fix cast to weak twice [pr] (#17548 )

also no gradient for weak target
no pm_fold_cast_const in full_rewrite_to_sink [PR] (#17547 )
2026-08-15 12:39:54 -04:00 · 2026-08-15 10:18:59 -04:00 · 2026-08-15 09:57:05 -04:00 · 2026-08-15 15:03:28 +03:00 · 2026-08-15 07:47:40 -04:00 · 2026-08-15 18:00:54 +09:00
63 changed files with 624 additions and 455 deletions
@@ -45,10 +45,6 @@ inputs:
    description: "Install qemu"
    required: false
    default: 'false'
-  docker-qemu:
-    description: "Setup docker to use qemu"
-    required: false
-    default: 'false'
 runs:
  using: "composite"
  steps:
@@ -286,9 +282,3 @@ runs:
        sudo mkdir -p /etc/OpenCL/vendors
        echo "/usr/lib/libRusticlOpenCL.so" | sudo tee /etc/OpenCL/vendors/rusticl.icd
        echo "RUSTICL_ENABLE=llvmpipe" >> "$GITHUB_ENV"
-
-    - name: Setup docker to use qemu
-      if: inputs.docker-qemu == 'true'
-      uses: docker/setup-qemu-action@v4
-      with:
-        platforms: arm64
@@ -656,7 +656,7 @@ jobs:
        with:
          key: compile-${{ matrix.backend }}
          deps: "testing_unit mesa"
-          docker-qemu: ${{ contains(matrix.dev, 'QCOMCL') }}
+          qemu: ${{ contains(matrix.dev, 'QCOMCL') }}
      - name: Test IMAGE
        shell: bash
        if: contains(matrix.dev, 'a630')
@@ -182,12 +182,12 @@ class GPTOSS:
    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)  # (B,N,H,D)/(B,N,KV,D)

-    if sliding:
-      attn = self._sliding_attention(xq, xk, xv, sinks)
-    elif getenv("HK_FLASH_ATTENTION"):
+    if getenv("HK_FLASH_ATTENTION"):
      from extra.thunder.amd.fa import flash_attention
-      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks)
+      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks, window=self.sliding_window if sliding else 0)
      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
+    elif sliding:
+      attn = self._sliding_attention(xq, xk, xv, sinks)
    else:
      xqm = xq.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
      xkm, xvm = xk.permute(0, 2, 1, 3).unsqueeze(2), xv.permute(0, 2, 1, 3).unsqueeze(2)
@@ -11,6 +11,7 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
+export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -11,6 +11,7 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
+export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -35,7 +35,7 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li
    return name

  for call in iter_kernel_calls(linear):
-    arg_uops = [b for b in call.src[1:] if b.op is not Ops.BIND]
+    arg_uops = [b for b in call.src[1:] if not b.is_bound_var]
    prg = to_program(call.src[0], Device[arg_uops[0].device].renderer)
    info = prg.arg
    functions[info.function_name] = prg.src[2].arg
@@ -122,7 +122,8 @@ def custom_mxfp4_gemm(C:UOp, A:UOp, B:UOp, scale_a:UOp, scale_b:UOp, *extra:UOp,
  groups_x, groups_y = UOp.special(ceildiv(N, tile_n), "gidx0"), UOp.special(ceildiv(M, tile_m), "gidx1")
  lds = UOp.placeholder((163840,), dtypes.uint8, 0, AddrSpace.LOCAL)
  sink = UOp.sink(C.base, A.base, B.base, scale_a.base, scale_b.base, *(x.base for x in extra), lds, threads, groups_x, groups_y,
-                  arg=KernelInfo(f"custom_mxfp4_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K)))
+                  arg=KernelInfo(f"mxfp4_gemm_{M}_{N}_{K}",
+                                 estimates=Estimates(ops=2*M*N*K, mem=(M*half_k+N*half_k)*A.dtype.itemsize+M*N*C.dtype.itemsize)))
  insts = build_kernel(M, N, K, tile_m, tile_n)
  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=x) for x in insts))))

@@ -79,7 +79,7 @@ if __name__ == "__main__":
  linear, var_vals = C.linear_with_vars()
  last_call = linear.src[-1]
  ast = last_call.src[0]
-  bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]
+  bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]

  src = compiled.asm["ptx"]
  # specify the shared memory here so we don't need to do it dynamically
@@ -3,7 +3,7 @@ from typing import cast, Any, Callable
 import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit
 assert sys.platform != 'win32'
 from dataclasses import dataclass
-from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_cmdbuf
+from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, encode_kernargs_clike, make_cmdbuf
 from tinygrad.runtime.support.hcq2 import make_binary_patch
 from tinygrad.uop.ops import sint, UOp
 from tinygrad.device import Compiled, BufferSpec, Buffer, Device
@@ -290,14 +290,14 @@ class AMDAllocator(HCQAllocator['AMDDevice']):
  def __init__(self, dev:AMDDevice):
    super().__init__(dev, supports_copy_from_disk=dev.has_copy_queue, supports_transfer=dev.has_copy_queue and not dev.is_usb())

-  def _alloc(self, size:int, options:BufferSpec) -> HCQ2Buffer:
+  def _alloc(self, size:int, options:BufferSpec) -> HCQBuffer:
    return self.dev.iface.alloc(size, host=options.host, uncached=options.uncached, cpu_access=options.cpu_access or not self.dev.has_copy_queue)

  def _do_free(self, opaque, options:BufferSpec): self.dev.iface.free(opaque)

-  def _do_map(self, buf:HCQ2Buffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)
+  def _do_map(self, buf:HCQBuffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)

-  def _do_unmap(self, buf:HCQ2Buffer): self.dev.iface.unmap(buf)
+  def _do_unmap(self, buf:HCQBuffer): self.dev.iface.unmap(buf)

@dataclass
 class AMDQueueDesc:
@@ -16,8 +16,9 @@ apt-get install -y python3-pip python3-venv git tmux rclone clang

 ### 1.2 Install Python deps
 ```bash
-python3 -m pip install --break-system-packages numpy tqdm wandb tiktoken sentencepiece
+python3 -m pip install --break-system-packages --ignore-installed typing-extensions numpy tqdm wandb tiktoken sentencepiece
 ```
+Note: `--ignore-installed typing-extensions` is needed because the base image ships typing-extensions 4.10.0 without a RECORD file, so pip cannot uninstall it.

 ### 1.3 Install ROCm dev headers
 The base image has ROCm runtime but NOT the HIP dev headers. Need:
@@ -85,36 +86,45 @@ Files downloaded (~85GB total, ~6 minutes):
 - `c4-validation-91205-samples.en_text_document.idx` (1.8 MB)
 - `LICENSE.txt`, `NOTICE.txt`

+**Wait for rclone to fully complete before starting training.** Starting training while the dataset is still downloading will read a truncated .bin file, causing `ValueError: all input arrays must have the same shape` in the dataloader. The stale `.index_cache` and `.blend_cache` files must also be deleted if this happens:
+```bash
+rm -f /raid/datasets/c4-8b/*.index_cache /raid/datasets/c4-8b/*.blend_cache
+```
+
 ## Phase 4: wandb Login
 ```bash
 wandb login
 ```
 Enter API key from https://wandb.ai/authorize

+Alternatively, pass the key directly:
+```bash
+wandb login <API_KEY>
+```
+
 ## Phase 5: Run Training

+Run training in tmux so it survives SSH disconnects:
+```bash
+tmux new-session -d -s train 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm WANDB=1 bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh 2>&1 | tee /root/train.log'
+```
+Attach with `tmux attach -t train`.
+
 ### 5.1 Smoke test (beam search, 2 layers, real data)
 Always run beam first to validate the pipeline:
 ```bash
-cd /root/tinygrad
-COMGR_PATH=/opt/rocm/lib/libamd_comgr.so \
-COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so \
-CC=/opt/rocm/core-7.14/lib/llvm/bin/clang \
-DEV=AMD:HIP \
-ROCM_PATH=/opt/rocm \
-  bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
+tmux new-session -d -s beam 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh 2>&1 | tee /root/beam.log'
 ```

+The beam test runs 10 training steps with 2 layers. Expected results:
+- ~0.29s per step after warmup
+- ~700K GFLOPS, ~7% MFU (low because only 2 layers)
+- ~380 GB VRAM used
+- Loss stable at ~12.55 with random init
+
 ### 5.2 Full training run
 ```bash
-cd /root/tinygrad
-COMGR_PATH=/opt/rocm/lib/libamd_comgr.so \
-COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so \
-CC=/opt/rocm/core-7.14/lib/llvm/bin/clang \
-DEV=AMD:HIP \
-ROCM_PATH=/opt/rocm \
-WANDB=1 \
-  bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh
+tmux new-session -d -s train 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm WANDB=1 bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh 2>&1 | tee /root/train.log'
 ```

 ## Environment Variable Reference
@@ -180,13 +190,18 @@ $ lspci -nn | grep AMD
 ```
 CPU flags include `hypervisor`. `dmesg` shows `Hypervisor detected: KVM`.

-### PCI device ID
-`lspci -v` shows device ID `0x75b0` and subsystem ID `0x75a0`:
-```
-83:00.0 Processing accelerators: ... Device 75b0
-    Subsystem: ... Device 75a0
-```
-tinygrad's `PCIIface` in `ops_amd.py` and `hive_reset.py` did not list `0x75b0`, so the GPU was not found. Adding `0x75b0` to the device ID list in both files fixes the detection.
+### Working path: amdgpu driver (KFDIface)
+The amdgpu driver loads on boot and binds to all 8 GPUs, creating `/dev/kfd` and 64 renderD nodes (`/dev/dri/renderD128` through `/dev/dri/renderD191`). tinygrad's `KFDIface` enumerates GPUs through `/sys/devices/virtual/kfd/kfd/topology/nodes` and uses `/dev/kfd` for ioctl. No PCI device ID patching is needed — the KFD path does not use `PCIIface` or `AMDev._run_discovery()`.
+
+This is the working configuration. No code changes to tinygrad are required.
+
+### PCIIface path (does not work on this VM)
+For reference, the `PCIIface` path was also explored but does not work in this KVM guest:
+
+- `PCIIface` in `ops_amd.py` does not list device ID `0x75b0`. Adding it allows PCI detection but `AMDev._run_discovery()` fails because the VRAM BAR reads all `0xFF`.
+- This was observed with the GPU unbound from any driver, after PCI reset, and with VFIO bound.
+- VFIO binding (`vfio-pci` with `enable_unsafe_noiommu_mode=1`) succeeded but VRAM BAR still reads all `0xFF`.
+- No IOMMU in guest — `dmesg` has no `AMD-Vi` entries, PCI devices have no `iommu_group` symlink.

 ### amdgpu driver behavior
 On first boot, amdgpu loaded and bound to all 8 GPUs. On one boot it failed to initialize:
@@ -198,24 +213,5 @@ On first boot, amdgpu loaded and bound to all 8 GPUs. On one boot it failed to i
 ```
 On a subsequent boot, amdgpu initialized successfully (SMU initialized, VRAM ready). After unbinding all 8 GPUs from amdgpu, `rmmod amdgpu` wedged the module (stuck in "Unloading" state in `/proc/modules`), requiring a full VM reboot.

-### `/dev/kfd`
-`/dev/kfd` exists when amdgpu is loaded. Opening it returns `OSError: [Errno 22] Invalid argument`.
-
-### VRAM BAR reads all 0xFF
-After amdgpu initializes the GPU and is then unbound, reading the VRAM BAR (via `/sys/bus/pci/devices/0000:83:00.0/resource0`) returns all `0xFF` at all offsets — including the discovery table at `vram_size - 64KB`. tinygrad's `AMDev._run_discovery()` fails with `AssertionError: discovery signatures mismatch`.
-
-A PCI reset (`echo 1 > /sys/bus/pci/devices/0000:83:00.0/reset`) did not change the VRAM contents — still all `0xFF`.
-
-VRAM was also all `0xFF` when read via `/dev/mem` at the BAR physical address (`0xa0000000000`).
-
-### VFIO attempt
-Bound the GPU to `vfio-pci` with `enable_unsafe_noiommu_mode=1`. The GPU bound successfully and `/dev/vfio/noiommu-0` appeared. Running tinygrad with `VFIO=1` still failed with the same `discovery signatures mismatch` — VRAM BAR still reads all `0xFF`.
-
-### No IOMMU in guest
-`dmesg` has no `AMD-Vi` entries. PCI devices have no `iommu_group` symlink.
-
 ### No fan control
 No `fan*` or `pwm*` hwmon entries exist. Only `temp*`, `power*`, `freq*` are exposed. GPU temps read 56-63°C, power ~265W per GPU.
-
-### Current status: NOT WORKING
-tinygrad's `PCIIface` finds the GPU (after adding `0x75b0`) but `AMDev._run_discovery()` fails because the VRAM discovery table reads all `0xFF`. This was observed with the GPU unbound from any driver, after PCI reset, and with VFIO bound.
@@ -110,7 +110,49 @@ def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
  return _sharded_empty(ref.shape, ref, axis)

@functools.cache
-def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink):
+def _windowed_lse(xq:Tensor, xk:Tensor, sinks, W:int) -> Tensor:
+  B, N, H, hd = xq.shape
+  H_KV = xk.shape[2]; R = H // H_KV; nb = N // W; sm = hd ** -0.5
+  q = xq.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
+  k = xk.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
+  k_prev = k.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
+  sc_d = (q @ k.transpose(-1, -2)) * sm
+  sc_p = (q @ k_prev.transpose(-1, -2)) * sm
+  li, lj = Tensor.arange(W).reshape(W, 1), Tensor.arange(W).reshape(1, W)
+  pv = (Tensor.arange(nb).reshape(nb, 1, 1) >= 1)
+  sc_d = (lj <= li).where(sc_d, -float("inf"))
+  sc_p = ((li < lj) & pv).where(sc_p, -float("inf"))
+  m = sc_d.max(-1, keepdim=True).maximum(sc_p.max(-1, keepdim=True))
+  if sinks is not None: m = m.maximum(sinks.reshape(1, H_KV, R, 1, 1, 1).float())
+  denom = (sc_d - m).exp().sum(-1, keepdim=True) + (sc_p - m).exp().sum(-1, keepdim=True)
+  if sinks is not None: denom = denom + (sinks.reshape(1, H_KV, R, 1, 1, 1).float() - m).exp()
+  return (m + denom.log()).reshape(B, H, N).unsqueeze(2)  # (B, H, 1, N), matches saved l_vec
+
+def _windowed_delta(xq:Tensor, xk:Tensor, xv:Tensor, do:Tensor, sinks, W:int) -> Tensor:
+  B, N, H, hd = xq.shape
+  H_KV = xk.shape[2]; R = H // H_KV; nb = N // W; sm = hd ** -0.5
+  q = xq.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
+  k = xk.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
+  v = xv.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
+  dob = do.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
+  k_prev = k.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
+  v_prev = v.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
+  sc_d = (q @ k.transpose(-1, -2)) * sm
+  sc_p = (q @ k_prev.transpose(-1, -2)) * sm
+  li, lj = Tensor.arange(W).reshape(W, 1), Tensor.arange(W).reshape(1, W)
+  pv = (Tensor.arange(nb).reshape(nb, 1, 1) >= 1)
+  sc_d = (lj <= li).where(sc_d, -float("inf"))
+  sc_p = ((li < lj) & pv).where(sc_p, -float("inf"))
+  m = sc_d.max(-1, keepdim=True).maximum(sc_p.max(-1, keepdim=True))
+  if sinks is not None: m = m.maximum(sinks.reshape(1, H_KV, R, 1, 1, 1).float())
+  e_d, e_p = (sc_d - m).exp(), (sc_p - m).exp()
+  denom = e_d.sum(-1, keepdim=True) + e_p.sum(-1, keepdim=True)
+  if sinks is not None: denom = denom + (sinks.reshape(1, H_KV, R, 1, 1, 1).float() - m).exp()
+  o = ((e_d / denom) @ v) + ((e_p / denom) @ v_prev)
+  delta = (dob * o).sum(-1)
+  return delta.reshape(B, H, N).unsqueeze(2)
+
+def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=0):
  def grad(dou:UOp, ker:UOp) -> tuple:
    do = Tensor(dou, device=dou.device)
    attn = Tensor(ker.src[1].after(ker), device=ker.src[1].device)
@@ -118,6 +160,8 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
    xq = Tensor(ker.src[3], device=ker.src[3].device)
    xk = Tensor(ker.src[4], device=ker.src[4].device)
    xv = Tensor(ker.src[5], device=ker.src[5].device)
+    if window:
+      l_vec = _windowed_lse(xq, xk, Tensor(ker.src[6], device=ker.src[6].device) if has_sink else None, window)

    dq = _sharded_empty((B, H, N, D), xq, axis=shard_axis_t)
    GROUP_SIZE = H_local // H_KV_local
@@ -128,8 +172,10 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
    # delta_vec = (do * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()
    delta_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
    delta_vec, dq = Tensor.custom_kernel(delta_vec, dq, attn, do, fxn=functools.partial(custom_fa_backward_pre, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:2]
+    if window:
+      delta_vec = _windowed_delta(xq, xk, xv, do, Tensor(ker.src[6], device=ker.src[6].device) if has_sink else None, window)

-    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:3]
+    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, window=window))[:3]

    if D == 64:
      dq = dq.reshape(B, H, N//16, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2).permute(0, 1, 2, 8, 9, 10, 11, 3, 4, 6, 7, 5, 12).reshape(B, H, N, D).transpose(1, 2)
@@ -149,7 +195,7 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
  return grad

 # TODO: remove write_flat once scheduler can remove reshapes between custom_kernel. TestCustomKernel.test_simple_reshape
-def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None):
+def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None, window:int=0):
  assert attn_mask is None, "attn_mask not supported"
  assert is_causal, "only causal attention supported"

@@ -176,18 +222,18 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
  attn = _sharded_empty((B, N, H * D), xq, axis=shard_axis) if write_flat else _sharded_empty_like(xq, axis=shard_axis)
  l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)

-  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink)
+  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=window)

  fwd_inputs = (attn, l_vec, xq, xk, xv) + ((sinks,) if has_sink else ())
-  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, 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, has_sink=has_sink), grad_fxn=grad)[:2]
+  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, 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, has_sink=has_sink, window=window), grad_fxn=grad)[:2]

  return attn, attn, l_vec

@functools.cache
-def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True):
+def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True, window:int=0):
  code = (pathlib.Path(__file__).parent / "fa_fwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}", f"-DWINDOW={window}"]

  Q_BLOCK_SIZE = 32
  NUM_WARPS = 8
@@ -247,10 +293,10 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
-def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, window:int=0):
  code = (pathlib.Path(__file__).parent / "fa_bwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DWINDOW={window}"]

  BLOCK_SIZE_KV = 256
  GROUP_SIZE = H // H_KV
@@ -360,7 +360,7 @@ class TestJitGraphSplit(unittest.TestCase):
    self.expect(f, inp, inp_cpu,
      graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
      multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
-      hcqgraph=[self.ji_graph(4)])
+      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()]) # cpu is hcq2 now, it does not join hcq graphs

  def test_jit_cpu_several(self):
    if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")
@@ -377,9 +377,9 @@ class TestJitGraphSplit(unittest.TestCase):
    inp = Tensor.randn(10, 10, device=Device.DEFAULT).realize()
    inp_cpu = Tensor.randn(10, 10, device="CPU").realize()
    self.expect(f, inp, inp_cpu,
-      graph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
-      multigraph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
-      hcqgraph=[self.ji_graph(5)])
+      graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
+      multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
+      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()])

  def test_jit_multidev(self):
    if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")
@@ -30,7 +30,7 @@ class TestLinearizer(unittest.TestCase):
    c = ((a.shrink(((0, 2),)) - a.shrink(((2, 4),))) - (b.shrink(((0, 2),)) - b.shrink(((2, 4),))))
    linear = c.schedule_linear()
    run_linear(linear)
-    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if s.op is not Ops.BIND]
+    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if not s.is_bound_var]
    assert len(rawbufs) == 3 and set(rawbufs[1:]) == {a.uop.base.realized, b.uop.base.realized}
    np_c = (np_a[:2] - np_a[2:]) - (np_b[:2] - np_b[2:])
    np.testing.assert_allclose(np_c, c.numpy(), atol=1e-4, rtol=1e-4)
@@ -411,7 +411,7 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
  last_call = linear.src[-1]
  ast = last_call.src[0]
  assert ast.op is Ops.SINK, f"helper_realized_ast expects a SINK {last_call}"
-  last_bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]
+  last_bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]
  # now all input buffers in last_call should be realized
  # create fresh buffers for the outputs
  bufs = [Buffer(x.device, x.size, x.dtype).allocate() if i < len(ast.src) else x for i,x in enumerate(last_bufs)]
@@ -1,9 +1,9 @@
 import unittest, random
 from tinygrad import Tensor, Device, nn, GlobalCounters, TinyJit, dtypes, Variable
-from tinygrad.uop.ops import Ops, UOp, AxisType
+from tinygrad.uop.ops import Ops, UOp, AxisType, graph_rewrite
 from tinygrad.helpers import getenv, prod, Context
 from tinygrad.nn.state import get_parameters
-from tinygrad.engine.realize import run_linear, compile_linear
+from tinygrad.engine.realize import run_linear, compile_linear, pm_beam, pm_compile
 import numpy as np
 from hypothesis import given, strategies as strat, settings
 from test.helpers import not_support_multi_device, needs_second_gpu, slow, call_is_graph, check_schedule, assert_kernel_count
@@ -79,9 +79,9 @@ class TestMultiTensor(unittest.TestCase):
  def test_shard_beam(self):
    cpu_2 = ("CPU:1", "CPU:2")
    src = Tensor.ones(16).shard(cpu_2, 0).realize()
-    pad = src.to(cpu_2[::-1]).schedule_linear().src[0]
-    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): prg = compile_linear(UOp(Ops.LINEAR, src=(pad,))).src[0].src[0]
-    self.assertNotEqual(prg.src[0].arg.applied_opts, ())
+    lin = UOp(Ops.LINEAR, src=(src.to(cpu_2[::-1]).schedule_linear().src[0],))
+    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): call = graph_rewrite(graph_rewrite(lin, pm_beam, ctx=1, walk=True), pm_compile, walk=True).src[0]
+    self.assertNotEqual(call.src[0].src[0].arg.applied_opts, ())

  def test_shard_same_device(self):
    X = Tensor.ones(256).contiguous().realize()
@@ -720,10 +720,11 @@ class TestOps(unittest.TestCase):
      return torch.autograd.grad(t ** c, t)[0].item()
    for x in [-math.inf, 0, 1, math.inf]:
      for c in [-1, 0, 0.3, 1, 2]:
-        tiny_out = get_tiny_gradient(x, c)
        torch_out = get_torch_gradient(x, c)
+        # the pow backward routes through exp2/log2, whose 0/inf behavior is undefined on WEBGPU
+        if Device.DEFAULT == "WEBGPU" and not math.isfinite(torch_out): continue
+        tiny_out = get_tiny_gradient(x, c)
        if math.isnan(tiny_out):
-          if Device.DEFAULT == "WEBGPU": continue # TODO: WEBGPU issue with nan
          assert math.isnan(torch_out)
        else:
          self.assertAlmostEqual(tiny_out, torch_out, msg=f"{x}, {c}")
@@ -749,6 +750,7 @@ class TestOps(unittest.TestCase):
  def test_exp2_log2_zero_times_negative(self):
    # gallivm's exp2/log2 have "undefined behavior with infs, 0s and nans", so exp2(log2(0)*y) returns 0 instead of inf
    helper_test_op(None, lambda x,y: (x.log2()*y).exp2(), lambda x,y: (x.log2()*y).exp2(), vals=[[0.0], [-0.7]], forward_only=True)
+  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "pow at 0 routes through exp2/log2, whose 0/inf behavior is undefined on WEBGPU")
  def test_pow_zero_const(self):
    helper_test_op(None, lambda x: x**0.3, vals=[[0.0]])
    helper_test_op(None, lambda x: x**0.0, vals=[[0.0]])
@@ -2,7 +2,7 @@ import unittest, pickle, types, tracemalloc
 import numpy as np
 from tinygrad import Tensor, Device, TinyJit, Variable, dtypes
 from tinygrad.helpers import GlobalCounters, ContextVar, Context, DEV
-from tinygrad.uop.ops import PatternMatcher, UPat, UOp
+from tinygrad.uop.ops import PatternMatcher, UPat, UOp, deconstruct_function

 class TestPickle(unittest.TestCase):
  def test_pickle_code_object(self):
@@ -11,6 +11,11 @@ class TestPickle(unittest.TestCase):
    fxn = types.FunctionType(pickle.loads(code_str), globals())
    self.assertEqual(fxn(2), 4)

+  def test_deconstruct_function_nested_comprehension(self):
+    # pre PEP 709, each comprehension is its own code object, so dtypes here is referenced two code objects deep
+    def fxn(): return [[dtypes.int for _ in range(2)] for _ in range(2)]
+    self.assertEqual(types.FunctionType(*deconstruct_function(fxn))(), fxn())
+
  def test_pickle_pattern_matcher(self):
    pm = PatternMatcher([(UPat.cvar('x'), lambda x: x*2)])
    sink = UOp.const(2)
@@ -0,0 +1,17 @@
+from tinygrad import Device, Tensor, TinyJit, dtypes
+from tinygrad.helpers import Timing, Context
+
+GPUS, DEPTH, SZ = 8, 4, 128 * 2**20
+WARMUP, ITERS = 3, 5
+devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(GPUS))
+bufs = tuple(Tensor.empty(SZ, dtype=dtypes.uint8, device=dev).contiguous().realize() for _ in range(DEPTH) for dev in devs)
+
+@TinyJit
+def all_to_all(*srcs:Tensor): return Tensor.realize(*(src.to(dst) for i,src in enumerate(srcs) for j,dst in enumerate(devs) if i % GPUS != j))
+
+if __name__ == "__main__":
+  with Context(ALL2ALL=1, JIT_BATCH_SIZE=0):
+    for i in range(-WARMUP, ITERS):
+      with Timing("ALL2ALL ", lambda ns: f" {SZ*GPUS*(GPUS-1)*DEPTH/ns:.2f} GB/s", enabled=i>=0):
+        all_to_all(*bufs)
+        for dev in devs: Device[dev].synchronize()
@@ -1,5 +1,5 @@
-import unittest, time
-from tinygrad import Tensor
+import unittest, time, itertools
+from tinygrad import Tensor, Context

 class TestScheduleScaling(unittest.TestCase):
  """Test that .schedule() scales linearly with graph size (no O(n^2) behavior)."""
@@ -130,5 +130,18 @@ class TestScheduleScaling(unittest.TestCase):
      return parts[0].cat(*parts[1:])
    self._assert_linear(concat_chain)

+  @Context(DEV="NULL:HIP:gfx1100")
+  def test_custom_kernel_assign_scaling(self):
+    from tinygrad.uop.ops import UOp, Ops, KernelInfo
+    from tinygrad.runtime.autogen.amd.rdna3.ins import s_nop
+    count = itertools.count(0)
+    def custom_kernel_assign(n):
+      def custom_asm(out):
+        return UOp(Ops.PROGRAM, src=(UOp.sink(out, arg=KernelInfo(f"fxn_{next(count)}")),
+                                     UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=s_nop(i)) for i in range(n*8)))))
+      call = Tensor.custom_kernel(Tensor.empty(1), fxn=custom_asm)[0]
+      return Tensor.cat(*[Tensor.empty(1).assign(call+i) for i in range(n)])
+    self._assert_linear(custom_kernel_assign, n_small=50, n_large=500)
+
 if __name__ == '__main__':
  unittest.main(verbosity=2)
@@ -86,7 +86,7 @@ def assert_jit_cache_len(fxn, expected_len):
  if linear is None or not linear.src:
    if expected_len != 0: raise KernelCountException(expected_len, 0)
    return
-  if expected_len and all(call_is_hcq(call) for call in linear.src): expected_len = 3 # HCQ2: merged same-queue calls + finalizer + bumps
+  if expected_len and all(call_is_hcq(call) for call in linear.src): expected_len = 4 # HCQ2: fence + reset + merged same-queue calls + finalizer
  if call_is_graph(linear.src[0]):
    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
    inner = linear.src[0].src[0].src[0]  # LINEAR UOp inside CUSTOM_FUNCTION
@@ -260,19 +260,6 @@ def _cond(cond, if_true, if_false):
 def _cond_hi16(cond, val: UOp) -> UOp: return _cond(cond, _hi16(val), val)
 def _apply_opsel(val: UOp, sel_bit: int, opsel: int) -> UOp: return _hi16(val) if opsel & (1 << sel_bit) else val

-def _set_lane_bit(old: UOp, lane: UOp, val: UOp, exec_mask: UOp) -> UOp:
-  """Set/clear a single bit in a mask based on lane index, respecting exec mask."""
-  if old.dtype in (dtypes.uint64, dtypes.int64):
-    dt = dtypes.uint64
-    mask = UOp.const(1, dt) << lane.cast(dt)
-    new_bit = _to_u32(val).cast(dt) << lane.cast(dt)
-    cleared = old.cast(dt) & (mask ^ UOp.const(0xFFFFFFFFFFFFFFFF, dt))
-    return _lane_active(exec_mask, lane).where(cleared | new_bit, old.cast(dt))
-  mask = _c(1) << lane.cast(dtypes.uint32)
-  new_bit = _to_u32(val) << lane.cast(dtypes.uint32)
-  cleared = old & (mask ^ _c(MASK32))
-  return _lane_active(exec_mask, lane).where(cleared | new_bit, old)
-
 def _val_to_u32(val: UOp) -> UOp:
  """Convert any value to uint32 for storage (bitcast floats, cast ints)."""
  if val.dtype == dtypes.uint32: return val
@@ -532,6 +519,19 @@ class _Ctx:
      return [self.wsgpr_dyn(reg, lo), self.wsgpr_dyn(reg + _c(1), hi)]
    return [self.wsgpr_dyn(reg, val)]

+  def wmask_lane_bit(self, reg: UOp, lane: UOp, val: UOp, exec_mask: UOp) -> list[UOp]:
+    """Set/clear bit `lane` of the mask at `reg` from val for exec-active lanes, preserving memory for inactive lanes"""
+    active, bit = _lane_active(exec_mask, lane), _to_u32(val)
+    if self.wave_size <= 32:
+      old = self.rsgpr_dyn(reg)
+      mask = _c(1) << lane.cast(dtypes.uint32)
+      return [self.wsgpr_dyn(reg, active.where((old & (mask ^ _c(MASK32))) | (bit << lane.cast(dtypes.uint32)), old))]
+    off = (lane & _c(31, dtypes.int)).cast(dtypes.uint32)
+    mask = _c(1) << off
+    def half(old: UOp, sel: UOp) -> UOp: return sel.where(active.where((old & (mask ^ _c(MASK32))) | (bit << off), old), old)
+    return [self.wsgpr_dyn(reg, half(self.rsgpr_dyn(reg), lane < _c(32, dtypes.int))),
+            self.wsgpr_dyn(reg + _c(1), half(self.rsgpr_dyn(reg + _c(1)), _c(32, dtypes.int) <= lane))]
+
  def rmask(self, reg: UOp) -> UOp:
    """Read a lane mask (VCC/EXEC). Combines lo/hi for wave64."""
    if self.wave_size > 32: return _u64(self.rsgpr_dyn(reg), self.rsgpr_dyn(reg + _c(1)))
@@ -718,9 +718,7 @@ class _Ctx:
        raw_stores.append(('vgpr_direct', self.vgpr.index(val[0].valid(active)).store(new_val)))
        continue
      if 'D0' in dest and '[laneId]' in dest:
-        old_vcc = self.rmask(_c(VCC_LO.offset))
-        new_vcc = _set_lane_bit(old_vcc, lane, val, exec_mask)
-        raw_stores.extend([('vcc', s) for s in self.wmask(_c(VCC_LO.offset), new_vcc)])
+        raw_stores.extend([('vcc', s) for s in self.wmask_lane_bit(_c(VCC_LO.offset), lane, val, exec_mask)])
      elif dest.startswith('D0'):
        dest_suffix = re.match(r'D0\.(\w+)', dest)
        if dest_suffix is not None:
@@ -1039,13 +1037,11 @@ def _compile_sdwa(inst: irc.VOP1_SDWA | irc.VOP2_SDWA | irc.VOP2_SDWA_SDST | irc
        result = _sdwa_write(old, result, dst_sel, dst_unused)
      stores.append(ctx.wvgpr_dyn(vdst_reg, lane, result, exec_mask))
    elif dest.startswith('VCC'):
-      old_vcc = ctx.rmask(_c(VCC_LO.offset))
-      stores.extend(ctx.wmask(_c(VCC_LO.offset), _set_lane_bit(old_vcc, lane, val, exec_mask)))
+      stores.extend(ctx.wmask_lane_bit(_c(VCC_LO.offset), lane, val, exec_mask))
  if vcc_val is not None:
    # Initialize sdst to 0 before lane loop (old value may be unrelated data), then set lane bits in loop
    init_stores = [ctx.wsgpr_dyn(sdst_off, _c(0)), ctx.wsgpr_dyn(sdst_off + _c(1), _c(0))]
-    old_sdst = ctx.rmask(sdst_off)
-    stores.extend(ctx.wmask(sdst_off, _set_lane_bit(old_sdst, lane, vcc_val, exec_mask)))
+    stores.extend(ctx.wmask_lane_bit(sdst_off, lane, vcc_val, exec_mask))
    if stores:
      return UOp.sink(*init_stores, UOp.sink(*stores).end(lane), *ctx.inc_pc())
    return UOp.sink(*init_stores, *ctx.inc_pc())
@@ -1,6 +1,6 @@
 import unittest, itertools, math
 from tinygrad import Tensor, dtypes, Context
-from tinygrad.dtype import DType, ConstType, truncate
+from tinygrad.dtype import DType, ConstType
 from tinygrad.uop.ops import Ops, UOp
 from test.helpers import full_rewrite
 import numpy as np
@@ -51,16 +51,13 @@ class TestWeakConstFolding(unittest.TestCase):
  def test_invalid_poison(self):
    self.assertTrue(UOp.invalid().alu(Ops.CDIV, UOp.const(0)).simplify().is_invalid)

-  def test_cast_commits_to_dtype_grid(self):
-    # committing a weak const to a stated width puts the value on that width's grid, same as storage packing and native compilers
-    v = 1/123008  # not representable in float16
-    out = UOp.const(v).cast(dtypes.half).simplify()
-    self.assertEqual((out.op, out.dtype, out.val), (Ops.CONST, dtypes.half, truncate[dtypes.half](v)))
-    self.assertNotEqual(out.val, v)
-    # the grid commit preserves the sign of zero
-    self.assertEqual(math.copysign(1, UOp.const(-0.0).cast(dtypes.half).simplify().val), -1)
-    # observable at tensor level: the const-folded comparison agrees with the committed value
-    self.assertTrue((Tensor(-3.2).cast(dtypes.float32) <= truncate[dtypes.float32](-3.2)).item())
+  def test_single_rounding_log10_backward(self):
+    # log10 backward folds log10(2)/log(2) = 1/log(10) in one rounding, not the double-rounded 1/float32(log(10))
+    x = Tensor([1.0, 2.0, 3.0])
+    ast = next(s.src[0] for s in x.log10().sum().gradient(x)[0].schedule_linear().src if s.src[0].op is Ops.SINK)
+    const = next(u.arg for u in full_rewrite(ast).toposort() if u.op is Ops.CONST and u.dtype is dtypes.float32)
+    # correctly rounded: within half a float32 ulp of the exact value (folding at float32 lands 0.66 ulp off)
+    self.assertLess(abs(const - 1/math.log(10)), 2**-26)

 class TestBinaryOpsConstFolding(unittest.TestCase):
  def test_add_literal_zero(self):
@@ -143,13 +143,13 @@ class TestModuloAndDivisionFolding(unittest.TestCase):
    self.assertIs(apply_rewrite(x_var_uop.cast(dtypes.weakint) % 10).render(simplify=False), x_var_uop.render(simplify=False))

  def test_full_graph_rewrite_division_with_remainder(self):
-    x_var_uop = UOp.variable('x', 7, 9)
+    x_var_uop = UOp.variable('x', 7, 9, param=True)
    optimized_sink = apply_rewrite(x_var_uop // 2)
    for x_value in range(7, 10):
      self.assertEqual(x_value // 2, evaluate_uop(optimized_sink, {'x': x_value}))

  def test_full_graph_rewrite_complex_mod_div_expression(self):
-    x_var_uop = UOp.variable('x', 1, 10)
+    x_var_uop = UOp.variable('x', 1, 10, param=True)
    optimized_sink = apply_rewrite(((x_var_uop * 5) % 3) // 2)
    for x_value in range(1, 11):
      original_result = ((x_value * 5) % 3) // 2
@@ -25,7 +25,7 @@ def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UO
  ))

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

 class TestValidIdxSimplification(unittest.TestCase):
@@ -157,7 +157,7 @@ class TestGraphRewrite(unittest.TestCase):
    self.assertEqual(nout.val, 3.0)

  def test_depth_2_fold(self):
-    v = UOp.variable("v", 0, 1, dtypes.float)
+    v = UOp.variable("v", 0, 1, dtypes.float, param=True)
    c1 = UOp.const(1.0)
    c2 = UOp.const(2.0)
    nout = graph_rewrite(v+c1+c2, simple_pm)
@@ -339,7 +339,7 @@ class TestUOpGraph(unittest.TestCase):
    self.assertEqual(len([x for x in uops if x.op is Ops.CAST]), 1)

  def test_depth_2_const_fold(self):
-    v = UOp.variable("tmp", 0, 1, dtypes.int)
+    v = UOp.variable("tmp", 0, 1, dtypes.int, param=True)
    c2 = UOp.const(2, dtypes.int)
    c4 = UOp.const(4, dtypes.int)
    vc = v+c2
@@ -3,10 +3,10 @@ import unittest, pickle, functools, math
 import z3

 from tinygrad.dtype import dtypes, ConstType, DType, Invalid
-from test.helpers import get_uops
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad.uop.spec import spec_shared, type_verify
-from tinygrad.uop.symbolic import sym, pm_fold_cast_const, commutative, pm_simplify_valid, pm_move_where_on_load
+from tinygrad.uop.symbolic import sym, commutative, pm_simplify_valid, pm_move_where_on_load
+from tinygrad.uop.weak import pm_cast_weak
 from tinygrad.uop.validate import uops_to_z3

 def check_uop_against_string(self, v:UOp, s:str):
@@ -16,7 +16,8 @@ def check_uop_against_string(self, v:UOp, s:str):
  s_eval = graph_rewrite(s_eval, commutative, name="cannonicalize eval")
  self.assertIs(s_eval, v, f"eval did not match simplified: {s_eval} != {v.render()} for {s}")

-def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint): return UOp.variable(name,min_val,max_val,dtype)
+def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint):
+  return UOp.variable(name, min_val, max_val, dtype, param=True)
 def uconst(val): return UOp.const(val)
 def usum(ops): return functools.reduce(lambda x,y: x+y, ops)
 def uand(ops): return functools.reduce(lambda x,y: x*y, ops)
@@ -35,7 +36,7 @@ class TestSymbolic(unittest.TestCase):
    self.assertEqual(solver.check(expr1 != expr2), z3.unsat, "simplified expression not equal to original")

  def helper_test_variable(self, v, n, m, s, test_z3:bool=True):
-    v_simplified = graph_rewrite(v, sym+pm_fold_cast_const, name="simplify symbolic uop")
+    v_simplified = graph_rewrite(v, sym+pm_cast_weak, name="simplify symbolic uop")
    if test_z3: self.check_equal_z3(v, v_simplified)
    nmin, nmax = v_simplified.vmin, v_simplified.vmax
    check_uop_against_string(self, v_simplified, s)
@@ -442,7 +443,7 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable(uand([uconst(1), Variable("a", 0, 1)]), 0, 1, "a")

  def test_masked_shr_fold(self):
-    x = UOp.variable('x', 0, 255, dtype=dtypes.uint32)
+    x = UOp.variable('x', 0, 255, dtype=dtypes.uint32, param=True)
    self.helper_test_variable((x & -4) >> 2, 0, 63, "(x>>2)")

  def test_bool_or_not_tautology(self):
@@ -483,12 +484,12 @@ class TestSymbolic(unittest.TestCase):

  def test_div_drop_small_terms(self):
    # from openpilot, shouldnt simplify
-    gidx0 = UOp.variable("gidx0", 0, 10)
-    gidx1 = UOp.variable("gidx1", 0, 10)
-    lidx0 = UOp.variable("lidx0", 0, 1)
-    lidx1 = UOp.variable("lidx1", 0, 1)
-    ridx1005 = UOp.variable("ridx1005", 0, 2)
-    ridx1006 = UOp.variable("ridx1006", 0, 2)
+    gidx0 = UOp.variable("gidx0", 0, 10, param=True)
+    gidx1 = UOp.variable("gidx1", 0, 10, param=True)
+    lidx0 = UOp.variable("lidx0", 0, 1, param=True)
+    lidx1 = UOp.variable("lidx1", 0, 1, param=True)
+    ridx1005 = UOp.variable("ridx1005", 0, 2, param=True)
+    ridx1006 = UOp.variable("ridx1006", 0, 2, param=True)
    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)")

@@ -997,7 +998,7 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable(cond.ne(False), 0, 1, "(x<2)")

  def test_bitcast_chain(self):
-    a = UOp.variable("a", 0, 3, dtype=dtypes.int32)
+    a = UOp.variable("a", 0, 3, dtype=dtypes.int32, param=True)
    self.assertIs(graph_rewrite(a.bitcast(dtypes.float32).bitcast(a.dtype), sym), a)

  def test_negation_in_where(self):
@@ -1012,22 +1013,6 @@ class TestSymbolic(unittest.TestCase):
    b = Variable("b", 0, 3)
    self.helper_test_variable(-a<-b, False, True, "(b<a)")

-  def test_where_cast(self):
-    s = Variable("s", 0, 3, dtypes.int)
-    cond = s < 2
-    a = Variable("a", 0, 3, dtypes.int)
-    b = Variable("b", 0, 3, dtypes.int)
-    expr = cond.where(a, b).cast(dtypes.half)
-
-    # TODO: copied from render, render does not support cast
-    glbl = UOp.param(0, dtypes.int, (1,))
-    uops = get_uops(UOp(Ops.STORE, src=(glbl.index(UOp.const(0, dtypes.int)), expr)).sink())
-    rewritten_uop = [uop for uop in uops if uop.op is Ops.STORE][0].src[1]
-
-    # the vars are now scalar PARAMs
-    pvar = {u.expr: u for u in rewritten_uop.toposort() if u.op is Ops.PARAM}
-    self.assertEqual(rewritten_uop, (pvar['s']<UOp.const(2, dtypes.int)).where(pvar['a'].cast(dtypes.half), pvar['b'].cast(dtypes.half)))
-
  def test_where_merge_branches(self):
    cond1 = Variable("s", 0, 10) < 6
    cond2 = Variable("s", 0, 10) > 2
@@ -1180,7 +1165,7 @@ class TestSymbolicVariables(unittest.TestCase):
    assert (a//4 + a//6).variables() == [a]

  def test_variable_min_eq_max_bind_folds(self):
-    b = Variable("x", 1, 1).bind(1)
+    b = UOp.variable("x", 1, 1).bind(1)
    s = b.simplify()
    self.assertEqual(s.op, Ops.CONST)
    self.assertEqual(s.val, 1)
@@ -1374,6 +1359,16 @@ class TestInvalidIndex(unittest.TestCase):
    c2 = UOp.const((1, Invalid, 1, 1))
    self.assertIs((c1+c2).simplify(), UOp.const((2, Invalid, Invalid, Invalid)))

+  def test_gated_load_keeps_index_valid(self):
+    # the load executes even on gated-off iterations: gated_given_valid must not erase its mask (PADTO OOB shape)
+    buf = UOp.param(0, dtypes.bool, (17,))
+    ridx = Variable("ridx", 0, 31)
+    cond = ridx < 17
+    load = buf.index(ridx.valid(cond))
+    out = graph_rewrite(cond.where(load.where(uconst(2), uconst(0)), UOp.invalid()), sym)
+    idx = next(u for u in out.toposort() if u.op is Ops.INDEX)
+    self.assertIs(idx.src[1].get_valid(), cond.simplify())
+
 class TestStoreLoadFolding(unittest.TestCase):
  """Tests for store(index, load(index)) -> NOOP rule. This rule matches patterns that EMERGE during simplification."""
  def test_store_load_folding(self):
@@ -1,10 +1,12 @@
 import unittest
-from tinygrad import dtypes, Variable
+from tinygrad import dtypes
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import Ops, UOp, AxisType
 from test.helpers import to_uops_list

+def Variable(name, nmin, nmax): return UOp.variable(name, nmin, nmax, param=True)
+
 class TestValidateOOB(unittest.TestCase):
  """Test z3 validation of index bounds for different ALU ops and patterns."""

@@ -305,10 +305,10 @@ class TestVizTree(unittest.TestCase):

  def test_tree_view(self):
    with save_viz() as viz:
-      a = UOp.variable("a",0,10)
-      b = UOp.variable("b",0,10)
-      c = UOp.variable("c",0,10)
-      d = UOp.variable("d",0,10)
+      a = UOp.variable("a",0,10,param=True)
+      b = UOp.variable("b",0,10,param=True)
+      c = UOp.variable("c",0,10,param=True)
+      d = UOp.variable("d",0,10,param=True)
      sink = UOp.sink(a+b, c+d)
      def tree_rewrite(): return graph_rewrite(sink, root, name="root")
      tree_rewrite()
@@ -540,6 +540,10 @@ class TestAssign(unittest.TestCase):
    c = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
    c[0:2].bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000], dtype=dtypes.uint32)).realize()
    np.testing.assert_allclose(c.numpy(), [4.0, 3.0, 3.0, 4.0])
+    # without .realize()
+    a = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
+    a.bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000, 0x40000000, 0x3f800000], dtype=dtypes.uint32))
+    np.testing.assert_allclose(a.numpy(), [4.0, 3.0, 2.0, 1.0])

  def test_assign_bitcast_different_size(self):
    # assign to a shape-changing bitcast view (only works on DISK currently)
@@ -76,6 +76,11 @@ class TestWeakPromotion(unittest.TestCase):
      committed = graph_rewrite((UOp.const(1).cast(dtypes.int32) + UOp.const(1.0)).cast(dtypes.float32), pm_lower_index_dtype, ctx={})
    self.assertEqual([u.dtype for u in committed.toposort() if u.op is Ops.ADD], [dtypes.float32])

+  def test_div_sub_operand_kept_weak(self):
+    a = Tensor.empty(4, dtype=dtypes.float32)
+    for t in (a / 1, a - 0):
+      self.assertEqual(t.uop.src[1].dtype, dtypes.weakfloat)
+
  def test_cast_weak_expression_commits_at_cast_floor(self):
    # the floor never narrows: a cast BELOW the default does not pull the compute width down with it
    with Context(DEFAULT_FLOAT=dtypes.float32):
@@ -88,6 +93,13 @@ class TestWeakPromotion(unittest.TestCase):
      out = Tensor(1.0, dtype=dtypes.float32, device="CPU") / denom
      self.assertAlmostEqual(out.item(), 1 / (70000 + 1e-5), places=10)

+  def test_stacked_weak_casts_convert_each_kind(self):
+    # each weak cast is a kind conversion: weakint truncates before weakfloat re-lifts (neither is only a marker)
+    x = Tensor([2.5, -3.7], dtype=dtypes.float32, device="CPU")
+    stacked = x.cast(dtypes.weakint).cast(dtypes.weakfloat)
+    self.assertIs(stacked.dtype, dtypes.weakfloat)
+    self.assertEqual(stacked.tolist(), [2.0, -3.0])
+
  def test_uop_scalar_const_lifts_kind(self):
    for dtype, value, out_dtype, const_dtype in ((dtypes.weakint, 1, dtypes.weakint, dtypes.weakint),
                                                 (dtypes.int32, 1, dtypes.int32, dtypes.weakint),
@@ -51,6 +51,10 @@ class TestTensorGradient(unittest.TestCase):
    with self.assertRaises(RuntimeError): x.sum().gradient(x)
    with self.assertRaises(RuntimeError): x.float().sum().gradient(x)

+  def test_const_target_raise(self):
+    t = Tensor(2.0)
+    with self.assertRaises(RuntimeError): (t * 2.0).gradient(t)
+
  def test_copy_to_device_gradient(self):
    t = Tensor([1.0, 2, 3]).realize()
    t.to("CPU:1").square().sum().backward()
@@ -100,7 +104,7 @@ class TestTensorGradient(unittest.TestCase):

  def test_implicit_broadcast_where_gradient(self):
    # WHERE with a bare ()-shape branch: the scalar's gradient counts the positions where it is selected
-    cond, x, w = Tensor([True, False, True]), Tensor([1.0, 2.0, 3.0]), Tensor(4.0)
+    cond, x, w = Tensor([True, False, True]), Tensor([1.0, 2.0, 3.0]), Tensor(4.0, dtype=dtypes.float32)
    dw = Tensor(cond.uop.alu(Ops.WHERE, x.uop, w.uop)).sum().gradient(w)[0]
    self.assertEqual(dw.shape, ())
    self.assertEqual(dw.item(), 1.0)
@@ -109,7 +113,7 @@ class TestTensorGradient(unittest.TestCase):

  def test_implicit_broadcast_alu_gradient(self):
    # MUL with a bare ()-shape src, no EXPAND in the graph
-    x, w = Tensor([1.0, 2.0, 3.0]), Tensor(2.0)
+    x, w = Tensor([1.0, 2.0, 3.0]), Tensor(2.0, dtype=dtypes.float32)
    m = x.uop.alu(Ops.MUL, w.uop)
    self.assertIs(m.src[1], w.uop)
    dw = Tensor(m).sum().gradient(w)[0]
@@ -118,7 +122,7 @@ class TestTensorGradient(unittest.TestCase):

  def test_implicit_broadcast_intermediate_accumulation(self):
    # s is used directly and through an implicit broadcast edge, each edge's gradient reduces to s's shape before they sum
-    x, p = Tensor([1.0, 2.0, 3.0]), Tensor(0.5)
+    x, p = Tensor([1.0, 2.0, 3.0]), Tensor(0.5, dtype=dtypes.float32)
    s = p.sin()
    z = Tensor(x.uop.alu(Ops.MUL, s.uop)).sum() + s
    dp = z.gradient(p)[0]
@@ -25,10 +25,10 @@ class TestHCQUnit(unittest.TestCase):
    cpu_call = UOp(Ops.PROGRAM, src=(UOp.sink(),)).call(UOp.new_buffer("CPU", 1, dtypes.float))
    gpu_devs = [d0]

-    # local MMIO: GPU works alone and with CPU in batch (cpu_support=True)
+    # CPU uses HCQ2 and is no longer batched into legacy HCQ graphs.
    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
+    assert HCQGraph.supports_uop(gpu_devs, cpu_call) is False
+    assert HCQGraph.supports_uop(gpu_devs + [cpu_dev], gpu_call) is False

    # USB MMIO: GPU-only still works, but CPU batching must be rejected (cpu_support=False)
    orig_view = d0.timeline_signal.base_buf.view
@@ -1,11 +1,16 @@
 import unittest
 import functools
-from tinygrad import Tensor, Variable, UOp
+from tinygrad import Tensor, Variable, UOp, function
 from tinygrad.uop.ops import KernelInfo
 from tinygrad.schedule import schedule_cache

-def custom_set0_kernel(A:UOp, num:int) -> UOp:
-  return A[0].set(num).sink(arg=KernelInfo(f"custom_set0_{num}"))
+def custom_add_kernel(A:UOp, B:UOp, num:int=0) -> UOp:
+  return A[0].set(B[0] + num).sink(arg=KernelInfo(f"custom_add_{num}"))
+
+def custom_add_backward(grad_output:UOp, _) -> tuple[None, UOp]:
+  grad = Tensor.invalids(*grad_output.shape, dtype=grad_output.dtype, device=grad_output.device)
+  grad = Tensor.custom_kernel(grad, Tensor(grad_output, device=grad_output.device), fxn=functools.partial(custom_add_kernel, num=0))[0]
+  return None, grad.uop

 class TestScheduleCache(unittest.TestCase):
  def test_bound_variable_reuses_cache(self):
@@ -25,27 +30,27 @@ class TestScheduleCache(unittest.TestCase):

  def test_custom_kernel(self):
    for i in range(4):
-      a = Tensor.empty(1)
-      a = Tensor.custom_kernel(a, fxn=functools.partial(custom_set0_kernel, num=i))[0]
+      a, b = Tensor.empty(1), Tensor.ones(1)
+      a = Tensor.custom_kernel(a, b, fxn=functools.partial(custom_add_kernel, num=i))[0]
      a.realize()
-      self.assertEqual(a.item(), i)
+      self.assertEqual(a.item(), i+1)

  def test_same_custom_function_reuses_cache(self):
    schedule_cache.clear()
-    fxn = functools.partial(custom_set0_kernel, num=10)
+    fxn = functools.partial(custom_add_kernel, num=10)

    # first run
-    a = Tensor.empty(1)
-    a = Tensor.custom_kernel(a, fxn=fxn)[0]
+    a, x = Tensor.empty(1), Tensor.ones(1)
+    a = Tensor.custom_kernel(a, x, fxn=fxn)[0]
    a.realize()
-    self.assertEqual(a.item(), 10)
+    self.assertEqual(a.item(), 11)
    cache_size_after_first = len(schedule_cache)

    # second run with same function should reuse cache
-    b = Tensor.empty(1)
-    b = Tensor.custom_kernel(b, fxn=fxn)[0]
+    b, x = Tensor.empty(1), Tensor.ones(1)
+    b = Tensor.custom_kernel(b, x, fxn=fxn)[0]
    b.realize()
-    self.assertEqual(b.item(), 10)
+    self.assertEqual(b.item(), 11)
    self.assertEqual(len(schedule_cache), cache_size_after_first)

  def test_simple(self):
@@ -65,5 +70,29 @@ class TestScheduleCache(unittest.TestCase):
      print(num)
    self.assertEqual(len(schedule_cache), start_len_schedule_cache)

+  def test_simple_precompile(self):
+    @function(precompile=True, precompile_backward=True)
+    def f(x:Tensor) -> Tensor:
+      out = Tensor.invalids(*x.shape, dtype=x.dtype, device=x.device)
+      out = Tensor.custom_kernel(out, x, fxn=functools.partial(custom_add_kernel, num=10), grad_fxn=custom_add_backward)[0]
+      return out + x
+
+    # warmup
+    x = Tensor.ones(1).realize()
+    out = f(x)
+    out.backward(x)
+    self.assertEqual(out.item(), 12)
+    self.assertEqual(x.grad.item(), 2)
+
+    # use the cache next time function is called
+    start_len_schedule_cache = len(schedule_cache)
+    for _ in range(3):
+      x = Tensor.ones(1).realize()
+      out = f(x)
+      out.backward(x)
+      self.assertEqual(out.item(), 12)
+      self.assertEqual(x.grad.item(), 2)
+    self.assertEqual(len(schedule_cache), start_len_schedule_cache)
+
 if __name__ == "__main__":
  unittest.main()
@@ -12,7 +12,7 @@ from tinygrad.dtype import dtypes, AddrSpace

 # import all pattern matchers here
 from tinygrad.codegen.gpudims import pm_add_gpudims
-from tinygrad.uop.symbolic import sym, symbolic_simple, symbolic, pm_fold_cast_const, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
+from tinygrad.uop.symbolic import sym, symbolic_simple, symbolic, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
 from tinygrad.uop.movement import mop_cleanup
 from tinygrad.codegen.decomp.dtype import pm_dtype_decomps
 from tinygrad.codegen.decomp.op import get_late_rewrite_patterns, get_simplifying_rewrite_patterns
@@ -301,7 +301,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
    sink = graph_rewrite(sink, pm_split_ranges+pm_flatten_range, ctx={}, name="split ranges")

    # symbolic (NOTE: this is a requirement for pm_simplify_ranges to be correct)
-    sink = graph_rewrite(sink, sym+pm_fold_cast_const+pm_flatten_range, name="initial symbolic")
+    sink = graph_rewrite(sink, sym+pm_flatten_range, name="initial symbolic")

    # optimize (schedule) the AST
    sink = graph_rewrite(sink, pm_flatten_range+pm_simplify_ranges, ctx={}, name="simplify ranges")
@@ -346,7 +346,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:

  # lower index dtype
  # NOTE: we need indexing_simplify to remove the cast to long using the Invalid
-  sink = graph_rewrite(sink, symbolic_simple+pm_fold_cast_const+pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")
+  sink = graph_rewrite(sink, symbolic_simple+pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")

  # final symbolic before decomp
  sink = graph_rewrite(sink, symbolic, name="final symbolic")
@@ -357,7 +357,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:

  # floordiv+mod / dtype decomp (early)
  supported_ops = tuple(ren.code_for_op.keys())
-  pm_decomp = symbolic_simple+pm_fold_cast_const+get_simplifying_rewrite_patterns(supported_ops)
+  pm_decomp = symbolic_simple+get_simplifying_rewrite_patterns(supported_ops)
  sink = graph_rewrite(sink, pm_decomp, name="early decompositions")

  # late decomps + move gates from unrenderable INVALID where
@@ -57,7 +57,7 @@ def add_gpudims(ctx:Renderer, s:UOp):

  # get the idxs
  ki: KernelInfo = s.arg
-  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int).cast(dtypes.weakint)]
+  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int, param=True).cast(dtypes.weakint)]
  elif ki.dont_use_locals:
    assert not local_dims, "can't use locals if there's no local dims"
    idxs = get_grouped_dims("idx", global_shape, ctx.global_max, reverse=True)
@@ -89,7 +89,7 @@ def add_gpudims(ctx:Renderer, s:UOp):

 pm_device_to_var = PatternMatcher([
  # the DEVICE axis is not a program axis, it's bound per device at launch. lower it to the _device_num variable (like SPECIAL for devices)
-  (UPat(Ops.RANGE, name="r"), lambda r: UOp.variable("_device_num", 0, r.vmax, dtype=r.dtype) if r.arg[-1] is AxisType.DEVICE else None),
+  (UPat(Ops.RANGE, name="r"), lambda r: UOp.variable("_device_num", 0, r.vmax, dtype=r.dtype, param=True) if r.arg[-1] is AxisType.DEVICE else None),
  # ENDs that closed a DEVICE range no longer close it
  (UPat(Ops.END, name="e"), lambda e: e.replace(src=(e.src[0],)+tuple(s for s in e.src[1:] if s.op is not Ops.PARAM))
   if any(s.op is Ops.PARAM and s.arg.name == '_device_num' for s in e.src[1:]) else None),
@@ -26,7 +26,7 @@ def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
    # check if idx is out of bound when X is on the wrong side of the bound: X in [c+1, vmax] or [vmin, c-1]
    lo, hi = (c + 1, X.vmax) if is_upper_bound else (X.vmin, c - 1)
    if lo <= hi:
-      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype)
+      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype, param=True)
      subs = [{X: fake}]
      # idx may not have X itself, so also substitute a term of X: v -> fake - (X - v)
      terms = list(X.split_uop(Ops.ADD))
@@ -3,7 +3,7 @@ from tinygrad.uop.ops import PatternMatcher, UPat, Ops
 from tinygrad.dtype import Invalid, dtypes

 def move_where_load(gate, l, a, w):
-  return l.replace(src=(l.src[0], l.vconst_like(0) if a.is_invalid else
+  return l.replace(src=(l.src[0], l.vconst_like(0) if a.is_invalid else l.const_like(a.val) if a.op is Ops.CONST else
    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(w.dtype)

 pm_move_gates_from_index = PatternMatcher([
@@ -1,7 +1,7 @@
 import itertools
 from typing import Callable
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, graph_rewrite, _substitute, range_start, AxisType
-from tinygrad.uop.symbolic import symbolic, pm_fold_cast_const, invalid_gate
+from tinygrad.uop.symbolic import symbolic, invalid_gate
 from tinygrad.helpers import partition
 from tinygrad.dtype import dtypes

@@ -32,7 +32,7 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
        s0, s1 = r0.src[0], r1.src[0]
        # do the merge
        new_range = r0.replace(src=(s0*s1,))
-        nidx = graph_rewrite(u, _substitute+symbolic+pm_fold_cast_const+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
+        nidx = graph_rewrite(u, _substitute+symbolic+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
                             name=f"check_merge_{r0.arg[0]}_{r1.arg[0]}")

        # check if it simplifies
@@ -137,7 +137,7 @@ def reduce_collapse(red:UOp, u:UOp, pm:PatternMatcher=pm_reduce_collapse) -> UOp
    for u in included:
      for s in u.src:
        if s in included or s in replaces or s.op in {Ops.CONST, Ops.PARAM, Ops.BUFFER}: continue
-        replaces[s] = UOp.variable(f'in{len(replaces)}', s.vmin, s.vmax, s.dtype)
+        replaces[s] = UOp.variable(f'in{len(replaces)}', s.vmin, s.vmax, s.dtype, param=True)
    collapse_fxn = u.substitute(replaces).reduce(r, arg=Ops.ADD)
    sink = graph_rewrite(collapse_fxn, pm, name="reduce_collapse")
    if not no_range(sink): return None
@@ -44,7 +44,7 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
    current_batch, current_batch_devs = [], []

  for si in linear.src:
-    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,))])
+    devs = dedup([Device[x] for b in si.src[1:] if not b.is_bound_var 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_uop(devs, si)
@@ -12,7 +12,7 @@ from tinygrad.codegen.opt.postrange import args_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_arg_uops(call:UOp) -> tuple[UOp, ...]: return tuple(s for s in call.src[1:] if not s.is_bound_var)

 def get_call_outs_ins(call:UOp) -> tuple[tuple[int, ...], tuple[int, ...]]:
  ast = call.src[0]
@@ -1,4 +1,5 @@
 import functools, time
+from dataclasses import replace
 from typing import Generic, TypeVar, Callable, cast, overload
 from tinygrad.helpers import Context, dedup, getenv, DEBUG
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, PatternMatcher, UPat
@@ -12,7 +13,7 @@ def add_to_ctx(ctx, x:UOp):
  return ret

 pm_ctx = PatternMatcher([
-  (UPat((Ops.BUFFER, Ops.BIND), name="x"), add_to_ctx),
+  (UPat(Ops.BUFFER, name="x"), add_to_ctx),
  (UPat((Ops.AFTER, Ops.CONTIGUOUS), name="x"),
   lambda ctx,x: add_to_ctx(ctx,x) if not x.op_in_backward_slice_with_self(Ops.PARAM) and x.op_in_backward_slice_with_self(Ops.BUFFER) else None),
 ])
@@ -23,6 +24,10 @@ def invalid_outputs(uret:UOp) -> set[UOp]:
  return {u.src[0].buf_uop for u in uret.backward_slice_with_self
          if u.op is Ops.STORE and u.src[1].base.is_invalid and not u.src[0].buf_uop.is_realized}

+def renumber_invalid_outputs(uret:UOp) -> UOp:
+  return uret.substitute({b:b.replace(arg=replace(b.arg, slot=i))
+                          for i,b in enumerate(x for x in uret.toposort(enter_calls=False) if x in invalid_outputs(uret))})
+
 ReturnType = TypeVar('ReturnType')
 class _function(Generic[ReturnType]):
  depth = 0
@@ -65,6 +70,7 @@ class _function(Generic[ReturnType]):
    # the BUFFERs that are left are the implicit inputs
    num_explicit = len(call_uops)
    uret = graph_rewrite(uret, pm_ctx, (call_uops, invalid_outputs(uret)), bottom_up=True, name="get_implicit_inputs")
+    uret = renumber_invalid_outputs(uret)
    name = getattr(self.fxn, '__qualname__', None) or type(self.fxn).__qualname__
    if not self.allow_implicit:
      implicit_buffers = [x for x in call_uops[num_explicit:] if x.op is Ops.BUFFER]
@@ -115,7 +115,8 @@ class ElementwiseMixin(CreationMixin):
    ```
    """
    a, b = self._broadcasted(x, reverse)
-    return a + (-b)
+    # alu, not +: _broadcasted already promoted these, and a second promote would cast -b (only a bare weak CONST is kept weak)
+    return a.alu(Ops.ADD, -b)

  def mul(self, x: Self | ConstType, reverse: bool = False) -> Self:
    """
@@ -245,8 +246,9 @@ class ElementwiseMixin(CreationMixin):
    if dtypes.is_int(a.dtype) and dtypes.is_int(b.dtype):
      if rounding_mode == "trunc": return a.alu(Ops.CDIV, b)
      if rounding_mode == "floor": return a.alu(Ops.FLOORDIV, b)
-      a = a.cast(dtypes.default_float)
-    d = a * b.reciprocal()
+    if dtypes.is_int(a.dtype) or a.dtype == dtypes.bool: a = a.cast(dtypes.default_float)
+    # alu, not *: _broadcasted already promoted these, and a second promote would cast 1/b (only a bare weak CONST is kept weak)
+    d = a.alu(Ops.MUL, b.reciprocal())
    if rounding_mode is None: return d
    if rounding_mode == "trunc": return d.trunc()
    if rounding_mode == "floor": return d.floor()
@@ -3,6 +3,7 @@ import math, dataclasses
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, all_metadata, broadcast_axes
 from tinygrad.helpers import argsort
 from tinygrad.dtype import sum_acc_dtype
+from tinygrad.function import renumber_invalid_outputs

 def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
  if op == Ops.ADD: return (ctx._broadcast_to(ret.src[0].shape),)
@@ -40,6 +41,7 @@ def call_gradient(ctx:UOp, k:UOp, needed:set[int]) -> tuple[UOp|None, ...]:
  # collect needed gradient bodies, compact unused params, create a single backward CALL
  grad_bodies = [(i, grads[p]) for i in needed if (p:=params.get(i)) is not None and p in grads]
  bwd_body = UOp.maketuple(*(gb for _, gb in grad_bodies)).substitute(fwd_subs, walk=True)
+  bwd_body = renumber_invalid_outputs(bwd_body)
  bwd_body, compact_args = _compact_params(bwd_body, (*args, *grad_args, *fwd_outs))
  bwd_call = bwd_body.call(*compact_args, name=(k.arg.name or "")+"_backward", precompile=k.arg.precompile_backward)
  gb_map = {i: idx for idx, (i, _) in enumerate(grad_bodies)}
@@ -460,6 +460,7 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
    """
    assert gradient is not None or self.shape == tuple(), "when no gradient is provided, backward must be called on a scalar tensor"
    if not (self.is_floating_point() and all(t.is_floating_point() for t in targets)): raise RuntimeError("only float Tensors have gradient")
+    if any(t.dtype in dtypes.weaks for t in targets): raise RuntimeError("cannot take gradient wrt a weak Tensor")
    from tinygrad.mixin.gradient import compute_gradient
    if gradient is None: gradient = self.const_like(1.0)
    target_uops = [t._uop for t in targets]
@@ -264,6 +264,7 @@ class ClangRenderer(CStyleLanguage):
  nan = '__builtin_nanf("")'

  # language options
+  barrier = "__atomic_thread_fence(__ATOMIC_SEQ_CST);"
  buffer_suffix = " restrict"
  type_map = {dtypes.bool:"_Bool", dtypes.half:"__fp16"}
  code_for_op = {**({k:v for k,v in CStyleLanguage.code_for_op.items() if k not in [Ops.EXP2, Ops.SIN, Ops.LOG2, Ops.TRUNC, Ops.RECIPROCAL]}),
@@ -142,7 +142,7 @@ base_rewrite = PatternMatcher([
  (UPat(Ops.IF, name="x"), lambda ctx,x: f"  br i1 {ctx[x.src[0]]}, label %ifbody_{ctx[x][1:]}, label %ifskip_{ctx[x][1:]}\nifbody_{ctx[x][1:]}:"),
  (UPat(Ops.ENDIF, name="x"), lambda ctx,x: f"  br label %ifskip_{ctx[x.src[0]][1:]}\nifskip_{ctx[x.src[0]][1:]}:"),

-  (UPat(Ops.BARRIER), lambda ctx: "")
+  (UPat(Ops.BARRIER), lambda ctx: "  fence seq_cst")
 ])

 class LLVMRenderer(Renderer):
@@ -238,8 +238,8 @@ class AMDLLVMRenderer(LLVMRenderer):
    (UPat(Ops.CAST, dtypes.fp8s, (UPat(dtype=dtypes.float),), name="x",), lambda ctx,x:
      f"  {ctx[x]} = call i8 @f32_to_fp8({ldt(x.src[0].dtype)}  {ctx[x.src[0]]}, i1 {'1' if x.dtype == dtypes.fp8e5m2 else '0'})"),
    (UPat(Ops.CAST, dtypes.float, (UPat.var("y", dtypes.fp8s),), name="x",), lambda ctx,x,y:
-      f"  {ctx[x.src[0]]}_i32 = zext i8 {ctx[x.src[0]]} to i32\n"
-      f"  {ctx[x]} = call float @llvm.amdgcn.cvt.f32.{'bf8' if y.dtype == dtypes.fp8e5m2 else 'fp8'}(i32 {ctx[x.src[0]]}_i32, i32 0)"),
+      f"  {ctx[x]}_i32 = zext i8 {ctx[x.src[0]]} to i32\n"
+      f"  {ctx[x]} = call float @llvm.amdgcn.cvt.f32.{'bf8' if y.dtype == dtypes.fp8e5m2 else 'fp8'}(i32 {ctx[x]}_i32, i32 0)"),
  ]) + base_rewrite
  extra_matcher = LLVMRenderer.extra_matcher + create_non_native_float_pats(dtypes.fp8s) + PatternMatcher([
    # amd llvm intrinsics llvm.log2/llvm.exp2 don't support double
@@ -137,7 +137,8 @@ class NIRRenderer(Renderer):
    (UPat(Ops.CAST, (dtypes.uchar, dtypes.ushort), src=(UPat.var("x", dtypes.floats),), name="c"), lambda x,c: x.cast(dtypes.int32).cast(c.dtype)),
    # load/store use pointer arithmetic, and the cast does nothing. NOTE: this doesn't apply to image indexing cause it's 1-D
    (UPat((Ops.INDEX, Ops.SHRINK), src=(UPat.var("buf"), UPat.var("off")), allow_any_len=True, name="x"), lambda x,buf,off: x.replace(
-      src=(buf,off.cast(dtypes.long))+x.src[2:]) if buf.addrspace != AddrSpace.REG and not is_image_shape(buf._shape) else None),
+      src=(buf,UOp.const(off.val, dtypes.long) if off.op is Ops.CONST else off.cast(dtypes.long))+x.src[2:])
+      if buf.addrspace != AddrSpace.REG and not is_image_shape(buf._shape) else None),
    # images need index to be int for nir (coordinates only: the INDEX keeps its access dtype)
    (UPat.var("buf").index(UPat.var("idx_y"), UPat.var("idx_x"), name="x"),
     lambda x,buf,idx_y,idx_x: x.replace(src=(buf, idx_y.cast(dtypes.int), idx_x.cast(dtypes.int)))),
@@ -1,11 +1,11 @@
 from __future__ import annotations
-import platform, sys, os, ctypes, ctypes.util, functools, mmap, threading, array, itertools
-from dataclasses import replace
-from typing import cast
-from tinygrad.helpers import to_mv, OSX, WIN, Context, mv_address, suppress_finalizing, unwrap, data64_le, partition
-from tinygrad.device import Buffer, BufferSpec, TinyELF
-from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, HCQArgsState, HCQSignal, HCQProgram, MMIOInterface
-from tinygrad.runtime.support.hcq import CLikeArgsState
+import platform, sys, os, ctypes, functools, mmap, threading, array, struct, time
+from dataclasses import dataclass, replace
+from typing import cast, Callable
+from tinygrad.helpers import to_mv, from_mv, OSX, WIN, Context, mv_address, suppress_finalizing, unwrap, data64_le, to_tuple
+from tinygrad.device import Buffer, BufferSpec, TinyELF, Program, Device
+from tinygrad.runtime.support.hcq import HCQBuffer, MMIOInterface
+from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, make_cmdbuf, make_signal
 from tinygrad.renderer.cstyle import ClangRenderer
 from tinygrad.renderer.llvmir import CPULLVMRenderer
 from tinygrad.renderer.nir import LVPRenderer
@@ -13,11 +13,15 @@ from tinygrad.renderer.isa.x86 import X86Renderer
 from tinygrad.runtime.support.elf import jit_loader
 from tinygrad.runtime.autogen import libc
 from tinygrad.codegen import do_to_program
+from tinygrad.engine.realize import pm_flatten_linear, get_call_arg_uops, get_runtime
 from tinygrad import UOp, dtypes
 from tinygrad.dtype import AddrSpace
-from tinygrad.uop.ops import sint, KernelInfo, Ops, UPat, PatternMatcher, graph_rewrite
+from tinygrad.uop.ops import KernelInfo, Ops, UPat, PatternMatcher, graph_rewrite

-MAX_ARGS, CMD_SIZE, RING_SLOTS = 63, 64, (16 << 10)
+MAX_ARGS, CMD_SIZE, RING_SLOTS, FUNCS = 63, 64, (16 << 10), (() if WIN else ('clock_gettime', 'sem_wait', 'sem_post'))
+
+# *****************
+# 1. workers

 def signal_prog():
  val = UOp.param(1, dtypes.int, (), vmin_vmax=(0, dtypes.int.max), name="value", addrspace=AddrSpace.ALU)
@@ -35,79 +39,86 @@ def timestamp_prog():
    val = ts.after(call)[0].load() * 1_000_000_000 + ts.after(call)[1].load()
  return UOp.param(0, dtypes.uint64, (1,))[0].store(val)

-def quit_prog():
-  fn = UOp.param(0, dtypes.uint64, (1 if WIN else 3,))
-  if WIN: return fn[0].load().call(UOp.const(0, dtypes.uint64), ret_dtype=dtypes.void) # ExitThread(0)
-  sem = UOp.param(1, dtypes.uint64, (1,))
-
-  close = fn[2].load().call(sem[0], ret_dtype=dtypes.void) # sem_close(sem)
-  return fn.after(close)[0].load().call(UOp.const(0, dtypes.uint64), ret_dtype=dtypes.void) # pthread_exit(0)
-
 def worker_prog():
  ring = UOp.param(0, dtypes.uint64, (RING_SLOTS * CMD_SIZE,), volatile=True)
-  wait, sem = UOp.param(1, dtypes.uint64, (1,), volatile=True), UOp.param(2, dtypes.uint64, (1,))
-  cur = UOp.range(2**64-1, 0, dtype=dtypes.uint64)
+  wait, done = UOp.param(1, dtypes.uint64, (1,), volatile=True), UOp.param(2, dtypes.uint64, (1,), volatile=True)
+  sem, cur = UOp.param(3, dtypes.uint64, (1,)), UOp.range(2**64-1, 0, dtype=dtypes.uint64) # sem is unused on windows, it has to come last

  # spin on windows, sem_wait to sleep on posix
  if WIN: ready = (v:=wait.after(lw:=UOp.loop(1), cur)[0].load()).end(lw, v <= cur)
  else: ready = (rv:=wait.after(lw:=UOp.loop(1), cur)[0].load().call(sem.after(cur)[0], ret_dtype=dtypes.int)).end(lw, rv != 0)

  entry = [ring.after(ready).index((cur % RING_SLOTS) * CMD_SIZE + i).load() for i in range(CMD_SIZE)]
-  return entry[0].call(*entry[1:], ret_dtype=dtypes.void).end(cur)
+  return done.after(entry[0].call(*entry[1:], ret_dtype=dtypes.void)).index(0).store(cur + 1).end(cur)

-def host_wait(ctx, dst:UOp, val:UOp) -> UOp:
-  return (cur:=dst.after(loop:=UOp.loop(next(ctx))).index(UOp.const(0, dtypes.int)).load()).end(loop, cur < val)
+@dataclass
+class CPUWorker: ring:Buffer; put:Buffer; sem:Buffer; sys:Buffer; done:Buffer; thread:threading.Thread  # noqa: E702

-pm_host_opsel = PatternMatcher([(UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))), host_wait)])
+# *****************
+# 2. queue encoders

-def encode_host_queue(q:UOp) -> UOp:
-  # TODO: subset of hcq2 for now
-  spins, (store,) = partition(graph_rewrite(q, pm_host_opsel, ctx=itertools.count(), walk=True, name="host opsel").src, lambda u: u.op is Ops.END)
-  assert store.op is Ops.INS and store.arg == "store", f"host queue cannot encode {store.op} {store.arg}"
-  return store.src[0].after(*spins).index(UOp.const(0, dtypes.int)).store(store.src[1])
+def cpu_cmd(devs:tuple[str, ...], prog, *args:UOp) -> UOp:
+  progs = [get_runtime(d, prog) if isinstance(prog, UOp) else cast(CPUDevice, Device[d]).prgs[prog] for d in devs]
+  addrs = tuple(UOp.const(p.addr, dtypes.uint64) for p in progs)
+  words = ((addrs[0] if len(addrs) == 1 else UOp(Ops.STACK, dtypes.uint64, addrs)),) + args
+  return UOp(Ops.INS, dtypes.void, words + (UOp.const(0, dtypes.uint64),) * (CMD_SIZE - len(words)), arg="cmd")

-class CPUComputeQueue(HWQueue):
-  def __init__(self, dev):
-    super().__init__()
-    self.dev = dev
-  def _cmd(self, prog, args=(), vals=()): return self.exec(prg:=self.dev.prgs[prog], prg.fill_kernargs(args, vals), None, None)
-  def memory_barrier(self): return self
-  def exec(self, prg:CPUProgram, args_state:HCQArgsState, global_size, local_size):
-    if (lvp:=isinstance(args_state, LVPArgsState)): self.bind_args_state(args_state)
-    args:list[sint|None] = [args_state.buf.va_addr] if lvp else [*[x.va_addr for x in args_state.bufs], *args_state.vals]
-    assert len(args) <= MAX_ARGS, f"CPU programs support at most {MAX_ARGS} arguments, got {len(args)}"
-    for tid in range(1 if lvp else (global_size or (1,))[0]):
-      if not lvp and 'core_id' in prg.runtimevars: args[prg.runtimevars['core_id']] = tid
-      self.q(prg, *[unwrap(x) for x in args], *([0] * (MAX_ARGS - len(args))))
-    return self
-  def wait(self, signal, value=0): return self._cmd(wait_prog, (signal.base_buf,), (value,))
-  def timestamp(self, signal): return self._cmd(timestamp_prog, (signal.base_buf.offset(8, 8), self.dev.func_table._buf.offset(0, 8)))
-  def signal(self, signal, value:sint=0): return self._cmd(signal_prog, (signal.base_buf,), (value,))
-  def _submit(self, dev):
-    dev.ensure_worker()
-    ring_view = dev.ring.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
-    for off in range(0, len(self._q), CMD_SIZE):
-      entry = [self._q[off].addr, *self._q[off+1:off+CMD_SIZE]]
-      ring_view[(base:=(dev.ring_pos % RING_SLOTS) * CMD_SIZE):base+CMD_SIZE] = array.array('Q', (int(x) & ((1<<64)-1) for x in entry))
-      dev.ring_pos += 1
-      if WIN: dev.sys.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = dev.ring_pos
-      else: assert libc.sem_post(dev.sem) == 0
+def cpu_exec(ctx:tuple[str, ...], call:UOp, prg:UOp) -> UOp:
+  args = [get_call_arg_uops(call)[i].getaddr(ctx) for i in prg.arg.globals] + [v.cast(dtypes.uint64) for v in prg.arg.vars]
+  if (core:=prg.arg.runtimevars.get('core_id')) is None: return cpu_cmd(ctx, prg, *args)

-class LVPArgsState(CLikeArgsState):
-  def __init__(self, buf, prg, bufs, vals=()): super().__init__(buf, prg, bufs, vals, [*data64_le(buf.va_addr + 12), (len(bufs) + len(vals)) * 2])
+  la = [cpu_cmd(ctx,prg,*args[:(cid:=(len(prg.arg.globals)+core))],UOp.const(t, dtypes.uint64),*args[cid+1:]) for t in range(prg.arg.global_size[0])]
+  return UOp(Ops.LINEAR, dtypes.void, tuple(la))
+
+pm_cpu_opsel = PatternMatcher([
+  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="prg"),), name="call", allow_any_len=True), cpu_exec),
+
+  (UPat(Ops.INS, arg="barrier"), lambda: UOp(Ops.NOOP, dtypes.void, ())),
+  (UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))),
+   lambda ctx, dst, val: cpu_cmd(ctx, wait_prog, dst.getaddr(ctx), val.cast(dtypes.uint64))),
+  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))),
+   lambda ctx, dst, val: cpu_cmd(ctx, signal_prog, dst.getaddr(ctx), val.cast(dtypes.uint64))),
+  (UPat(Ops.INS, arg="timestamp", src=(UPat(name="dst"),)),
+   lambda ctx, dst: cpu_cmd(ctx, timestamp_prog, dst.getaddr(ctx), *(() if WIN else (make_signal(ctx, tag="func:clock_gettime").getaddr(ctx),)))),
+])
+
+def encode_queue(q:UOp) -> UOp:
+  devs, queue = to_tuple(q.arg[0]), q.arg[1]
+  lin = graph_rewrite(q, pm_cpu_opsel+pm_flatten_linear, ctx=devs, walk=True, name=f"{queue} opsel")
+
+  cnt = sum(len(ins.src) for ins in lin.src) // CMD_SIZE
+  assert cnt < RING_SLOTS, f"submit of {cnt} entries doesn't fit the ring"
+  cmdbuf = make_cmdbuf(lin, devs, buf=UOp.placeholder((cnt*CMD_SIZE,), dtypes.uint64, next(UOp.unique_num), device=devs).rtag("cmdbuf"))
+  ring = UOp.placeholder((ring_words:=RING_SLOTS*CMD_SIZE,), dtypes.uint64, 0, device=devs, volatile=True).rtag(f"{queue}_ring")
+  put, done, sem, sysbuf = (make_signal(devs, tag=f"{queue}_{name}") for name in ("put", "done", "sem", "sys"))
+
+  # submits are serialized on the submitter, so they can bump put without atomics
+  ran = done.after(l:=UOp.loop(next(UOp.unique_num))).index(0).load()
+  room = ran.end(l, put.index(0).load() - ran > RING_SLOTS - cnt) # wait until cnt entries fit in the ring
+  base = ((put.after(room).index(0).load() % RING_SLOTS) * CMD_SIZE).cast(dtypes.int)
+  e = UOp.range(cnt, next(UOp.unique_num), dtype=dtypes.int, src=(cmdbuf, ring))
+  copy = UOp.group(*[ring.index((base + e*CMD_SIZE + w) % ring_words).store(cmdbuf.index(e*CMD_SIZE + w).load()) for w in range(CMD_SIZE)])
+
+  # wake the worker after each entry, keeping the post with the stores stops it from hoisting out of the loop
+  wake = copy.end(e) if WIN else make_signal(devs, tag="func:sem_post").after(copy).index(0).load().call(sem.index(0), ret_dtype=dtypes.void).end(e)
+  bumped = put.after(wake).index(0).store(put.index(0).load() + cnt)
+  return sysbuf.after(bumped).index(0).store(put.index(0).load() + cnt) if WIN else bumped
+
+# *****************

 # NOTE: MAP_JIT is added to mmap module in python 3.13
 MAP_JIT = 0x0800

-class CPUProgram(HCQProgram['CPUDevice']):
+class CPUProgram(Program['CPUDevice']):
  rt_lib = None
  try: rt_lib = ctypes.CDLL(ctypes.util.find_library('System' if OSX else 'kernel32') if OSX or WIN else 'libgcc_s.so.1')
  except OSError: pass

  def __init__(self, dev:CPUDevice, obj:TinyELF):
-    self.signature, self.runtimevars = obj.signature, {name:slot for name,slot,*_ in obj.signature if name == 'core_id'}
+    self.dev, self.name, self.signature = dev, obj.name, obj.signature
+    self.runtimevars = {name:slot for name,slot,*_ in obj.signature if name == 'core_id'}
+    self.lvp = obj.target.renderer == "LVP"

-    LVP = obj.target.renderer == "LVP"
    if sys.platform == "win32": # mypy doesn't understand when WIN is used here
      PAGE_EXECUTE_READWRITE, MEM_COMMIT, MEM_RESERVE = 0x40, 0x1000, 0x2000
      ctypes.windll.kernel32.VirtualAlloc.restype = ctypes.c_void_p
@@ -117,7 +128,7 @@ class CPUProgram(HCQProgram['CPUDevice']):
      ctypes.windll.kernel32.GetCurrentProcess.restype = ctypes.c_void_p
      proc = ctypes.windll.kernel32.GetCurrentProcess()
      ctypes.windll.kernel32.FlushInstructionCache(ctypes.c_void_p(proc), ctypes.c_void_p(self.addr), ctypes.c_size_t(len(obj.lib)))
-      self.fxn = ctypes.CFUNCTYPE(None)(self.addr)
+      self.fxn = ctypes.CFUNCTYPE(None, ctypes.c_void_p)(self.addr) if self.lvp else ctypes.CFUNCTYPE(None)(self.addr)
    else:
      # On apple silicon with SPRR enabled (it always is in macos) RWX pages are unrepresentable: https://blog.svenpeter.dev/posts/m1_sprr_gxf/
      # MAP_JIT allows us to easily flip pages from RW- to R-X and vice versa. It is a noop on intel cpus. (man pthread_jit_write_protect_np)
@@ -125,7 +136,7 @@ class CPUProgram(HCQProgram['CPUDevice']):
      self.addr = mv_address(self.mem)

      if OSX: unwrap(CPUProgram.rt_lib).pthread_jit_write_protect_np(False)
-      lib = jit_loader(obj.lib, base=ctypes.addressof(ctypes.c_void_p.from_buffer(self.mem)), link_libs=['m']) if LVP else obj.lib
+      lib = jit_loader(obj.lib, base=ctypes.addressof(ctypes.c_void_p.from_buffer(self.mem)), link_libs=['m']) if self.lvp else obj.lib
      self.mem.write(lib)
      if OSX: unwrap(CPUProgram.rt_lib).pthread_jit_write_protect_np(True)

@@ -138,15 +149,30 @@ class CPUProgram(HCQProgram['CPUDevice']):
        # msync should be a universal POSIX way to do this
        libc.msync(ctypes.c_void_p(self.addr), len(lib), libc.MS_SYNC | libc.MS_INVALIDATE)

-      self.fxn = ctypes.CFUNCTYPE(None)(self.addr)
+      self.fxn = ctypes.CFUNCTYPE(None, ctypes.c_void_p)(self.addr) if self.lvp else ctypes.CFUNCTYPE(None)(self.addr)

-    super().__init__(LVPArgsState if LVP else HCQArgsState, dev, obj, kernargs_alloc_size=12+256 if LVP else 0)
+  def __call__(self, *bufs:HCQBuffer, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1),
+               vals:tuple[int|None, ...]=(), wait:bool=False, timeout:int|None=None) -> float|None:
+    st = time.perf_counter()
+    if self.lvp:
+      lvp_args = bytearray(12 + (len(bufs) + len(vals)) * 8)
+      addr = mv_address(lvp_args)
+      struct.pack_into(f'<3I{len(bufs)}Q', lvp_args, 0, *data64_le(addr+12), (len(bufs)+len(vals))*2, *[b.va_addr for b in bufs])
+      for v,(off,dt) in zip(vals, TinyELF.iter_sig(self.signature[-len(vals):], len(bufs)*8)): struct.pack_into(f'<{dt.fmt}', lvp_args, 12+off, v)
+      self.fxn(addr)
+    else:
+      args = [*[cast(int, b.va_addr) for b in bufs], *cast(tuple[int, ...], vals)]
+      assert len(args) <= MAX_ARGS, f"CPU programs support at most {MAX_ARGS} arguments, got {len(args)}"
+      for tid in range(global_size[0]):
+        if 'core_id' in self.runtimevars: args[self.runtimevars['core_id']] = tid
+        self.fxn(*[ctypes.c_uint64(x) for x in args])
+    return time.perf_counter() - st if wait else None

  @suppress_finalizing
  def __del__(self):
    if sys.platform == 'win32': ctypes.windll.kernel32.VirtualFree(ctypes.c_void_p(self.addr), ctypes.c_size_t(0), 0x8000) #0x8000 - MEM_RELEASE

-class CPUAllocator(HCQAllocator):
+class CPUAllocator(HCQAllocator['CPUDevice']):
  def __init__(self, dev:CPUDevice): super().__init__(dev, supports_copy_from_disk=False, supports_transfer=False)
  def _alloc(self, size:int, options:BufferSpec) -> HCQBuffer:
    if options.external_ptr is not None: addr, buf = options.external_ptr, None
@@ -154,68 +180,55 @@ class CPUAllocator(HCQAllocator):
    else: addr = mv_address(buf:=mmap.mmap(-1, size, mmap.MAP_ANON | mmap.MAP_SHARED, mmap.PROT_READ | mmap.PROT_WRITE))
    return HCQBuffer(va:=addr, sz:=size, meta=buf, view=MMIOInterface(va, sz, fmt='B'), owner=self.dev)
  def _as_buffer(self, src) -> memoryview: return to_mv(src.va_addr, src.size)
+  def _copyin(self, dest:HCQBuffer, src:memoryview):
+    self.dev.synchronize()
+    ctypes.memmove(int(dest.va_addr), from_mv(src), len(src))
+  def _copyout(self, dest:memoryview, src:HCQBuffer):
+    self.dev.synchronize()
+    ctypes.memmove(from_mv(dest), int(src.va_addr), len(dest))
  def _do_map(self, buf:HCQBuffer):
    if buf.view is None or not isinstance(buf.view, MMIOInterface): raise RuntimeError("Cannot map buffer without view to cpu")
    return HCQBuffer(buf.view.addr, buf.size, view=buf.view, owner=buf.owner)
  def _unmap(self, mb): pass  # CPU _do_map returns a view wrapper, nothing to release

-class CPUDevice(HCQCompiled):
-  pm_lower = PatternMatcher([
-    (UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),)), encode_host_queue)])
-
-  pm_bufferize = PatternMatcher([
-    (UPat(Ops.PARAM, tag="sentinel_signal"), lambda ctx: ctx[0].signal("sentinel", (1 << 64) - 1)),
-    (UPat(Ops.PARAM, tag="timeline_signal"), lambda ctx: ctx[0].signal("timeline")),
-    (UPat(Ops.PARAM, tag="timeline_value"), lambda ctx: ctx[0].signal("value", 1)),
-    (UPat(Ops.PARAM, tag="signal", name="b"), lambda ctx, b: ctx[0].signal(b.arg.slot)),
-  ])
-
-  @functools.cache
-  def signal(self, name:str|int, init_value:int=0) -> Buffer:
-    (buf:=Buffer(self.device, 1, dtypes.uint64, preallocate=True)).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = init_value
-    return buf
+class CPUDevice(HCQ2Compiled):
+  wait_timeout_ms, has_copy_queue = 30000, False
+  pm_lower = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),)), encode_queue)])

  def __init__(self, device:str=""):
-    super().__init__(device, CPUAllocator(self), [ClangRenderer, CPULLVMRenderer, LVPRenderer, X86Renderer], CPUProgram, HCQSignal,
-      functools.partial(CPUComputeQueue, self), arch={'amd64':'x86_64', 'aarch64':'arm64'}.get(m:=platform.machine().lower(), m)+",native")
+    super().__init__(device, CPUAllocator(self), [ClangRenderer, CPULLVMRenderer, LVPRenderer, X86Renderer], CPUProgram,
+      arch={'amd64':'x86_64', 'aarch64':'arm64'}.get(m:=platform.machine().lower(), m)+",native")

-    self.ring_pos = 0
+    self.pm_bufferize = PatternMatcher(
+      [(UPat(Ops.PARAM, tag=f"COMPUTE:0_{n}"), lambda ctx, n=n: getattr(ctx[0].worker, n)) for n in ("ring", "put", "sem", "sys", "done")] +
+      [(UPat(Ops.PARAM, tag=f"func:{f}"), lambda ctx, f=f: ctx[0].func_ptr(f)) for f in FUNCS]) + self.pm_bufferize

-    # posix uses sem to put cpus into sleep
-    self.sem_addr = 0
-    if not WIN:
-      self.sem = libc.sem_open(sem_name:=f"/tinygrad-{os.getpid()}-{id(self):x}".encode(), os.O_CREAT|os.O_EXCL, 0o600, 0) # type: ignore[call-arg]
-      self.sem_addr = unwrap(ctypes.cast(self.sem, ctypes.c_void_p).value)
-      if self.sem_addr == ctypes.c_void_p(-1).value or libc.sem_unlink(sem_name): raise OSError(ctypes.get_errno(), "semaphore")
-
-    # TODO: move to hcq2
    with Context(EMULATED_DTYPES="", TRACK_MATCH_STATS=0):
-      prgs = {f: f().sink(arg=KernelInfo(f.__name__), tag=1) for f in (signal_prog, wait_prog, timestamp_prog, quit_prog, worker_prog)}
-      self.prgs = {f: self.runtime(do_to_program(v, ClangRenderer(replace(self.renderer.target, renderer="CLANG"))).to_elf()) for f,v in prgs.items()}
+      clang = ClangRenderer(replace(self.renderer.target, renderer="CLANG"))
+      self.prgs:dict[Callable, CPUProgram] = {f: CPUProgram(self, do_to_program(f().sink(arg=KernelInfo(f.__name__), tag=1), clang).to_elf())
+                                              for f in (signal_prog, wait_prog, timestamp_prog, worker_prog)}

-  @functools.cached_property
-  def ring(self) -> Buffer: return Buffer(self.device, RING_SLOTS * CMD_SIZE, dtypes.uint64, preallocate=True)
-  @functools.cached_property
-  def sys(self) -> Buffer: return Buffer(self.device, 1, dtypes.uint64, preallocate=True)
-  @functools.cached_property
-  def sem_buf(self) -> Buffer: return Buffer(self.device, 1, dtypes.uint8, options=BufferSpec(external_ptr=self.sem_addr), preallocate=True)
+  def func_ptr(self, name:str) -> Buffer: return self.func_table.view(1, dtypes.uint64, FUNCS.index(name)*8).ensure_allocated()

-  # TODO: move to hcq2 infra
  @functools.cached_property
  def func_table(self) -> Buffer:
-    fns = ([0, ctypes.windll.kernel32.ExitThread, 0, 0] if WIN else  # type: ignore[attr-defined]
-           [libc.dll.clock_gettime, libc.dll.pthread_exit, libc.dll.sem_wait, libc.dll.sem_close])
-    addrs = array.array('Q', [unwrap(ctypes.cast(f, ctypes.c_void_p).value) if f else 0 for f in fns])
-    (ft:=Buffer(self.device, len(fns), dtypes.uint64, preallocate=True)).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[:] = addrs
+    lib = ctypes.windll.kernel32 if sys.platform == "win32" else libc.dll # type: ignore[attr-defined]
+    (ft:=Buffer(self.device, len(FUNCS), dtypes.uint64, preallocate=True))._buf.cpu_view().view(fmt='Q')[:] = \
+      array.array('Q', [unwrap(ctypes.cast(getattr(lib, f), ctypes.c_void_p).value) for f in FUNCS])
    return ft

-  @functools.cache
-  def ensure_worker(self):
-    threading.Thread(target=cast(CPUProgram, self.prgs[worker_prog]).fxn, daemon=True, args=[ctypes.c_uint64(x) for x in
-      [self.ring._buf.va_addr, self.sys._buf.va_addr if WIN else self.func_table._buf.va_addr+16, self.sem_addr]]).start()
+  @functools.cached_property
+  def worker(self) -> CPUWorker:
+    ring, put, sysbuf, done = (Buffer(self.device, sz, dtypes.uint64, preallocate=True) for sz in (RING_SLOTS*CMD_SIZE, 1, 1, 1))
+    addr, hsem = 0, None

-  def finalize(self):
-    if self.ring_pos == 0: return # the worker starts with the first submit
-    ft = self.func_table._buf
-    CPUComputeQueue(self)._cmd(quit_prog, (ft.offset(8, 8),) if WIN else (ft.offset(8, 24), self.sem_buf._buf)).submit(self)
-    self.ring_pos = 0
+    # sem are posix-only
+    if not WIN:
+      hsem = libc.sem_open(nm:=f"/tinygrad-{os.getpid()}-{id(ring):x}".encode(), os.O_CREAT|os.O_EXCL, 0o600, 0) # type: ignore[call-arg]
+      if (addr:=unwrap(ctypes.cast(hsem, ctypes.c_void_p).value)) == ctypes.c_void_p(-1).value or libc.sem_unlink(nm):
+        raise OSError(ctypes.get_errno(), "semaphore")
+    sem = Buffer(self.device, 1, dtypes.uint64, options=BufferSpec(external_ptr=addr), preallocate=True)
+
+    worker_args = [ring._buf.va_addr, sysbuf._buf.va_addr if WIN else self.func_ptr('sem_wait')._buf.va_addr, done._buf.va_addr, addr]
+    (worker:=threading.Thread(target=self.prgs[worker_prog].fxn, daemon=True, args=[ctypes.c_uint64(x) for x in worker_args])).start()
+    return CPUWorker(ring, put, sem, sysbuf, done, worker)
@@ -49,7 +49,8 @@ class DiskDevice(Compiled):
    DiskDevice._tried_io_uring_init = True

    if sys.platform == 'linux' and not hasattr(sys, "getandroidapilevel"):
-      fd = libc.syscall(io_uring.NR_io_uring_setup, 4096, ctypes.byref(p:=io_uring.struct_io_uring_params()))
+      p = io_uring.struct_io_uring_params(flags=io_uring.IORING_SETUP_SQPOLL, sq_thread_idle=0xffffffff)
+      fd = libc.syscall(io_uring.NR_io_uring_setup, 4096, ctypes.byref(p))
      if fd < 0: return

      sq_ptr = libc.mmap(0, p.sq_off.array + p.sq_entries * 4, mmap.PROT_READ | mmap.PROT_WRITE, mmap.MAP_SHARED | MAP_POPULATE, fd, 0)
@@ -67,6 +68,7 @@ class DiskDevice(Compiled):
        kring_mask=u32ptr(sq_ptr+p.cq_off.ring_mask), cqes=ctypes.cast(cq_ptr+p.cq_off.cqes, ctypes.POINTER(io_uring.struct_io_uring_cqe)))

      DiskDevice.io_uring = io_uring.struct_io_uring(ring_fd=fd, sq=sqdesc, cq=cqdesc) # type: ignore
+      libc.syscall(io_uring.NR_io_uring_enter, fd, 0, 0, io_uring.IORING_ENTER_SQ_WAKEUP)

 class DiskBuffer:
  def __init__(self, device:DiskDevice, size:int, offset=0):
@@ -124,7 +126,6 @@ class DiskAllocator(Allocator):
        # Send sqe
        DiskDevice.io_uring.sq.array[sqe_index] = sqe_index
        DiskDevice.io_uring.sq.ktail[0] = tail + 1
-        libc.syscall(io_uring.NR_io_uring_enter, DiskDevice.io_uring.ring_fd, 1, 1, io_uring.IORING_ENTER_GETEVENTS)

        reqs.append((copy_batch, copied_in, minor_offset, real_copy_size:=min(sqe.len - minor_offset, size - copied_in)))
        next_read_offset += sqe.len
@@ -3,10 +3,10 @@
 # works to test the tensor cores, and all the uops in general
 # this is the (living) definition of uops
 from typing import Any, TYPE_CHECKING
-import pickle, base64, itertools, time, sys, functools
+import pickle, base64, itertools, time, sys, functools, ctypes
 from dataclasses import replace
 from tinygrad.dtype import bitcast, DType, dtypes, AddrSpace, truncate, storage_fmt_for_dtype, to_storage_scalar, from_storage_scalar
-from tinygrad.helpers import all_same, getenv, flatten, Target, IMAGE, is_image_shape, cpu_profile
+from tinygrad.helpers import all_same, getenv, flatten, Target, IMAGE, is_image_shape, cpu_profile, mv_address
 from tinygrad.device import Buffer, Compiled, Compiler, Allocator, Program, TinyELF
 from tinygrad.codegen.opt import tc
 from tinygrad.uop.ops import exec_alu, python_alu, Ops, UOp, GroupOp
@@ -134,6 +134,13 @@ class PythonProgram(Program['PythonDevice']):
                               for k in range(len(src_values))], j, u.dtype) for j in range(load_sz)]
          else:
            values[u] = load(src_values, 0, u.dtype)
+        elif u.op is Ops.CALL:
+          assert u.dtype is dtypes.void
+          cfunc = ctypes.CFUNCTYPE(None, *[ctypes.c_uint64] * (len(src_values)-1))
+          values[u] = []
+          for args,gate in zip(zip(*src_values), exec_masks[-1]):
+            call_args = [(mv_address(x[0]) + x[1]*dt.itemsize) if isinstance(x, tuple) else x for x,dt in zip(args, src_dtypes)]
+            values[u].append(cfunc(call_args[0])(*call_args[1:]) if gate else None)
        elif u.op is Ops.WMMA:
          first_src_dtype = u.src[0].dtype
          assert isinstance(first_src_dtype, DType) # mypy
@@ -1,4 +1,4 @@
-import ctypes, struct, platform, pathlib, subprocess, sys
+import ctypes, struct, platform, pathlib, shutil, subprocess, sys, tarfile, tempfile
 from tinygrad.device import Compiler
 from tinygrad.helpers import DEBUG, system, fetch, unwrap
 from tinygrad.runtime.support.compiler_mesa import disas_adreno
@@ -11,10 +11,13 @@ class QCOMCompiler(Compiler):
  def __init__(self, arch:str):
    assert arch.split(',')[0] == "a630", "only a630 supported"
    if platform.machine() == "aarch64": self.arch, self.chip_id, self.llvm_inst = arch, 0x6030001, llvm_qcom.cl_compiler_create_llvm_instance()
-    else: self.arch, self.chip_id, self.compiler_process = arch, 0x6030001, subprocess.Popen(
-      (f"docker run --rm -i --platform linux/aarch64 -e PYTHONPATH=/ -e QEMU_CPU=max,pauth=off -v {pathlib.Path(__file__).parents[2]}:/tinygrad "
-       f"-v {fetch('https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so')}:/lib/libllvm-qcom.so python:3.12-slim "
-       f"python /tinygrad/runtime/support/compiler_qcom.py {arch}").split(), stdout=subprocess.PIPE, stdin=subprocess.PIPE, bufsize=0)
+    else:
+      self.arch, self.chip_id, self.fs = arch, 0x6030001, tempfile.TemporaryDirectory()
+      with tarfile.open(fetch('https://git.tinygrad.win/sirhcm/images/releases/download/v2/qcomcl.tar.gz')) as t: t.extractall(fs:=self.fs.name)
+      if (qemu:=shutil.which("qemu-aarch64-static")): argv = f"{qemu} -cpu max,pauth=off -L {fs} {fs}/usr/bin/python3 {__file__} {arch}"
+      else: argv = (f"docker run --rm -i --platform linux/aarch64 -v {fs}/usr:/usr -v {pathlib.Path(__file__).parents[2]}:/tinygrad "
+                    f"-e PYTHONPATH=/ -e QEMU_CPU=max,pauth=off gcr.io/distroless/static python3 /tinygrad/runtime/support/compiler_qcom.py {arch}")
+      self.compiler_process = subprocess.Popen(argv.split(), stdout=subprocess.PIPE, stdin=subprocess.PIPE, bufsize=0)
    super().__init__(f"compile_qcomcl_{arch}")

  def __del__(self): llvm_qcom.cl_compiler_destroy_llvm_instance(self.llvm_inst) if platform.machine() == "aarch64" else self.compiler_process.kill()
@@ -517,8 +517,7 @@ class HCQCompiled(Compiled, Generic[SignalType]):
    if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()

 class HCQBuffer:
-  def __init__(self, va_addr:sint, size:int, meta:Any=None, _base:HCQBuffer|None=None, view:MMIOInterface|None=None,
-               owner:HCQCompiled|None=None):
+  def __init__(self, va_addr:sint, size:int, meta:Any=None, _base:HCQBuffer|None=None, view:MMIOInterface|None=None, owner:Any=None):
    self.va_addr, self.size, self.meta, self._base, self.view = va_addr, size, meta, _base, view
    self._devs, self.owner = ([owner] if owner is not None else []), owner
    self._mappings:dict[HCQCompiled, HCQBuffer] = {} # mapping to the other devices
@@ -1,5 +1,5 @@
 from __future__ import annotations
-from typing import cast, Callable, TypeVar, Generic, Any, Sequence
+from typing import cast, Callable, TypeVar, Generic, Any, Sequence, Iterable
 import struct, functools, time, collections, itertools, decimal, statistics
 from dataclasses import replace, dataclass
 from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, dedup, pluralize, JIT_BATCH_SIZE, unwrap, PROFILE
@@ -7,9 +7,9 @@ from tinygrad.helpers import to_tuple, round_up, partition, data64_le, panic, Co
 from tinygrad.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer, DepsTracker
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEntry, ProfileGraphEvent
 from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, rewrite_group, GroupOp
-from tinygrad.uop.symbolic import symbolic, pm_fold_cast_const
+from tinygrad.uop.symbolic import symbolic
 from tinygrad.dtype import dtypes, truncate
-from tinygrad.runtime.support.hcq import MMIOInterface
+from tinygrad.runtime.support.hcq import MMIOInterface, HCQBuffer
 from tinygrad.runtime.support.memory import BumpAllocator
 from tinygrad.renderer import Renderer, Estimates
 from tinygrad.engine.realize import to_program, get_call_arg_uops, get_call_name, get_call_outs_ins, estimate_uop
@@ -22,7 +22,7 @@ HCQDeviceType = TypeVar('HCQDeviceType', bound='HCQ2Compiled')

 HCQ_RUNTIME_DEV = ContextVar("HCQ_RUNTIME_DEV", "CPU")

-HCQ_DEVS = frozenset(("AMD",))
+HCQ_DEVS = frozenset(("AMD", "CPU"))
 HCQ_P2P_DEVS = HCQ_DEVS | frozenset(("CPU",))
 HCQ_CACHE_TAGS = frozenset(("program", "systems", "template"))

@@ -87,16 +87,19 @@ def encode_kernargs_clike(call:UOp, prg:UOp, devs:str|tuple[str, ...]) -> UOp:
 def replace_call_buffers(ctx:tuple[list[UOp], dict[UOp, int]], call:UOp) -> UOp|None:
  bufs, slots = ctx
  for s in call.src[1:]:
-    if s.op not in (Ops.PARAM, Ops.BIND) and slots.setdefault(s, len(bufs)) == len(bufs): bufs.append(s)
-  return call.replace(src=call.src[:1] + tuple(s if s.op in (Ops.PARAM, Ops.BIND) else s.param_like(slots[s]) for s in call.src[1:]))
+    if s.op is not Ops.PARAM and not s.is_bound_var and slots.setdefault(s, len(bufs)) == len(bufs): bufs.append(s)
+  return call.replace(src=call.src[:1] + tuple(s if s.op is Ops.PARAM or s.is_bound_var else s.param_like(slots[s]) for s in call.src[1:]))
 pm_replace_buffers = PatternMatcher([(UPat(Ops.CALL, name="call"), replace_call_buffers)])

 # *****************
 # 1.1. prep: staging copies

-def _need_staging(a, b): return all_devices_in(a.device, HCQ_DEVS) and not all_devices_in(b.device, HCQ_P2P_DEVS)
+def _need_staging(a, b): return all_devices_in(a.device, HCQ_DEVS - {"CPU"}) and not all_devices_in(b.device, HCQ_P2P_DEVS)

-def hcq_call_devs(call:UOp) -> Any|None: return next((b.device for b in call.src[1:] if all_devices_in(b.device, HCQ_DEVS)), None)
+def _get_enqueue_devs(call:UOp) -> Any|None:
+  if not (bufs:=call.src[1:]) or not all(all_devices_in(b.device, HCQ_P2P_DEVS) for b in bufs): return None
+  devs = min(bufs, key=lambda b: to_tuple(b.device)[0].startswith("CPU")).device # prio to enqueue on not CPU device
+  return devs if all_devices_in(devs, HCQ_DEVS) else None

 def stage_copy(dst:UOp, src:UOp) -> UOp|None:
  if not (_need_staging(src, dst) or _need_staging(dst, src)): return None
@@ -105,7 +108,7 @@ def stage_copy(dst:UOp, src:UOp) -> UOp|None:
  return UOp(Ops.LINEAR, src=(src.copy_to_device("CPU").call(stage, src), stage.copy_to_device(dst.device).call(dst, stage)))

 def kernel_copy(call:UOp, dst:UOp, src:UOp) -> UOp|None:
-  if (devs:=hcq_call_devs(call)) is None or Device[(dev:=to_tuple(devs)[0])].has_copy_queue: return None
+  if (devs:=_get_enqueue_devs(call)) is None or Device[(dev:=to_tuple(devs)[0])].has_copy_queue: return None
  d, s = (UOp.param(i, dst.dtype, (n:=dst.max_numel(),), device=devs) for i in range(2))
  ast = d.index(r:=UOp.range(n, 0)).store(s.index(r).load()).end(r).sink(arg=KernelInfo(name="copy"), tag=1)
  return call.replace(src=(to_program(ast, Device[dev].renderer), dst, src))
@@ -136,7 +139,7 @@ def _get_deps(ctx:DepsTracker, bufs_by_lane:list[list[Any]], write, key:tuple[tu

 def _build_wait_cmds(slots:dict[str, int], dep_lanes:list[tuple[tuple, int, int]], devices:tuple[str, ...], queue:str) -> tuple[list[UOp], set[int]]:
  # opt1: same-queue ops are fifo-ordered
-  if devices[0].split(":")[0] in {"AMD", "QCOM"} or queue.startswith("COPY"):
+  if devices[0].split(":")[0] in {"AMD", "QCOM", "CPU"} or queue.startswith("COPY"):
    dep_lanes = [(dep, dlane, lane) for dep, dlane, lane in dep_lanes if (dep[0][dlane], dep[1]) != (devices[lane], queue)]

  # opt2: keep latest dep per (dep device, queue, cur lane)
@@ -159,7 +162,7 @@ def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[t
    for b in itertools.chain.from_iterable(_get_call_bufs_by_lane(call, devices)):
      for bd in to_tuple(b.device): dev_bufs[bd][id(b)] = b

-  n, fences, fins, signal_tags = len(batch_info), [], [], set()
+  n, fences, resets, fins, signal_tags = len(batch_info), [], [], [], set()
  for _, devgroup in itertools.groupby(sorted(dev_bufs), key=lambda d: d.split(":")[0]):
    devs = tuple(devgroup)

@@ -173,16 +176,17 @@ def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[t
    fin_submit = make_submit(*waits, UOp(Ops.INS, arg="store", src=(tl_signal, tl_value.index(0))), devs=devs, queue="COMPUTE:0")
    epoch = (epoch_slot:=tl_value.after(fin_submit).index(0)).load()

-    # fence once per device group on this schedule's previous epoch, then reset any queue signals used by the group
+    # fence once per device group on this schedule's previous epoch
    qs = dedup([qn for bdevs, qn in batch_info if set(bdevs) & set(devs)])
    sched_epoch = make_signal(devs, next(UOp.unique_num))

    wait_device_epoch = (done:=tl_signal.after(loop:=UOp.loop(0)).index(0).load()).end(loop, done < sched_epoch.index(0).load())
-    resets = [make_signal(devs, slots[q]).after(wait_device_epoch).index(0).store(0) for q in qs]
+    fences.append(make_call("hcq_fence", UOp.sink(wait_device_epoch), HCQInfo(devs)))

-    fences.append(make_call("hcq_fence", UOp.sink(*(resets or [wait_device_epoch])), HCQInfo(devs)))
+    # queues of other groups wait on these signals, so reset them only after every group reached its epoch
+    if qs: resets.append(make_call("hcq_reset", UOp.sink(*[make_signal(devs, slots[q]).index(0).store(0) for q in qs]), HCQInfo(devs)))
    fins.append(make_call("hcq_finalizer", UOp.sink(epoch_slot.store(epoch + 1), sched_epoch.after(fin_submit).index(0).store(epoch)), HCQInfo(devs)))
-  return fences, fins, signal_tags
+  return fences + resets, fins, signal_tags

 def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]], profile:bool) -> list[UOp]:
  batch_info = [(devices, "COMPUTE:0" if call.src[0].op is Ops.PROGRAM else "COPY:0") for call, devices in batch]
@@ -229,7 +233,7 @@ def sched_hcq_batches(l:UOp, profile:bool) -> UOp:
  srcs:list[UOp] = []
  batch:list[tuple[UOp, tuple[str, ...]]] = []
  for call in l.src:
-    if (devs:=hcq_call_devs(call)) is not None: batch.append((call, to_tuple(devs)))
+    if (devs:=_get_enqueue_devs(call)) is not None: batch.append((call, to_tuple(devs)))
    else: srcs, batch = srcs + _finalize_batch(batch, profile) + [call], []
  return l.replace(src=tuple(srcs + _finalize_batch(batch, profile)))

@@ -354,9 +358,13 @@ pm_split_patches = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION

 # *****************

+def _rank_ranges(uops:Iterable[UOp]) -> dict[UOp, UOp]:
+  return {r: r.replace(arg=(i,)+r.arg[1:]) for i,r in enumerate(sorted([u for u in uops if u.op is Ops.RANGE], key=lambda r: r.arg))}
+
 def replace_params(call:UOp) -> UOp|None:
  body, variables, param_ops = call.src[0], call.src[0].variables(), {Ops.PARAM, Ops.MSTACK}
-  args = dedup([s for u in body.toposort(gate=lambda u: u.op not in param_ops) for s in u.src if s.op in param_ops and s not in variables])
+  tops = body.toposort(gate=lambda u: u.op not in param_ops)
+  args = dedup([s for u in tops for s in u.src if s.op in param_ops and s not in variables])

  patched, refhold = partition(call.src[1:], lambda x: x.src[0] in args)
  by_root = {p.src[0]: p for p in patched}
@@ -368,7 +376,7 @@ def replace_params(call:UOp) -> UOp|None:
  refhold += [a for a in addrs if a not in held and all(b.op is not Ops.PARAM or b.tag is not None for b in unwrap_mstack(a))]

  sub = {(b:=u.without_after): UOp.param(i, u.dtype, shape=b.shape, device=HCQ_RUNTIME_DEV.value, volatile=b.op is Ops.PARAM and b.arg.volatile)
-         for i,u in enumerate(c_args)} | {v: v.replace(arg=replace(v.arg, slot=-1)) for v in variables if v.op is Ops.PARAM}
+         for i,u in enumerate(c_args)} | {v: v.replace(arg=replace(v.arg, slot=-1)) for v in variables if v.op is Ops.PARAM} | _rank_ranges(tops)
  info = replace(call.arg.aux, inputs=next((i for i,u in enumerate(c_args) if u.without_after.tag == "inputs"), None))
  return call.replace(src=(body.substitute(sub).replace(arg="hcq_args"), *c_args, *refhold),
                      arg=replace(call.arg, aux=info)) # TODO: call.after(*refhold)?
@@ -437,7 +445,7 @@ def hcq_compile(linear:UOp, input_uops:list[UOp]|None, profile:bool) -> UOp:
    linear = graph_rewrite(linear, pm_encode_cmdbufs+pm_pack_placeholders, walk=True, name="encode and pack", enter_calls=True)

    # patches and runtime uops
-    linear = graph_rewrite(linear, pm_early_simplify+symbolic+pm_fold_cast_const, bottom_up=False, name="simplify patches", enter_calls=True)
+    linear = graph_rewrite(linear, pm_early_simplify+symbolic, bottom_up=False, name="simplify patches", enter_calls=True)
    linear = graph_rewrite(linear, pm_split_patches, walk=True, name="split patches")

    # and compile it
@@ -469,7 +477,7 @@ def fold_binary(buf:UOp, blob:UOp) -> UOp:
 def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
  for off,val in zip(off.src, val.src):
    for b,v in zip((bs:=mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)), val.src if val.op is Ops.STACK else (val,)*len(bs)):
-      data = struct.pack(f'<{v.dtype.fmt}', truncate[v.dtype](v.val))
+      data = struct.pack(f'<{v.dtype.fmt}', truncate[v.dtype]((v.src[0] if v.op is Ops.CAST else v).val))
      b.ensure_allocated().as_memoryview(force_zero_copy=True, no_sync=True).cast('B')[(bo:=off.val*buf.dtype.itemsize):bo+len(data)] = data
  return UOp(Ops.NOOP)

@@ -484,7 +492,7 @@ def resolve_getaddr(buf:UOp, g:UOp) -> UOp:

 pm_resolve_patches = PatternMatcher([
  # multi
-  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
+  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat.any(UPat(Ops.CONST), UPat(Ops.CAST, src=(UPat(Ops.CONST),)))], name="op"), push_stack),
  (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),

  # getaddr
@@ -510,7 +518,7 @@ def hcq_link(linear:UOp, cache=True) -> UOp:
  bufs = {(j,i):a for j,c in enumerate(linear.src) for i,a in enumerate(c.src[1:], 1)
          if a.op is Ops.AFTER and unwrap_mstack(a.src[0])[0].tag in HCQ_CACHE_TAGS}
  linear = linear.substitute({x:link_buf_cache[k] for a in bufs.values() if (k:=link_buf_key(a)) in link_buf_cache for x in (a, a.src[0])}, walk=True)
-  linear = graph_rewrite(linear, pm_resolve_patches+symbolic+pm_fold_cast_const+pm_assert_no_afters, bpm=pm_bufferize, ctx=cache, bottom_up=False,
+  linear = graph_rewrite(linear, pm_resolve_patches+symbolic+pm_assert_no_afters, bpm=pm_bufferize, ctx=cache, bottom_up=False,
                         name="resolve patches")
  for (j,i),a in bufs.items(): link_buf_cache.setdefault(link_buf_key(a), linear.src[j].src[i])
  if cache: link_linear_cache[linear_key] = linear
@@ -521,6 +529,7 @@ def hcq_link(linear:UOp, cache=True) -> UOp:

 class HCQ2Compiled(Compiled):
  timestamp_divider: float = 1000.0
+  wait_timeout_ms: float = 30000.0

  def __init__(self, device:str, allocator:HCQAllocator, compilers:list[type[Renderer]], runtime, can_recover:bool=False, arch=None):
    self.device_id:int = int(device.split(":")[1]) if ":" in device else 0
@@ -577,9 +586,10 @@ class HCQ2Compiled(Compiled):
    sig = self.signal("timeline").as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
    tl = self.signal("value", 1).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
    timeout = timeout if timeout is not None and self.can_recover else None
-    st = time.perf_counter()
-    while sig[0] < tl[0] - 1:
-      if time.perf_counter() - st > (timeout or 3000) / 1000: self.on_device_hang()
+    st, done = time.perf_counter(), sig[0]
+    while done < tl[0] - 1:
+      if done != (done:=sig[0]): st = time.perf_counter()
+      elif time.perf_counter() - st > (timeout or self.wait_timeout_ms) / 1000: self.on_device_hang()
    if self.prof_ents: self.collect_prof()

  def on_device_hang(self): raise RuntimeError(f"{self.device} hang detected")
@@ -617,17 +627,17 @@ class HCQ2Buffer:
    return HCQ2Buffer(self.va_addr+offset, meta=self.meta, view=(self.view.view(offset=offset, size=size) if self.view is not None else None))

 class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
-  def _as_buffer(self, buf:HCQ2Buffer) -> memoryview:
+  def _as_buffer(self, buf:HCQBuffer) -> memoryview:
    return unwrap(buf.view).mv

-  def _map(self, buf:HCQ2Buffer) -> HCQ2Buffer:
+  def _map(self, buf:HCQBuffer) -> HCQBuffer:
    if not hasattr(self, '_do_map'): raise NotImplementedError("map failed: no method implemented")
    return self._do_map(buf)

  def _do_unmap(self, mb): self.dev.iface.free(mb)

  @suppress_finalizing
-  def _free(self, buf:HCQ2Buffer, options:BufferSpec|None=None):
+  def _free(self, buf:HCQBuffer, options:BufferSpec|None=None):
    if options is not None and options.external_ptr is not None: return
    self.dev.synchronize()
    if hasattr(self, '_do_free'): self._do_free(buf, options)
@@ -636,4 +646,4 @@ class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
    self.dev.synchronize()
    self._do_unmap(mb)

-  def _offset(self, buf, size:int, offset:int) -> HCQ2Buffer: return buf.offset(offset=offset, size=size)
+  def _offset(self, buf, size:int, offset:int) -> HCQBuffer: return buf.offset(offset=offset, size=size)
@@ -8,14 +8,13 @@ from tinygrad.helpers import DEBUG, cpu_profile, TracingKey, SPEC, pluralize, SC

 # unwrap VIEW/CAST/etc to find the actual data source (kernel output, buffer, or multi-device op)
 def _unwrap_src(s: UOp) -> UOp:
-  while len(s.src) and s.op not in {Ops.AFTER, Ops.BUFFER, Ops.PARAM, Ops.MSELECT, Ops.MSTACK, Ops.BIND}: s = s.src[0]
+  while len(s.src) and s.op not in {Ops.AFTER, Ops.BUFFER, Ops.PARAM, Ops.MSELECT, Ops.MSTACK}: s = s.src[0]
  return s

-# a buffer state is AFTER | BUFFER | PARAM. MSELECT/MSTACK join per-device states, BIND is not a buffer dependency
+# a buffer state is AFTER | BUFFER | PARAM. MSELECT/MSTACK join per-device states
 def _states(s: UOp) -> list[UOp]:
  s = _unwrap_src(s)
  if s.op in {Ops.MSELECT, Ops.MSTACK}: return [st for ss in s.src for st in _states(ss)]
-  if s.op is Ops.BIND: return []
  assert s.op in {Ops.AFTER, Ops.BUFFER, Ops.PARAM}, f"input to kernel must resolve to a buffer state, not {s.op}"
  return [s]

@@ -71,7 +70,7 @@ def create_schedule(sched_sink:UOp) -> UOp:
      else:
        k = rk.src[0] if rk.op is Ops.END else rk
        assert k.op is Ops.CALL, f"unexpected op in queue: {k.op}"
-        buf_uops = tuple(_unwrap_src(s).buf_uop for s in k.src[1:] if s.op is not Ops.BIND)
+        buf_uops = tuple(_unwrap_src(s).buf_uop for s in k.src[1:] if not s.is_bound_var)
        linearized.append(k.src[0].call(*buf_uops))
      for x in children.get(rk, []):
        in_degree[x] -= 1
@@ -100,7 +99,8 @@ pm_post_sched_cache = PatternMatcher([

 def resolve_linear_call(linear_call:UOp):
  linear = graph_rewrite(linear_call.src[0], pm_post_sched_cache, ctx=({}, linear_call.src[1:]), walk=True, name="params to buffers")
-  binds = {f"p{i}":x.src[0] for i,x in enumerate(linear_call.src[1:]) if x.op is Ops.BIND}
+  # map the call body params back to the original Variables stored in the call args
+  binds = {f"p{i}":x.src[0].replace(op=Ops.PARAM) for i,x in enumerate(linear_call.src[1:]) if x.is_bound_var}
  return linear.substitute({v:binds[v.expr] for v in linear.variables() if v.expr in binds}, enter_calls=True, name="resolve scalar params")

 pm_resolve_linear_call = PatternMatcher([
@@ -184,13 +184,13 @@ def create_linear_with_vars(big_sink:UOp) -> tuple[UOp, dict[str, int]]:

  # vars used in the schedule
  used_vars = set().union(*[{v.expr for v in si.src[0].variables()} for si in linear.src])
-  # get var_vals
+  # get var_vals from the bound Variables in the call args
  var_vals: dict[str, int] = {}
  for b in big_sink.src[1:]:
-    if b.op is Ops.BIND:
-      nm = b.src[0].expr
+    if b.is_bound_var:
+      v, val = b.unbind()
+      nm = v.expr
      if nm not in used_vars: continue
-      val = b.src[1].val
      if var_vals.get(nm, val) != val: raise RuntimeError(f"bind mismatch on {nm}, {var_vals[nm]} != {val}")
      var_vals[nm] = val

@@ -24,7 +24,7 @@ class IndexingContext:


 ALWAYS_CONTIGUOUS: set[Ops] = {Ops.CONTIGUOUS, Ops.AFTER, Ops.BUFFER,
-                      Ops.CONST, Ops.BIND, Ops.MSELECT, Ops.MSTACK, Ops.PARAM,
+                      Ops.CONST, Ops.MSELECT, Ops.MSTACK, Ops.PARAM,
                      Ops.LOAD, Ops.CALL, Ops.FUNCTION}

 def realize(ctx:IndexingContext, tr:UOp) -> None: ctx.realize_map[tr] = None
@@ -35,7 +35,7 @@ def realize_srcs(ctx:IndexingContext, rb:UOp) -> None:

 def realize_store_after_src(ctx:IndexingContext, dest:UOp, src:UOp):
  # you don't usually have to do this for assign unless there's a WAR hazard like TestAssign.test_assign_double_diamond_reduce
-  if dest.base in src.backward_slice_with_self: ctx.realize_map[src] = None
+  if dest.base in src.toposort(enter_calls=False): ctx.realize_map[src] = None

 def realize_custom_kernel_srcs(ctx:IndexingContext, c:UOp) -> None:
  for s in c.src[1:]:
@@ -69,7 +69,9 @@ def broadcast_rngs(x:UOp, src:UOp, rngs:tuple[UOp, ...]) -> tuple[UOp, ...]:

 # TODO: srcs contain (real data srcs, something else, ranges) and the boundary is confusing. see range_start
 def data_srcs(op:Ops, src:tuple[UOp, ...]) -> tuple[UOp, ...]:
-  if op in {Ops.PARAM, Ops.BUFFER, Ops.RANGE, Ops.SPECIAL, Ops.BIND}: return ()
+  if op in {Ops.PARAM, Ops.BUFFER, Ops.RANGE, Ops.SPECIAL}: return ()
+  # the store of a bound Variable only carries the input value, it has no data srcs
+  if op is Ops.STORE and src[0].is_variable: return ()
  if op in GroupOp.Movement|{Ops.INDEX, Ops.STAGE, Ops.REDUCE, Ops.AFTER, Ops.END}: return src[:1]
  return src

@@ -4,7 +4,7 @@ import itertools
 from tinygrad.dtype import dtypes, AddrSpace, Invalid, to_dtype, strong_dtype
 from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, KernelInfo, ParamArg, shape_to_shape_arg
 from tinygrad.uop.ops import graph_rewrite, sint, AxisType, BottomUpGate, rewrite_group, identity_element
-from tinygrad.uop.symbolic import symbolic, pm_fold_cast_const
+from tinygrad.uop.symbolic import symbolic
 from tinygrad.uop.movement import mop_cleanup
 from tinygrad.helpers import prod, getenv, dedup, all_int, DEBUG, SPLIT_REDUCEOP, DEBUG_RANGEIFY, VIZ, MAX_KERNEL_BUFFERS, SPEC
 from tinygrad.helpers import PCONTIG, FLOAT16, OPENPILOT_HACKS, argsort, partition, get_single_element
@@ -60,7 +60,7 @@ pm_mops = PatternMatcher([
 # 0. do some cleanup rewrites, mostly copied from the old stuff

 def fix_store_hazard(target:UOp, src:UOp):
-  if (base:=target.base) not in src.backward_slice_with_self: return None
+  if (base:=target.base) not in src.toposort(enter_calls=False): return None
  # PERMUTE and FLIP reorder indices, SHRINK can have overlapping regions when dest is also shrunk
  unsafe = {Ops.PERMUTE, Ops.FLIP} | ({Ops.SHRINK} if target.op_in_backward_slice_with_self(Ops.SHRINK) else set())
  reaches_base: dict[UOp, bool] = {}
@@ -461,11 +461,12 @@ pm_add_buffers = pm_mops+pm_flatten_bufferize+PatternMatcher([
 class LocalAddBufferContext:
  dg:int = 0
  map:dict = field(default_factory=dict)
-  vars:dict = field(default_factory=dict)
  range:int = 0
  opts:tuple|None = None

 def debuf(ctx:LocalAddBufferContext, buf:UOp):
+  # Variables (ALU buffers with a value range) are scalar symbolic values, not real buffers: they become ALU params with no slot
+  if buf.is_variable: return buf.replace(op=Ops.PARAM)
  param = UOp(Ops.PARAM, src=(UOp.const(prod(buf.max_shape)),),
              arg=ParamArg(ctx.dg, buf.dtype, addrspace=buf.addrspace, device=buf.device))
  ret = param.reshape(buf.max_shape)
@@ -475,10 +476,6 @@ def debuf(ctx:LocalAddBufferContext, buf:UOp):
  ctx.dg += 1
  return ret

-def unbind_kernel(ctx:LocalAddBufferContext, b:UOp):
-  ctx.vars[b] = None
-  return b.src[0]
-
 def handle_after(ctx:LocalAddBufferContext, after:UOp):
  if after.addrspace == AddrSpace.LOCAL: return None
  buf = after.buf_uop
@@ -502,8 +499,7 @@ to_define_global = PatternMatcher([
  (UPat(Ops.STORE, name="x"), find_bufs),
  (UPat((Ops.BUFFER, Ops.MSTACK, Ops.MSELECT), name="buf"), debuf),
  (UPat(Ops.PARAM, name="v"), lambda v:
-   UOp.variable(v.arg.name, v.arg.vmin_vmax[0], v.arg.vmin_vmax[1], v.dtype, multiple_of=v.arg.multiple_of)
-   if v.arg.name is not None and v.arg.vmin_vmax is not None else None),
+   v.replace(arg=replace(v.arg, slot=-1)) if v.arg.name is not None and v.arg.vmin_vmax is not None and v.arg.slot != -1 else None),

  # this renumbers the params
  (UPat(Ops.PARAM, name="buf"), lambda ctx, buf:
@@ -512,7 +508,8 @@ to_define_global = PatternMatcher([
  # ALU params are scalar symbolic values, not buffers.
  (UPat(Ops.INDEX, src=(UPat(Ops.PARAM, name="v"),)), lambda v: v if v.addrspace == AddrSpace.ALU else None),

-  (UPat(Ops.BIND, name="b"), unbind_kernel),
+  # bound Variables are stores into Variable buffers: strip the store, the buffer becomes an ALU param via debuf
+  (UPat(Ops.AFTER, name="b"), lambda b: b.src[0] if b.is_bound_var else None),
  (UPat(Ops.AFTER, name="after"), handle_after),

  # remove device from local BUFFERIZE
@@ -541,13 +538,16 @@ pm_add_param_range_tags = PatternMatcher([
 def split_store(x:UOp) -> UOp|None:
  # if we have any open ranges here, we don't split. open DEVICE ranges are fine, they are bound per device at launch
  if any(r.arg[-1] is not AxisType.DEVICE for r in x.ranges): return None
+  # the store of a bound Variable is an input value, not a kernel
+  st = x.src[0] if x.op is Ops.END else x
+  if st.op is Ops.STORE and st.src[0].is_variable: return None

  # local kernel rewrite
  lctx = LocalAddBufferContext()
  ret = graph_rewrite(x, to_define_global+pm_flatten_range+rangeify_codegen, ctx=lctx, name="kernel split", bottom_up=True)

  # create the Kernel. NOTE: buffers can be on different devices here now, they are compiled to SDMA copies later by schedule
-  return ret.sink(arg=KernelInfo(opts_to_apply=lctx.opts)).call(*lctx.map.values(), *lctx.vars.keys())
+  return ret.sink(arg=KernelInfo(opts_to_apply=lctx.opts)).call(*lctx.map.values())

 split_kernels = PatternMatcher([
  (UPat((Ops.STORE, Ops.END), name="x"), split_store),
@@ -584,7 +584,7 @@ def get_kernel_graph(sink:UOp) -> UOp:
  tsink, rctx = run_rangeify(tsink, bool(DEBUG_RANGEIFY))

  tsink = graph_rewrite(tsink,
-                        symbolic+pm_fold_cast_const+pm_reduce_simplify+pm_const_buffer_folding+pm_remove_bufferize+pm_no_indexing_calls,
+                        symbolic+pm_reduce_simplify+pm_const_buffer_folding+pm_remove_bufferize+pm_no_indexing_calls,
                        name="symbolic+reduce_collapse+debuf")
  tsink = graph_rewrite(tsink, pm_limit_bufs, ctx=rctx, name="limit buffers")

@@ -109,7 +109,7 @@ def _precompiled_output_redirect(s:UOp, t:UOp) -> UOp|None:
 def transform_precompiled_call(c:UOp) -> UOp|None:
  if not c.arg.precompile: return None
  assert c.src[0].op is Ops.TUPLE, f"expected TUPLE body for precompiled FUNCTION, got {c.src[0].op}"
-  input_buffers = tuple(x.contiguous() if x.op not in {Ops.AFTER, Ops.BIND} else x for x in c.src[1:])
+  input_buffers = tuple(x.contiguous() if x.op is not Ops.AFTER else x for x in c.src[1:])

  # add the outputs to the call
  srcs = c.src[0].src
@@ -176,6 +176,8 @@ pm_early_transform_tensor_graph = PatternMatcher([
 ])

 def finalize_after(ctx:AllocCtx, x:UOp):
+  # bound Variables are call inputs, not assigns: they stay in the graph and pm_replace_buf turns them into call args
+  if x.is_bound_var: return None
  # untagged: record as an assign for the call body
  if x.tag is None:
    ctx.assigns.append(x)
@@ -196,7 +198,7 @@ def finalize_after(ctx:AllocCtx, x:UOp):

 def replace_input_buffer(ctx:AllocCtx, b:UOp):
  ctx.replacements.append(b)
-  if b.op is Ops.BIND: return b.param_like(len(ctx.replacements)-1)
+  if b.is_bound_var or b.is_variable: return b.param_like(len(ctx.replacements)-1)
  return UOp.param(len(ctx.replacements)-1, b.dtype, b.shape, b.device,
                   addrspace=b.addrspace if b.addrspace is not None else AddrSpace.GLOBAL)

@@ -214,8 +216,8 @@ pm_replace_buf = PatternMatcher([
  # replace SHRINK with PARAM
  (UPat(Ops.SHRINK, src=(UPat(Ops.BUFFER),), name="b", allow_any_len=True), replace_input_view),
  (UPat(Ops.BITCAST, src=(UPat.any(UPat(Ops.SHRINK, src=(UPat(Ops.BUFFER),), allow_any_len=True), UPat(Ops.BUFFER)),), name="b"), replace_input_view),
-  # strip value from BIND for cache key normalization, so different values hit same cache
-  (UPat(Ops.BIND, src=(UPat(Ops.PARAM), UPat(Ops.CONST)), name="b"), replace_input_buffer),
+  # strip the stored value from bound Variables for cache key normalization, so different values hit same cache
+  (UPat(Ops.AFTER, name="b"), lambda ctx,b: replace_input_buffer(ctx, b) if b.is_bound_var else None),
 ])

@rewrite_group(lambda _,ret: f"Callify {pluralize('Buffer', len(ret[1]))}")
@@ -451,12 +453,11 @@ class Tensor(RandMixin):
      return self
    # STORE+AFTER: STORE is the write effect (void), AFTER wraps the view for correct shape/ranging
    assign = self.uop.after(self.uop.store(x.uop))
-    if (base := self.uop.base).op in {Ops.BUFFER, Ops.AFTER} and self.uop is not base and not self.uop.has_buffer_identity():
+    ib = self.uop
+    while not ib.has_buffer_identity() and ib.op in GroupOp.Movement|{Ops.BITCAST, Ops.DETACH}: ib = ib.src[0]
+    if ib is not self.uop and ib.has_buffer_identity(after_ok=True):
      # view assign: replace at the buffer-identity level (e.g. RESHAPE(BUFFER)) so @function's substitution catches it
-      ib = self.uop
-      while not ib.has_buffer_identity() and ib is not base: ib = ib.src[0]
-      assigned_ib = ib.after(assign)
-      _apply_map_to_tensors({ib: assigned_ib}, name="Embed View Assign")
+      _apply_map_to_tensors({ib: ib.after(assign)}, name="Embed View Assign")
    else:
      # simple assign
      self.uop = assign
@@ -741,7 +742,7 @@ class Tensor(RandMixin):
    the reference frames (`ref_frames`).
    """
    ref_frames = [x.contiguous() for x in ref_frames or []]
-    assert frame_pos.op is Ops.BIND, "frame_pos must be a bound Variable"
+    assert frame_pos.is_bound_var, "frame_pos must be a bound Variable"
    srcs = (out:=Tensor.empty(*shape, device=self.device, dtype=self.dtype), self.contiguous(), state.contiguous(), *ref_frames)
    fn = UOp(Ops.CUSTOM_FUNCTION, src=(frame_pos.src[0], *[UOp.const(s, dtypes.int) for s in shape]), arg="encdec")
    return Tensor(out.uop.after(fn.call(*[s.uop for s in srcs], frame_pos)))
@@ -13,9 +13,6 @@ class FastEnum(IntEnum):
 class Ops(FastEnum):
  # ** 1 -- defines/special **

-  # BIND pairs a symbolic PARAM with a concrete value
-  BIND = auto()
-
  # this is a RANGE for GPU dimensions, similar to symbolic shapes but not exactly
  SPECIAL = auto()

@@ -150,9 +150,6 @@ def dtype_from_uop(op:Ops, src:tuple[UOp,...], arg:Any) -> DType|None:
    case Ops.STACK:
      if len(src) == 0: return dtypes.void
      return promo_dtype(src)
-    case Ops.BIND:
-      assert src[0].dtype == src[1].dtype, f"bind dtype mismatch {src[0].dtype} != {src[1].dtype}"
-      return src[0].dtype
    case Ops.WMMA:
      # WMMA output dtype is the accumulator dtype (src[2])
      return src[2].dtype
@@ -377,7 +374,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):

      # some ops init the shape
      case Ops.GETADDR: return ()
-      case Ops.BIND | Ops.RANGE | Ops.SPECIAL: return ()
+      case Ops.RANGE | Ops.SPECIAL: return ()
      case Ops.BINARY: return (len(self.arg),)
      case Ops.BUFFER:
        if len(self.src): return self.src[0].as_shape
@@ -520,10 +517,12 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
    if self.op is Ops.CONST: return self
    if self.op is Ops.SINK and all(s.op is Ops.CONST or (s.op is Ops.STACK and len(s.src) == 0) for s in self.src): return self
    # late import!
-    from tinygrad.uop.symbolic import symbolic, pm_fold_cast_const
+    from tinygrad.uop.symbolic import symbolic
    with Context(TRACK_MATCH_STATS=0 if not tracked else TRACK_MATCH_STATS.value):
-      return graph_rewrite(self, symbolic+pm_fold_cast_const, name="simplify")
-  def ssimplify(self) -> UOp|ConstType: return ret.val if (ret:=self.simplify()).op is Ops.CONST else ret
+      return graph_rewrite(self, symbolic, name="simplify")
+  def ssimplify(self) -> UOp|ConstType:
+    if (ret := self.simplify()).op is Ops.CAST and ret.src[0].op is Ops.CONST: return ret.dtype.const(ret.src[0].val)
+    return ret.val if ret.op is Ops.CONST else ret
  def _eval(self, dtype, expected_type:Type[T]) -> T:
    assert self.dtype in dtype, f"eval with wrong dtype {self}"
    vmin, vmax = (simple_self:=self.simplify())._min_max
@@ -799,11 +798,9 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
      case Ops.PAD | Ops.SHRINK: src_args = list(zip(*arg))
      case Ops.PERMUTE | Ops.FLIP: src_args = []
      case Ops.STACK:
-        # arg is the other srcs; all are cast to the promoted dtype, spec requires STACK srcs to match its dtype
        srcs = (self,)+tuple(arg)
        dtype = cast(DType, dtype_from_uop(Ops.STACK, srcs, None))
-        # TODO: why cast here?
-        return UOp(Ops.STACK, dtype, tuple(u if u.base.is_invalid else u.cast(dtype) for u in srcs))
+        return UOp(Ops.STACK, dtype, tuple(u if u.base.is_invalid else UOp.const(u.val, dtype) if u.op is Ops.CONST else u.cast(dtype) for u in srcs))
      case _: raise RuntimeError(f"{op} is not a MovementOp")
    usrcs = [shape_to_shape_arg(arg) for arg in src_args]
    if len(usrcs) == 0: return UOp(op, src=(self,), arg=arg)
@@ -960,8 +957,8 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
    if self.op is Ops.UNSHARD: return self.src[0].realized
    # only these can be realized
    if self.op not in (Ops.BUFFER, Ops.MSTACK): return None
-    # LOCAL/REG scratch buffers are never realized
-    if self.op is Ops.BUFFER and self.addrspace in (AddrSpace.LOCAL, AddrSpace.REG): return None
+    # LOCAL/REG scratch buffers are never realized, and Variables (ALU) have no real storage
+    if self.op is Ops.BUFFER and self.addrspace in (AddrSpace.LOCAL, AddrSpace.REG, AddrSpace.ALU): return None
    # an unbacked intermediate BUFFER (directly or as an MSTACK source) is not realized
    if any(b.op is Ops.BUFFER and buffers.get(b) is None for b in self.backward_slice_with_self): return None
    # NOTE: this is used by the JIT to determine which inputs we capture
@@ -972,29 +969,41 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
  # *** uop Variable stuff ***

  @staticmethod
-  def variable(name:str, min_val:PyConst, max_val:PyConst, dtype:DType=dtypes.weakint, multiple_of:int=1) -> UOp:
-    return UOp(Ops.PARAM, src=(shape_to_shape_arg(()),),
-               arg=ParamArg(-1, dtype, name=name, vmin_vmax=(min_val, max_val), multiple_of=multiple_of, addrspace=AddrSpace.ALU))
+  def variable(name:str, min_val:PyConst, max_val:PyConst, dtype:DType=dtypes.weakint, multiple_of:int=1, param:bool=False) -> UOp:
+    # a Variable is a 0-d BUFFER in the ALU addrspace; binding it is storing a CONST into it
+    # param=True creates the kernel-side form directly: an ALU PARAM (what the BUFFER becomes inside kernels)
+    arg = ParamArg(-1, dtype, name=name, vmin_vmax=(min_val, max_val), multiple_of=multiple_of, addrspace=AddrSpace.ALU)
+    return UOp(Ops.PARAM if param else Ops.BUFFER, src=(shape_to_shape_arg(()),), arg=arg)
+  @property
+  def is_variable(self) -> bool:
+    # a Variable is a 0-d BUFFER in the ALU addrspace that carries a value range (it becomes a PARAM inside kernels)
+    return self.op is Ops.BUFFER and isinstance(self.arg, ParamArg) and \
+           self.arg.vmin_vmax is not None and self.arg.addrspace is AddrSpace.ALU and self._shape == ()
+  @property
+  def is_bound_var(self) -> bool:
+    # a bound Variable is bind()'s AFTER(var, STORE(var, CONST))
+    return self.op is Ops.AFTER and self.src[0].is_variable and self.src[1].op is Ops.STORE and \
+           self.src[1].src[0] is self.src[0] and self.src[1].src[1].op is Ops.CONST and len(self.src) == 2
  @property
  def expr(self) -> str:
-    assert self.op is Ops.PARAM
+    assert self.op in {Ops.PARAM, Ops.BUFFER}
    return unwrap(self.arg.name)
  def bind(self, val:int|UOp):
-    assert self.op is Ops.PARAM and self.addrspace is AddrSpace.ALU, f"op is {self.op}, need PARAM"
+    assert self.is_variable, f"op is {self.op}, need Variable"
    uval = self.const_like(val) if isinstance(val, int) else val
    assert self.vmin <= uval.vmin and uval.vmax <= self.vmax, f"bind {val} not in range [{self.vmin}, {self.vmax}]"
    assert uval.divides(self.arg.multiple_of) is not None, f"bind {val} not divisible by {self.arg.multiple_of}"
-    return UOp(Ops.BIND, src=(self, uval))
+    return self.after(self.store(uval))
  def unbind(self) -> tuple[Variable, int]:
-    assert self.op is Ops.BIND and self.src[0].op is Ops.PARAM and self.src[1].op is Ops.CONST, f"can't unbind {self}"
-    return self.src[0], self.src[1].val
+    assert self.is_bound_var, f"can't unbind {self}"
+    return self.src[0], self.src[1].src[1].val
  def unbind_all(self) -> tuple[UOp, dict[Variable, int]]:
    ret:dict[Variable, int] = {}
    return graph_rewrite(self, pm_unbind, ctx=ret), ret
  def variables(self) -> list[Variable]:
-    return sorted({x if x.op is Ops.PARAM else UOp.variable("_device_num", 0, x.vmax, dtype=x.dtype)
-                   for x in self.backward_slice_with_self if (x.op is Ops.RANGE and x.arg[-1] is AxisType.DEVICE) or x.op is Ops.PARAM
-                   and x.arg.addrspace is AddrSpace.ALU}, key=lambda v: v.expr)
+    return sorted({x if x.op in {Ops.PARAM, Ops.BUFFER} else UOp.variable("_device_num", 0, x.vmax, dtype=x.dtype, param=True)
+                   for x in self.backward_slice_with_self if (x.op is Ops.RANGE and x.arg[-1] is AxisType.DEVICE) or
+                   (x.op is Ops.PARAM and x.arg.addrspace is AddrSpace.ALU) or x.is_variable}, key=lambda v: v.expr)

  # *** uop symbolic stuff ***

@@ -1005,7 +1014,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
    if self.op is Ops.STACK: return math.gcd(*[x.const_factor() for x in self.src])
    if self.op is Ops.ADD: return math.gcd(self.src[0].const_factor(), self.src[1].const_factor())
    if self.op is Ops.MUL: return self.src[0].val if self.src[0].op is Ops.CONST else self.src[1].val if self.src[1].op is Ops.CONST else 1
-    if self.op is Ops.PARAM and self.arg.multiple_of is not None: return self.arg.multiple_of
+    if self.op in (Ops.PARAM, Ops.BUFFER) and isinstance(self.arg, ParamArg) and self.arg.multiple_of is not None: return self.arg.multiple_of
    return 1
  def divides(self, v:int) -> UOp|None:
    if v==1: return self
@@ -1017,7 +1026,8 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
    if self.op is Ops.MUL:
      if (d0:=self.src[0].divides(v)) is not None: return d0 * self.src[1]
      if (d1:=self.src[1].divides(v)) is not None: return self.src[0] * d1
-    if self.op is Ops.PARAM and self.arg.multiple_of is not None: return self // v if self.arg.multiple_of%v == 0 else None
+    if self.op in (Ops.PARAM, Ops.BUFFER) and isinstance(self.arg, ParamArg) and self.arg.multiple_of is not None:
+      return self // v if self.arg.multiple_of%v == 0 else None
    return None # generic None if we aren't sure
  def pop_const(self, op=Ops.ADD) -> tuple[UOp, PyConst]:  # NOTE: assume Invalid ALU is resolved
    return (self.src[0], self.src[1].val) if self.op is op and self.src[1].op is Ops.CONST else (self, identity_element(op, self.dtype))
@@ -1082,9 +1092,9 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
    # float has NAN issue and we use explicit NAN in transcendental
    if self.op is Ops.WHERE and dtypes.is_int(self.dtype): return min(self.src[1].vmin, self.src[2].vmin), max(self.src[1].vmax, self.src[2].vmax)
    # NOTE: returned UOp is assumed to be CONST
-    if self.op is Ops.PARAM and self.arg.vmin_vmax is not None: return self.arg.vmin_vmax
+    if self.op in (Ops.PARAM, Ops.BUFFER) and isinstance(self.arg, ParamArg) and self.arg.vmin_vmax is not None: return self.arg.vmin_vmax
    if self.op in (Ops.RANGE, Ops.SPECIAL) and self.dtype is not dtypes.void: return 0, (self.src[0]-1).vmax
-    if self.op is Ops.BIND: return self.src[0]._min_max # ignore the bound value
+    if self.op is Ops.AFTER: return self.src[0]._min_max
    if self.op is Ops.STACK: return min(x.vmin for x in self.src), max(x.vmax for x in self.src)
    if self.op is Ops.CONST and self.val is not Invalid: return self.val, self.val
    if self.op is Ops.INDEX: return self.src[0]._min_max
@@ -1101,7 +1111,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
  def _sym_fxn(self):
    from tinygrad.uop.render import _render_with_splits, renderer_infer
    sself = self.simplify()
-    varnames = tuple(dedup(x.expr for x in sself.toposort() if x.op is Ops.PARAM and x.arg.addrspace == AddrSpace.ALU))
+    varnames = tuple(dedup(x.expr for x in sself.toposort() if (x.op is Ops.PARAM and x.arg.addrspace == AddrSpace.ALU) or x.is_variable))
    # TODO: sanitize varnames, or don't use naked eval while staying fast
    ret = _render_with_splits(list(sself.toposort()), renderer_infer, {sself})
    lines = [f"  {k}={v}" for k,v in ret.items() if k != "ast"] + [f"  return {ret['ast']}"]
@@ -1156,7 +1166,10 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
    src: tuple[UOp, ...] = (UOp(Ops.NOOP) if shape is None else shape_to_shape_arg(shape),)
    return UOp(Ops.PARAM, src=src, arg=ParamArg(slot, dtype, vmin_vmax, multiple_of, name, addrspace, axis, device, volatile))
  def param_like(self, slot:int):
-    if self.op is Ops.BIND: return self.src[0].replace(arg=replace(self.src[0].arg, slot=slot, name=f"p{slot}"))
+    # Variables become ALU params in the call body; the stored value (if bound) stays in the call args
+    if self.is_bound_var or self.is_variable:
+      b = self.src[0] if self.op is Ops.AFTER else self
+      return UOp(Ops.PARAM, src=b.src, arg=replace(b.arg, slot=slot, name=f"p{slot}"))
    addrspace = self.addrspace if self.addrspace is not None else AddrSpace.GLOBAL
    return UOp.param(slot, self.dtype, self.shard_shape if self.axis is not None else self._shape, self.device, addrspace=addrspace, axis=self.axis)

@@ -1410,9 +1423,9 @@ class UPat(OpMixin):
    return res

 def deconstruct_function(fxn:Callable) -> tuple:
-  new_globals = {k:v for k,v in fxn.__globals__.items() if k in fxn.__code__.co_names}
-  for co in fxn.__code__.co_consts:
-    if isinstance(co, types.CodeType): new_globals.update({k:v for k,v in fxn.__globals__.items() if k in co.co_names})
+  # globals can be referenced from arbitrarily nested code objects (comprehensions/lambdas, pre PEP 709)
+  def names(co:types.CodeType) -> set: return set(co.co_names).union(*(names(c) for c in co.co_consts if isinstance(c, types.CodeType)))
+  new_globals = {k:v for k,v in fxn.__globals__.items() if k in names(fxn.__code__)}
  # NOTE: optional round trip through pickle!
  assert fxn.__closure__ is None, "closures are not supported in pattern matchers"
  ret = fxn.__code__, new_globals, fxn.__name__, fxn.__defaults__
@@ -1754,7 +1767,7 @@ def do_unbind(ctx:dict[Variable, int], x:UOp):
  v,i = x.unbind()
  ctx[v] = i
  return v
-pm_unbind = PatternMatcher([(UPat(Ops.BIND, name="x"), do_unbind)])
+pm_unbind = PatternMatcher([(UPat(Ops.AFTER, name="x"), lambda ctx,x: do_unbind(ctx,x) if x.is_bound_var else None)])

 # ctx is source UOp for which we are finding a contiguous view for. used in contiguous_view_offset
 pm_contiguous_view_offset = PatternMatcher([
@@ -33,12 +33,13 @@ def strip_binary_parens(x:UOp, left:str, right:str, code_for_op) -> str:

 renderer = PatternMatcher([
  (UPat(Ops.PARAM, name="x"), lambda x: x.arg.name if x.arg.name is not None else f"p{x.arg.slot}"),
+  (UPat(Ops.BUFFER, name="x"), lambda x: x.arg.name if isinstance(x.arg, ParamArg) and x.arg.name is not None else f"b{x.arg.slot}"),
+  (UPat(Ops.AFTER, name="x"), lambda ctx,x: ctx[x.src[0]]),
  (UPat((Ops.SPECIAL), name="x"), lambda x: x.arg),
  (UPat(Ops.RANGE, dtypes.void, name="x"), lambda x: f"loop{x.arg[0]}"),
  (UPat(Ops.RANGE, name="x"), lambda x: f"r{range_str(x)}"),
  (UPat(Ops.CONST, name="x"), lambda x: str(x.val)),
  (UPat(Ops.CAST, name="x"), lambda ctx,x: f"({str(x.dtype)[7:]})({ctx[x.src[0]]})"),
-  (UPat(Ops.BIND, name="x"), lambda ctx,x: ctx[x.src[0]]),
  (UPat(Ops.NEG, name="x"), lambda ctx,x: f"(-{ctx[x.src[0]]})"),
  (UPat(Ops.RECIPROCAL, name="x"), lambda ctx,x: f"(1/{ctx[x.src[0]]})"),
  (UPat(Ops.MAX, name="x"), lambda ctx,x: f"max({ctx[x.src[0]]}, {ctx[x.src[1]]})"),
@@ -141,9 +141,8 @@ spec_tensor = PatternMatcher([
   (isinstance(buf.dtype, DType) and matches_dtype(buf.src[0], dtypes.weakint) and is_device(buf.arg.device))
   if isinstance(buf.arg, ParamArg) and buf.addrspace is AddrSpace.GLOBAL else None),

-  # Tensor variable bindings
-  (UPat(Ops.BIND, (dtypes.int, dtypes.long, dtypes.weakint,), (UPat(Ops.PARAM), UPat.cvar(dtype=(dtypes.int,dtypes.long,dtypes.weakint,))), arg=None),
-   lambda: True),
+  # a Variable is a 0-d ALU BUFFER with a value range and no device
+  (UPat(Ops.BUFFER, src=(UPat(),), name="buf"), lambda buf: buf.arg.device is None if buf.is_variable else None),

  # custom function
  (UPat(Ops.CUSTOM_FUNCTION, name="x"), lambda x: isinstance(x.arg, str)),
@@ -241,9 +240,6 @@ spec_full = PatternMatcher([

  # all loads/stores
  (UPat((Ops.LOAD, Ops.STORE)), lambda: True),
-
-  # while BIND is being casted
-  (UPat(Ops.BIND, (dtypes.int, dtypes.weakint), (UPat(), UPat()), arg=None), lambda: True),
 ])+spec_tensor+spec_program+spec_hcq

 # ***** kernel graph spec *****
@@ -251,17 +247,16 @@ spec_full = PatternMatcher([
 spec_kernel_graph = PatternMatcher([
  # sink
  (UPat(Ops.SINK, dtypes.void), lambda: True),
-  # bind
-  (UPat(Ops.BIND), lambda: True),
-  # const + stack to make vconsts
+  # the store of a bound Variable binds it: AFTER(BUFFER, STORE(BUFFER, CONST)) in call args
+  (UPat(Ops.STORE, dtypes.void, (UPat(Ops.BUFFER, name="b"), UPat(Ops.CONST))), lambda b: b.is_variable),
+  # const + stack to make vconsts and shape args
  (UPat(Ops.CONST, src=()), lambda: True),
-  (UPat(Ops.STACK, src=()), lambda: True),
-  (UPat(Ops.STACK, src=UPat((Ops.CONST, Ops.BIND, Ops.PARAM))), lambda: True),
+  (UPat(Ops.STACK, name="s"), lambda s: all(x.op in (Ops.CONST, Ops.PARAM) or x.is_variable or x.is_bound_var for x in s.src) or None),
  # linear for more kernels (TODO: we should enter non sink calls)
  #(UPat(Ops.LINEAR), lambda: True),
  # param is outside buffer, buffer is local buffer
  (UPat(Ops.PARAM, name="x"), lambda x: isinstance(x.arg, ParamArg)),
-  (UPat(Ops.BUFFER, name="x"), lambda x: isinstance(x.arg, ParamArg) and x.addrspace == AddrSpace.GLOBAL),
+  (UPat(Ops.BUFFER, name="x"), lambda x: isinstance(x.arg, ParamArg) and x.addrspace in (AddrSpace.GLOBAL, AddrSpace.ALU)),
  # RESHAPE/BITCAST are NOOPs in the kernel graph (do we need them?)
  (UPat((Ops.RESHAPE, Ops.BITCAST)), lambda: True),
  # mstack/mselect
@@ -23,8 +23,12 @@ def fold_bitcast(root:UOp, c:UOp) -> UOp|None:
  if c.dtype.fmt is None or root.dtype.fmt is None or c.dtype.itemsize != root.dtype.itemsize: return None
  return root.const_like(bitcast(c.val, c.dtype, root.dtype))

+# const folding works for CONST, STACK, and casted CONST
+const_folding_pat = UPat.any(UPat((Ops.CONST, Ops.STACK)), UPat(Ops.CAST, src=(UPat(Ops.CONST),)))
+
 def const_arg(u:UOp) -> ConstType|tuple[ConstType, ...]|None:
  if u.op is Ops.CONST: return u.val
+  if u.op is Ops.CAST and u.src[0].op is Ops.CONST: return u.dtype.const(u.src[0].val)
  if u.op is Ops.STACK and all(s.op is Ops.CONST for s in u.src): return tuple(s.val for s in u.src)
  return None

@@ -96,10 +100,6 @@ pm_remove_invalid = PatternMatcher([
   if any(x.is_invalid for x in s.src) else None),
 ])

-# the one rule that collapses the pair CAST(dt, CONST(v)) into a typed CONST
-# TODO: delete this once CONST has no dtype
-pm_fold_cast_const = PatternMatcher([(UPat(Ops.CAST, name="root", src=(UPat.cvar("c"),)), lambda root, c: root.const_like(c.val))])
-
 symbolic_simple = pm_data_invalid + PatternMatcher([
  # ** self folding **
  (UPat.var("x") + 0, lambda x: x),    # x+0 -> x
@@ -136,10 +136,10 @@ symbolic_simple = pm_data_invalid + PatternMatcher([
  (UPat.var("x", dtype=dtypes.ints+(dtypes.bool, dtypes.weakint)) != UPat.var("x"),
   lambda x: x.const_like(False, dtypes.bool)), # x != x -> False (only ints)
  # ** constant folding **
-  (UPat(GroupOp.Unary, src=(UPat((Ops.CONST, Ops.STACK)),), name="a"), fold_const_alu),
+  (UPat(GroupOp.Unary, src=(const_folding_pat,), name="a"), fold_const_alu),
  # NOTE: THREEFRY(const,const) folds via its decomposition
-  (UPat(GroupOp.Binary-{Ops.THREEFRY}, src=(UPat((Ops.CONST, Ops.STACK)),)*2, name="a"), fold_const_alu),
-  (UPat(GroupOp.Ternary, src=(UPat((Ops.CONST, Ops.STACK)),)*3, name="a"), fold_const_alu),
+  (UPat(GroupOp.Binary-{Ops.THREEFRY}, src=(const_folding_pat,)*2, name="a"), fold_const_alu),
+  (UPat(GroupOp.Ternary, src=(const_folding_pat,)*3, name="a"), fold_const_alu),
  # bool MUL is AND, ADD/MAX is OR. prevents other rules to rewrite bool ADD/MUL incorrectly
  (UPat.var('x', dtype=dtypes.bool) * UPat.var('y', dtype=dtypes.bool), lambda x,y: x&y),
  (UPat.var('x', dtype=dtypes.bool) + UPat.var('y', dtype=dtypes.bool), lambda x,y: x|y),
@@ -245,7 +245,7 @@ symbolic = symbolic_simple+commutative+PatternMatcher([
  # complementary zero branches under the same condition select directly
  (UPat.var("c").where(UPat.var("t"), 0) + UPat.var("c").where(0, UPat.var("f")), lambda c,t,f: c.where(t, f)),
  # ALU/variable min==max -> CONST
-  (UPat({Ops.CMPLT, Ops.CMPNE, Ops.FLOORDIV, Ops.FLOORMOD, Ops.PARAM, Ops.BIND, Ops.SPECIAL}, name="x"),
+  (UPat({Ops.CMPLT, Ops.CMPNE, Ops.FLOORDIV, Ops.FLOORMOD, Ops.PARAM, Ops.AFTER, Ops.SPECIAL}, name="x"),
   lambda x: x.const_like(x.vmin) if x.vmin == x.vmax else None),
  (UPat(Ops.RANGE, src=(UPat(Ops.CONST,)), name="x"), lambda x: x.const_like(x.vmin) if x.vmin == x.vmax else None),
  # max folding
@@ -330,14 +330,14 @@ def uop_given_valid(valid:UOp, uop:UOp, try_simplex=True) -> UOp:
  for i,(expr,v) in enumerate(bounds.items()):
    v0, v1 = (expr.vmin if v[0] is None else v[0], expr.vmax if v[1] is None else v[1])
    # try checking the whole clause
-    all_candidates.append((expr, UOp.variable(f"fake{i}", v0, v1, expr.dtype)))
+    all_candidates.append((expr, UOp.variable(f"fake{i}", v0, v1, expr.dtype, param=True)))

    if try_simplex:
      # every candidate is a set of constrained UOp based on valid, and if every item in a set simplifies the uop into a same output, we rewrite uop
      candidates = [[all_candidates[-1]]]
      if expr.op is Ops.ADD and v0 == 1 and all(u.op in GroupOp.Irreducible for u in expr.split_uop(Ops.ADD)):
        # if the constraint is a simplex: X0 + X1 + ... > 0, we can check if all Xi > 0 simplify into the same output
-        candidates.append([(Xi, UOp.variable(f"fake{i}", 1, Xi.vmax, Xi.dtype)) for Xi in expr.split_uop(Ops.ADD)])
+        candidates.append([(Xi, UOp.variable(f"fake{i}", 1, Xi.vmax, Xi.dtype, param=True)) for Xi in expr.split_uop(Ops.ADD)])

      for candidate in candidates:
        # if every branch in candidate gives the same simplified uop, we can rewrite the uop
@@ -407,7 +407,8 @@ pm_move_where_on_load = PatternMatcher([
 ])

 def gated_given_valid(cond:UOp, x:UOp, i:UOp) -> UOp|None:
-  if x.dtype is not dtypes.weakint: return None
+  # pure index math only: a LOAD in x executes even where cond is false, so its INDEX valid must survive the assumption
+  if x.dtype is not dtypes.weakint or x.op_in_backward_slice_with_self(Ops.INDEX): return None
  # Skip if x contains DIV/MOD AND IMAGE mode is enabled -> image index e.g. openpilot
  if IMAGE.value > 0 and x.op_in_backward_slice_with_self(Ops.CDIV, Ops.CMOD, Ops.FLOORDIV, Ops.FLOORMOD): return None
  return cond.where(uop_given_valid(cond, x, try_simplex=False), i)
@@ -432,9 +433,6 @@ sym = symbolic+pm_simplify_valid+PatternMatcher([
  # reorder ALU/VECTORIZE
  (UPat(GroupOp.ALU, src=(UPat(Ops.STACK, src=UPat(name='x')), UPat(Ops.STACK, src=UPat(name='y'))), name='alu'),
   lambda x,y,alu: UOp(Ops.STACK, src=(UOp(alu.op, src=(x,y)),))),
-  # ** where **
-  # push cast to branches
-  (UPat.var("s").where(UPat.var("a"), UPat.var("b")).cast().named("cast"), lambda s,a,b,cast: s.where(a.cast(cast.dtype), b.cast(cast.dtype))),
  # ** pow **
  ((UPat(Ops.POW, name="p"), lambda p: xpow(*p.src))),
  # ** load/store folding **
@@ -37,6 +37,7 @@ z3_renderer = PatternMatcher([
  # variables
  (UPat(Ops.SPECIAL, name="x"), lambda x,ctx: create_bounded(x.arg, 0, ctx[1][x.src[0]]-1, ctx[0])),
  (UPat(Ops.PARAM, name="x"), lambda x,ctx: create_bounded(x.arg.name, x.vmin, x.vmax, ctx[0])),
+  (UPat(Ops.BUFFER, name="x"), lambda x,ctx: create_bounded(x.arg.name, x.vmin, x.vmax, ctx[0]) if x.is_variable else None),
  (UPat(Ops.RANGE, name="x"), lambda x,ctx: create_bounded(x.render(simplify=False), 0, ctx[1][x.src[0]]-1, ctx[0])),
  # loads are variables bounded by the min/max of the dtype. non-pointer INDEX is also a LOAD
  (UPat((Ops.LOAD, Ops.INDEX), dtypes.ints+(dtypes.weakint,), name="x"), lambda x,ctx:
@@ -60,7 +61,7 @@ z3_renderer = PatternMatcher([

 def uops_to_z3(solver:z3.Solver, *uops: UOp) -> list[z3.ExprRef]:
  # gate on upstream memory addressing, but keep INDEX as an unknown LOAD
-  lst = list(UOp.sink(*uops).toposort(gate=lambda x: x.op not in {Ops.AFTER, Ops.BUFFER, Ops.SHRINK} and \
+  lst = list(UOp.sink(*uops).toposort(gate=lambda x: x.op not in {Ops.AFTER, Ops.SHRINK} and (x.op is not Ops.BUFFER or x.is_variable) and \
                                      (x.dtype in dtypes.ints+(dtypes.bool, dtypes.weakint) or x.op is Ops.SINK)))[:-1]
  z3map: dict[UOp, z3.ExprRef] = {}
  for u in lst:
@@ -12,18 +12,18 @@ def lower_weak_node(u:UOp) -> UOp|None:
  if src == u.src or any(s.dtype in dtypes.weaks for s in src[start:]): return None
  dt = strong_dtype(least_upper_dtype(select_dtype(u), *(s.dtype for s in src)) if u.op in GroupOp.Binary
                    else unwrap(dtype_from_uop(u.op, src, u.arg)))
-  return u.replace(dtype=None, src=src[:start]+tuple(s if s.base.is_invalid else s.cast(dt) for s in src[start:])).cast(u.dtype)
+  return u.replace(dtype=None, src=src[:start]+tuple(s if s.base.is_invalid else commit_weak(s, dt) for s in src[start:])).cast(u.dtype)

 pm_lower_weak = PatternMatcher([
  (UPat(Ops.CONST, dtype=dtypes.weaks, name="u"), lambda u: UOp.const(u.val, select_dtype(u)).cast(u.dtype)),
-  # two stacked weak casts are a weakint value used as weakfloat (or vice versa): resolve the inner one at the outer kind's default.
+  # two stacked weak casts are two kind conversions: each resolves at its own kind's default
  # a SINGLE weak cast is never rewritten here, each consumer absorbs it on its own edge (see lower_weak_srcs)
  (UPat(Ops.CAST, dtype=dtypes.weaks, src=(UPat(Ops.CAST, dtype=dtypes.weaks, src=(UPat.var("x"),)),), name="u"),
-   lambda u,x: x.cast(select_dtype(u)).cast(u.dtype) if x.dtype not in dtypes.weaks else None),
+   lambda u,x: x.cast(select_dtype(u.src[0])).cast(select_dtype(u)).cast(u.dtype) if x.dtype not in dtypes.weaks else None),
  # Binary can widen from the bounds, all other nodes derive from the lowered sources.
  # a weakfloat Unary (sin/exp2/...) must resolve here, before the transcendental decomposition
  (UPat(GroupOp.Binary|GroupOp.Unary|{Ops.WHERE, Ops.RANGE, Ops.STACK, Ops.SPECIAL}, name="u"), lower_weak_node),
-  (UPat(Ops.PARAM, dtype=dtypes.weakint, name="u"),
+  (UPat((Ops.PARAM, Ops.BUFFER), dtype=dtypes.weakint, name="u"),
    lambda u: u.replace(dtype=None, arg=replace(u.arg, dtype=select_dtype(u))).cast(dtypes.weakint) if u.addrspace == AddrSpace.ALU else None),
 ])

@@ -40,7 +40,7 @@ def lower_weak_srcs(ctx:dict[UOp, UOp]|None, u:UOp) -> UOp|None:
  return None if ret is u else ret

 def commit_weak(s:UOp, dt:DType) -> UOp:
-  # a bare weak CONST commits directly (the value stays mathematical, emission truncates), a weak non-const src takes the demand cast
+  # a CONST commits directly at dt (the value stays mathematical, emission truncates), a non-const src takes the cast
  return UOp.const(s.val, dt) if s.op is Ops.CONST else s.cast(dt)

 def commit_weak_srcs(u:UOp) -> UOp|None:
@@ -65,9 +65,13 @@ def cast_weak_srcs(c:UOp, u:UOp) -> UOp|None:

 pm_cast_weak = PatternMatcher([
  (UPat(Ops.CAST, name="c", src=(UPat(GroupOp.ALU, dtype=dtypes.weaks, name="u"),)), cast_weak_srcs),
+  (UPat(Ops.CAST, name="c", src=(UPat(Ops.CONST, dtype=dtypes.weaks, name="u"),)), lambda c,u: commit_weak(u, c.dtype)),
 ])

 pm_lower_index_dtype = pm_commit_weak+pm_cast_weak+PatternMatcher([
+  # a CAST between two concrete dtypes over a CONST is a value conversion: evaluate it once, at the width the CAST states
+  # TODO: delete this once CONST has no dtype
+  (UPat(Ops.CAST, dtypes.all, name="root", src=(UPat.cvar("c", dtypes.all),)), lambda root, c: root.const_like(c.val)),
  (UPat(GroupOp.All, name="u"),
   lambda ctx,u: lower_weak_srcs(ctx, u) if u.dtype not in dtypes.weaks and any(s.dtype in dtypes.weaks for s in u.src) else None),
  # a valid index into an n-element buffer lives in [0,n): a gated long index narrows when n-1 fits int32 (out-of-gate wraps, discarded)
Author	SHA1	Message	Date
chenyuandGitHub	5ca87f1bac	fix cast to weak twice [pr] (#17548 ) also no gradient for weak target	2026-08-15 12:39:54 -04:00
chenyuandGitHub	26cbadd69a	no pm_fold_cast_const in full_rewrite_to_sink [PR] (#17547 )	2026-08-15 10:18:59 -04:00
chenyuandGitHub	fae893753b	no pm_fold_cast_const in get_kernel_graph [pr] (#17546 ) * no pm_fold_cast_const in get_kernel_graph [pr] * maybe	2026-08-15 09:57:05 -04:00
nimlgenandGitHub	c17849a1f8	disk: neable polling (#17538 )	2026-08-15 15:03:28 +03:00
chenyuandGitHub	539a03343a	no casted const from sub and div [pr] (#17543 )	2026-08-15 07:47:40 -04:00
qazalandGitHub	a57569349c	renumber invalids before callify (#17542 ) * renumber invalids before callify * change * Revert "change" This reverts commit 6f4df1541e79721a85ee3f5801114454f264c973. * renumber in tensor * scope renumber_invalid_outputs * cleanup	2026-08-15 18:00:54 +09:00
qazalandGitHub	5c43a89fb1	precompile_backward tests for sched_cache (#17544 ) * work * back * work * keep +	2026-08-15 15:25:31 +09:00
qazalandGitHub	e6f5bb9c09	simple test for Invalid clone cache miss regression (#17541 ) * simple test for Invalid clone cache miss regression * xfail * _	2026-08-15 11:00:21 +09:00
chenyuandGitHub	4b0525e594	no pm_fold_cast_const in UOp.simplify and hcq2 [pr] (#17540 )	2026-08-14 21:39:47 -04:00
chenyuandGitHub	64ccbde3bb	clean up STACK with a const [PR] (#17539 )	2026-08-14 21:07:01 -04:00
chenyuandGitHub	6ea665ed66	remove pm_fold_cast_const from dtype decomp [pr] (#17536 ) * remove pm_fold_cast_const from dtype decomp [pr] * fix	2026-08-14 16:24:51 -04:00
nimlgenandGitHub	0725acc392	reenable hcq2 ci (#17532 )	2026-08-14 23:13:18 +03:00
wozeparrotandGitHub	4a1f32977c	gptoss: default GROUPED_MOE=1 (#17537 )	2026-08-14 12:24:37 -07:00
chenyuandGitHub	13c381b0c0	remove pm_fold_cast_const from initial symbolic [pr] (#17535 ) interestingly it gives more accurate numerics when composing const like log10	2026-08-14 14:14:40 -04:00
chenyuandGitHub	ac7067ac60	fix deconstruct_function for python 3.11 (#17534 )	2026-08-14 13:38:34 -04:00
chenyuandGitHub	80169c6758	remove where push cast to branches from sym [pr] (#17533 ) * remove where push cast to branches from sym [pr] not really needed and one less place that generates casted weak const when it's not needed * fix	2026-08-14 12:57:56 -04:00
nimlgenandGitHub	adacaa3e17	hcq2 fix hangs (#17529 )	2026-08-14 16:23:59 +03:00
chenyuandGitHub	89ab344c42	fix assign into bitcast with no explicit realize (#17531 )	2026-08-14 09:23:01 -04:00
chenyuandGitHub	25c3bd027b	remove pm_fold_cast_const in simplify_merge_adjacent [pr] (#17530 )	2026-08-14 09:10:05 -04:00
nimlgenandGitHub	6b35220622	cpu hcq2 (#17503 ) * cpu hcq2 * temp * slop * test with backpressure * x * x * x * x * x * x * Dx * save reverts * um? * x * x * call from py * x? * x * submitters gone * x * x * z * Dx * Dx * x * x * fixes * repl * x * f * for now keep hcqbuffer	2026-08-14 15:06:03 +03:00
George HotzandGitHub	b1859805b1	remove Ops.BIND (#17511 ) * remove Ops.BIND * param arg * simplify that * simplify * cleaner * props, not functions * param and buffer can share	2026-08-13 23:52:16 -07:00
qazalandGitHub	faba071b1d	don't enter CALL body in assign fixups (#17527 ) * fix python time regression in mxfp4 * this saves even more time * s_nop test * itertools count * cleanup	2026-08-14 15:05:04 +09:00
qazalandGitHub	81dc8ec232	Revert "amd: fix ALL2ALL speed on amdgpu (gpt) (#17518 )" (#17526 ) This reverts commit `cc6d33bde7`.	2026-08-14 10:39:14 +09:00
geohot	95ca5081fe	hotfix: update extra/runbook_digitalocean_mi350x	2026-08-13 17:26:39 -07:00
sirhcmandGitHub	303d1677b3	qcomcl: use qemu for compilation (#17524 ) Autogen / In-tree Autogen (comgr 2) (push) Failing after 15s Details Autogen / In-tree Autogen (push) Failing after 17s Details Deploy Docs / deploy (push) Successful in 2m42s Details Unit Tests / Docs (push) Successful in 2m59s Details Unit Tests / Torch Backend Tests (push) Successful in 5m2s Details Unit Tests / Python Backend (push) Successful in 3m48s Details Unit Tests / Linters (push) Successful in 2m55s Details Unit Tests / Null Tests (push) Successful in 6m47s Details Unit Tests / Unit Tests (push) Successful in 6m36s Details Unit Tests / SPEC=2 (1) (push) Successful in 6m21s Details Unit Tests / SPEC=2 (2) (push) Successful in 6m1s Details Unit Tests / Fuzzing (push) Successful in 3m29s Details Unit Tests / CL IMAGE Tests (push) Successful in 5m48s Details Unit Tests / openpilot Compile Tests (push) Successful in 6m56s Details Unit Tests / ONNX (CPU) Tests (push) Successful in 9m49s Details Unit Tests / Optimization Tests (push) Successful in 7m35s Details Unit Tests / Test LLM (push) Successful in 3m33s Details Unit Tests / Models (push) Successful in 4m59s Details Unit Tests / Linux (DSP) (push) Successful in 5m19s Details Unit Tests / Linux (DEV=CL) (push) Successful in 5m58s Details Unit Tests / Linux (DEV=CPU:CLANG) (push) Successful in 7m0s Details Unit Tests / Linux (DEV=CPU:LLVM) (push) Successful in 5m54s Details Unit Tests / Linux (DEV=CPU:LVP) (push) Successful in 6m12s Details Unit Tests / Linux (DEV=CPU:X86) (push) Successful in 5m44s Details Unit Tests / Linux (DEV=WEBGPU) (push) Successful in 6m25s Details Unit Tests / AMD ASM IDE (push) Successful in 5m42s Details Unit Tests / Linux (am) (push) Successful in 7m18s Details Unit Tests / Linux (amd gfx1100) (push) Successful in 7m22s Details Unit Tests / Linux (amdllvm gfx1100) (push) Successful in 6m30s Details Unit Tests / Linux (amd gfx1201) (push) Successful in 7m11s Details Unit Tests / Linux (amdllvm gfx1201) (push) Successful in 6m50s Details Unit Tests / Linux (amd gfx950) (push) Successful in 8m18s Details Unit Tests / Linux (amdllvm gfx950) (push) Successful in 6m41s Details Unit Tests / Linux (nv) (push) Successful in 7m14s Details Unit Tests / Linux (ptx) (push) Successful in 3m55s Details Unit Tests / Compile-only (DEV=NULL:IR3:a630) (push) Successful in 2m41s Details Unit Tests / Compile-only (DEV=NULL:NAK:sm_120) (push) Successful in 1m59s Details Unit Tests / Compile-only (DEV=NULL:QCOMCL:a630) (push) Successful in 4m34s Details Unit Tests / hcq2 (push) Failing after 1m11s Details Autogen / In-tree Autogen (macos) (push) Canceled after 0s Details Benchmarks / Mac pytest (push) Canceled after 0s Details Benchmarks / LLM (DEV=AMD) (push) Canceled after 0s Details Benchmarks / LLM (DEV=METAL) (push) Canceled after 0s Details Benchmarks / LLM (DEV=NV) (push) Canceled after 0s Details Benchmarks / HLB-CIFAR10 (DEV=AMD) (push) Canceled after 0s Details Benchmarks / HLB-CIFAR10 (DEV=METAL) (push) Canceled after 0s Details Benchmarks / HLB-CIFAR10 (DEV=NV) (push) Canceled after 0s Details Benchmarks / MLPerf (AMD) (push) Canceled after 0s Details Benchmarks / MLPerf (NV) (push) Canceled after 0s Details Benchmarks / Stable Diffusion (DEV=AMD) (push) Canceled after 0s Details Benchmarks / Stable Diffusion (DEV=METAL) (push) Canceled after 0s Details Benchmarks / Stable Diffusion (DEV=NV) (push) Canceled after 0s Details Benchmarks / Tests (DEV=AMD) (push) Canceled after 0s Details Benchmarks / Tests (DEV=METAL) (push) Canceled after 0s Details Benchmarks / Tests (DEV=NV) (push) Canceled after 0s Details Benchmarks / UsbGPU Benchmark (push) Canceled after 0s Details Benchmarks / comma Benchmark (0.11.2) (push) Canceled after 0s Details Benchmarks / comma Benchmark (0.11.0) (push) Canceled after 0s Details Benchmarks / DSP Benchmark (push) Canceled after 0s Details Benchmarks / UsbGPU Benchmark (comma) (push) Canceled after 0s Details Benchmarks / PCI Driver Benchmark (DEV=AMD) (push) Canceled after 0s Details Benchmarks / PCI Driver Benchmark (DEV=NV) (push) Canceled after 0s Details Benchmarks / LLVM Speed (push) Canceled after 0s Details Platform Tests / MacOS (unit) (push) Canceled after 0s Details Platform Tests / MacOS (unit, mock) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=METAL) (1) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=METAL) (2) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=CPU:CLANG) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=CPU:LLVM) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=CPU:LVP) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=WEBGPU) (push) Canceled after 0s Details Platform Tests / Windows (DEV=CPU:CLANG) (push) Canceled after 0s Details Platform Tests / Windows (DEV=CPU:LLVM) (push) Canceled after 0s Details Platform Tests / Windows (DEV=CPU:X86) (push) Canceled after 0s Details Platform Tests / Windows (DEV=WEBGPU) (push) Canceled after 0s Details	2026-08-13 20:16:02 -04:00
George HotzandGitHub	673c6463f9	disable HCQ2 for AMD CI (#17523 )	2026-08-13 13:14:40 -07:00
wozeparrotandGitHub	849074f0db	gptoss: use fa swa (#17522 )	2026-08-13 10:39:22 -07:00
George HotzandGitHub	0252cb8fa7	remove anchors from CI flow (#17521 ) Autogen / In-tree Autogen (comgr 2) (push) Successful in 41s Details Deploy Docs / deploy (push) Successful in 46s Details Autogen / In-tree Autogen (push) Successful in 47s Details Unit Tests / Linters (push) Successful in 2m23s Details Unit Tests / Docs (push) Successful in 3m32s Details Unit Tests / Python Backend (push) Successful in 3m50s Details Unit Tests / Unit Tests (push) Successful in 3m55s Details Unit Tests / SPEC=2 (2) (push) Successful in 4m10s Details Unit Tests / Null Tests (push) Successful in 4m25s Details Unit Tests / SPEC=2 (1) (push) Successful in 4m28s Details Unit Tests / Torch Backend Tests (push) Successful in 4m47s Details Unit Tests / Fuzzing (push) Successful in 2m32s Details Unit Tests / Linux (DSP) (push) Successful in 2m1s Details Unit Tests / Models (push) Successful in 2m30s Details Unit Tests / CL IMAGE Tests (push) Successful in 3m29s Details Unit Tests / Test LLM (push) Successful in 2m40s Details Unit Tests / openpilot Compile Tests (push) Successful in 4m25s Details Unit Tests / Linux (DEV=CL) (push) Successful in 4m41s Details Unit Tests / Linux (DEV=CPU:LVP) (push) Successful in 3m14s Details Unit Tests / ONNX (CPU) Tests (push) Successful in 6m24s Details Unit Tests / Optimization Tests (push) Successful in 6m13s Details Unit Tests / Linux (DEV=CPU:LLVM) (push) Successful in 3m32s Details Unit Tests / Linux (DEV=CPU:X86) (push) Successful in 3m33s Details Unit Tests / Linux (DEV=CPU:CLANG) (push) Successful in 4m30s Details Unit Tests / AMD ASM IDE (push) Successful in 2m44s Details Unit Tests / Linux (DEV=WEBGPU) (push) Successful in 4m16s Details Unit Tests / hcq2 (push) Successful in 3m13s Details Unit Tests / Linux (amdllvm gfx1100) (push) Successful in 3m55s Details Unit Tests / Linux (amd gfx1100) (push) Successful in 4m24s Details Unit Tests / Linux (amd gfx1201) (push) Successful in 4m23s Details Unit Tests / Linux (am) (push) Successful in 4m45s Details Unit Tests / Linux (amdllvm gfx1201) (push) Successful in 4m6s Details Unit Tests / Compile-only (DEV=NULL:QCOMCL:a630) (push) Failing after 24s Details Unit Tests / Linux (amdllvm gfx950) (push) Successful in 3m51s Details Unit Tests / Compile-only (DEV=NULL:NAK:sm_120) (push) Successful in 1m39s Details Unit Tests / Compile-only (DEV=NULL:IR3:a630) (push) Successful in 2m20s Details Unit Tests / Linux (amd gfx950) (push) Successful in 5m42s Details Unit Tests / Linux (ptx) (push) Successful in 4m1s Details Unit Tests / Linux (nv) (push) Successful in 5m13s Details Autogen / In-tree Autogen (macos) (push) Canceled after 0s Details Benchmarks / Mac pytest (push) Canceled after 0s Details Benchmarks / LLM (DEV=AMD) (push) Canceled after 0s Details Benchmarks / LLM (DEV=METAL) (push) Canceled after 0s Details Benchmarks / LLM (DEV=NV) (push) Canceled after 0s Details Benchmarks / HLB-CIFAR10 (DEV=AMD) (push) Canceled after 0s Details Benchmarks / HLB-CIFAR10 (DEV=METAL) (push) Canceled after 0s Details Benchmarks / HLB-CIFAR10 (DEV=NV) (push) Canceled after 0s Details Benchmarks / MLPerf (AMD) (push) Canceled after 0s Details Benchmarks / MLPerf (NV) (push) Canceled after 0s Details Benchmarks / Stable Diffusion (DEV=AMD) (push) Canceled after 0s Details Benchmarks / Stable Diffusion (DEV=METAL) (push) Canceled after 0s Details Benchmarks / Stable Diffusion (DEV=NV) (push) Canceled after 0s Details Benchmarks / Tests (DEV=AMD) (push) Canceled after 0s Details Benchmarks / Tests (DEV=METAL) (push) Canceled after 0s Details Benchmarks / Tests (DEV=NV) (push) Canceled after 0s Details Benchmarks / UsbGPU Benchmark (push) Canceled after 0s Details Benchmarks / comma Benchmark (0.11.2) (push) Canceled after 0s Details Benchmarks / comma Benchmark (0.11.0) (push) Canceled after 0s Details Benchmarks / DSP Benchmark (push) Canceled after 0s Details Benchmarks / UsbGPU Benchmark (comma) (push) Canceled after 0s Details Benchmarks / PCI Driver Benchmark (DEV=AMD) (push) Canceled after 0s Details Benchmarks / PCI Driver Benchmark (DEV=NV) (push) Canceled after 0s Details Benchmarks / LLVM Speed (push) Canceled after 0s Details Platform Tests / MacOS (unit) (push) Canceled after 0s Details Platform Tests / MacOS (unit, mock) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=METAL) (1) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=METAL) (2) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=CPU:CLANG) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=CPU:LLVM) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=CPU:LVP) (push) Canceled after 0s Details Platform Tests / MacOS (DEV=WEBGPU) (push) Canceled after 0s Details Platform Tests / Windows (DEV=CPU:CLANG) (push) Canceled after 0s Details Platform Tests / Windows (DEV=CPU:LLVM) (push) Canceled after 0s Details Platform Tests / Windows (DEV=CPU:X86) (push) Canceled after 0s Details Platform Tests / Windows (DEV=WEBGPU) (push) Canceled after 0s Details * ci: remove yaml anchors from test.yml for gitea actions compatibility Gitea Actions does not support YAML anchors/aliases, which causes the workflow to fail parsing. Replace the &linux/linux anchor with a plain runs-on: ubuntu-24.04. ci: keep runner selection, inline expression instead of anchors Instead of replacing the anchored runs-on with a plain ubuntu-24.04 (which drops the namespace-profile-tinygrad routing for collaborator PRs), inline the full ${{ }} expression at every job. No YAML anchors, works with runners that can't parse them (gitea runner), and identical behavior on GitHub Actions.	2026-08-13 10:17:22 -07:00
qazalandGitHub	cc6d33bde7	amd: fix ALL2ALL speed on amdgpu (gpt) (#17518 )	2026-08-13 16:24:45 +09:00
qazalandGitHub	16c5ff2490	add external_benchmark_all2all.py (#17507 ) * add external_benchmark_all2all.py * mv * more minimal * less * fix space	2026-08-13 15:50:43 +09:00
wozeparrotandGitHub	1b7f040984	fa: paas through window (#17517 )	2026-08-13 14:45:57 +08:00
qazalandGitHub	39d144546e	fix mxfp4 mem estimate (#17515 ) * add mem estimates * rename * move	2026-08-13 14:48:15 +09:00