("c", dtype=dtypes.bool).logical_not()).or_casted(),),
-    allow_any_len=True, name="l").or_casted()), lambda c,l,a: l.replace(src=(l.src[0], a.src[0] if a.op is Ops.CAST and a.src[0].dtype == l.dtype
-                                                                             else a.cast(l.dtype))+l.src[2:]).cast(a.dtype)),
+  (UPat.var("gate").where(UPat(Ops.LOAD, src=(UPat(), UPat(), UPat.var("gate")), name="l").or_casted(), UPat.var("a")), lambda gate,l,a:
+   l.replace(src=(l.src[0], a.src[0] if a.op is Ops.CAST and a.src[0].dtype == l.dtype else a.cast(l.dtype), l.src[2])).cast(a.dtype)),
+  (UPat.var("gate").where(UPat.var("a"), UPat(Ops.LOAD,
+    src=(UPat(), UPat(), UPat.var("gate", dtype=dtypes.bool).logical_not()), name="l").or_casted()), lambda gate,l,a:
+   l.replace(src=(l.src[0], a.src[0] if a.op is Ops.CAST and a.src[0].dtype == l.dtype else a.cast(l.dtype), l.src[2])).cast(a.dtype)),
 ])
 
 # *** Ops.REDUCE -> Ops.DEFINE_ACC ***
@@ -317,12 +314,12 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
     topo = inp.toposort()
     ended_ranges = flatten([x.ended_ranges for x in topo if x.op is Ops.END])
     input_ranges = tuple([x for x in topo if x.op is Ops.RANGE and x not in reduce_range and x not in ended_ranges])
-    identity = red.const(red.dtype, identity_element(red.arg, red.dtype.scalar()))
+    identity = red.const(red.dtype, identity_element(red.arg[0], red.dtype.scalar()))
     acc = UOp.placeholder((1,), red.dtype, ctx.acc_num, AddrSpace.REG)
     acc_init = acc.after(*input_ranges).index(UOp.const(dtypes.weakint, 0)).store(identity)
     lst = [acc.after(acc_init, *reduce_range).index(UOp.const(dtypes.weakint, 0))] + lst  # put acc as the first element
     ctx.acc_num += 1
-  ret = functools.reduce(lambda x,y: x.alu(red.arg, y), lst)
+  ret = functools.reduce(lambda x,y: x.alu(red.arg[0], y), lst)
   if len(reduce_range) == 0: return ret
   end = acc.index(UOp.const(dtypes.weakint, 0)).store(ret).end(*reduce_range).rtag("mergeable")
   return acc.after(end).index(UOp.const(dtypes.weakint, 0))
@@ -358,12 +355,16 @@ pm_reduce = PatternMatcher([
 
 # add loads
 
+def add_load(idx:UOp):
+  if isinstance(idx.dtype, PtrDType): return None
+  assert isinstance(idx.src[0].dtype, PtrDType), f"param is not PtrDType {idx.src[0].dtype}"
+  return idx.replace(dtype=idx.src[0].dtype).load(dtype=idx.dtype.base)
+
 pm_add_loads = PatternMatcher([
   # add loads to non ptr index
-  (UPat(Ops.INDEX, name="idx"), lambda idx: None if isinstance(idx.dtype, PtrDType) else
-    idx.replace(dtype=idx.src[0].dtype).load(dtype=idx.dtype.base)),
+  (UPat(Ops.INDEX, name="idx"), add_load),
   # remove loads from stores
-  (UPat(Ops.STORE, src=(UPat(Ops.LOAD), UPat(name="val")), name="s"), lambda s,val: s.replace(src=(s.src[0].src[0], val))),
+  (UPat(Ops.STORE, src=(UPat(Ops.LOAD),), allow_any_len=True, name="s"), lambda s: s.replace(src=(s.src[0].src[0],)+s.src[1:])),
 ])
 
 # make images

tinygrad/codegen/late/expander.py

@@ -45,7 +45,7 @@ def do_expand(root:UOp):
     else:
       # non-UNROLL input
       if root.op in range_start and i >= range_start[root.op]:
-        # for any range args of STORE/REDUCE, pass them through
+        # for any range args of REDUCE/WMMA/END/etc., pass them through
         new_srcs.append(src)
       elif root.op is Ops.INDEX and i >= 1 and not isinstance(root.dtype, PtrDType):
         new_srcs.append(src)

tinygrad/codegen/late/gater.py

@@ -0,0 +1,13 @@
+# this transforms Invalid into gated load/stores
+
+from tinygrad.uop.ops import PatternMatcher, UPat
+from tinygrad.dtype import Invalid, dtypes
+
+pm_move_gates_from_index = PatternMatcher([
+  (UPat.var("buf").index(UPat.var("gate").where(UPat.var("idx"), UPat(arg=Invalid))).or_casted(name="cast").load(name="l"),
+   lambda buf,gate,idx,cast,l: buf.index(idx, ptr=True).cast(cast.dtype).load(l.const_like(0), gate, dtype=l.dtype)),
+  (UPat.var("buf").index(UPat.var("gate").where(UPat.var("idx"), UPat(arg=Invalid))).or_casted(name="cast").store(UPat.var("data")),
+   lambda buf,gate,idx,cast,data: buf.index(idx, ptr=True).cast(cast.dtype).store(data, gate)),
+  # remove hanging weakint casts
+  (UPat.var("buf").index(UPat.var("idx", dtypes.ints).cast()), lambda buf,idx: buf.index(idx, ptr=True)),
+])

tinygrad/codegen/opt/postrange.py

@@ -219,7 +219,7 @@ class Scheduler:
   def _apply_tc_opt(self, use_tensor_cores:int, axis:int, tc_select:int, opt_level:int) -> None|list[UOp]:
     if not (reduceops := self.reduceops): raise KernelOptError("no reduce ops for TensorCore")
     reduceop = reduceops[0]
-    if use_tensor_cores and reduceop.arg is Ops.ADD:
+    if use_tensor_cores and reduceop.arg[0] is Ops.ADD:
       mul = reduceop.src[0] if reduceop.src[0].op is not Ops.CAST else reduceop.src[0].src[0]
       if mul.op is not Ops.MUL: return None
       in0, in1 = mul.src
@@ -305,7 +305,7 @@ class Scheduler:
 
             # preserve extra reduces
             reduce_ranges = [x for x in UOp.sink(*reduceop.src[1:]).toposort() if x.op is Ops.RANGE and x.arg[0] not in tc_reduce_axes]
-            if len(reduce_ranges): tc_uop = UOp(Ops.REDUCE, tc_uop.dtype, (tc_uop,)+tuple(reduce_ranges), Ops.ADD)
+            if len(reduce_ranges): tc_uop = UOp(Ops.REDUCE, tc_uop.dtype, (tc_uop,)+tuple(reduce_ranges), (Ops.ADD, ()))
             self.ast = self.ast.substitute({reduceop: tc_uop})
           self.tensor_core = tc
           return axes
@@ -317,7 +317,7 @@ class Scheduler:
   @property
   def reduceop(self) -> UOp|None:
     if not (red := self.reduceops): return None
-    return UOp(Ops.REDUCE_AXIS, red[0].dtype, red[0].src, (red[0].arg, ()))
+    return UOp(Ops.REDUCE, red[0].dtype, red[0].src, red[0].arg)
   @property
   def bufs(self) -> list[UOp]: return [x for x in self.ast.toposort() if x.op is Ops.INDEX][::-1]
   @property

tinygrad/codegen/opt/search.py

@@ -1,12 +1,13 @@
-import functools, math, time, multiprocessing, traceback, signal, atexit
+import math, time, multiprocessing, traceback, signal, atexit
 from dataclasses import replace
-from tinygrad.uop.ops import sym_infer, AxisType, pyrender, UOp, Ops
-from tinygrad.device import Device, Buffer, Compiler
+from tinygrad.uop.ops import sym_infer, AxisType, UOp
+from tinygrad.uop.render import pyrender
+from tinygrad.device import Device, Buffer
 from tinygrad.helpers import prod, flatten, DEBUG, CACHELEVEL, diskcache_get, diskcache_put, getenv, Context, colored, time_to_str, unwrap
 from tinygrad.helpers import IGNORE_BEAM_CACHE
 from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 from tinygrad.tensor import Tensor
-from tinygrad.engine.realize import CompiledRunner
+from tinygrad.engine.realize import get_runtime
 from tinygrad.codegen import to_program
 from tinygrad.codegen.opt.postrange import Scheduler
 
@@ -34,25 +35,24 @@ def get_test_global_size(global_size, max_global_size, var_vals):
         break
   return test_global_size, input_size / prod(test_global_size)
 
-def _time_program(prg:UOp, lib:bytes, var_vals:dict[str, int], rawbufs:list[Buffer], early_stop:float|None=None,
+def _time_program(prg:UOp, var_vals:dict[str, int], rawbufs:list[Buffer], early_stop:float|None=None,
                   allow_test_size:int=True, max_global_size:int|None=65536, clear_l2=False, cnt=3, name="test", dev_timeout=False) -> list[float]:
   timeout = int(early_stop * 1e3) if dev_timeout and early_stop is not None and early_stop < math.inf else None
   factor = 1
-  info = prg.arg
   if allow_test_size and max_global_size is not None:
-    global_size, factor = get_test_global_size(info.global_size, max_global_size, var_vals)
-    prg = prg.replace(arg=replace(info, global_size=tuple(global_size)))
-  if len(prg.src) <= 4 or prg.src[4].op is not Ops.BINARY: prg = prg.replace(src=prg.src + (UOp(Ops.BINARY, arg=lib),))
-  try: car = CompiledRunner(prg, prg.src[1].arg)
+    global_size, factor = get_test_global_size(prg.arg.global_size, max_global_size, var_vals)
+    prg = prg.replace(arg=replace(prg.arg, global_size=tuple(global_size)))
+  try: rt = get_runtime(prg.src[1].arg, prg)
   except AssertionError: return [math.inf] * cnt
+  global_size, local_size = prg.arg.launch_dims(var_vals)
+  bufs = [rawbufs[i]._buf for i in prg.arg.globals]
   tms = []
-  input_bufs = [rawbufs[i] for i in car.p.globals]
   for _ in range(cnt):
     if clear_l2:
       if hasattr(dev:=Device[prg.src[1].arg], 'invalidate_caches'): dev.invalidate_caches()
       else:
         with Context(DEBUG=0, BEAM=0, CAPTURING=0, TRACK_MATCH_STATS=0): Tensor.ones(1024,1024).contiguous().realize(do_update_stats=False)
-    tms.append(unwrap(car(input_bufs, var_vals, wait=True, timeout=timeout))*factor)
+    tms.append(unwrap(rt(*bufs, global_size=global_size, local_size=local_size, vals=prg.arg.vals(var_vals), wait=True, timeout=timeout))*factor)
     if early_stop is not None and early_stop < min(tms): break
   return tms
 
@@ -61,22 +61,21 @@ def timeout_handler(signum, frame):
   if DEBUG >= 2: print("*** BEAM COMPILE TIMEOUT")
   raise TimeoutException()
 
-def _try_compile(x:tuple[int,Scheduler], compiler:Compiler) -> tuple[int, tuple[UOp, bytes, float]|None]:
+def _try_compile(x:tuple[int,Scheduler]) -> tuple[int, tuple[UOp, float]|None]:
   if hasattr(signal, "alarm"):
     signal.signal(getattr(signal, 'SIGALRM'), timeout_handler)
     # set timeout
     signal.alarm(getenv("BEAM_TIMEOUT_SEC", 10))
   ret = None
   try:
+    st = time.perf_counter()
     prg = to_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].ren)
+    et = time.perf_counter() - st
     uops = prg.src[2].src
     if len(uops) >= (uops_max:=getenv("BEAM_UOPS_MAX", 3000)) > 0:
       if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too many uops. {len(uops)=}, {uops_max=}")
       raise RuntimeError("too many uops")
-    st = time.perf_counter()
-    prog = prg.src[4].arg if len(prg.src) > 4 and prg.src[4].op is Ops.BINARY else compiler.compile(prg.src[3].arg)
-    et = time.perf_counter() - st
-    ret = (prg, prog, et)
+    ret = (prg, et)
   except RuntimeError:
     if DEBUG >= 4: traceback.print_exc()
   except Exception as e:
@@ -150,12 +149,11 @@ def beam_search(s:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True
     while not exiting:
       candidates: list[Scheduler] = flatten([get_kernel_actions(si, include_0=False).values() for si,_ in beam])
       timed: list[tuple[Scheduler, float]] = []
-      _compile_fn = functools.partial(_try_compile, compiler=dev.compiler)
       least_compute_ops = math.inf
-      for i,proc in (map(_compile_fn, enumerate(candidates)) if beam_pool is None else beam_pool.imap_unordered(_compile_fn, enumerate(candidates))):
+      for i, proc in ((map if beam_pool is None else beam_pool.imap_unordered)(_try_compile, enumerate(candidates))):
         if proc is None: continue
-        prg, lib, compile_et = proc
-        if lib in seen_libs: continue
+        prg, compile_et = proc
+        if (lib:=prg.src[4].arg) in seen_libs: continue
         # filter out kernels that use 1000x more compute than the smallest
         estimates = prg.src[0].arg.estimates
         least_compute_ops = min(this_compute_ops:=sym_infer(estimates.ops if estimates is not None else 0, var_vals), least_compute_ops)
@@ -163,7 +161,7 @@ def beam_search(s:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True
           if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too much compute. {this_compute_ops} when least is {least_compute_ops}")
           continue
         seen_libs.add(lib)
-        try: tms = _time_program(prg, lib, var_vals, rawbufs, early_stop=beam[0][1]*3 if len(beam) else 1.0,
+        try: tms = _time_program(prg, var_vals, rawbufs, early_stop=beam[0][1]*3 if len(beam) else 1.0,
                                  allow_test_size=allow_test_size, clear_l2=hasattr(dev, 'invalidate_caches'),
                                  dev_timeout=getenv("BEAM_DEV_TIMEOUT", 1))
         except Exception as e:

tinygrad/codegen/simplify.py

@@ -14,10 +14,11 @@ def flatten_range(r:UOp) -> UOp|None:
 
 pm_flatten_range = PatternMatcher([
   # real ranges only
-  (UPat((Ops.REDUCE, Ops.STORE, Ops.END), name="r"), flatten_range),
+  (UPat((Ops.REDUCE, Ops.END), name="r"), flatten_range),
 ])
 
-def count_divmod(x:UOp) -> int: return sum(u.op in {Ops.IDIV, Ops.MOD} for u in x.backward_slice)
+# index/range arithmetic uses FLOORDIV/FLOORMOD prior to late rewrite
+def count_divmod(x:UOp) -> int: return sum(u.op in {Ops.FLOORDIV, Ops.FLOORMOD} for u in x.backward_slice)
 def simplify_merge_adjacent(u:UOp) -> UOp|None:
   reduce_ranges = [x.ranges for x in u.backward_slice_with_self if x.op is Ops.REDUCE]
   # on END we only want to merge adjacent ranges, on REDUCE we want to try all combinations
@@ -75,14 +76,14 @@ pm_split_ranges = PatternMatcher([
 def no_range(u:UOp) -> bool: return not any(x.op is Ops.RANGE for x in u.backward_slice_with_self)
 
 def reduce_unparented(red:UOp) -> UOp|None:
-  if red.arg not in {Ops.ADD, Ops.MAX, Ops.MUL}: return None
+  if red.arg[0] not in {Ops.ADD, Ops.MAX, Ops.MUL}: return None
   assert all(x.op is Ops.RANGE for x in red.src[1:]), "some reduce srcs aren't ranges"
   reduce_parented, reduce_unparented = partition(red.src[1:], lambda x: x in red.src[0].ranges)
   if len(reduce_unparented) == 0: return None
   ret = red.replace(src=(red.src[0],)+tuple(reduce_parented)) if len(reduce_parented) or red.dtype != red.src[0].dtype else red.src[0]
-  if red.arg is Ops.ADD:
+  if red.arg[0] is Ops.ADD:
     for r in reduce_unparented: ret = ret * r.src[0].cast(ret.dtype.scalar()).broadcast(ret.dtype.count)
-  if red.arg is Ops.MUL:
+  if red.arg[0] is Ops.MUL:
     for r in reduce_unparented: ret = ret ** r.src[0].cast(ret.dtype.scalar()).broadcast(ret.dtype.count)
   return ret
 
@@ -145,12 +146,12 @@ def reduce_load_collapse(red:UOp, u:UOp) -> UOp|None: return reduce_collapse(red
 
 # remove REDUCE without loads (generic arange opt / indexing).
 pm_reduce_simplify = pm_reduce_unparented + PatternMatcher([
-  (UPat(Ops.REDUCE, src=(UPat.var("u"),), allow_any_len=True, arg=Ops.ADD, name="red"), reduce_collapse),
+  (UPat(Ops.REDUCE, src=(UPat.var("u"),), allow_any_len=True, arg=(Ops.ADD, ()), name="red"), reduce_collapse),
 ])
 # remove REDUCE on load, comes from indexing a tensor with another tensor
 def no_load(u:UOp) -> bool: return not any(x.op is Ops.INDEX for x in u.backward_slice_with_self)
 pm_load_collapse = PatternMatcher([
-  (UPat(Ops.REDUCE, arg=Ops.ADD, src=(UPat.var("u"), UPat()), name="red"), reduce_load_collapse),
+  (UPat(Ops.REDUCE, arg=(Ops.ADD, ()), src=(UPat.var("u"), UPat()), name="red"), reduce_load_collapse),
   # we want to make sure we dont do math on a loaded index since that can cause overflow, this undoes the rule in pm_reduce_load_collapse
   ((UPat.var("x", dtypes.weakint)+UPat.var("y"))<UPat.var("c"), lambda x,y,c: x < c-y if no_load(y) and no_load(c) and not no_load(x) else None),
 ])

tinygrad/device.py

@@ -5,7 +5,7 @@ from typing import Any, Generic, TypeVar, Iterator, Generator, TYPE_CHECKING
 import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re, atexit, pickle, decimal
 from tinygrad.helpers import BENCHMARKS, CI, OSX, LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored
 from tinygrad.helpers import Context, CCACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, suppress_finalizing
-from tinygrad.helpers import select_by_name, select_first_inited, DEV, EMULATED_DTYPES, IMAGE, FLOAT16, TracingKey, size_to_str, Target, VIZ
+from tinygrad.helpers import select_by_name, select_first_inited, DEV, EMULATED_DTYPES, IMAGE, FLOAT16, TracingKey, size_to_str, Target
 from tinygrad.helpers import pluralize
 from tinygrad.dtype import DType, PtrDType, dtypes, _to_np_dtype
 if TYPE_CHECKING: from tinygrad.renderer import Renderer
@@ -371,9 +371,8 @@ if PROFILE:
     with open(fn:=temp("profile.pkl", append_user=True), "wb") as f: pickle.dump(cpu_events+Compiled.profile_events+Buffer.profile_events, f)
 
     PROFILE.value = 0
-    if VIZ > 0:
-      from tinygrad.uop.ops import launch_viz
-      launch_viz("PROFILE", fn)
+    from tinygrad.uop.ops import launch_viz
+    launch_viz("PROFILE", fn)
 
 def enumerate_devices_str() -> Generator[str, None, None]:
   from tinygrad import Tensor, Device

tinygrad/engine/jit.py

@@ -1,12 +1,12 @@
 from typing import TypeVar, Generic, Callable, Any
 import functools, collections
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import flatten, merge_dicts, DEBUG, Context, BEAM, getenv, colored, JIT, JIT_BATCH_SIZE, dedup, pluralize, VIZ
+from tinygrad.helpers import flatten, merge_dicts, DEBUG, Context, BEAM, getenv, JIT, JIT_BATCH_SIZE, dedup, pluralize, VIZ
 from tinygrad.device import Buffer, Compiled, Device, MultiBuffer
 from tinygrad.dtype import DType, dtypes
 from tinygrad.uop.ops import UOp, PatternMatcher, Variable, sym_infer, Ops, buffers, track_rewrites, graph_rewrite
-from tinygrad.engine.realize import capturing, CompiledRunner, Runner, Estimates, compile_linear, run_linear, get_runner, graph_cache, estimate_uop
-from tinygrad.engine.realize import unwrap_multi, resolve_params
+from tinygrad.engine.realize import capturing, Estimates, compile_linear, run_linear, graph_cache, estimate_uop, get_runtime
+from tinygrad.engine.realize import unwrap_multi, resolve_params, get_call_arg_uops, get_call_outs_ins
 from tinygrad.schedule.memory import memory_plan_rewrite, _collect_bufs
 from tinygrad.nn.state import get_parameters
 from tinygrad.schedule.rangeify import mop_cleanup
@@ -25,7 +25,7 @@ def prune_linear(linear:UOp, needed:set[UOp]) -> tuple[UOp, UOp]:
 def create_graph_call(batch:list[UOp]) -> UOp:
   # all external inputs are PARAMs
   input_list = dedup(u for si in batch for b in si.src[1:] for u in b.toposort() if u.op is Ops.PARAM)
-  cf = UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(UOp(Ops.LINEAR, src=tuple(batch)), *input_list), arg="graph")
+  cf = UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(UOp(Ops.LINEAR, src=tuple(batch)),), arg="graph")
   return cf.call(*input_list, metadata=tuple(m for si in batch for m in si.arg.metadata))
 
 def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
@@ -48,8 +48,8 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
     devs = dedup([Device[x] for b in si.src[1:] if b.op is not Ops.BIND for x in (b.device if isinstance(b.device, tuple) else (b.device,))])
     graph_t = graph_class(devs[0]) if devs[0].graph is not None else None
 
-    can_graph = graph_t is not None and graph_t.supports_exec_item(devs, si)
-    can_extend = can_graph and graph_t is not None and (not current_batch_devs or graph_t.supports_exec_item(current_batch_devs, si)) \
+    can_graph = graph_t is not None and graph_t.supports_uop(devs, si)
+    can_extend = can_graph and graph_t is not None and (not current_batch_devs or graph_t.supports_uop(current_batch_devs, si)) \
       and (max_batch_size == 0 or len(current_batch) < max_batch_size)
     if not can_extend and current_batch: flush_batch()
 
@@ -59,14 +59,6 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
   if current_batch: flush_batch()
   return linear.replace(src=tuple(new_src))
 
-def _call_outs_ins(call:UOp) -> tuple[set[int], set[int]]:
-  non_bind = [s for s in call.src[1:] if s.op is not Ops.BIND]
-  ast = call.src[0]
-  if ast.op is Ops.PROGRAM: return set(ast.arg.outs), set(ast.arg.ins)
-  if ast.op in (Ops.COPY, Ops.BUFFER_VIEW): return {0}, {1}
-  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return {0}, set(range(1, len(non_bind)))
-  return set(), set()
-
 def _copy_input(u:UOp) -> UOp:
   run_linear(UOp(Ops.LINEAR, src=(u.copy_to_device(u.device).call(new:=UOp.new_buffer(u.device, u.arg, u.dtype), u, metadata=()),)))
   return new
@@ -95,17 +87,17 @@ def _check_no_non_tensor_return(ret):
 
 def graph_class(dev): return dev.graph.func if isinstance(dev.graph, functools.partial) else dev.graph
 
-class GraphRunner(Runner):
+class GraphRunner:
   def __init__(self, linear:UOp, input_uops:tuple[UOp, ...]=()):
     self.linear = linear.src[0]
     self.calls: list[tuple[int, UOp, list[Buffer], dict[str, int]]] = []
-    self.progs: list[CompiledRunner|None] = []
+    self.runtimes: list[Any|None] = []
     self.uop_replace: list[list[tuple[int, int]]] = []
     for call in self.linear.src:
-      replace = [(p, b.arg) for p, b in enumerate(b for b in call.src[1:] if b.op is not Ops.BIND) if b.op is Ops.PARAM]
+      replace = [(p, b.arg) for p, b in enumerate(get_call_arg_uops(call)) if b.op is Ops.PARAM]
       for dev_idx, (bufs, device_vars) in enumerate(unwrap_multi(call, resolve_params(call, input_uops))):
         self.calls.append((dev_idx, call.src[0], [b.ensure_allocated() for b in bufs], device_vars))
-        self.progs.append(get_runner(bufs[0].device, call.src[0]) if call.src[0].op is Ops.PROGRAM else None)
+        self.runtimes.append(get_runtime(bufs[0].device, call.src[0]) if call.src[0].op is Ops.PROGRAM else None)
         self.uop_replace.append(replace)
 
     self.var_vals_replace:dict[int, list[tuple[int, int]]] = {}
@@ -114,20 +106,20 @@ class GraphRunner(Runner):
 
     def is_sym_dim(dim) -> bool: return not all(isinstance(d, (int, float)) for d in dim)
 
-    crs = [(j, p, self.calls[j][3]) for j,p in enumerate(self.progs) if isinstance(p, CompiledRunner)]
-    self.vars = sorted({v.expr for _,p,dv in crs for v in p.p.vars if v.expr not in dv | p.p.runtimevars})
-    self.symbolic_dims = dedup(tuple(d) for _,p,_ in crs for d in (p.p.local_size, p.p.global_size) if d and is_sym_dim(d))
+    crs = [(j, self.calls[j][1].arg, self.calls[j][3]) for j in range(len(self.calls)) if self.calls[j][1].op is Ops.PROGRAM]
+    self.vars = sorted({v.expr for _,p,dv in crs for v in p.vars if v.expr not in dv | p.runtimevars})
+    self.symbolic_dims = dedup(tuple(d) for _,p,_ in crs for d in (p.local_size, p.global_size) if d and is_sym_dim(d))
 
     def find_symbolic_dim(dim): return self.symbolic_dims.index(tuple(dim)) if dim is not None and tuple(dim) in self.symbolic_dims else None
 
     for j,p,dv in crs:
-      if (replace:=[(i, self.vars.index(v.expr)) for i, v in enumerate(p.p.vars) if v.expr not in dv | p.p.runtimevars]):
+      if (replace:=[(i, self.vars.index(v.expr)) for i, v in enumerate(p.vars) if v.expr not in dv | p.runtimevars]):
         self.var_vals_replace[j] = replace
-      global_dim_idx, local_dim_idx = find_symbolic_dim(p.p.global_size), find_symbolic_dim(p.p.local_size)
+      global_dim_idx, local_dim_idx = find_symbolic_dim(p.global_size), find_symbolic_dim(p.local_size)
       if global_dim_idx is not None or local_dim_idx is not None:
         self.launch_dims_replace[j] = (global_dim_idx, local_dim_idx)
-        assert p.p.local_size is not None
-        self.launch_dims_base[j] = (tuple(p.p.global_size), tuple(p.p.local_size))
+        assert p.local_size is not None
+        self.launch_dims_base[j] = (tuple(p.global_size), tuple(p.local_size))
 
     estimates = sum((estimate_uop(call) for call in self.linear.src), Estimates())
 
@@ -135,7 +127,9 @@ class GraphRunner(Runner):
     self.w_dependency_map: dict[int, list[tuple[int, int, Any]]] = collections.defaultdict(list)
     self.r_dependency_map: dict[int, list[tuple[int, int, Any]]] = collections.defaultdict(list)
 
-    super().__init__(colored(f"<batched {len(self.calls)}>", "cyan"), self.calls[0][2][0].device.split(":")[0], estimates.simplify())
+    self.device, self.estimates = self.calls[0][2][0].device.split(":")[0], estimates.simplify()
+
+  def __call__(self, input_uops:tuple[UOp, ...], var_vals:dict[str, int], wait=False) -> float|None: raise NotImplementedError("override this")
 
   def updated_vars(self, var_vals: dict[str, int]):
     vals = [var_vals[v] for v in self.vars]
@@ -168,17 +162,17 @@ class GraphRunner(Runner):
 
   @staticmethod
   def _all_devs(batch_devs:list[Compiled], new_call:UOp) -> list[Compiled]:
-    return dedup(batch_devs + [Device[x] for b in new_call.src[1:] if b.op is not Ops.BIND
+    return dedup(batch_devs + [Device[x] for b in get_call_arg_uops(new_call)
                  for x in (b.device if isinstance(b.device, tuple) else (b.device,))])
 
   @staticmethod
-  def supports_exec_item(batch_devs:list[Compiled], new_call:UOp) -> bool:
+  def supports_uop(batch_devs:list[Compiled], new_call:UOp) -> bool:
     return new_call.src[0].op is Ops.PROGRAM and len(GraphRunner._all_devs(batch_devs, new_call)) == 1
 
 # a marker for your graph supporting multiple devices of the same type
 class MultiGraphRunner(GraphRunner):
   @staticmethod
-  def supports_exec_item(batch_devs:list[Compiled], new_call:UOp) -> bool:
+  def supports_uop(batch_devs:list[Compiled], new_call:UOp) -> bool:
     # Devices must be the same type
     return new_call.src[0].op in (Ops.PROGRAM, Ops.COPY) and len(dedup([type(d) for d in GraphRunner._all_devs(batch_devs, new_call)])) == 1
 
@@ -197,9 +191,9 @@ class CapturedJit(Generic[ReturnType]):
     out: set[UOp] = set()
     for call in self.linear.toposort():
       if call.op is not Ops.CALL: continue
-      non_bind = [s for s in call.src[1:] if s.op is not Ops.BIND]
-      outs, ins = _call_outs_ins(call)
-      out |= {non_bind[k] for k in outs - ins if non_bind[k].op in (Ops.BUFFER, Ops.BUFFER_VIEW)}
+      arg_uops = get_call_arg_uops(call)
+      outs, ins = get_call_outs_ins(call)
+      out |= {arg_uops[k] for k in set(outs) - set(ins) if arg_uops[k].op in (Ops.BUFFER, Ops.BUFFER_VIEW)}
     return out
 
   def __call__(self, input_uops:list[UOp], var_vals:dict[str, int]) -> ReturnType:

tinygrad/engine/realize.py

@@ -1,9 +1,9 @@
-from typing import cast, Iterator
+from __future__ import annotations
+from typing import cast, Iterator, Any
 import time, random, itertools, math, contextlib, weakref
 from dataclasses import dataclass, replace, field
-from tinygrad.helpers import colored, DEBUG, GlobalCounters, ansilen, NOOPT, all_int, Metadata, TRACEMETA, TracingKey
-from tinygrad.helpers import BEAM, DEVECTORIZE, size_to_str, time_to_str, VALIDATE_WITH_CPU, cpu_profile, PROFILE, ProfilePointEvent, cpu_events
-from tinygrad.helpers import prod, EMULATED_DTYPES, flatten
+from tinygrad.helpers import colored, DEBUG, GlobalCounters, ansilen, all_int, TRACEMETA, prod, flatten
+from tinygrad.helpers import BEAM, size_to_str, time_to_str, VALIDATE_WITH_CPU, PROFILE, ProfilePointEvent, cpu_events
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, sym_infer, buffers, graph_rewrite, ProgramInfo
 from tinygrad.device import Device, Buffer, MultiBuffer
@@ -11,50 +11,74 @@ from tinygrad.renderer import Estimates
 from tinygrad.codegen import to_program
 from tinygrad.codegen.opt.postrange import bufs_from_ast
 
+# **************** Helpers ****************
+
+def get_call_arg_uops(call:UOp) -> tuple[UOp, ...]: return tuple(s for s in call.src[1:] if s.op is not Ops.BIND)
+
+def get_call_outs_ins(call:UOp) -> tuple[tuple[int, ...], tuple[int, ...]]:
+  ast = call.src[0]
+  if ast.op is Ops.PROGRAM: return tuple(ast.arg.outs), tuple(ast.arg.ins)
+  if ast.op in (Ops.COPY, Ops.BUFFER_VIEW): return (0,), (1,)
+  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return (0,), tuple(range(1, len(get_call_arg_uops(call))))
+  return (), ()
+
+def get_call_name(call:UOp, bufs:list[Buffer], var_vals:dict[str, int]|None=None) -> str:
+  def _uop_sz_to_str(uop:UOp) -> str: return size_to_str(sym_infer(prod(uop.shape) * uop.dtype.itemsize, var_vals or {}))
+
+  ast, arg_uops = call.src[0], get_call_arg_uops(call)
+  if ast.op is Ops.PROGRAM: return ast.arg.name
+  if ast.op is Ops.BUFFER_VIEW: return colored(f"view {_uop_sz_to_str(arg_uops[0]):>10} @ {ast.arg[1] * arg_uops[1].dtype.itemsize:<10d}", "yellow")
+  if ast.op is Ops.COPY: return colored(f"copy {_uop_sz_to_str(arg_uops[0]):>10}, {bufs[0].device[:7]:>7s} <- {bufs[1].device[:7]:7s}", "yellow")
+  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return colored(f"enc/dec {_uop_sz_to_str(arg_uops[0])}", "yellow")
+  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph": return colored(f"batched {len(ast.src[0].src)}", "cyan")
+  raise NotImplementedError("get_call_name is not implemented")
+
 # **************** Stat ****************
 
 def estimate_uop(call:UOp) -> Estimates:
-  if call.src[0].op is Ops.SINK: call = pm_compile.rewrite(call)
-
   ast = call.src[0]
   if ast.op is Ops.PROGRAM: return ast.src[0].arg.estimates or Estimates()
   if ast.op is Ops.COPY or (ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec"):
     nbytes = prod(call.src[1].shape) * call.src[1].dtype.itemsize
     return Estimates(lds=nbytes, mem=nbytes)
-  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph":
-    return runner.estimates if (runner:=graph_cache.get(ast)) is not None else Estimates()
+  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph": return get_graph_runtime(ast).estimates
   return Estimates()
 
-def update_stats(display_name:str, device:str, estimates:Estimates, var_vals:dict[str, int], et:float|None, buf_count:int,
-                 jit=False, metadata:tuple[Metadata, ...]=(), first_run=False):
+first_run_cache:set[bytes] = set()
+@contextlib.contextmanager
+def track_stats(ctx:ExecContext, call:UOp, device:str, bufs:list[Buffer], var_vals:dict[str, int]):
+  if PROFILE:
+    outputs, inputs = get_call_outs_ins(call)
+    cpu_events.append(ProfilePointEvent(device, "exec", len(cpu_events), {"metadata": call.arg.metadata, "var_vals": var_vals,
+      "bufs": [b.trace_num for b in bufs], "name": get_call_name(call, bufs, var_vals), "outputs": outputs, "inputs": inputs}))
+  et: list[float|None] = [None]
+  if DEBUG >= 2: st = time.perf_counter()
+  yield et
+  if not ctx.do_update_stats: return
+
+  if DEBUG >= 2 and et[0] is None:
+    Device[device].synchronize()
+    et[0] = time.perf_counter() - st
+
+  estimates = estimate_uop(call)
   GlobalCounters.kernel_count += 1
   GlobalCounters.global_ops += (op_est:=sym_infer(estimates.ops, var_vals))
   GlobalCounters.global_mem += (mem_est:=sym_infer(estimates.mem, var_vals))
-  if et is not None: GlobalCounters.time_sum_s += et
+  if et[0] is not None: GlobalCounters.time_sum_s += et[0]
   if DEBUG >= 2:
+    display_name = get_call_name(call, bufs, var_vals)
     lds_est = sym_infer(estimates.lds, var_vals)
-    header_color = 'magenta' if jit else ('green' if first_run else None)
-    ptm = colored(time_to_str(et, w=9), "yellow" if et > 0.01 else None) if et is not None else ""
-    flops, membw, ldsbw = op_est/(et or 1e-20), mem_est/(et or 1e-20), lds_est/(et or 1e-20)
+    header_color = 'magenta' if ctx.jit else ('green' if call.src[0].key not in first_run_cache else None)
+    ptm = colored(time_to_str(et[0], w=9), "yellow" if et[0] > 0.01 else None) if et[0] is not None else ""
+    flops, membw, ldsbw = op_est/(et[0] or 1e-20), mem_est/(et[0] or 1e-20), lds_est/(et[0] or 1e-20)
     flops_str = f"{flops*1e-9:7.0f} GFLOPS" if flops < 1e14 else colored(f"{flops*1e-12:7.0f} TFLOPS", 'green')
     mem_str = f"{membw*1e-9:4.0f}|{ldsbw*1e-9:<6.0f} GB/s" if membw < 1e13 and ldsbw < 1e15 else \
       colored(f"{membw*1e-12:4.0f}|{ldsbw*1e-12:<6.0f} TB/s", 'green')
     print(f"{colored(f'*** {device[:7]:7s} {GlobalCounters.kernel_count:4d}', header_color)}"+
-      f" {display_name+' '*(46-ansilen(display_name))} arg {buf_count:2d} mem {GlobalCounters.mem_used/1e9:6.2f} GB"+
-      ("" if et is None else f" tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({flops_str} {mem_str})")+
-      f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in metadata] if metadata else ''}")
-
-# **************** Runners ****************
-
-class Runner:
-  def __init__(self, display_name:str, device:str, estimates=Estimates()):
-    self.first_run, self.display_name, self.device, self.estimates = True, display_name, device, estimates
-  @property
-  def dev(self): return Device[self.device]
-  def exec(self, rawbufs:list[Buffer], var_vals:dict[str, int]|None=None) -> float|None:
-    return self(rawbufs, {} if var_vals is None else var_vals)
-  def __call__(self, rawbufs:list[Buffer], var_vals:dict[str, int], wait=False) -> float|None:
-    raise NotImplementedError("override this")
+      f" {display_name+' '*(46-ansilen(display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:6.2f} GB"+
+      ("" if et[0] is None else f" tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({flops_str} {mem_str})")+
+      f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in call.arg.metadata] if call.arg.metadata else ''}")
+    first_run_cache.add(call.src[0].key)
 
 local_size_cache: dict[bytes, tuple[int, ...]] = {}
 def optimize_local_size(call:UOp, prg:UOp) -> UOp|None:
@@ -78,43 +102,24 @@ def optimize_local_size(call:UOp, prg:UOp) -> UOp|None:
   new_global = tuple(g//l if g%l == 0 else g/l for g,l in zip(prg.arg.global_size, local_size))
   return call.replace(src=(prg.replace(arg=replace(prg.arg, global_size=new_global, local_size=local_size)), *call.src[1:]))
 
-class CompiledRunner(Runner):
-  def __init__(self, prg:UOp, device:str):
-    info: ProgramInfo = prg.arg
-    sink = prg.src[0]
-    if DEBUG >= 3 and sink.arg.applied_opts: print(sink.arg.applied_opts)
-    if DEBUG >= 4: print(prg.src[3].arg)
-    if len(prg.src) <= 4 or prg.src[4].op is not Ops.BINARY:
-      with cpu_profile(TracingKey(f"compile {info.name}", (info.function_name,)), "TINY"):
-        lib = Device[device].compiler.compile_cached(prg.src[3].arg)
-      prg = prg.replace(src=prg.src + (UOp(Ops.BINARY, arg=lib),))
-    self.prg:UOp = prg
-    self.p:ProgramInfo = info
-    if DEBUG >= 7: Device[device].compiler.disassemble(prg.src[4].arg)
-    self._prg = Device[device].runtime(info.function_name, prg.src[4].arg, *info.aux, runtimevars=info.runtimevars)
-    super().__init__(info.name, device, sink.arg.estimates or Estimates())
+# **************** runtime cache ****************
 
-  def __call__(self, rawbufs:list[Buffer], var_vals:dict[str, int]|None=None, wait=False, timeout:int|None=None) -> float|None:
-    if var_vals is None: var_vals = {}
-    global_size, local_size = self.p.launch_dims(var_vals)
-    return self._prg(*[x._buf for x in rawbufs], global_size=tuple(global_size), local_size=tuple(local_size) if local_size else None,
-                     vals=tuple(var_vals[k.expr] if k.expr not in self.p.runtimevars else None for k in self.p.vars), wait=wait, timeout=timeout)
+runtime_cache: dict[tuple[bytes, str], Any] = {}
+def get_runtime(device:str, ast:UOp):
+  assert ast.op is Ops.PROGRAM and isinstance(ast.arg, ProgramInfo), "get_runtime should only be called with a PROGRAM ast"
+  if (runtime:=runtime_cache.get(key:=(ast.key, device))) is None:
+    if DEBUG >= 3 and ast.src[0].arg.applied_opts: print(ast.src[0].arg.applied_opts)
+    if DEBUG >= 4: print(ast.src[3].arg)
+    if DEBUG >= 7: Device[device].compiler.disassemble(ast.src[4].arg)
+    runtime = runtime_cache[key] = Device[device].runtime(ast.arg.function_name, ast.src[4].arg, *ast.arg.aux, runtimevars=ast.arg.runtimevars)
+  return runtime
 
-# **************** method cache ****************
-
-method_cache: dict[tuple[str, type, bytes, tuple, bool], CompiledRunner] = {}
-def get_runner(device:str, ast:UOp) -> CompiledRunner:
-  # TODO: this should be all context relevant to rendering
-  context = (NOOPT.value, DEVECTORIZE.value, EMULATED_DTYPES.value)
-  ckey = (device, type(Device[device].compiler), ast.key, context, False)
-  if cret:=method_cache.get(ckey): return cret
-  bkey = (device.split(":")[0], type(Device[device].compiler), ast.key, context, True)
-  if bret:=method_cache.get(bkey):
-    method_cache[ckey] = ret = CompiledRunner(bret.prg, device)
-  else:
-    prg = to_program(ast, Device[device].renderer)
-    method_cache[ckey] = method_cache[bkey] = ret = CompiledRunner(prg, device)
-  return ret
+graph_cache:weakref.WeakKeyDictionary[UOp, Any] = weakref.WeakKeyDictionary()
+def get_graph_runtime(ast:UOp, input_uops:tuple[UOp, ...]|None=None):
+  assert ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph", "get_graph_runtime should only be called with a graph ast"
+  if (runtime:=graph_cache.get(ast)) is None and input_uops is not None:
+    graph_cache[ast] = runtime = Device[ast.device if isinstance(ast.device, str) else ast.device[0]].graph(ast, input_uops=input_uops)
+  return runtime
 
 # **************** run linear ****************
 
@@ -130,21 +135,7 @@ class ExecContext:
 def _resolve(b:UOp, inputs:tuple[UOp, ...]) -> UOp:
   if b.op in (Ops.BUFFER_VIEW, Ops.MSELECT) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg], *b.src[1:]))
   return inputs[b.arg] if b.op is Ops.PARAM else b
-def resolve_params(call:UOp, inputs:tuple[UOp, ...]) -> list[UOp]: return [_resolve(b, inputs) for b in call.src[1:] if b.op is not Ops.BIND]
-
-@contextlib.contextmanager
-def track_stats(ctx:ExecContext, call:UOp, device:str, display_name:str, bufs:list[Buffer], var_vals:dict[str, int],
-                outputs=(0,), inputs=(1,), first_run=False):
-  if PROFILE: cpu_events.append(ProfilePointEvent(device, "exec", len(cpu_events), {"metadata": call.arg.metadata, "var_vals": var_vals,
-                                                  "bufs": [b.trace_num for b in bufs], "name": display_name, "outputs": outputs, "inputs": inputs}))
-  timing: list[float|None] = [None]
-  if DEBUG >= 2: st = time.perf_counter()
-  yield timing
-  if not ctx.do_update_stats: return
-  if DEBUG >= 2 and timing[0] is None:
-    Device[device].synchronize()
-    timing[0] = time.perf_counter() - st
-  update_stats(display_name, device, estimate_uop(call), var_vals, timing[0], len(bufs), jit=ctx.jit, metadata=call.arg.metadata, first_run=first_run)
+def resolve_params(call:UOp, inputs:tuple[UOp, ...]) -> list[UOp]: return [_resolve(b, inputs) for b in get_call_arg_uops(call)]
 
 def unwrap_multi(call:UOp, resolved:list[UOp]) -> Iterator[tuple[list[Buffer], dict[str, int]]]:
   bufs = [b.buffer for b in resolved]
@@ -157,16 +148,13 @@ def exec_view(ctx:ExecContext, call, ast):
   resolved = resolve_params(call, ctx.input_uops)
   bufs = [cast(Buffer, b.buffer) for b in resolved]
   bv = bufs[1].view(resolved[0].arg, ast.dtype, ast.arg[1]*bufs[1].dtype.itemsize)
-  with track_stats(ctx, call, bv.device, colored(f"view {bv.nbytes:8d} @ {bv.offset:<10d}", "yellow"), [bv, bufs[1]], ctx.var_vals):
-    buffers[resolved[0]] = bv
+  with track_stats(ctx, call, bv.device, [bv, bufs[1]], ctx.var_vals): buffers[resolved[0]] = bv
 
 def exec_copy(ctx:ExecContext, call, ast):
   for bufs, device_vars in unwrap_multi(call, resolve_params(call, ctx.input_uops)):
     dest, src = bufs[0].ensure_allocated(), bufs[1].ensure_allocated()
-    xfer = hasattr(dest.allocator,'_transfer') and dest.allocator.supports_transfer and dest.device.split(":")[0] == src.device.split(":")[0]
-    name = colored(f"{'xfer' if xfer else 'copy'} {size_to_str(bufs[0].nbytes):>10}, {dest.device[:7]:>7s} <- {src.device[:7]:7s}", "yellow")
-    with track_stats(ctx, call, dest.device, name, [dest, src], ctx.var_vals):
-      if xfer:
+    with track_stats(ctx, call, dest.device, [dest, src], ctx.var_vals):
+      if hasattr(dest.allocator,'_transfer') and dest.allocator.supports_transfer and dest.device.split(":")[0] == src.device.split(":")[0]:
         dest.allocator._transfer(dest._buf, src._buf, dest.nbytes, src_dev=src.allocator.dev, dest_dev=dest.allocator.dev) # type:ignore[attr-defined]
       elif src.device.startswith("DISK") and getattr(src.allocator.dev, 'fd', None) is not None \
            and hasattr(dest.allocator, 'copy_from_disk') and src.nbytes >= 4096 and dest.allocator.supports_copy_from_disk:
@@ -178,35 +166,31 @@ def exec_copy(ctx:ExecContext, call, ast):
 def exec_kernel(ctx:ExecContext, call, ast):
   for bufs, device_vars in unwrap_multi(call, resolve_params(call, ctx.input_uops)):
     var_vals = {**ctx.var_vals, **device_vars}
-    prg = get_runner(bufs[0].device, ast)
-    prg_bufs = [bufs[i].ensure_allocated() for i in prg.p.globals]
-
-    with track_stats(ctx, call, prg.device, prg.display_name, prg_bufs, var_vals,
-                     outputs=tuple(prg.p.outs), inputs=tuple(prg.p.ins), first_run=prg.first_run) as timing:
-      timing[0] = prg(prg_bufs, var_vals, wait=DEBUG >= 2)
-      prg.first_run = False
+    prg_bufs = [bufs[i].ensure_allocated() for i in ast.arg.globals]
+    rt = get_runtime(device:=bufs[0].device, ast)
+    global_size, local_size = ast.arg.launch_dims(var_vals)
+    with track_stats(ctx, call, device, prg_bufs, var_vals) as tm:
+      tm[0] = rt(*[b._buf for b in prg_bufs], global_size=global_size, local_size=local_size, vals=ast.arg.vals(var_vals), wait=DEBUG>=2)
 
 def exec_validate(ctx:ExecContext, call, ast):
   import numpy as np
   for bufs, device_vars in unwrap_multi(call, resolve_params(call, ctx.input_uops)):
-    cpu_bufs, dev_bufs = bufs[:len(bufs)//2], bufs[len(bufs)//2:]
-    cpu_prg = get_runner("CPU", ast.src[0])
-    cpu_prg([cpu_bufs[i].ensure_allocated() for i in cpu_prg.p.globals], {**ctx.var_vals, **device_vars}, wait=False)
-    for i in cpu_prg.p.outs: np.testing.assert_allclose(dev_bufs[i].ensure_allocated().numpy(), cpu_bufs[i].numpy(), rtol=1e-3, atol=1e-3)
+    bufs, dev_bufs = bufs[:len(bufs)//2], bufs[len(bufs)//2:]
+    var_vals = {**ctx.var_vals, **device_vars}
+    cpu_rt = get_runtime("CPU", prg:=to_program(ast.src[0], Device["CPU"].renderer))
+    global_size, local_size = prg.arg.launch_dims(var_vals)
+    cpu_rt(*[bufs[i].ensure_allocated()._buf for i in prg.arg.globals], global_size=global_size, local_size=local_size, vals=prg.arg.vals(var_vals))
+    for i in prg.arg.outs: np.testing.assert_allclose(dev_bufs[i].ensure_allocated().numpy(), bufs[i].numpy(), rtol=1e-3, atol=1e-3)
 
 def exec_encdec(ctx:ExecContext, call, ast):
   bufs = [cast(Buffer, b.buffer).ensure_allocated() for b in resolve_params(call, ctx.input_uops)]
   shape, pos_var = tuple(s.arg for s in ast.src if s.op is Ops.CONST), ast.variables()[0].expr
-  with track_stats(ctx, call, bufs[0].device, colored(f"enc/dec {size_to_str(bufs[0].nbytes)}", "yellow"), bufs, ctx.var_vals):
+  with track_stats(ctx, call, bufs[0].device, bufs, ctx.var_vals):
     bufs[0].allocator._encode_decode(bufs[0]._buf, bufs[1]._buf, bufs[2]._buf, [x._buf for x in bufs[3:]], shape, ctx.var_vals[pos_var])
 
-graph_cache:weakref.WeakKeyDictionary[UOp, Runner] = weakref.WeakKeyDictionary()
-def exec_graph(ctx:ExecContext, call, cf):
-  bufs = flatten([b.bufs if isinstance(b, MultiBuffer) else [b] for b in (u.buffer for u in resolve_params(call, ctx.input_uops))])
-  if (runner:=graph_cache.get(cf)) is None:
-    graph_cache[cf] = runner = Device[cf.device if isinstance(cf.device, str) else cf.device[0]].graph(cf, input_uops=ctx.input_uops)
-  with track_stats(ctx, call, runner.device, runner.display_name, bufs, ctx.var_vals) as t:
-    t[0] = runner(bufs, ctx.var_vals, wait=DEBUG >= 2, input_uops=ctx.input_uops) # type: ignore[call-arg]
+def exec_graph(ctx:ExecContext, call, ast):
+  rt = get_graph_runtime(ast, ctx.input_uops)
+  with track_stats(ctx, call, rt.device, [], ctx.var_vals) as t: t[0] = rt(ctx.input_uops, ctx.var_vals, wait=DEBUG>=2) # type: ignore[call-arg]
 
 # flatten LINEAR-in-LINEAR: any nested LINEAR child gets inlined into its parent's src
 pm_flatten_linear = PatternMatcher([
@@ -215,7 +199,7 @@ pm_flatten_linear = PatternMatcher([
 ])
 
 def _validate(call:UOp, sink:UOp) -> UOp:
-  params = tuple(p for p in call.src[1:] if p.op is not Ops.BIND)
+  params = get_call_arg_uops(call)
   shadows = tuple(UOp.new_buffer(("CPU",)*len(p.device) if isinstance(p.device, tuple) else "CPU", prod(p.max_shape), p.dtype.base) for p in params)
   copies = tuple(p.copy_to_device(s.device).call(s, p) for s, p in zip(shadows, params))
   return UOp(Ops.LINEAR, src=copies + (call, UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(sink,), arg="validate").call(*shadows, *params)))
@@ -241,7 +225,7 @@ pm_exec = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.COPY, name="ast"),), name="call", allow_any_len=True), exec_copy),
   (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="ast"),), name="call", allow_any_len=True), exec_kernel),
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="encdec", name="ast"),), name="call", allow_any_len=True), exec_encdec),
-  (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="graph", name="cf"),), name="call", allow_any_len=True), exec_graph),
+  (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="graph", name="ast"),), name="call", allow_any_len=True), exec_graph),
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="validate", name="ast"),), name="call", allow_any_len=True), exec_validate),
 ])
 

tinygrad/gradient.py

@@ -8,7 +8,7 @@ def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
   def broadcast_to_input(x): return x.reshape(x.shape+(1,)*(len(ret.src[0].shape)-len(x.shape))).expand(ret.src[0].shape)
   if op == Ops.ADD: return (broadcast_to_input(ctx),)
   if op == Ops.MAX:
-    assert ret.op is Ops.REDUCE_AXIS, "only works on REDUCE_AXIS"
+    assert ret.op is Ops.REDUCE, "only works on REDUCE"
     mask = ret.src[0].eq(broadcast_to_input(ret)).cast(ctx.dtype)
     count = mask._rop(Ops.ADD, ret.arg[1])
     return ((mask/broadcast_to_input(count)) * broadcast_to_input(ctx),)
@@ -61,7 +61,7 @@ pm_gradient = PatternMatcher([
     ((x>y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)), (x<y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)))),
   (UPat(Ops.MUL, name="ret"), lambda ctx, ret: (ret.src[1]*ctx, ret.src[0]*ctx)),
   (UPat(Ops.WHERE, name="ret"), lambda ctx, ret: (None, ret.src[0].where(ctx, ctx.const_like(0)), ret.src[0].where(ctx.const_like(0), ctx))),
-  (UPat(Ops.REDUCE_AXIS, name="ret"), lambda ctx, ret: reduce_gradient(ctx, ret, ret.arg[0])),
+  (UPat(Ops.REDUCE, name="ret"), lambda ctx, ret: reduce_gradient(ctx, ret, ret.arg[0])),
   (UPat(Ops.CONTIGUOUS), lambda ctx: (ctx,)),
   (UPat(Ops.CONTIGUOUS_BACKWARD), lambda ctx: (ctx.contiguous(),)),
   (UPat(Ops.RESHAPE, name="ret"), lambda ctx, ret: (ctx.reshape(ret.src[0].shape), None)),
@@ -75,8 +75,8 @@ pm_gradient = PatternMatcher([
   (UPat(Ops.COPY, name="ret"), lambda ctx, ret: (ctx.copy_to_device(ret.src[0].device), None)),
   (UPat(Ops.MULTI, name="ret"), lambda ctx, ret: ctx.shard(ret.device, ret.axis).src),
   (UPat(Ops.TUPLE), lambda ctx: ctx.src),
-  # NOTE: this is only correct when the KERNEL has a single output
-  (UPat(Ops.AFTER), lambda ctx: (ctx, ctx)),
+  (UPat(Ops.AFTER, src=(UPat.var("d"), UPat(Ops.CALL, name="k"))), lambda ctx, d, k:
+    (ctx, UOp.maketuple(*(ctx if i == k.src.index(d)-1 else UOp(Ops.NOOP) for i in range(len(k.src)-1))))),
   # there's no gradient for bitcast
   (UPat(Ops.BITCAST), lambda: (None,)),
 ])
@@ -113,7 +113,11 @@ def compute_gradient(root:UOp, root_grad:UOp, targets:set[UOp]) -> dict[UOp, UOp
     assert len(lgrads) == len(t0.src), f"got {len(lgrads)} gradient, expected {len(t0.src)}"
     for k,v in zip(t0.src, lgrads):
       if v is None: continue
-      if k in grads and grads[k].op is not Ops.NOOP: grads[k] = grads[k] + v
+      if k in grads and grads[k].op is not Ops.NOOP:
+        if v.op is Ops.TUPLE and grads[k].op is Ops.TUPLE:
+          grads[k] = UOp.maketuple(*(p + n if (p.op is not Ops.NOOP and n.op is not Ops.NOOP) else
+                                     n if p.op is Ops.NOOP else p for p, n in zip(grads[k].src, v.src)))
+        else: grads[k] = grads[k] + v
       else: grads[k] = v
       if len(forward_metadata:=all_metadata.get(t0, ())):
         backward_metadata = tuple(dataclasses.replace(x, backward=True) for x in forward_metadata)

tinygrad/helpers.py

@@ -64,6 +64,9 @@ def next_power2(x): return 1 if x == 0 else 1 << (x - 1).bit_length()
 # cstyle div and mod
 def cdiv(x:int, y:int) -> int: return abs(x)//abs(y)*(1,-1)[x*y<0] if y != 0 else 0
 def cmod(x:int, y:int) -> int: return x-cdiv(x,y)*y
+# python floor div and mod
+def floordiv(x:int, y:int) -> int: return x//y if y != 0 else 0
+def floormod(x:int, y:int) -> int: return x-floordiv(x,y)*y
 def lo32(x:Any) -> Any: return x & 0xFFFFFFFF # Any is sint
 def hi32(x:Any) -> Any: return x >> 32 # Any is sint
 def data64(data:Any) -> tuple[Any, Any]: return (data >> 32, data & 0xFFFFFFFF) # Any is sint
@@ -238,7 +241,7 @@ SPLIT_REDUCEOP, NO_MEMORY_PLANNER, LRU = ContextVar("SPLIT_REDUCEOP", 1), Contex
 RING, ALL2ALL, ALLREDUCE_CAST = ContextVar("RING", 1), ContextVar("ALL2ALL", 0), ContextVar("ALLREDUCE_CAST", 1)
 CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
 VALIDATE_WITH_CPU, DISABLE_FAST_IDIV = ContextVar("VALIDATE_WITH_CPU", 0), ContextVar("DISABLE_FAST_IDIV", 0)
-CORRECT_DIVMOD_FOLDING, FUSE_OPTIM = ContextVar("CORRECT_DIVMOD_FOLDING", 0), ContextVar("FUSE_OPTIM", 0)
+FUSE_OPTIM = ContextVar("FUSE_OPTIM", 0)
 ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE = ContextVar("ALLOW_DEVICE_USAGE", 1), ContextVar("MAX_BUFFER_SIZE", 0)
 MAX_KERNEL_BUFFERS = ContextVar("MAX_KERNEL_BUFFERS", 0)
 EMULATED_DTYPES = ContextVar("EMULATED_DTYPES", "")
@@ -441,24 +444,28 @@ def _ensure_downloads_dir() -> pathlib.Path:
     return downloads_dir
   return pathlib.Path(cache_dir) / "downloads"
 
-def fetch(url:str, name:pathlib.Path|str|None=None, subdir:str|None=None, gunzip:bool=False,
-          allow_caching=not getenv("DISABLE_HTTP_CACHE"), headers:dict[str, str]={}) -> pathlib.Path:
+def fetch(url:str, name:pathlib.Path|str|None=None, subdir:str|None=None, gunzip:bool=False, allow_caching=not getenv("DISABLE_HTTP_CACHE"),
+          headers:dict[str, str]={}, sha256:str|None=None) -> pathlib.Path:
   import urllib.request
   if url.startswith(("/", ".")): return pathlib.Path(url)
   if name is not None and (isinstance(name, pathlib.Path) or '/' in name): fp = pathlib.Path(name)
   else:
     hh = "_"+hashlib.md5(("\n".join(f"{k.strip()}:{v.strip()}" for k,v in sorted(headers.items()))).encode("utf-8")).hexdigest() if headers else ""
     fp = _ensure_downloads_dir() / (subdir or "") / ((name or hashlib.md5(url.encode('utf-8')).hexdigest()) + hh + (".gunzip" if gunzip else ""))
-  if not fp.is_file() or not allow_caching:
+  if not fp.is_file() or not allow_caching or (sha256 and hashlib.sha256(fp.read_bytes()).hexdigest() != sha256):
     (_dir := fp.parent).mkdir(parents=True, exist_ok=True)
     with urllib.request.urlopen(urllib.request.Request(url, headers={"User-Agent": "tinygrad 0.12.0", **headers}), timeout=10) as r:
       assert r.status in {200, 206}, r.status
       length = int(r.headers.get('content-length', 0)) if not gunzip else None
       readfile = gzip.GzipFile(fileobj=r) if gunzip else r
       progress_bar:tqdm = tqdm(total=length, unit='B', unit_scale=True, desc=f"{url}", disable=CI)
+      h = hashlib.sha256() if sha256 else None
       with tempfile.NamedTemporaryFile(dir=_dir, delete=False) as f:
-        while chunk := readfile.read(16384): progress_bar.update(f.write(chunk))
+        while chunk := readfile.read(16384):
+          if h: h.update(chunk)
+          progress_bar.update(f.write(chunk))
         f.close()
+        if h and (actual_sha256:=h.hexdigest()) != sha256: raise RuntimeError(f"fetch sha mismatch, expected {sha256} but got {actual_sha256}")
         pathlib.Path(f.name).rename(fp)
       progress_bar.update(close=True)
       if length and (file_size:=os.stat(fp).st_size) < length: raise RuntimeError(f"fetch size incomplete, {file_size} < {length}")

tinygrad/mixin/__init__.py

@@ -6,7 +6,7 @@ from tinygrad.mixin.movement import MovementMixin
 from tinygrad.mixin.reduce import ReduceMixin
 from tinygrad.uop import Ops
 from tinygrad.uop.ops import _broadcast_shape, resolve, smax, smin, identity_element
-from tinygrad.dtype import ConstType, DTypeLike, Invalid, InvalidType, PtrDType, PyConst, dtypes, least_upper_dtype, sum_acc_dtype, to_dtype
+from tinygrad.dtype import ConstType, DType, DTypeLike, Invalid, InvalidType, PtrDType, PyConst, dtypes, least_upper_dtype, sum_acc_dtype, to_dtype
 from tinygrad.helpers import all_int, argfix, ceildiv, flatten, flat_to_grouped, make_tuple, prod, resolve_pool_pads, round_up
 
 if TYPE_CHECKING:
@@ -191,6 +191,36 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
     """
     return self._tri(self.shape[-2], self.shape[-1], diagonal+1, self.device).where(self.zeros_like(), self)
 
+  # ***** random *****
+
+  @staticmethod
+  def _threefry_random_bits(key, counts0, counts1):
+    x = (counts1.cast(dtypes.uint64) << 32) | counts0.cast(dtypes.uint64)
+    x = x.threefry((key[1]._broadcast_to(x.shape).cast(dtypes.uint64) << 32) | key[0]._broadcast_to(x.shape).cast(dtypes.uint64))
+    return (x & 0xffffffff).cast(dtypes.uint32).cat(((x >> 32) & 0xffffffff).cast(dtypes.uint32))
+
+  @classmethod
+  def random_bits(cls, key:Self, counter:Self, num:int) -> Self:
+    low, high = counter[0:1], counter[1:2]
+    bits = []
+    for i in range(0, num, dtypes.uint32.max):
+      chunk_num = min(num - i, dtypes.uint32.max)
+      c_low = low + (i & 0xffffffff)
+      c_high = high + (i >> 32) + (c_low < low).cast(dtypes.uint32)
+      new_key = cls._threefry_random_bits(key, c_low, c_high)
+      counts0 = cls.arange(ceildiv(chunk_num, 2), device=key.device, dtype=dtypes.uint32)
+      counts1 = counts0 + ceildiv(chunk_num, 2)
+      bits.append(cls._threefry_random_bits(new_key, counts0, counts1)[:chunk_num])
+    return bits[0].cat(*bits[1:])
+
+  @staticmethod
+  def _bits_to_rand(bits, shape:tuple[int, ...], dtype:DType):
+    _, nmant = dtypes.finfo(dtype)
+    uint_dtype = {1: dtypes.uint8, 2: dtypes.uint16, 4: dtypes.uint32, 8: dtypes.uint64}[dtype.itemsize]
+    uint_bits = bits.bitcast(uint_dtype)
+    float_one_bits = uint_bits.ones_like(dtype=dtype).bitcast(uint_dtype)
+    return uint_bits.rshift(dtype.bitsize - nmant).bitwise_or(float_one_bits).bitcast(dtype)[:prod(shape)].sub(1).reshape(shape)
+
   def _pad_constant(self, pX, value:float) -> Self:
     # shrink first for negative pads, then pad with only non-negative values
     pX = tuple((0, 0) if p is None else p for p in pX)

tinygrad/mixin/dtype.py

@@ -7,6 +7,8 @@ class DTypeMixin:
 
   def cast(self, dtype:DType) -> Self: raise NotImplementedError
 
+  def bitcast(self, dtype:DType) -> Self: raise NotImplementedError
+
   def element_size(self) -> int:
     """
     Returns the size in bytes of an individual element in the tensor.

tinygrad/mixin/elementwise.py

@@ -181,7 +181,7 @@ class ElementwiseMixin(DTypeMixin, CreationMixin):
     return self._binop(Ops.IDIV, x, reverse)
 
   def mod(self, x: Self | ConstType, reverse: bool = False) -> Self:
-    return self._binop(Ops.MOD, x, reverse)
+    return self._binop(Ops.FLOORMOD, x, reverse)
 
   def div(self, x: Self | ConstType, reverse: bool = False) -> Self:
     lhs, rhs = self._broadcasted(x, reverse)
@@ -206,7 +206,7 @@ class ElementwiseMixin(DTypeMixin, CreationMixin):
     return self.div(x)
 
   def __floordiv__(self, x: Self | ConstType) -> Self:
-    return self.idiv(x)  # TODO: idiv is trunc div, not floordiv
+    return self._binop(Ops.FLOORDIV, x, False)
 
   def __mod__(self, x: Self | ConstType) -> Self:
     return self.mod(x)
@@ -233,7 +233,7 @@ class ElementwiseMixin(DTypeMixin, CreationMixin):
     return self.div(x, True)
 
   def __rfloordiv__(self, x: Self | ConstType) -> Self:
-    return self.idiv(x, True)
+    return self._binop(Ops.FLOORDIV, x, True)
 
   def __rand__(self, x: Self | ConstType) -> Self:
     return self.bitwise_and(x, True)

tinygrad/nn/onnx.py

@@ -644,7 +644,7 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
   def BitwiseOr(x:Tensor,y:Tensor): return x | y
   def BitwiseXor(x:Tensor,y:Tensor): return x ^ y
   def BitwiseNot(x:Tensor): return ~x
-  def Mod(x:Tensor,y:Tensor,fmod=0): return x - x.div(y, rounding_mode="trunc") * y if fmod else x % y
+  def Mod(x:Tensor,y:Tensor,fmod=0): return x.fmod(y) if fmod else x % y
 
   # ***** Casting Ops *****
   # NOTE: saturate only applies to FP8 types

tinygrad/renderer/cstyle.py

@@ -44,12 +44,11 @@ base_rewrite = PatternMatcher([
   # default const render
   (UPat(Ops.CONST, name="x"), lambda ctx,x: str(x.arg)),
   # new load/store
-  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var('idx')), allow_any_len=True),
+  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var('idx'))),
    lambda ctx,buf,idx: f"({ctx[buf]}+{strip_parens(ctx[idx]) if idx.arg == Ops.ADD else ctx[idx]})"),
-  (UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat.var("gate"))).or_casted("bidx"), UPat.var("var"))),
-   lambda ctx,bidx,var,gate: f"({ctx[gate]}?*{ctx[bidx]}:{ctx[var]})"),
   (UPat(Ops.LOAD, src=(UPat.var('bidx'),)), lambda ctx,bidx: f"(*{ctx[bidx]})"),
-  (UPat(Ops.STORE, src=(UPat.var('bidx'), UPat.var("var"))), lambda ctx,bidx,var: f"*{ctx[bidx]} = {ctx[var]};"),
+  (UPat(Ops.LOAD, src=(UPat.var("bidx"), UPat.var("var"), UPat.var("gate"))), lambda ctx,bidx,var,gate: f"({ctx[gate]}?*{ctx[bidx]}:{ctx[var]})"),
+  (UPat(Ops.STORE, src=(UPat.var('bidx'), UPat.var("var")), allow_any_len=True), lambda ctx,bidx,var: f"*{ctx[bidx]} = {ctx[var]};"),
   # alu/gep
   # TODO: look for left-associative
   (UPat(GroupOp.ALU, name="x"), lambda ctx,x: ctx.code_for_op[x.op](
@@ -302,11 +301,11 @@ class OpenCLRenderer(CStyleLanguage):
     (UPat(Ops.CONST, dtypes.bfloat16, name="x"),
       lambda ctx,x: f"{(struct.unpack('I', struct.pack('f', float_to_bf16(x.arg)))[0] >> 16)}u"),
     # load/store image (OpenCL)
-    (UPat(Ops.LOAD, dtype=dtypes.float.vec(4), src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2)), UPat.var("gate")), UPat.var("var"))),
+    (UPat(Ops.LOAD, dtype=dtypes.float.vec(4), src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2))), UPat.var("var"), UPat.var("gate"))),
       lambda ctx,buf,idx,var,gate: f"({ctx[gate]}?read_imagef({ctx[buf]}, smp, {ctx[idx]}):{ctx[var]})"),
     (UPat(Ops.LOAD, dtype=dtypes.float.vec(4), src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2))),)),
       lambda ctx,buf,idx: f"read_imagef({ctx[buf]}, smp, {ctx[idx]})"),
-    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2)), allow_any_len=True),
+    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2))),
                           UPat.var("var", dtypes.float.vec(4))), allow_any_len=True),
       lambda ctx,buf,idx,var: f"write_imagef({ctx[buf]}, {ctx[idx]}, {ctx[var]});"),
   ]) + base_rewrite
@@ -536,6 +535,7 @@ class HIPRenderer(CStyleLanguage):
     if any(dt.scalar() == dtypes.half for dt in used_dtypes): prefix.append("#define half _Float16")
     if any(dt.scalar() in dtypes.fp8s for dt in used_dtypes):
       prefix += ["typedef unsigned char hip_bf8;", "typedef unsigned char hip_fp8;"]
+    if any(u.op is Ops.CAST and u.dtype in dtypes.fp8s and u.src[0].dtype == dtypes.float for u in uops):
       prefix.append("""static inline __attribute__((device)) unsigned char f32_to_fp8(float v, int is_bf8) {
   v = (((*(unsigned*)&v)&0x7F800000)!=0x7F800000)?__builtin_amdgcn_fmed3f(v,is_bf8?57344.0f:448.0f,is_bf8?-57344.0f:-448.0f) : v;
   return (unsigned char)(is_bf8?__builtin_amdgcn_cvt_pk_bf8_f32(v,v,0,false):__builtin_amdgcn_cvt_pk_fp8_f32(v,v,0,false));\n}""")

tinygrad/renderer/llvmir.py

@@ -76,14 +76,14 @@ base_rewrite = PatternMatcher([
   # memory load/store
   (UPat(Ops.INDEX, name="x"), lambda ctx,x:
    f"  {ctx[x]} = getelementptr inbounds {ldt(x.dtype.base)}, {ldt(x.src[0].dtype)} {ctx[x.src[0]]}, {ldt(x.src[1].dtype)} {ctx[x.src[1]]}"),
-  (UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat(), UPat(), UPat.var("mask"))).or_casted("idx"), UPat.var("alt")), allow_any_len=True, name="x"),
+  (UPat(Ops.LOAD, src=(UPat.var("idx"), UPat.var("alt"), UPat.var("mask")), name="x"),
    lambda ctx,x,idx,alt,mask:
    f"  br label {ctx[x]}_entry\n{ctx[x][1:]}_entry:\n"
    f"  br i1 {ctx[mask]}, label {ctx[x]}_load, label {ctx[x]}_exit\n{ctx[x][1:]}_load:\n"
    f"  {ctx[x]}_yes = load {ldt(x.dtype)}, {ldt(idx.dtype)} {ctx[idx]}\n"
    f"  br label {ctx[x]}_exit\n{ctx[x][1:]}_exit:\n"
    f"  {ctx[x]} = phi {ldt(x.dtype)} [{ctx[x]}_yes, {ctx[x]}_load], [{ctx[alt]}, {ctx[x]}_entry]"),
-  (UPat(Ops.LOAD, src=(UPat.var('idx'),), allow_any_len=True, name="x"),
+  (UPat(Ops.LOAD, src=(UPat.var('idx'),), name="x"),
    lambda ctx,x,idx: f"  {ctx[x]} = load {ldt(x.dtype)}, {ldt(idx.dtype)} {ctx[idx]}"),
   (UPat(Ops.STORE, name="x"), lambda ctx,x: f"  store {ldt(x.src[1].dtype)} {ctx[x.src[1]]}, {ldt(x.src[0].dtype)} {ctx[x.src[0]]}"),