no pm_fold_cast_const in UOp.simplify and hcq2 [pr]

* remove pm_fold_cast_const from dtype decomp [pr]

* fix

.github/actions/setup-tinygrad/action.yml

@@ -41,10 +41,6 @@ inputs:
     description: "Install LLVM?"
     required: false
     default: 'false'
-  tinydreno:
-    description: "Install tinydreno"
-    required: false
-    default: 'false'
   qemu:
     description: "Install qemu"
     required: false
@@ -277,12 +273,6 @@ runs:
       shell: bash
       run: brew install llvm@20
 
-    # *** tinydreno ***
-    - name: Install tinydreno (linux)
-      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
-
     # *** OpenCL ***
     - name: Install rusticl
       if: inputs.opencl == 'true'

.github/workflows/platform.yml

@@ -179,35 +179,3 @@ jobs:
       - name: Run test_tiny
         shell: bash
         run: python -m pytest -n=auto test/test_tiny.py --durations=20
-
-
-  qcomclcompiletests:
-    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-qcomcl
-          deps: testing_unit
-          tinydreno: 'true'
-      - name: Set env
-        shell: bash
-        run: printf "DEV=NULL:QCOMCL:a630\nNULL_ALLOW_COPYOUT=1" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-      - name: Run test_ops (IMAGE)
-        shell: bash
-        env:
-          IMAGE: 1
-          DEV: "NULL:QCOMCL:a630,IMAGE_PITCH_ALIGNMENT=64"
-        run: |
-          DEBUG=4 python test/backend/test_ops.py TestOps.test_gemm | grep read_imagef
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20

.github/workflows/test.yml

@@ -21,7 +21,7 @@ concurrency:
 jobs:
   docs:
     name: Docs
-    runs-on: &linux ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 10
     env:
       CHECK_OOB: 0
@@ -61,7 +61,7 @@ jobs:
 
   torchbackend:
     name: Torch Backend Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -88,7 +88,7 @@ jobs:
 
   bepython:
     name: Python Backend
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -126,7 +126,7 @@ jobs:
 
   linter:
     name: Linters
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 10
 
     steps:
@@ -157,7 +157,7 @@ jobs:
 
   nulltest:
     name: Null Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
 
     steps:
@@ -191,7 +191,7 @@ jobs:
 
   unittest:
     name: Unit Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
 
     steps:
@@ -228,7 +228,7 @@ jobs:
       matrix:
         group: [1, 2]
     name: SPEC=2 (${{ matrix.group }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -244,7 +244,7 @@ jobs:
 
   fuzzing:
     name: Fuzzing
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 10
     steps:
     - name: Checkout Code
@@ -260,7 +260,7 @@ jobs:
 
   testopenclimage:
     name: CL IMAGE Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
       - name: Checkout Code
@@ -280,7 +280,7 @@ jobs:
 
   testopenpilot:
     name: openpilot Compile Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
       - name: Checkout Code
@@ -309,7 +309,7 @@ jobs:
 
   testonnxcpu:
     name: ONNX (CPU) Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
 
     steps:
@@ -328,7 +328,7 @@ jobs:
 
   testoptim:
     name: Optimization Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     steps:
       - name: Checkout Code
@@ -360,7 +360,7 @@ jobs:
 
   testllm:
     name: Test LLM
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     env:
       CHECK_OOB: 0
@@ -387,7 +387,7 @@ jobs:
 
   testmodels:
     name: Models
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
       - name: Checkout Code
@@ -407,7 +407,7 @@ jobs:
 
   testdsp:
     name: Linux (DSP)
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -435,7 +435,7 @@ jobs:
           - 'WEBGPU'
 
     name: Linux (DEV=${{ matrix.dev }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     steps:
     - name: Checkout Code
@@ -461,7 +461,7 @@ jobs:
 
   testamdasm:
     name: AMD ASM IDE
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     env:
       DEV: MOCKKFD+AMD
@@ -507,7 +507,7 @@ jobs:
 
   hcq2:
     name: hcq2
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 5
     steps:
       - name: Checkout Code
@@ -521,10 +521,7 @@ jobs:
       - name: Run HCQ2 tests
         run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/test_tiny.py
       - name: Run HCQ2 multi-device tests
-        run: |
-          HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_multitensor.py \
-            TestMultiTensor.test_simple_add TestMultiTensor.test_shard_reduce \
-            TestMultiTensor.test_backward_sum TestMultiTensor.test_matmul_shard_0_0
+        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest -n=auto test/backend/test_multitensor.py
       - name: Run HCQ2 JIT tests
         run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_jit.py
       - name: Run HCQ2 unit tests
@@ -532,7 +529,7 @@ jobs:
 
   testmockam:
     name: Linux (am)
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     env:
       DEV: MOCKPCI+AMD
@@ -568,7 +565,7 @@ jobs:
         arch: [gfx1100, gfx1201, gfx950]
 
     name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     env:
       DEV: MOCKKFD+AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
@@ -606,7 +603,7 @@ jobs:
         backend: [ptx, nv]
 
     name: Linux (${{ matrix.backend }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     env:
       FORWARD_ONLY: 1
@@ -640,10 +637,17 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        backend: [ir3, nak]
-    name: Compile-only (${{ matrix.backend }})
-    runs-on: *linux
+        dev:
+          - 'NULL:IR3:a630'
+          - 'NULL:QCOMCL:a630'
+          - 'NULL:NAK:sm_120'
+    name: Compile-only (DEV=${{ matrix.dev }})
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
+    env:
+      NULL_ALLOW_COPYOUT: 1
+      DEV: ${{ matrix.dev }}${{ contains(matrix.dev, 'a630') && ',IMAGE_PITCH_ALIGNMENT=64' || '' }}
+      IMAGE: ${{ contains(matrix.dev, 'a630') && '1' || '0' }}
     steps:
       - name: Checkout Code
         uses: actions/checkout@v6
@@ -652,21 +656,14 @@ jobs:
         with:
           key: compile-${{ matrix.backend }}
           deps: "testing_unit mesa"
-      - name: Set env
+          qemu: ${{ contains(matrix.dev, 'QCOMCL') }}
+      - name: Test IMAGE
         shell: bash
-        run: printf "NULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'DEV=NULL:IR3:a630' || matrix.backend == 'nak' && 'DEV=NULL:NAK:sm_120' }}" >> $GITHUB_ENV
+        if: contains(matrix.dev, 'a630')
+        run: DEBUG=7 python3 test/backend/test_ops.py TestOps.test_gemm | grep isam
       - name: Run test_ops
         shell: bash
         run: |
           python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
           DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
           python -m pytest -n=auto test/backend/test_ops.py --durations=20
-      - name: Run test_ops (IMAGE)
-        if: matrix.backend == 'ir3'
-        shell: bash
-        env:
-          IMAGE: 1
-          DEV: "NULL:IR3:a630,IMAGE_PITCH_ALIGNMENT=64"
-        run: |
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_gemm | grep image_load
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20

examples/mlperf/models/gpt_oss.py

@@ -12,7 +12,7 @@ from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
 from tinygrad.uop.ops import Ops, UOp
 from extra.models.llama import apply_rotary_emb
 from extra.llama_kernels.rmsnorm import rmsnorm
-from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8
+from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8, asm_gemm, can_use_asm_gemm
 from extra.gemm.moe_gemm import grouped_mx_gemm
 from extra.gemm.moe_routing import route, dispatch, combine
 
@@ -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)
@@ -263,7 +263,11 @@ class GPTOSS:
                         w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
       h, *_ = self.run_layer(h, freqs_cis, mask_full, i % 2 == 0, attn_kwargs, ffn_kwargs, save=save)
 
-    logits = self.norm(h) @ self.output.T
+    h_normed = self.norm(h)
+    pad = (-self.dim) % 256
+    h_padded, w_padded = h_normed.pad((None, None, (0, pad))), self.output.pad(((0, 0), (0, pad)))
+    if ASM_GEMM and can_use_asm_gemm(h_padded, w_padded.T): logits = asm_gemm(h_padded, w_padded.T)
+    else: logits = h_normed @ self.output.T
     return logits
 
 def _get_pads(uop:UOp) -> list[UOp]:

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

@@ -46,7 +46,7 @@ export DATA_SEED=${DATA_SEED:-5760}
 export JITBEAM=${JITBEAM:-3}
 export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
 
-export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
+export FAKEDATA=${FAKEDATA:-$([[ "$DEV" == NULL:* ]] && echo 1 || echo 0)} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
   export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
 fi

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

@@ -1,8 +1,8 @@
 #!/usr/bin/env bash
 
 export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
+export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
+export PATH="$ROCM_PATH/bin:$PATH"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000

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

@@ -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}

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

@@ -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}

extra/export_model.py

@@ -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

extra/gemm/cdna_asm_gemm.py

@@ -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))))
 

extra/gemm/triton_nv_matmul.py

@@ -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

extra/hcq2/ops_amd2.py

@@ -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
@@ -290,12 +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:HCQBuffer): self.dev.iface.unmap(buf)
 
 @dataclass
 class AMDQueueDesc:
@@ -388,15 +390,24 @@ class KFDIface:
     return hcqbuf
 
   def free(self, mem):
+    self._unmap(mem)
+    if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
+    kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
+
+  def unmap(self, mem):
+    self._unmap(mem)
+    if getattr(mem, '_owns_kfd_handle', False): kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
+
+  def _unmap(self, mem):
     gpus = (ctypes.c_int32 * 1)(self.gpu_id)
     stm = kfd.AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(gpus), n_devices=1)
     assert stm.n_success == 1
-    if mem.owner == self.dev:
-      if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
-      kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
 
   def map(self, mem):
-    if mem.owner is not None and mem.owner._is_cpu(): return self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)
+    if mem.owner is not None and mem.owner._is_cpu():
+      mapped = self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)
+      mapped._owns_kfd_handle = True
+      return mapped
 
     c_gpus = (ctypes.c_int32 * 1)(self.gpu_id)
     stm = kfd.AMDKFD_IOC_MAP_MEMORY_TO_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(c_gpus), n_devices=1)
@@ -468,6 +479,7 @@ class PCIIface(PCIIfaceBase):
 
   def require_profile_mode(self): return True
   def is_wgp_active(self, xcc, se, sa, wgp) -> bool: return True # TODO: account for WGP disablement on some asics.
+  def unmap(self, mem): self.free(mem)
 
   def _compute_props(self):
     self.ip_versions = self.dev_impl.ip_ver

extra/runbook_digitalocean_mi350x.md

@@ -0,0 +1,217 @@
+# Runbook: Llama 3 8B Training on DigitalOcean MI350X
+
+## Machine Specs
+- 8x MI350X GPUs (gfx950, device ID 75b0), 288GB VRAM each
+- 2TB RAM, 192 CPUs, 2TB disk
+- ROCm 7.14 at `/opt/rocm` (NOT `/opt/rocm-7.1.1` like the submission scripts assume)
+- Python 3.12
+
+## Phase 1: System Setup
+
+### 1.1 Install packages
+```bash
+apt-get update
+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 --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:
+```bash
+apt-get install -y amdrocm-core-dev
+```
+This installs `hip/hip_runtime.h` at `/opt/rocm/core-7.14/include/hip/hip_runtime.h`.
+The symlink `/opt/rocm/include` → `/opt/rocm/core-7.14/include` makes it available at `/opt/rocm/include/hip/hip_runtime.h`.
+
+### 1.4 Configure ROCm comgr
+ROCm 7.14 ships comgr 3.3 at `/opt/rocm/lib/libamd_comgr.so`. tinygrad's DLL loader needs explicit env vars to find it (it searches for `libcomgr.so*` by default, not `libamd_comgr.so*`). Set these in the run command:
+```bash
+export COMGR_PATH=/opt/rocm/lib/libamd_comgr.so
+export COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so
+```
+Also add ROCm libs to ldconfig so comgr's shared library dependencies resolve:
+```bash
+cat > /etc/ld.so.conf.d/rocm.conf << 'EOF'
+/opt/rocm/lib
+/opt/rocm/lib/llvm/lib
+/opt/rocm/lib/rocm_sysdeps/lib
+EOF
+ldconfig
+```
+
+### 1.5 Install geohot tmux config
+```bash
+curl -sL https://raw.githubusercontent.com/geohot/configuration/master/.tmux.conf -o ~/.tmux.conf
+```
+
+### 1.6 Reload amdgpu driver
+tinygrad's HCQ backend needs `/dev/kfd` which is created by the amdgpu kernel driver.
+If the driver was unloaded, reload it:
+```bash
+modprobe amdgpu
+ls /dev/kfd  # should exist
+```
+
+## Phase 2: Clone tinygrad
+```bash
+cd /root
+git clone https://github.com/tinygrad/tinygrad.git
+cd tinygrad
+python3 -m pip install --break-system-packages -e .
+```
+
+## Phase 3: Download C4 Dataset
+
+The C4 data is on the MLCommons Cloudflare R2 bucket in Megatron-LM indexed format.
+
+```bash
+rclone config create mlc-training s3 provider=Cloudflare \
+  access_key_id=76ea42eadb867e854061a1806220ee1e \
+  secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 \
+  endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
+
+mkdir -p /raid/datasets/c4-8b
+rclone copy mlc-training:mlcommons-training-wg-public/llama3_1/datasets/c4/llama3_1_8b/ /raid/datasets/c4-8b/ -P
+```
+
+Files downloaded (~85GB total, ~6 minutes):
+- `c4-train.en_6_text_document.bin` (79 GB)
+- `c4-train.en_6_text_document.idx` (870 MB)
+- `c4-validation-91205-samples.en_text_document.bin` (159 MB)
+- `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
+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
+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
+
+| Variable | Value | Why |
+|---|---|---|
+| `COMGR_PATH` | `/opt/rocm/lib/libamd_comgr.so` | tinygrad's DLL loader needs explicit path to find comgr 3.3 |
+| `COMGR_3_PATH` | `/opt/rocm/lib/libamd_comgr.so` | comgr 3.x uses a separate `comgr_3` module with its own path var |
+| `CC` | `/opt/rocm/core-7.14/lib/llvm/bin/clang` | System clang doesn't know gfx950; must use ROCm's bundled clang |
+| `DEV` | `AMD:HIP` | Force HIPRenderer (comgr-based) over HIPCCRenderer (hipcc subprocess) |
+| `ROCM_PATH` | `/opt/rocm` | Script defaults to `/opt/rocm-7.1.1` which doesn't exist |
+| `WANDB` | `1` | Enable wandb logging (off by default) |
+
+## Architecture
+
+| Component | Source file |
+|---|---|
+| Model | `examples/mlperf/models/flat_llama.py` — FlatTransformer, FP8 MXFP4 weights, fused QKV, flash attention |
+| Trainer | `examples/mlperf/model_train.py` → `train_llama3()` |
+| Optimizer | `examples/mlperf/optim.py` — GradAccClipAdamW, master weights, FP8 re-quant |
+| LR schedule | `examples/mlperf/lr_schedulers.py` — CosineAnnealingLRWithWarmup |
+| Dataloader | `examples/mlperf/dataloader.py` — Megatron-LM indexed bin format |
+| ASM GEMM | `extra/gemm/cdna_asm_gemm.py` — gfx950 MFMA assembly, MXFP4 |
+| Flash attention | `extra/thunder/amd/fa.py` |
+| Fused kernels | `extra/llama_kernels/` — rmsnorm, silu, quantize, fused_ce |
+| GPU driver | `tinygrad/runtime/ops_amd.py` — HCQ, direct KFD ioctl |
+| Renderer | `tinygrad/renderer/cstyle.py` — HIPRenderer for gfx950 |
+| comgr compiler | `tinygrad/runtime/support/compiler_amd.py` — HIPCompiler using comgr 3.3 |
+
+## Troubleshooting
+
+### `'hip/hip_runtime.h' file not found`
+Install `amdrocm-core-dev`:
+```bash
+apt-get install -y amdrocm-core-dev
+```
+
+### `'gfx950' is not a recognized processor` + LLVM crash
+System clang doesn't know gfx950. Set `CC=/opt/rocm/core-7.14/lib/llvm/bin/clang`.
+
+### `comgr not available: try setting COMGR_PATH?`
+Add ROCm libs to ldconfig and set `COMGR_PATH` and `COMGR_3_PATH`:
+```bash
+# /etc/ld.so.conf.d/rocm.conf should contain /opt/rocm/lib paths
+ldconfig
+```
+
+### `comgr not available: try setting COMGR_3_PATH?`
+comgr 3.x uses a separate module. Set `COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so` too.
+
+### `No such file or directory: 'clang'`
+Install clang: `apt-get install -y clang` (for CPU compilation).
+For gfx950 HIP compilation, comgr (not clang) is used — ensure the ROCm 7.14 comgr 3.3 is properly loaded via `COMGR_PATH` and `COMGR_3_PATH`.
+
+## Appendix: KVM Virtualization Observations
+
+### Virtualization detection
+```
+$ systemd-detect-virt
+kvm
+$ lspci -nn | grep AMD
+83:00.0 ... Device [1002:75b0]
+```
+CPU flags include `hypervisor`. `dmesg` shows `Hypervisor detected: KVM`.
+
+### 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:
+```
+[  799.780369] amdgpu 0000:83:00.0: Failed to alloc msi vectors
+[  799.781476] amdgpu 0000:83:00.0: sw_init of IP block <vega20_ih> failed -22
+[  799.782724] amdgpu 0000:83:00.0: amdgpu_device_ip_init failed
+[  799.793885] amdgpu 0000:83:00.0: Fatal error during GPU init
+```
+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.
+
+### 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.

extra/thunder/amd/fa.py

@@ -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

test/backend/test_custom_kernel.py

@@ -422,9 +422,8 @@ class TestCustomKernel(unittest.TestCase):
       return Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
     GlobalCounters.reset()
     y = run(x[0]).realize()
-    # it's copying the input and the output
-    # TODO: subbuffer usage has runtime specific behavior, this will be fixed after the removal of SLICE.
-    assert_kernel_count(2 if y.device in ("CL", "WEBGPU") else 1)
+    # backends that support contiguous views don't launch extra kernels
+    assert_kernel_count(2 if x[0].uop.contiguous_view() is None else 1)
     self.assertEqual(y.tolist(), [1, 2, 3, 4])
 
   @Context(DEV="CPU")

test/backend/test_jit.py

@@ -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")

test/backend/test_linearizer.py

@@ -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)]

test/backend/test_multitensor.py

@@ -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()

test/backend/test_pickle.py

@@ -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)

test/external/external_benchmark_all2all.py

@@ -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()

test/external/external_test_schedule_scaling.py

@@ -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)

test/helpers.py

@@ -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

test/mockgpu/amd/emu.py

@@ -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())

test/models/test_whisper.py

@@ -74,6 +74,7 @@ class TestWhisper(unittest.TestCase):
         err
       )
 
+  @slow
   def test_transcribe_file1(self):
     self.assertEqual(transcribe_file(self.model, self.enc, TEST_FILE_1),  TRANSCRIPTION_1)
 
@@ -89,6 +90,7 @@ class TestWhisper(unittest.TestCase):
     self.assertEqual(TRANSCRIPTION_1,  transcriptions[0])
     self.assertEqual(TRANSCRIPTION_2,  transcriptions[1])
 
+  @slow
   def test_transcribe_batch21(self):
     waveforms = [load_file_waveform(TEST_FILE_2), load_file_waveform(TEST_FILE_1)]
     transcriptions = transcribe_waveform(self.model, self.enc, waveforms)

test/null/test_const_folding.py

@@ -1,6 +1,6 @@
 import unittest, itertools, math
 from tinygrad import Tensor, dtypes, Context
-from tinygrad.dtype import DType, ConstType, 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):

test/null/test_graph_rewrite.py

@@ -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

test/null/test_memory_planner.py

@@ -27,10 +27,15 @@ def _make_linear(buffer_lists, copies=None):
     calls.append(UOp(Ops.CALL, src=(src0, *bufs)))
   return UOp(Ops.LINEAR, src=tuple(calls))
 
+def _get_planned_view(buf:UOp) -> tuple[UOp, int, int]|None:
+  view = buf.src[0] if buf.op is Ops.BITCAST else buf
+  if view.op is not Ops.SHRINK or view.src[0].op is not Ops.BUFFER: return None
+  return (arena:=view.src[0]), view.src[1].val * arena.dtype.itemsize, view.src[2].val * arena.dtype.itemsize
+
 def _get_arena(buf, linear, result):
   for orig_si, new_si in zip(linear.src, result.src):
     for orig, new in zip(orig_si.src[1:], new_si.src[1:]):
-      if orig is buf and new.op is Ops.SLICE: return new.src[0]
+      if orig is buf and (planned:=_get_planned_view(new)) is not None: return planned[0]
   return None
 
 def check_assign(buffer_lists, copies=None):
@@ -41,8 +46,8 @@ def check_assign(buffer_lists, copies=None):
   replace_map: dict[int, tuple[UOp, int, int]] = {}
   for orig_si, new_si in zip(linear.src, result.src):
     for orig, new in zip(orig_si.src[1:], new_si.src[1:]):
-      if new.op is Ops.SLICE and id(orig) not in replace_map:
-        replace_map[id(orig)] = (new.src[0], new.src[1].val * new.src[0].dtype.itemsize, new.arg * new.dtype.itemsize)
+      if (planned:=_get_planned_view(new)) is not None and id(orig) not in replace_map:
+        replace_map[id(orig)] = planned
 
   # verify pinned buffers are not planned
   for buf in held_bufs:

test/null/test_simplify_valid_idx.py

@@ -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):

test/null/test_uop_graph.py

@@ -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

test/null/test_uop_symbolic.py

@@ -3,7 +3,6 @@ import unittest, pickle, functools, math
 import z3
 
 from tinygrad.dtype import dtypes, ConstType, DType, Invalid
-from 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
@@ -16,7 +15,8 @@ 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)
@@ -442,7 +442,7 @@ 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 +483,12 @@ 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)")
 
@@ -948,6 +948,11 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(cond.where(u0, u1), 0, 1, "((a<2)!=True)")
     self.helper_test_variable(cond.where(u0, u1).where(u0, u1), 0, 1, "(a<2)")
 
+  def test_equivalent_const_max(self):
+    x = Variable("x", -10, 10)
+    self.helper_test_variable((x < 0).where(0, x), 0, 10, "x.maximum(0)")
+    self.helper_test_variable((0 < x).where(x, 0), 0, 10, "x.maximum(0)")
+
   def test_where_combine(self):
     cond = Variable("x", 0, 3) < 2
     a = Variable("a", 0, 3)
@@ -992,7 +997,7 @@ 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):
@@ -1007,22 +1012,6 @@ 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
@@ -1175,7 +1164,7 @@ 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)
@@ -1369,6 +1358,16 @@ 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):

test/null/test_validate_oob.py

@@ -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."""
 

test/null/test_viz.py

@@ -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()

test/opt/test_kernel_opts.py

@@ -239,6 +239,15 @@ class TestKernelOpts(unittest.TestCase):
     helper_linearizer_opt(a.sum().exp(), [[Opt(OptOps.PADTO, 0, 32)],])
     helper_linearizer_opt(a.sum(0).exp(), [[Opt(OptOps.PADTO, 1, 32)],])
 
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  @unittest.expectedFailure
+  def test_padto_group_full_unroll_sum(self):
+    a = Tensor.ones(2, 28, 4096, dtype=dtypes.bfloat16).realize()
+    out = ((a * 0.5).float().square()).sum(axis=(0, 2))
+    opts_to_apply = [Opt(OptOps.GROUPTOP, 1, 256), Opt(OptOps.PADTO, 3, 32), Opt(OptOps.UNROLL, 2, 0), Opt(OptOps.UPCAST, 0, 7)]
+    helper_linearizer_opt(out, [opts_to_apply], check_default_opt=False)
+
   def test_padto_sum(self):
     N = 18
     # NOTE: this setup prevents 17 * 17 contiguous merged into one dimension

test/unit/test_assign.py

@@ -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)

test/unit/test_hcq_graph.py

@@ -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

test/unit/test_llm_server.py

@@ -13,12 +13,18 @@ V_TOKS = UOp.variable("toks", 1, 32)  # 32 is the default chunk_size in generate
 class TestTransformerGenerate(unittest.TestCase):
   def test_warmup(self):
     model, calls = Transformer(TEST_CONFIG), []
-    def generate(tokens):
+    def generate(tokens, **kwargs):
       calls.append(tokens)
       yield from (1, 2)
     with patch.object(model, "generate", generate): model.warmup()
     self.assertEqual(calls, [[0], [0]])
 
+  def test_warmup_then_generate_with_default_chunk(self):
+    # warmup must not capture JIT graphs that generate()'s default chunk_size then rejects
+    model = Transformer(TEST_CONFIG)
+    model.warmup()
+    self.assertIsInstance(next(model.generate([5, 6, 7, 8])), int)
+
   def test_first_recurrent_generate_before_state_init(self):
     model = Transformer(TEST_CONFIG)
     model.has_recurrent_block = True
@@ -38,6 +44,15 @@ class TestTransformerGenerate(unittest.TestCase):
       next(model.generate([1, 2, 3, 4, 5, 42, 10]))
     self.assertEqual(calls, [((1, 1), V_START_POS.bind(5)), ((1, 1), V_START_POS.bind(6))])
 
+  def test_recurrent_divergent_prompt_restarts(self):
+    model, calls = Transformer(TEST_CONFIG), []
+    model.has_recurrent_block, model._cached_tokens = True, [1, 2, 9]
+    def mock_call(self, tokens, start_pos, temperature):
+      calls.append(start_pos)
+      return Tensor([[42]])
+    with patch.object(Transformer, '__call__', mock_call): next(model.generate([1, 2, 10, 11]))
+    self.assertEqual(calls[0], V_START_POS.bind(0))
+
   def test_template_starts_reasoning(self):
     router = StreamRouter(reasoning=True)
     self.assertEqual(list(router.route("reasoning</think>answer")),
@@ -178,6 +193,12 @@ class TestTransformerGenerate(unittest.TestCase):
     # with temperature=2.0, we should see at least 2 distinct outputs across 5 runs
     self.assertGreater(len(runs), 1, "high temperature should produce varied outputs")
 
+  def test_recurrent_temperature_high_produces_variety(self):
+    model = Transformer(TEST_CONFIG)
+    model.has_recurrent_block = True
+    outputs = {model.forward(Tensor([[1]]), 0, Tensor([2.0])).item() for _ in range(5)}
+    self.assertGreater(len(outputs), 1)
+
   def test_temperature_passed_to_forward(self):
     """Temperature from generate should be passed through to __call__."""
     model = Transformer(TEST_CONFIG)

test/unit/test_metal_graph.py

@@ -1,5 +1,4 @@
 import unittest
-from unittest.mock import MagicMock
 from tinygrad import Device
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.dtype import dtypes
@@ -11,36 +10,27 @@ class TestMetalGraph(unittest.TestCase):
     self.MetalGraph = MetalGraph
     self.dev = Device[Device.DEFAULT]
 
-  def metal_buf(self, offset):
-    buf = MagicMock()
-    if offset > 0:
-      buf.op = Ops.SLICE
-      src = MagicMock()
-      src.dtype = dtypes.uint8
-      buf.src = (src, UOp.const(offset))
-      buf.dtype = dtypes.uint8
-    else:
-      buf.op = Ops.BUFFER
-    buf.device = Device.DEFAULT
-    return buf
+  def metal_buf(self, offset, bitcast=False):
+    size = 4 if bitcast else 1
+    buf = UOp.new_buffer(Device.DEFAULT, offset+size, dtypes.uint8)
+    if offset: buf = buf[offset:offset+size]
+    return buf.bitcast(dtypes.float32) if bitcast else buf
 
-  def call(self, *bufs):
-    c = MagicMock()
-    c.src = (MagicMock(op=Ops.PROGRAM),) + tuple(bufs)
-    return c
+  def supports_uop(self, *bufs):
+    return self.MetalGraph.supports_uop([self.dev], UOp(Ops.PROGRAM, src=(UOp.sink(),)).call(*bufs))
 
   def test_supports_uop_normal_offset(self):
-    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF))) is True
+    assert self.supports_uop(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF)) is True
 
   def test_supports_uop_overflow_offset(self):
-    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(0x100000000))) is False
+    assert self.supports_uop(self.metal_buf(0), self.metal_buf(0x100000000)) is False
 
-  def test_supports_uop_nonmetal_buf(self):
-    # non-SLICE ops should not be checked for offset
-    buf = MagicMock()
-    buf.op = Ops.BUFFER
-    buf.device = Device.DEFAULT
-    self.MetalGraph.supports_uop([self.dev], self.call(buf))
+  def test_supports_uop_non_view_buf(self):
+    assert self.supports_uop(self.metal_buf(0)) is True
+
+  def test_supports_uop_bitcast(self):
+    assert self.supports_uop(self.metal_buf(0xFFFFFFFF, bitcast=True)) is True
+    assert self.supports_uop(self.metal_buf(0x100000000, bitcast=True)) is False
 
 if __name__ == "__main__":
   unittest.main()

tinygrad/codegen/__init__.py

@@ -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")
@@ -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

tinygrad/codegen/gpudims.py

@@ -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),

tinygrad/codegen/late/coalesce.py

@@ -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))

tinygrad/codegen/late/gater.py

@@ -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([

tinygrad/codegen/opt/tc.py

@@ -1,6 +1,7 @@
 import math, functools
 from dataclasses import dataclass
 from tinygrad.dtype import DType, dtypes
+from tinygrad.uop.ops import PatternMatcher, UOp, UPat, Ops
 
 @dataclass(frozen=True)
 class TensorCore: # D = A * B + C, A is (M x K), B is (K x N), C and D are (M x N)
@@ -135,6 +136,42 @@ amd_cdna4 = amd_cdna_1616128 + amd_cdna_161632 + amd_cdna_161616
 
 def get_amd(arch): return {"gfx942": amd_cdna3, "gfx950": amd_cdna4, "gfx1200": amd_rdna4, "gfx1201": amd_rdna4}.get(arch, amd_rdna3)
 
+pm_validate_wmma_rdna3 = PatternMatcher([
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.int32), lambda x: x.replace(
+    src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
+    if x.src[0].dtype == dtypes.int8 and x.src[0].max_numel() == 16 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.half), lambda x: UOp(Ops.STACK, src=tuple(x.replace(
+      src=(x.src[0], x.src[1], UOp(Ops.STACK, src=tuple(x.src[2].index(UOp.const(j//2, dtypes.int16))
+      if j%2 == 0 else UOp.const(0.0, x.src[2].dtype)
+      for j in range(x.max_numel()*2)))),
+      arg=(*x.arg[:4], None)).index(UOp.const(i*2, dtypes.int16))
+      for i in range(x.max_numel()))) if x.max_numel() == 8 else None),
+  (UPat(Ops.WMMA, name="x"), lambda x: x.replace(
+    src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
+    if x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 16 else None),
+])
+
+pm_validate_wmma_rdna4 = PatternMatcher([
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.bfloat16), lambda x: x.replace(
+    dtype=dtypes.uint16,
+    src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2].bitcast(dtypes.uint16)))
+      .bitcast(dtypes.bfloat16) if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
+    if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None)
+])
+
+pm_validate_wmma_cdna = PatternMatcher([
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
+    if x.arg[0][2] == 128 and x.src[0].dtype.itemsize <= 8 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
+    if x.max_numel() == 4 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 4 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64), x.src[2]))
+    if x.max_numel() == 4 and x.src[0].dtype in dtypes.fp8_ocp and x.src[0].max_numel() == 8 else None),
+])
 # ***** Apple Metal *****
 
 metal = [TensorCore(dims=(8,8,8), threads=32, elements_per_thread=(2,2,2), dtype_in=di, dtype_out=do,

tinygrad/codegen/simplify.py

@@ -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

tinygrad/engine/jit.py

@@ -44,9 +44,7 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
     current_batch, current_batch_devs = [], []
 
   for si in linear.src:
-    if si.src[0].op is Ops.SLICE: continue
-
-    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)
@@ -180,7 +178,7 @@ class CapturedJit(Generic[ReturnType]):
       if call.op is not Ops.CALL: continue
       arg_uops = get_call_arg_uops(call)
       outs, ins = get_call_outs_ins(call)
-      out |= {arg_uops[k] for k in set(outs) - set(ins) if arg_uops[k].op in (Ops.BUFFER, Ops.SLICE)}
+      out |= {b for k in set(outs) - set(ins) if (b:=u if (cv:=(u:=arg_uops[k]).contiguous_view()) is None else cv[0]).op is Ops.BUFFER}
     return out
 
   def __call__(self, input_uops:list[UOp], var_vals:dict[str, int]) -> ReturnType:

tinygrad/engine/realize.py

@@ -4,20 +4,20 @@ import time, random, itertools, math, contextlib, weakref, array
 from dataclasses import dataclass, replace, field
 from tinygrad.helpers import colored, DEBUG, GlobalCounters, ansilen, all_int, prod, flatten, Context, getenv, to_tuple
 from tinygrad.helpers import BEAM, size_to_str, time_to_str, VALIDATE_WITH_CPU, PROFILE, ProfilePointEvent, cpu_events, wait_cond
-from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, AxisType, sym_infer, buffers, graph_rewrite
-from tinygrad.device import Device, Buffer, MultiBuffer
+from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, AxisType, sym_infer, graph_rewrite
+from tinygrad.device import Device, Buffer, MultiBuffer, ProfileGraphEntry
 from tinygrad.renderer import Estimates
 from tinygrad.codegen import to_program
 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]
   if ast.op is Ops.PROGRAM: return tuple(ast.arg.outs), tuple(ast.arg.ins)
-  if ast.op in (Ops.COPY, Ops.SLICE): return (0,), (1,)
+  if ast.op is Ops.COPY: return (0,), (1,)
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return (0,), tuple(range(1, len(get_call_arg_uops(call))))
   return (), ()
 
@@ -27,9 +27,6 @@ def get_call_name(call:UOp, bufs:Sequence[Buffer|UOp], var_vals:dict[str, int]|N
 
   ast, arg_uops = call.src[0], get_call_arg_uops(call)
   if ast.op is Ops.PROGRAM: return ast.arg.name
-  if ast.op is Ops.SLICE:
-    offset = ast.src[1].val * arg_uops[1].dtype.itemsize
-    return colored(f"view {_uop_sz_to_str(arg_uops[0]):>10} @ {offset:<10d}", "yellow")
   if ast.op is Ops.COPY: return colored(f"copy {_uop_sz_to_str(arg_uops[0]):>10}, {_dev_str(bufs[0]):>7s} <- {_dev_str(bufs[1]):7s}", "yellow")
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return colored(f"enc/dec {_uop_sz_to_str(arg_uops[0])}", "yellow")
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph": return colored(f"batched {len(ast.src[0].src)}", "cyan")
@@ -140,7 +137,7 @@ class ExecContext:
   cache: bool = True
 
 def _resolve(b:UOp, inputs:tuple[UOp, ...]) -> UOp:
-  if b.op in (Ops.SLICE, Ops.MSELECT) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg.slot], *b.src[1:]))
+  if b.op in (Ops.MSELECT, Ops.SHRINK) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg.slot], *b.src[1:]))
   if b.op is Ops.MSTACK: return b.replace(src=tuple(_resolve(x, inputs) for x in b.src))
   return inputs[b.arg.slot] if b.op is Ops.PARAM else b
 def resolve_params(call:UOp, inputs:tuple[UOp, ...]) -> list[UOp]: return [_resolve(b, inputs) for b in get_call_arg_uops(call)]
@@ -154,13 +151,6 @@ def unwrap_multi(call:UOp, resolved:list[UOp]) -> Iterator[tuple[list[Buffer], d
                    for x in call.src[0].toposort())
     for j, per_dev in enumerate(zip(*[cast(MultiBuffer, b).bufs for b in bufs])): yield list(per_dev), {"_device_num": j} if has_dnum else {}
 
-def exec_view(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
-  resolved = resolve_params(call, ctx.input_uops)
-  bufs = [cast(Buffer, b.buffer) for b in resolved]
-  bv = bufs[1].view(resolved[0].max_numel(), ast.dtype, ast.src[1].val*bufs[1].dtype.itemsize)
-  with track_stats(ctx, call, bv.device, [bv, bufs[1]], ctx.var_vals): buffers[resolved[0]] = bv
-  return None
-
 def exec_copy(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
   for bufs, device_vars in unwrap_multi(call, resolve_params(call, ctx.input_uops)):
     dest, src = bufs[0].ensure_allocated(), bufs[1].ensure_allocated()
@@ -210,27 +200,30 @@ def exec_graph(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
   return t[0]
 
 def exec_hcq(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
-  if (inputs:=call.arg.aux.inputs) is not None:
+  if (info:=call.arg.aux).inputs is not None:
     bufs = [_resolve(ctx.input_uops[i], ctx.input_uops).buffer for i in call.arg.aux.input_idxs]
-    table = call.src[1+inputs].buffer
+    table = call.src[1+info.inputs].buffer
     for j,dev in enumerate(call.arg.aux.device):
       addrs = array.array('Q', [(b.bufs[j] if isinstance(b, MultiBuffer) else b).get_buf(dev).va_addr for b in bufs])
       mv = (table.bufs[j] if isinstance(table, MultiBuffer) else table).ensure_allocated()._buf.cpu_view().view(fmt='Q')
       wait_cond(lambda: mv[0], value=0, timeout_ms=ctx.timeout or getenv("HCQDEV_WAIT_TIMEOUT_MS", 30000), msg=f"{dev} hang detected")
       mv[:len(addrs)] = addrs
 
-  exec_kernel(replace(ctx, update_stats=False), call, ast)
+  exec_kernel(replace(ctx, update_stats=DEBUG>=3), call, ast)
 
-  tms:list[float|None] = []
-  for e in (aux:=call.arg.aux).prof: cast(Any, Device[e.device]).prof_ents[e.st_id] = e
-  for d in [cast(Any, Device[x]) for x in aux.device]:
-    with track_stats(ctx, call, d.device, [], ctx.var_vals) as et:
-      if ctx.wait:
-        d.synchronize(timeout=ctx.timeout)
-        ts = [d.signal(i)._buf.cpu_view().view(fmt='Q')[0] for e in aux.prof if e.device == d.device for i in (e.st_id, e.en_id)]
-        if ts: et[0] = float(max(ts)-min(ts))/d.timestamp_divider/1e6
-      tms += et
-  return tms[0]
+  tms = []
+  for devices,name,estimates,prof in info.kernels:
+    for device in devices:
+      tm = None
+      if prof:
+        (d:=cast(Any, Device[device])).prof_ents[prof[0]] = ProfileGraphEntry(device, name, *prof)
+        if ctx.wait:
+          d.synchronize(timeout=ctx.timeout)
+          st, en = (d.signal(x)._buf.cpu_view().view(fmt='Q')[0] for x in prof)
+          tms.append(tm:=float(en-st)/d.timestamp_divider/1e6)
+      stat_call = call.replace(arg=replace(call.arg, name=name, aux=replace(info, estimates=estimates, kernels=())))
+      with track_stats(ctx, stat_call, device, [], ctx.var_vals) as et: et[0] = tm
+  return max(tms) if tms else None
 
 # flatten LINEAR-in-LINEAR: any nested LINEAR child gets inlined into its parent's src
 pm_flatten_linear = PatternMatcher([
@@ -261,7 +254,6 @@ pm_optimize_local_size = PatternMatcher([
 ])
 
 pm_exec = PatternMatcher([
-  (UPat(Ops.CALL, src=(UPat(Ops.SLICE, name="ast"),), name="call", allow_any_len=True), exec_view),
   (UPat(Ops.CALL, src=(UPat(Ops.COPY, name="ast"),), name="call", allow_any_len=True), exec_copy),
   (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="ast"),), name="call", allow_any_len=True), exec_kernel),
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="encdec", name="ast"),), name="call", allow_any_len=True), exec_encdec),
@@ -276,7 +268,7 @@ def compile_linear(linear:UOp, beam:int|None=None, validate=False, input_uops:li
   if validate: linear = graph_rewrite(linear, pm_validate, name="validate", walk=True)
   if (beam_val:=BEAM.value if beam is None else beam) >= 1: linear = graph_rewrite(linear, pm_beam, ctx=beam_val, walk=True)
   linear = graph_rewrite(linear, pm_compile, name="precompile kernels", walk=True)
-  if getenv("HCQ2"): linear = hcq_compile(linear, input_uops, bool(PROFILE) if profile is None else profile)
+  if getenv("HCQ2"): linear = hcq_compile(linear, input_uops, bool(PROFILE or DEBUG >= 2) if profile is None else profile)
   return graph_rewrite(linear, pm_optimize_local_size, name="optimize local size", walk=True)
 
 def link_linear(linear:UOp, cache=True) -> UOp: return hcq_link(linear, cache=cache) if getenv("HCQ2") else linear

tinygrad/function.py

@@ -12,7 +12,7 @@ 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),
 ])

tinygrad/helpers.py

@@ -1,9 +1,9 @@
 from __future__ import annotations
 import time
 START_TIME = time.perf_counter()
-import os, functools, platform, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass, gc
+import os, functools, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass, gc
 from collections import defaultdict
-import subprocess, shutil, math, types, copyreg, inspect, importlib, decimal, itertools, difflib
+import shutil, math, types, copyreg, inspect, importlib, decimal, itertools, difflib
 from dataclasses import dataclass, field, replace
 from typing import ClassVar, Iterable, Any, TypeVar, Callable, Sequence, TypeGuard, Iterator, Generic, Generator, cast, overload
 
@@ -13,8 +13,7 @@ U = TypeVar("U")
 def prod(x:Iterable[T]) -> T|int: return functools.reduce(operator.mul, x, 1)
 
 # NOTE: helpers is not allowed to import from anything else in tinygrad
-OSX, WIN = platform.system() == "Darwin", sys.platform == "win32"
-ARCH_X86 = any(x in platform.processor() for x in ("Intel", "i386", "x86_64"))
+OSX, WIN = sys.platform == "darwin", sys.platform == "win32"
 BASEDIR = pathlib.Path(__file__).parent
 
 # fix colors on Windows, https://stackoverflow.com/questions/12492810/python-how-can-i-make-the-ansi-escape-codes-to-work-also-in-windows
@@ -231,7 +230,7 @@ class _DEV(ContextVar):
 
 DEV, DEBUG, BEAM, NOOPT = _DEV("DEV", ""), ContextVar("DEBUG", 0), ContextVar("BEAM", 0), ContextVar("NOOPT", 0)
 IMAGE, FLOAT16, OPENPILOT_HACKS = ContextVar("IMAGE", 0), ContextVar("FLOAT16", 0), ContextVar("OPENPILOT_HACKS", 0)
-JIT, JIT_BATCH_SIZE = ContextVar("JIT", 2 if OSX and ARCH_X86 else 1), ContextVar("JIT_BATCH_SIZE", 32)
+JIT, JIT_BATCH_SIZE = ContextVar("JIT", 1), ContextVar("JIT_BATCH_SIZE", 32)
 CHUNK_SIZE = 2**20  # TinyFS content-addressed store: blob chunk + hash-tree node granularity
 WINO, CAPTURING, TRACEMETA, NO_COLOR = ContextVar("WINO", 0), ContextVar("CAPTURING", 1), ContextVar("TRACEMETA", 1), ContextVar("NO_COLOR", 0)
 TRAINING = ContextVar("TRAINING", 0)
@@ -454,9 +453,9 @@ def _ensure_downloads_dir() -> pathlib.Path:
   if pathlib.Path("/etc/tinybox-release").is_file():
     # try creating dir with sudo
     if not (downloads_dir := pathlib.Path("/raid/downloads")).exists():
-      subprocess.run(["sudo", "mkdir", "-p", downloads_dir], check=True)
-      subprocess.run(["sudo", "chown", "tiny:root", downloads_dir], check=True)
-      subprocess.run(["sudo", "chmod", "775", downloads_dir], check=True)
+      system(f"sudo mkdir -p {downloads_dir}")
+      system(f"sudo chown tiny:root {downloads_dir}")
+      system(f"sudo chmod 775 {downloads_dir}")
     return downloads_dir
   return pathlib.Path(cache_dir) / "downloads"
 
@@ -497,6 +496,7 @@ def fetch_fw(path:str, name:str, sha256:str) -> bytes:
 # *** Exec helpers
 
 def system(cmd:str, **kwargs) -> str:
+  import subprocess
   st = time.perf_counter()
   try: ret = subprocess.check_output(cmd.split(), stderr=subprocess.STDOUT, **kwargs).decode().strip()
   except subprocess.CalledProcessError as e:

tinygrad/llm/model.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 import functools, itertools, pathlib
 from dataclasses import dataclass, replace
-from tinygrad import Tensor, nn, UOp, TinyJit, getenv, function
+from tinygrad import Tensor, nn, UOp, TinyJit, getenv, function, dtypes
 from tinygrad.nn import Linear
 from tinygrad.llm.gguf import gguf_load
 from tinygrad.uop.ops import resolve
@@ -187,8 +187,8 @@ class TransformerBlock(FFNBlock):
 
   def _init_state(self, x:Tensor):
     if not hasattr(self, "cache_kv"):
-      # TODO: how is the dtype of this determined?
-      self.cache_kv = Tensor.empty(2, x.shape[0], self.config.n_kv_heads, self.config.max_context, self.config.head_dim, device=x.device)
+      self.cache_kv = Tensor.empty(2, x.shape[0], self.config.n_kv_heads, self.config.max_context, self.config.head_dim,
+                                   dtype=dtypes.default_float, device=x.device)
       self.freqs_cis = precompute_freqs_cis(self.config.rope_dim, self.config.max_context, self.config.rope_theta, device=x.device)
 
 class MLATransformerBlock(FFNBlock):
@@ -261,13 +261,14 @@ class GatedDeltaNetBlock(FFNBlock):
   def _attention(self, x:Tensor, start_pos:int|UOp) -> Tensor:
     B, T, _ = x.shape
     assert T == 1, "GatedDeltaNetBlock currently only supports T=1"
+    is_kda = hasattr(self, "ssm_g_a")
 
     # input processing
     x = x.half()
-    out_gate = self.ssm_g_b(self.ssm_g_a(x)) if hasattr(self, "ssm_g_a") else self.attn_gate(x)
+    out_gate = self.ssm_g_b(self.ssm_g_a(x)) if is_kda else self.attn_gate(x)
     out_gate = out_gate.reshape(B, 1, self.num_v_heads, self.head_v_dim)
     beta = self.ssm_beta(x).sigmoid().reshape(B, self.num_v_heads, 1, 1)
-    alpha = self.ssm_f_b(self.ssm_f_a(x)) if hasattr(self, "ssm_f_a") else self.ssm_alpha(x)
+    alpha = self.ssm_f_b(self.ssm_f_a(x)) if is_kda else self.ssm_alpha(x)
     alpha = ((alpha.float() + self.ssm_dt["bias"]).softplus().reshape(B, self.num_v_heads, -1) *
              self.ssm_a.reshape(1, self.num_v_heads, -1)).exp().unsqueeze(-2)
 
@@ -291,14 +292,13 @@ class GatedDeltaNetBlock(FFNBlock):
 
     # output
     core_attn_out = self.ssm_norm((recurrent_state@q).squeeze(-1).reshape(B, 1, self.num_v_heads, self.head_v_dim))
-    out_gate = out_gate.sigmoid() if hasattr(self, "ssm_g_a") else out_gate.silu()
+    out_gate = out_gate.sigmoid() if is_kda else out_gate.silu()
     return self.ssm_out((core_attn_out * out_gate).reshape(B, 1, -1).cast(x.dtype))
 
   # recurrent state can't be partially reused after divergence, force a full rebuild
   def _state_reset_ops(self):
     return [self.conv_state.assign(self.conv_state.const_like(0)),
             self.recurrent_state.assign(self.recurrent_state.const_like(0))] if hasattr(self, "conv_state") else []
-  def _reusable_prefix_len(self, prefix_len:int, cached_len:int) -> int: return 0 if prefix_len != cached_len else prefix_len
 
   def _init_state(self, x):
     if not hasattr(self, "conv_state"):
@@ -326,7 +326,8 @@ class Transformer:
   def forward(self, tokens:Tensor, start_pos:int|UOp, temperature:Tensor) -> Tensor:
     x = self.token_embd(tokens).float()                   # (B, T, D)
     for block in self.blk: x = block(x, start_pos)
-    logits = self.output(self.output_norm(x))[:, -1, :]
+    # only run the output projection on the last token
+    logits = self.output(self.output_norm(x[:, -1:]))[:, -1, :]
     # Gumbel-max trick: argmax(logits/temp - log(-log(uniform))) is equivalent to sampling from softmax(logits/temp)
     return (logits / temperature.maximum(1e-12) - (Tensor.rand_like(logits).maximum(1e-12).log().neg()).log()).argmax(-1, keepdim=True)
 
@@ -420,6 +421,10 @@ class Transformer:
     for _ in range(2): list(zip(range(2), self.generate([0])))
 
   def get_start_pos(self, tokens:list[int]) -> int:
+    # recurrent state can't be partially reused after divergence: reuse it only when tokens extend the cached prefix
+    if self.has_recurrent_block:
+      return len(self._cached_tokens) if self._cached_tokens and len(self._cached_tokens) < len(tokens) \
+        and tokens[:len(self._cached_tokens)] == self._cached_tokens else 0
     prefix_len = sum(1 for _ in itertools.takewhile(lambda ab: ab[0] == ab[1], zip(tokens[:-1], self._cached_tokens)))
     return min(block._reusable_prefix_len(prefix_len, len(self._cached_tokens)) for block in self.blk)
 

tinygrad/nn/state.py

@@ -1,4 +1,4 @@
-import json, math, pathlib, zipfile, pickle, tarfile, struct, functools, io, zlib
+import json, math, pathlib, struct, functools, io, zlib
 from collections import OrderedDict
 from typing import Any, Callable, BinaryIO, Iterable, cast
 from tinygrad.tensor import Tensor
@@ -219,6 +219,7 @@ def load_state_dict(model, state_dict:dict[str, Tensor], strict=True, verbose=Tr
 
 @accept_filename
 def zip_extract(t: Tensor) -> dict[str, Tensor]:
+  import zipfile
   files: dict[str, Tensor] = {}
   with zipfile.ZipFile(TensorIO(t), "r") as myzip:
     # sadly, the extra length needs to be read from the local header of each file.
@@ -249,6 +250,7 @@ def tar_extract(t: Tensor) -> dict[str, Tensor]:
   tensors = nn.state.tar_extract(Tensor(pathlib.Path("archive.tar")))
   ```
   """
+  import tarfile
   with tarfile.open(fileobj=TensorIO(t), mode="r") as tar:
     return {member.name:t[member.offset_data:member.offset_data+member.size] for member in tar if member.type == tarfile.REGTYPE}
 
@@ -303,6 +305,7 @@ def torch_load(t:Tensor) -> dict[str, Tensor]:
                "FloatTensor": None, "Parameter": Parameter}
   whitelist = {"torch", "collections", "numpy", "_codecs"}  # NOTE: this is not for security, only speed
   class Dummy: pass
+  import pickle, zipfile, tarfile
   class TorchPickle(pickle.Unpickler):
     def find_class(self, module, name):
       module_root = module.split(".")[0]

tinygrad/renderer/cstyle.py

@@ -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]}),

tinygrad/renderer/llvmir.py

@@ -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
@@ -279,43 +279,9 @@ exit: %packed = phi i32 [%packed_bf8, %do_bf8], [%packed_fp8, %do_fp8]\n  %trunc
       (UPat(Ops.WMMA, name="wmma"), lambda ctx, wmma, rdna4=AMDLLVMRenderer.is_rdna4(target.arch), cdna=self.is_cdna:
         render_wmma_amd(ctx, wmma, cdna, rdna4))
     ])
-    if self.is_cdna:
-      self.extra_matcher += PatternMatcher([
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
-          if x.arg[0][2] == 128 and x.src[0].dtype.itemsize <= 8 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-          if x.max_numel() == 4 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 4 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64), x.src[2]))
-          if x.max_numel() == 4 and x.src[0].dtype in dtypes.fp8_ocp and x.src[0].max_numel() == 8 else None),
-      ])
-    if target.arch in {"gfx1100", "gfx1151"}:
-      self.extra_matcher += PatternMatcher([
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.int32), lambda x: x.replace(
-          src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
-          if x.src[0].dtype == dtypes.int8 and x.src[0].max_numel() == 16 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.half), lambda x: UOp(Ops.STACK, src=tuple(x.replace(
-          src=(x.src[0], x.src[1], UOp(Ops.STACK, src=tuple(x.src[2].index(UOp.const(j//2, dtypes.int16))
-            if j%2 == 0 else UOp.const(0.0, x.src[2].dtype)
-            for j in range(x.max_numel()*2)))),
-          arg=(*x.arg[:4], None)).index(UOp.const(i*2, dtypes.int16))
-          for i in range(x.max_numel()))) if x.max_numel() == 8 else None),
-        (UPat(Ops.WMMA, name="x"), lambda x: x.replace(
-          src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-          if x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 16 else None),
-      ])
-    if target.arch in {"gfx1200", "gfx1201"}:
-      self.extra_matcher += PatternMatcher([
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.bfloat16), lambda x: x.replace(
-          dtype=dtypes.uint16,
-          src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2].bitcast(dtypes.uint16)))
-            .bitcast(dtypes.bfloat16) if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-          if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None)
-      ])
+    if self.is_cdna: self.extra_matcher += tc.pm_validate_wmma_cdna
+    if target.arch in {"gfx1100", "gfx1151"}: self.extra_matcher += tc.pm_validate_wmma_rdna3
+    if target.arch in {"gfx1200", "gfx1201"}: self.extra_matcher += tc.pm_validate_wmma_rdna4
 
   def supported_dtypes(self): return {d for d in super().supported_dtypes()
                                       if (d not in dtypes.fp8_ocp or self.target.arch == "gfx950") and d not in dtypes.fp8_fnuz}

tinygrad/renderer/nir.py

@@ -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)))),

tinygrad/runtime/autogen/__init__.py

@@ -1,4 +1,4 @@
-import glob, importlib, os, pathlib, shutil, subprocess, tarfile, tempfile
+import glob, importlib, os, pathlib, subprocess
 from tinygrad.helpers import fetch, flatten, system, getenv
 
 root = (here:=pathlib.Path(__file__).parent).parents[2]
@@ -31,6 +31,7 @@ def load(name, files, **kwargs):
   if not (f:=(root/(path:=kwargs.pop("path", __name__)).replace('.','/')/f"{name}.py")).exists() or getenv('REGEN'):
     files, kwargs['args'] = files() if callable(files) else files, args() if callable(args:=kwargs.get('args', [])) else args
     if (srcs:=kwargs.pop('srcs', None)):
+      import tempfile, tarfile
       srcpath = (td:=tempfile.TemporaryDirectory(f"autogen-src-{name.replace('/','-')}")).name + "/"
       for src in (srcs if isinstance(srcs, list) else [srcs]):
         if 'tar' in src:
@@ -157,7 +158,7 @@ def __getattr__(nm):
           *[f"python3 src/compiler/nir/nir_{s}_h.py --outdir gen" for s in ["intrinsics", "intrinsics_indices"]]]), cwd=path, shell=True, check=True),
   srcs="https://gitlab.freedesktop.org/mesa/mesa/-/archive/mesa-25.2.7/mesa-25.2.7.tar.gz",
   dll=f"'tinymesa_cpu' if DEV.renderer == 'LVP' else 'tinymesa', {tinymesa_path}, emsg='pip install tinymesa==25.2.7.2'",
-  prolog=["from tinygrad.helpers import DEV", "import gzip, base64, platform, sysconfig, os"],
+  prolog=["from tinygrad.helpers import DEV", "import gzip, base64, sysconfig, os"],
   epilog=lambda path: [system(f"{root}/extra/mesa/lvp_nir_options.sh {path}")])
     case "libclang":
       return load("libclang",

tinygrad/runtime/autogen/mesa.py

@@ -5,7 +5,7 @@ from typing import Literal, TypeAlias
 from tinygrad.runtime.support.c import _IO, _IOW, _IOR, _IOWR
 from tinygrad.runtime.support import c
 from tinygrad.helpers import DEV
-import gzip, base64, platform, sysconfig, os
+import gzip, base64, sysconfig, os
 dll = c.DLL('mesa', 'tinymesa_cpu' if DEV.renderer == 'LVP' else 'tinymesa', os.path.join(sysconfig.get_paths()['platlib'], 'tinymesa'), emsg='pip install tinymesa==25.2.7.2')
 class struct_u_printf_info(c.Struct): pass
 u_printf_info: TypeAlias = struct_u_printf_info

tinygrad/runtime/graph/hcq.py

@@ -6,7 +6,6 @@ from tinygrad.device import Buffer, BufferSpec, Compiled, Device, MultiBuffer, P
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import UOp, Ops, Variable
 from tinygrad.engine.jit import GraphRunner, MultiGraphRunner
-from tinygrad.runtime.ops_rdma import RDMACopyQueue
 
 class HCQGraph(MultiGraphRunner):
   def __init__(self, *args, **kwargs):
@@ -50,7 +49,7 @@ class HCQGraph(MultiGraphRunner):
 
     self.comp_queues: dict[HCQCompiled, HWQueue] = {dev: unwrap(dev.hw_compute_queue_t)() for dev in self.devices}
     self.copy_queues: dict[tuple[HCQCompiled, int], HWQueue] = {} # lazy allocation, keyed by (device, queue_idx)
-    self.rdma_queues: dict[tuple[HCQCompiled, HCQCompiled], RDMACopyQueue] = {} # lazy allocation, keyed by device pair
+    self.rdma_queues: dict[tuple[HCQCompiled, HCQCompiled], "RDMACopyQueue"] = {} # lazy allocation, keyed by device pair
     self.num_copy_queues: int = getenv("HCQ_NUM_SDMA", min(len(self.devices), 8) if ALL2ALL >= 1 else 1)
     self.num_rdma_ops: dict[tuple[HCQCompiled, HCQCompiled], int] = collections.defaultdict(int)
 
@@ -104,6 +103,7 @@ class HCQGraph(MultiGraphRunner):
       elif is_rdma:
         enqueue_queue = self.comp_queues[enqueue_dev]
         rdma_key = (cast(HCQCompiled, Device[bufs[0].device]).rdma_dev(), enqueue_dev.rdma_dev())
+        from tinygrad.runtime.ops_rdma import RDMACopyQueue
         self.rdma_queues.setdefault(rdma_key, RDMACopyQueue(enqueue_dev.rdma_dev()))
       else:
         assert (enqueue_dev.hw_copy_queue_t is not None), "device must implement a copy queue"

tinygrad/runtime/graph/metal.py

@@ -113,5 +113,6 @@ class MetalGraph(GraphRunner):
   @staticmethod
   def supports_uop(batch_devs, new_call:UOp) -> bool:
     # Metal ICB replay encodes offsets as uint32; reject if any Metal buffer offset exceeds 32-bit range.
-    if any(b.op is Ops.SLICE and b.src[1].val * b.src[0].dtype.itemsize > 0xFFFFFFFF for b in new_call.src[1:]): return False
+    for shrink in [s for src in new_call.src[1:] if (s:=src.src[0] if src.op is Ops.BITCAST else src).op is Ops.SHRINK]:
+      if shrink.src[1].val * shrink.src[0].dtype.itemsize > 0xFFFFFFFF: return False
     return GraphRunner.supports_uop(batch_devs, new_call)

tinygrad/runtime/ops_cpu.py

@@ -1,11 +1,11 @@
 from __future__ import annotations
-import platform, sys, os, ctypes, 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)

tinygrad/runtime/ops_python.py

@@ -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

tinygrad/runtime/support/compiler_qcom.py

@@ -1,6 +1,6 @@
-import ctypes, struct
+import ctypes, struct, platform, pathlib, shutil, subprocess, sys, tarfile, tempfile
 from tinygrad.device import Compiler
-from tinygrad.helpers import DEBUG, system
+from tinygrad.helpers import DEBUG, system, fetch, unwrap
 from tinygrad.runtime.support.compiler_mesa import disas_adreno
 # see https://github.com/sirhcm/tinydreno
 from tinygrad.runtime.autogen import llvm_qcom
@@ -10,10 +10,17 @@ def _read_lib(lib, off) -> int: return struct.unpack("I", lib[off:off+4])[0]
 class QCOMCompiler(Compiler):
   def __init__(self, arch:str):
     assert arch.split(',')[0] == "a630", "only a630 supported"
-    self.arch, self.chip_id, self.llvm_inst = arch, 0x6030001, llvm_qcom.cl_compiler_create_llvm_instance()
+    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.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)
+  def __del__(self): llvm_qcom.cl_compiler_destroy_llvm_instance(self.llvm_inst) if platform.machine() == "aarch64" else self.compiler_process.kill()
 
   def __reduce__(self): return QCOMCompiler, (self.arch,)
 
@@ -25,6 +32,10 @@ class QCOMCompiler(Compiler):
     return handle
 
   def compile(self, src) -> bytes:
+    if platform.machine() != "aarch64":
+      unwrap(self.compiler_process.stdin).write(struct.pack("I", len(src.encode())) + src.encode())
+      if (lib:=unwrap(self.compiler_process.stdout).read(struct.unpack("I", unwrap(self.compiler_process.stdout).read(4))[0])): return lib
+      raise RuntimeError("QCOM Compilation Error")
     ch = self.checked(llvm_qcom.cl_compiler_compile_source(self.llvm_inst, self.chip_id, llvm_qcom.CL_MODE_64BIT, b"", 0, 0, 0, src.encode(), 0,
                                                            llvm_qcom.CL_SRC_STR, None))
     if DEBUG >= 8: print(system("llvm-dis", input=ctypes.string_at((comp:=ch.contents.compiled.contents).llvm_bitcode, comp.llvm_bitcode_size)))
@@ -36,3 +47,13 @@ class QCOMCompiler(Compiler):
     return ret
 
   def disassemble(self, lib: bytes): disas_adreno(lib[(ofs:=_read_lib(lib, 0xc0)):ofs+_read_lib(lib, 0x100)], self.chip_id)
+
+if __name__ == "__main__":
+  compiler = QCOMCompiler(sys.argv[1])
+  while (amt:=sys.stdin.buffer.read(4)):
+    try: lib = compiler.compile(sys.stdin.buffer.read(struct.unpack("I", amt)[0]).decode())
+    except Exception as e:
+      lib = b""
+      print(e, file=sys.stderr, flush=True)
+    sys.stdout.buffer.write(struct.pack("I", len(lib)) + lib)
+    sys.stdout.buffer.flush()

tinygrad/runtime/support/hcq.py

@@ -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

tinygrad/runtime/support/hcq2.py

@@ -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"))
 
@@ -33,13 +33,13 @@ class HCQInfo:
 
   input_idxs:tuple[int, ...] = () # indexes into input_uops used by this call
   inputs:int|None = None
-  prof:tuple[ProfileGraphEntry, ...] = () # st_id/en_id are timestamp signal slots until collect
+  kernels:tuple[tuple[tuple[str, ...], str, Estimates, tuple[int, ...]], ...] = ()
 
 def all_devices_in(d:Any, c:frozenset[str]) -> bool: return {x.split(":")[0] for x in to_tuple(d)} <= c
 
 def unwrap_mstack(u):
   if u.op is Ops.MSTACK: return tuple(x for s in u.src for x in unwrap_mstack(s))
-  return unwrap_mstack(u.src[0]) if u.op in {Ops.MSELECT, Ops.SLICE} else (u,)
+  return unwrap_mstack(u.src[0]) if u.op is Ops.MSELECT else (u,)
 
 def is_value_known_at_link(val:UOp) -> bool:
   runtime_reads = [u for u in val.toposort() if u.op in (Ops.LOAD, Ops.INDEX)]
@@ -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))
@@ -136,19 +139,19 @@ 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)
   latest = {((dep[0][dlane], dep[1]), lane): (dep, dlane) for dep, dlane, lane in sorted(dep_lanes, key=lambda x: x[0][2])}
-  deps:dict[tuple, list[int|None]] = collections.defaultdict(lambda: [None]*len(devices))
-  for (_, lane), (dep, dlane) in latest.items(): deps[dep][lane] = dlane
+  deps:dict[tuple, dict[int, list[int]]] = collections.defaultdict(lambda: collections.defaultdict(list))
+  for (_, lane), (dep, dlane) in latest.items(): deps[dep][lane].append(dlane)
 
   waits = []
-  for (ddevs, dqueue, dtag), lanes in deps.items():
-    sig = UOp.mstack(*[make_signal(d, tag="sentinel_signal") if dl is None else make_signal(ddevs[dl], slots[dqueue])
-                       for dl, d in zip(lanes, devices)])
-    waits.append(UOp(Ops.INS, arg="wait", src=(sig, UOp.const(dtag + 1, dtypes.uint64))))
+  for (ddevs, dqueue, dtag), by_lane in deps.items():
+    for ls in itertools.zip_longest(*(by_lane[lane] for lane in range(len(devices)))):
+      s = UOp.mstack(*[make_signal(d, tag="sentinel_signal") if dl is None else make_signal(ddevs[dl], slots[dqueue]) for dl, d in zip(ls, devices)])
+      waits.append(UOp(Ops.INS, arg="wait", src=(s, UOp.const(dtag + 1, dtypes.uint64))))
   return waits, {dtag for _, _, dtag in deps}
 
 def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[tuple[tuple[str, ...], str]],
@@ -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]
@@ -199,10 +203,10 @@ def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]], profile:bool) -> li
     signal_tags |= cur_signal_tags
 
   # build fences and finalizers
-  fences, finalizers, finalizer_signal_tags = _build_finalizers(batch, batch_info, deps_tracker, slots)
+  fences, fins, finalizer_signal_tags = _build_finalizers(batch, batch_info, deps_tracker, slots)
   signal_tags |= finalizer_signal_tags
 
-  src, prof = [], []
+  src, kerns = [], []
   for tag, ((call, _), (devices, queue), q) in enumerate(zip(batch, batch_info, call_waits)):
     # first queue use, sync prior device work with the device timeline
     if batch_info.index((devices, queue)) == tag:
@@ -212,24 +216,24 @@ def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]], profile:bool) -> li
     # and make hcq call
     name, info = get_call_name(call, get_call_arg_uops(call)), HCQInfo(devices, estimate_uop(call))
     ts_ids = [next(UOp.unique_num) for _ in range(2)] if profile else []
-    prof += [ProfileGraphEntry(d, name, *ts_ids) for d in devices if ts_ids]
+    kerns.append((devices, name, info.estimates, tuple(ts_ids)))
 
     ts_ins = [UOp(Ops.INS, arg="timestamp", src=(make_signal(devices, s),)) for s in ts_ids]
     q += ts_ins[:1] + [call.replace(arg=replace(call.arg, aux=info))] + ts_ins[1:]
 
     # signal the queue if someone waits for us
     if tag in signal_tags: q += [UOp(Ops.INS, arg="store", src=(make_signal(devices, slots[queue]), UOp.const(tag + 1, dtypes.uint64)))]
-    src.append(make_call(name, make_submit(*q, devs=devices, queue=queue).sink(), info))
+    src.append(make_call(f"submit {name}", make_submit(*q, devs=devices, queue=queue).sink(), info))
 
   # append batch timestamps to finalizers
-  finalizers = [f.replace(arg=replace(f.arg, aux=replace(a:=f.arg.aux, prof=tuple(e for e in prof if e.device in a.device)))) for f in finalizers]
-  return fences + src + finalizers
+  fins = [f.replace(arg=replace(f.arg, aux=replace(a:=f.arg.aux, kernels=tuple(x for x in kerns if set(x[0]) & set(a.device))))) for f in fins]
+  return fences + src + fins
 
 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)?
@@ -377,14 +385,15 @@ pm_replace_params = PatternMatcher([
 
 # *****************
 
-def resolve_getaddr_slice(bv:UOp, g:UOp) -> UOp:
+def resolve_getaddr_view(bv:UOp, g:UOp) -> UOp:
   base = bv.src[0].after(*g.src[0].src[1:] if g.src[0].op is Ops.AFTER else ())
-  itemsize = bv.src[0].dtype.itemsize if bv.src[0].without_after.op in (Ops.BUFFER, Ops.SLICE, Ops.MSTACK, Ops.MSELECT) else bv.dtype.itemsize
+  if bv.op is Ops.BITCAST: return UOp(Ops.GETADDR, src=(base,), arg=g.arg)
+  itemsize = bv.src[0].dtype.itemsize if bv.src[0].without_after.op in (Ops.BUFFER, Ops.MSTACK, Ops.MSELECT) else bv.dtype.itemsize
   return UOp(Ops.GETADDR, src=(base,), arg=g.arg) + UOp.const(bv.src[1].val * itemsize, dtypes.uint64)
 
 pm_early_simplify = PatternMatcher([
-  (UPat(Ops.GETADDR, src=(UPat.any(sl:=UPat(Ops.SLICE, name="bv"), sl.after(allow_any_len=True)),), name="g"), resolve_getaddr_slice),
-  (UPat(Ops.INDEX, src=(UPat(Ops.SLICE, name="bv"),), allow_any_len=True, name="x"),
+  (UPat(Ops.GETADDR, src=(UPat((Ops.SHRINK, Ops.BITCAST), name="bv").or_after(),), name="g"), resolve_getaddr_view),
+  (UPat(Ops.INDEX, src=(UPat(Ops.SHRINK, name="bv"),), allow_any_len=True, name="x"),
    lambda bv,x: x.replace(src=(bv.src[0], x.src[1] + bv.src[1].cast(x.src[1].dtype), *x.src[2:]))),
 ])
 
@@ -402,7 +411,7 @@ def pack_hcq_placeholders(call:UOp) -> UOp|None:
       sizes[b.tag] = offs[b] + b.max_numel()
   counts = collections.Counter(b.tag for b in bufs)
   bases = {b.tag:UOp.placeholder((sizes[b.tag],), b.dtype, next(UOp.unique_num), device=b.device).rtag(b.tag) for b in bufs if counts[b.tag] > 1}
-  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(offs.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
+  subs = {b:bases[b.tag][(off:=offs.get(b, 0)):off+b.max_numel()] for b in bufs if b.tag in bases}
   return call.replace(src=(call.src[0].substitute(subs, walk=True), *call.src[1:])) if subs else None
 pm_pack_placeholders = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), pack_hcq_placeholders)])
@@ -436,7 +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
@@ -468,7 +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)
 
@@ -483,7 +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
@@ -493,7 +502,7 @@ pm_resolve_patches = PatternMatcher([
   (UPat(name="buf").index(UPat(Ops.RANGE), allow_any_len=True)
     .store(UPat.any(UPat(Ops.BINARY, name="blob"), UPat(Ops.BINARY, name="blob").bitcast()).index(UPat(Ops.RANGE), allow_any_len=True).load())
     .end(UPat(Ops.RANGE)), fold_binary),
-  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").index(UPat(Ops.STACK, name="off")).store(UPat(Ops.STACK, name="val")), fold_const_store),
+  (UPat({Ops.BUFFER, Ops.MSTACK}, name="buf").index(UPat(Ops.STACK, name="off")).store(UPat(Ops.STACK, name="val")), fold_const_store),
 ])
 
 pm_assert_no_afters = PatternMatcher([(UPat(Ops.AFTER, name="a"), lambda a: panic(RuntimeError, f"AFTER left at hcq_link: {a.src[0].op}"))])
@@ -509,7 +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
@@ -520,6 +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
@@ -576,9 +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")
@@ -616,21 +627,23 @@ 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)
 
   def _unmap(self, mb):
     self.dev.synchronize()
-    self.dev.iface.free(mb)
+    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)

tinygrad/runtime/support/system.py

@@ -1,7 +1,8 @@
 from __future__ import annotations
-import os, mmap, array, functools, ctypes, select, contextlib, dataclasses, sys, itertools, struct, socket, subprocess, time, enum, atexit
+import os, mmap, array, functools, ctypes, ctypes.util, select, contextlib, dataclasses, sys, itertools, struct, socket
+import subprocess, time, enum, atexit
 from tinygrad.helpers import round_up, getenv, OSX, temp, ceildiv, unwrap, fetch, system, _ensure_downloads_dir, DEBUG, flatten, pluralize
-from tinygrad.runtime.autogen import libc, pci, vfio, iokit, corefoundation
+from tinygrad.runtime.autogen import libc, pci, vfio
 from tinygrad.runtime.support.hcq import FileIOInterface, MMIOInterface, HCQBuffer, hcq_filter_visible_devices
 from tinygrad.runtime.support.memory import VirtMapping, AddrSpace, BumpAllocator
 from tinygrad.runtime.support.usb import USB3, CustomASM24Controller, USBMMIOInterface
@@ -55,6 +56,7 @@ class _System:
   def pci_scan_bus(self, vendor:int, devices:tuple[tuple[int, tuple[int, ...]], ...], base_class:int|None=None) -> list[str]:
     all_devs = []
     if OSX:
+      from tinygrad.runtime.autogen import iokit, corefoundation
       def read_prop(svc, key) -> int:
         cfkey = corefoundation.CFStringCreateWithCString(None, key.encode(), corefoundation.kCFStringEncodingUTF8)
         cfdata = ctypes.cast(iokit.IORegistryEntryCreateCFProperty(svc, ctypes.cast(cfkey, iokit.CFStringRef), None, 0), corefoundation.CFDataRef)

tinygrad/schedule/__init__.py

@@ -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
 

tinygrad/schedule/indexing.py

@@ -23,8 +23,8 @@ class IndexingContext:
     return UOp.range(s, next(self.range_idx), axistype) if resolve(s!=1) else UOp.const(0)
 
 
-ALWAYS_CONTIGUOUS: set[Ops] = {Ops.CONTIGUOUS, Ops.AFTER, Ops.BUFFER, Ops.SLICE,
-                      Ops.CONST, Ops.BIND, Ops.MSELECT, Ops.MSTACK, Ops.PARAM,
+ALWAYS_CONTIGUOUS: set[Ops] = {Ops.CONTIGUOUS, Ops.AFTER, Ops.BUFFER,
+                      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
@@ -34,12 +34,8 @@ def realize_srcs(ctx:IndexingContext, rb:UOp) -> None:
     if s.base.op not in ALWAYS_CONTIGUOUS: ctx.realize_map[s] = None
 
 def realize_store_after_src(ctx:IndexingContext, dest:UOp, src:UOp):
-  # don't realize SLICE when it's the direct source of STORE+AFTER — the target buffer is the output
-  if src.op is Ops.SLICE and src in ctx.realize_map \
-     and not dest.op_in_backward_slice_with_self(Ops.SHRINK, Ops.PERMUTE, Ops.FLIP, Ops.PAD):
-    del ctx.realize_map[src]
   # 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:]:
@@ -73,8 +69,10 @@ 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 GroupOp.Movement|{Ops.INDEX, Ops.SLICE, Ops.STAGE, Ops.REDUCE, Ops.AFTER, Ops.END}: return src[:1]
+  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
 
 def create_bufferize_and_index_srcs(ctx:IndexingContext, x:UOp) -> list[UOp]:
@@ -84,7 +82,7 @@ def create_bufferize_and_index_srcs(ctx:IndexingContext, x:UOp) -> list[UOp]:
   for i, s in enumerate(x.src):
     new_src = s
     src_rngs = broadcast_rngs(x, s, ctx.range_map[x][0]) if x in ctx.range_map else ()
-    if s.op in {Ops.PARAM, Ops.BUFFER, Ops.SLICE, Ops.MSTACK, Ops.MSELECT, Ops.AFTER}:
+    if s.op in {Ops.PARAM, Ops.BUFFER, Ops.MSTACK, Ops.MSELECT, Ops.AFTER}:
       if x in ctx.range_map and i < data_src_count: new_src = new_src.index(*src_rngs)
     elif s in ctx.realize_map:
       realized_ranges = ctx.realize_map[s]

tinygrad/schedule/memory.py

@@ -52,11 +52,9 @@ def memory_plan_rewrite(linear:UOp, held_bufs:set[UOp]|None=None) -> UOp:
     peaks[_key(buf)] = (max(peaks[_key(buf)][0], offsets[buf] + buf.max_numel() * buf.dtype.itemsize), peaks[_key(buf)][1])
   arena_sizes = {key: round_up(peak, block_size) for key, (peak, _) in peaks.items()}
 
-  # build replace_map: each buffer becomes a SLICE into a shared per-device-lane arena
+  # build replace_map: each buffer becomes a SHRINK/BITCAST into a shared per-device-lane arena
   arenas = {key: UOp.new_buffer(key[0], sz, dtypes.int8) for key, sz in arena_sizes.items()}
-  replace_map:dict[UOp, UOp] = {}
-  for buf_uop, offset in offsets.items():
-    replace_map[buf_uop] = UOp(Ops.SLICE, buf_uop.dtype, (arenas[_key(buf_uop)], UOp.const(offset)), buf_uop.max_numel())
+  replace_map = {buf_uop:arenas[_key(buf_uop)][offset:offset+buf_uop.nbytes()].bitcast(buf_uop.dtype) for buf_uop, offset in offsets.items()}
 
   if DEBUG >= 1 and (omem:=sum(nbytes.values()) / 1e6) != (nmem:=sum(arena_sizes.values()) / 1e6):
     print(f"memory reduced from {omem:.2f} MB -> {nmem:.2f} MB, {len(first_appearance)} -> {len(arenas)} bufs")

tinygrad/schedule/rangeify.py

@@ -6,11 +6,11 @@ from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, K
 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.movement import mop_cleanup
-from tinygrad.helpers import prod, getenv, dedup, all_int, DEBUG, SPLIT_REDUCEOP, VIZ, MAX_KERNEL_BUFFERS, SPEC
+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
 from tinygrad.codegen.simplify import pm_flatten_range, pm_reduce_simplify
 from tinygrad.codegen.opt import Opt
-from tinygrad.schedule.indexing import BufferizeOpts, IndexingContext, apply_movement_op
+from tinygrad.schedule.indexing import run_rangeify, BufferizeOpts, IndexingContext, apply_movement_op
 from tinygrad.schedule.multi import multi_pm
 from tinygrad.schedule.allreduce import create_allreduce_function
 
@@ -39,16 +39,7 @@ pm_fold_moved_after = PatternMatcher([
 def _mop_index(r:UOp, idx:UOp):
   idxs = idx.src[1:]
   if len(idxs) == len(r.shape):
-    ret = r.src[0].index(*apply_movement_op(r.op, r.src[0].shape, r.marg, idxs), dtype=idx.dtype, arg=idx.arg)
-    if r.op is Ops.PAD:
-      # insert 0 for PAD with where
-      # TODO: does this need simplify
-      a = UOp.const(True)
-      for s in ret.src[1:]:
-        if s.op is Ops.WHERE and s.src[2].op is Ops.CONST and s.src[2].arg == Invalid:
-          a = a & s.src[0]
-      ret = a.where(ret, ret.const_like(0))
-    return ret
+    return r.src[0].index(*apply_movement_op(r.op, r.src[0].shape, r.marg, idxs), dtype=idx.dtype, arg=idx.arg)
   if r.op is Ops.RESHAPE:
     src_prefix = len(r.src[0].shape) - len(r.shape[len(idxs):])
     if src_prefix >= 0 and r.src[0].shape[src_prefix:] == r.shape[len(idxs):]:
@@ -69,7 +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] = {}
@@ -470,11 +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)
@@ -484,10 +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
@@ -511,8 +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:
@@ -521,7 +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
@@ -550,13 +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),
@@ -576,101 +567,9 @@ def convert_copy_to_store(ctx, copy:UOp, existing_buf:UOp|None=None):
   # reshape back to input
   return buf.after(buf.store(input_src)).reshape(copy.shape)
 
-def convert_contig_to_store(ctx, copy:UOp):
-  input_src = copy.src[0]
-  # create the output buffer
-  buf = UOp(Ops.BUFFER, src=(shape_to_shape_arg(input_src.max_shape),), arg=ParamArg(next(ctx), copy.dtype, device=copy.device))
-  # reshape back to input
-  view = buf.shrink_to(input_src.shape)
-  return view.after(view.store(input_src))
-
 pm_copy_to_store = PatternMatcher([
   (UPat(name="existing_buf").store(UPat(Ops.COPY, name="copy")), convert_copy_to_store),
   (UPat(Ops.COPY, name="copy"), convert_copy_to_store),
-  (UPat(Ops.CONTIGUOUS, name="copy"), convert_contig_to_store),
-])
-
-# **** simple rangeify ****
-
-from tinygrad.helpers import all_same
-from tinygrad.uop.ops import _broadcast_shape
-
-def expand_broadcast(x:UOp):
-  shapes = [u._shape for u in x.src]
-  if any(s is None for s in shapes) or all_same(shapes): return None
-  shape = _broadcast_shape(*shapes)
-  return x.replace(src=tuple([u.expand(shape) for u in x.src]))
-
-pm_expand_broadcast = PatternMatcher([
-  # expand broadcasts first
-  (UPat(GroupOp.Binary|GroupOp.Ternary|{Ops.STORE}, name="x"), expand_broadcast),
-])
-
-def expand_coeff(sink:UOp) -> dict[UOp,int]:
-  coeff: dict[UOp,int] = {sink: 1}
-  contig: dict[UOp,int] = {}
-  for u in reversed(list(sink.toposort())):
-    c = 1 if u.op is Ops.STORE else coeff.get(u, 0)
-    # symbolic coeffs mark on vmax, an extra CONTIGUOUS is always safe
-    if (c > 1 if isinstance(c, int) else c.vmax > 1) and u.op in (GroupOp.Elementwise | {Ops.REDUCE}) and u.device is not None:
-      contig[u] = c
-      c = 1
-    coeff[u] = c
-    mult = prod(u.shape) // prod(u.src[0].shape) if u.op is Ops.EXPAND else 1
-    for s in u.src: coeff[s] = coeff.get(s, 0) + c * (mult if s is u.src[0] else 1)
-  return contig
-
-def rangeify_on_reduce(ctx, inp:UOp, red:UOp, idx:UOp|None=None):
-  if red.arg[1] == 0: return None
-  if idx is None and len(red.shape) > 0: return None
-  # TODO: is AxisType.REDUCE a real thing?
-  rngs = [UOp.range(s, next(ctx), AxisType.REDUCE) for s in inp.shape[:red.arg[1]]]
-  return inp.index(*rngs, *(idx.src[1:] if idx is not None else ())).reduce(*rngs, arg=(red.arg[0], 0))
-
-def rangeify_on_store(ctx, x:UOp):
-  if x.shape == (): return None
-  rngs = [UOp.range(s, next(ctx)) for s in x.shape]
-  return x.src[0].index(*rngs).store(x.src[1].index(*rngs)).end(*rngs)
-
-def rangeify_on_stage(ctx, x:UOp):
-  if x.src[0].shape == (): return None
-  # size 1 dims don't get ranges, they are reshaped out and back in
-  if all_int(x.shape) and 0 < len(sq := tuple(s for s in x.shape if s != 1)) < len(x.shape):
-    return rangeify_on_stage(ctx, x.src[0].reshape(sq).bufferize(arg=x.arg)).reshape(x.shape)
-  rngs = [UOp.range(s, next(ctx)) for s in x.shape]
-  return x.replace(src=(x.src[0].index(*rngs), *rngs))
-
-def index_on_stack(stack:UOp, idx:UOp):
-  srcs = [s.index(*idx.src[2:]) for s in stack.src]
-  r0 = idx.src[1]
-  ret = srcs[-1]
-  for k in range(len(srcs)-2, -1, -1): ret = r0.eq(k).where(srcs[k], ret)
-  return ret
-
-pm_simple_rangeify = PatternMatcher([
-  # INDEX without src is nothing
-  (UPat(Ops.INDEX, src=(UPat.var('x'),)), lambda x: x),
-  # STAGE on shape () is nothing
-  (UPat(Ops.STAGE, src=(UPat.var('x'),)), lambda x: x if x.shape == () else None),
-  # if INDEX is on STAGE with the same ranges, remove the pair
-  (UPat(Ops.STAGE, allow_any_len=True, name="s").index(allow_any_len=True, name="i"),
-   lambda s,i: s.src[0] if s.src[1:] == i.src[1:] else None),
-  # reshape of a single element shaped value to scalar is an index
-  (UPat(Ops.RESHAPE, name="x"), lambda x: x.src[0].index(0) if x.marg == () and x.src[0].shape == (1,) else None),
-  # handle movement ops on INDEX
-  (UPat(GroupOp.Movement, name="r").index(name="idx", allow_any_len=True), _mop_index),
-  (UPat(Ops.STACK, name="stack").index(name="idx", allow_any_len=True), index_on_stack),
-  # pass index through elementwise
-  (UPat(GroupOp.Elementwise, name="b").index(name="idx", allow_any_len=True),
-   lambda b,idx: b.replace(src=tuple(s.index(*idx.src[1:]) for s in b.src))),
-])
-
-pm_range_creation = PatternMatcher([
-  # reduce/store are what creates ranges
-  (UPat(Ops.REDUCE, src=(UPat.var('inp'),), name="red").index(name="idx", allow_any_len=True), rangeify_on_reduce),
-  (UPat(Ops.REDUCE, src=(UPat.var('inp'),), name="red"), rangeify_on_reduce),
-  (UPat(Ops.STORE, name="x"), rangeify_on_store),
-  (UPat(Ops.STAGE, name="x"), rangeify_on_stage),
 ])
 
 @rewrite_group(new_ctx=False)
@@ -679,61 +578,15 @@ def get_kernel_graph(sink:UOp) -> UOp:
   if OPENPILOT_HACKS: tsink = graph_rewrite(tsink, pm_fold_moved_after, ctx={}, name="fold moved afters")
   tsink = graph_rewrite(tsink, pm_mops+earliest_rewrites, bottom_up=True, name="earliest rewrites")
 
-  # convert movement ops to ranges
-  #tsink, rctx = run_rangeify(tsink, bool(DEBUG_RANGEIFY))
-  tsink = graph_rewrite(tsink, pm_expand_broadcast, bottom_up=True, name="expand broadcast")
-
-  # mark ops that would be recomputed (expand coeff > 1) as CONTIGUOUS, like the realize map in run_rangeify
-  contig = expand_coeff(tsink)
-  subs: dict[UOp, UOp] = {}
-  for u in tsink.toposort():
-    u2 = u.replace(src=tuple(subs.get(s, s) for s in u.src))
-    subs[u] = u2.alu(Ops.STAGE, arg=BufferizeOpts(u2.device)) if u in contig else u2
-  tsink = subs[tsink]
-
-  # add buffers on copy
   tsink = graph_rewrite(tsink, pm_copy_to_store, ctx=itertools.count(0), bottom_up=True, name="convert copy to store")
 
-  # simple rangeify
-  tsink = graph_rewrite(tsink, pm_range_creation+pm_simple_rangeify, ctx=itertools.count(0), bottom_up=True, name="simple rangeify")
+  # convert movement ops to ranges
+  tsink, rctx = run_rangeify(tsink, bool(DEBUG_RANGEIFY))
 
-  # for each index on a stage without children, try to merge the stage into the consumer kernel
-  while 1:
-    indexes: dict[UOp, list[UOp]] = {}
-    consumers: dict[UOp, list[UOp]] = {}
-    for u in tsink.toposort():
-      if u.op is Ops.INDEX and u.src[0].op is Ops.STAGE:
-        indexes.setdefault(u.src[0], []).append(u)
-      for s in u.src: consumers.setdefault(s, []).append(u)
-    # the ranges wrapping u: REDUCE ranges on the path up, plus the enclosing END/STAGE nest
-    def nest_ranges(u:UOp) -> set[UOp]:
-      ret: set[UOp] = set()
-      stack, seen = [u], set()
-      while len(stack):
-        if (x := stack.pop()) in seen: continue
-        seen.add(x)
-        if x.op is Ops.REDUCE: ret.update(*[er.ranges for er in x.ended_ranges])
-        elif x.op in {Ops.END, Ops.STAGE}:
-          ret.update(*[er.ranges for er in x.ended_ranges])
-          continue
-        stack.extend(consumers.get(x, []))
-      return ret
-    subs = {}
-    for k,v in indexes.items():
-      # don't move REDUCE ranges up (real?)
-      if len(v) != 1 or not all(all([r.arg[-1] == AxisType.WEAK for r in s.ranges]) for s in v[0].src[1:]): continue
-      # merging must not add iteration multiplicity around range-bound computation in the stage body:
-      # every range of the enclosing kernel nest must be used by the index, unless the body has no inner loops
-      if not nest_ranges(v[0]) <= set().union(*[s.ranges for s in v[0].src[1:]]) and \
-         any(x.op is Ops.REDUCE for x in k.src[0].toposort(gate=lambda x: x.op is not Ops.STAGE)): continue
-      for old_r, new_r in zip(k.src[1:], v[0].src[1:]):
-        subs[old_r] = new_r
-    if not len(subs): break
-    tsink = tsink.substitute(subs)
-    tsink = graph_rewrite(tsink, pm_simple_rangeify, bottom_up=True, name=f"merge kernels ({len(subs)})")
-
-  tsink = graph_rewrite(tsink, 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")
+  tsink = graph_rewrite(tsink,
+                        symbolic+pm_fold_cast_const+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")
 
   if VIZ: graph_rewrite(tsink, PatternMatcher([]), name="View Rangeify")
 

tinygrad/tensor.py

@@ -23,6 +23,7 @@ class AllocCtx:
   bases: set[UOp] = field(default_factory=set)
   assigns: list[UOp] = field(default_factory=list)
   replacements: list[UOp] = field(default_factory=list)
+  views: set[UOp] = field(default_factory=set)
 
 def tag_uop(ctx:AllocCtx, x:UOp):
   if x.tag is not None: return None
@@ -63,40 +64,37 @@ def replace_contig_with_store_after(u:UOp):
 def replace_store_after_with_contig(u:UOp, src:UOp):
   assigned_to = u
   while assigned_to.op in {Ops.BITCAST, Ops.AFTER, Ops.UNSHARD}: assigned_to = assigned_to.src[0].base
-  if assigned_to.op not in {Ops.BUFFER, Ops.SLICE}: return src.contiguous(tag=u.tag)
+  if assigned_to.op is not Ops.BUFFER: return src.contiguous(tag=u.tag)
 
 def _make_buffer_view(src:UOp) -> UOp|None:
-  """If movement ops on src collapse to a contiguous range, return SLICE. Otherwise None."""
-  if (offset := src.contiguous_view_offset()) is None: return None
-  buf = src.base
-  if buf.op is Ops.SLICE:
-    byte_offset = buf.src[1].val * buf.src[0].dtype.itemsize + offset * src.dtype.itemsize
-    buf = buf.src[0]
-    if byte_offset % buf.dtype.itemsize != 0: return None
-    offset = byte_offset // buf.dtype.itemsize
-  return UOp(Ops.SLICE, src.dtype, (buf, UOp.const(offset)), src.numel())
+  if (cv := src.contiguous_view()) is None: return None
+  (buf, offset), size = cv, src.max_numel() * src.element_size() // cv[0].element_size()
+  if buf.op is not Ops.BUFFER: return None
+  # NB: make offset a UOp.variable here to do the offset computation in the kernels
+  return buf[offset:offset+size].bitcast(src.dtype)
 
-def contiguous_mops_to_view(c:UOp, src:UOp):
-  """MOPS(BUFFER) → SLICE when movement ops collapse to a contiguous range."""
+def contiguous_mops_to_view(ctx:AllocCtx, c:UOp, src:UOp):
+  """MOPS(BUFFER) → SHRINK when movement ops collapse to a contiguous range."""
   buf = src.base
-  if buf.op not in {Ops.BUFFER, Ops.SLICE, Ops.UNSHARD}: return None
-  if src.op is Ops.RESHAPE and src.src[0].op in {Ops.BUFFER, Ops.SLICE} and c.op is not Ops.BITCAST: return None
-  if c.op is not Ops.BITCAST and src.op is Ops.BUFFER: return None
+  while buf.op is Ops.BITCAST: buf = buf.src[0].base
+  if buf.op not in {Ops.BUFFER, Ops.UNSHARD}: return None
 
   # no symbolic shape
   if not all_int(c.shape): return None
 
   if buf.op is not Ops.UNSHARD and (view := _make_buffer_view(src)) is not None:
-    view = (view.replace(dtype=c.dtype, arg=c.numel()) if c.op is Ops.BITCAST else view).reshape(c.shape)
-    return c.replace(src=(view,)) if c.op is Ops.COPY else view
+    ctx.views.add(view)
+    view = view.reshape(c.shape)
+    return c.replace(src=(view,)+c.src[1:]) if c.op in {Ops.COPY, Ops.STORE} else view
 
-  # for UNSHARD tensors, use multi_pm to resolve per-shard movement ops, then create SLICE on the resolved result
+  # for UNSHARD tensors, use multi_pm to resolve per-shard movement ops, then create SHRINK on the resolved result
   if not isinstance(c.device, str):
     from tinygrad.schedule.multi import multi_pm
     resolved = graph_rewrite(src, multi_pm, name="multi_buffer_view")
     if resolved.op is not Ops.UNSHARD: return None
     if (view := _make_buffer_view(resolved.src[0])) is None: return None
-    return view.reshape(resolved.src[0].shape).unshard(resolved.arg, resolved.src[1:]).contiguous(tag=c.tag)
+    ctx.views.add(view)
+    return view.reshape(resolved.src[0].shape).unshard(resolved.arg, resolved.src[1:])
 
   return None
 
@@ -111,7 +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
@@ -151,8 +149,9 @@ pm_early_transform_tensor_graph = PatternMatcher([
   # resolve TUPLE+GETTUPLE (for precompiled calls)
   (UPat(Ops.GETTUPLE, src=(UPat(Ops.TUPLE, name="t"),), name="g"), lambda g,t: t.src[g.arg]),
 
-  # fold MOPS+BITCAST over BUFFER/SLICE into SLICE when movement ops collapse to contiguous range
-  (UPat((Ops.BITCAST, Ops.COPY, Ops.CONTIGUOUS), src=(UPat(GroupOp.Movement|{Ops.BUFFER}, name="src"),), name="c"), contiguous_mops_to_view),
+  # fold MOPS+BITCAST over BUFFER into SHRINK when movement ops collapse to contiguous range
+  (UPat((Ops.COPY, Ops.CONTIGUOUS), src=(UPat(GroupOp.Movement|{Ops.BITCAST}, name="src"),), name="c"), contiguous_mops_to_view),
+  (UPat(Ops.STORE, src=(UPat(Ops.BITCAST, name="src"), UPat()), name="c", allow_any_len=True), contiguous_mops_to_view),
 
   # remove contiguous on movement ops before a copy on disk
   (UPat(GroupOp.Movement-{Ops.SHRINK, Ops.RESHAPE}, name="x").f(Ops.CONTIGUOUS).f(Ops.COPY, name="copy"), lambda x,copy:
@@ -177,6 +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)
@@ -197,10 +198,12 @@ 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)
 
+def replace_input_view(ctx:AllocCtx, b:UOp): return replace_input_buffer(ctx, b) if b in ctx.views else None
+
 pm_finalize_call = PatternMatcher([
   (UPat(Ops.AFTER, name="x"), finalize_after),
   (UPat(Ops.COPY, name="x"), lambda ctx,x: ctx.assigns.append(x) if isinstance(x.device, str) and x.device.startswith(("DISK", "TINYFS")) else None),
@@ -210,10 +213,11 @@ pm_replace_buf = PatternMatcher([
   # replace BUFFER with PARAM for cache key normalization
   (UPat(Ops.BUFFER, src=(UPat(),), name="b"), lambda ctx,b:
    replace_input_buffer(ctx, b) if isinstance(b.arg, ParamArg) and b.addrspace is AddrSpace.GLOBAL else None),
-  # replace SLICE with PARAM. this rewrite is bottom up so BUFFERs we don't need won't be in the input
-  (UPat(Ops.SLICE, src=(UPat(Ops.BUFFER), UPat(Ops.CONST, dtype=dtypes.weakint)), name="b"), replace_input_buffer),
-  # 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),
+  # 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 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]))}")
@@ -229,7 +233,7 @@ def transform_to_call(big_sink:UOp) -> tuple[UOp, dict[UOp, UOp]]:
   big_sink = graph_rewrite(big_sink, add_tags, ctx=ctx, bottom_up=True, name="number the uops")
 
   # here we can break the tensor graph. this is the only place you need to maintain numbered tags
-  big_sink = graph_rewrite(big_sink, pm_early_transform_tensor_graph, name="early transform tensor graph")
+  big_sink = graph_rewrite(big_sink, pm_early_transform_tensor_graph, ctx=ctx, name="early transform tensor graph")
 
   # here we construct the final buffer_map: as-built nodes -> their final storage. values are never keys
   graph_rewrite(big_sink, pm_finalize_call, ctx=ctx, name="finalize call")
@@ -449,12 +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
@@ -739,7 +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)))

tinygrad/uop/__init__.py

@@ -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()
 
@@ -93,7 +90,7 @@ class Ops(FastEnum):
   CONTIGUOUS = auto(); CONTIGUOUS_BACKWARD = auto(); DETACH = auto()
 
   # buffer ops
-  STAGE = auto(); COPY = auto(); SLICE = auto(); MSELECT = auto(); MSTACK = auto(); CUSTOM_FUNCTION = auto()
+  STAGE = auto(); COPY = auto(); MSELECT = auto(); MSTACK = auto(); CUSTOM_FUNCTION = auto()
 
   # the core 6 movement ops! these only exist in the tensor graph
   RESHAPE = auto(); PERMUTE = auto(); EXPAND = auto(); PAD = auto(); FLIP = auto()

tinygrad/uop/ops.py

@@ -45,8 +45,7 @@ axis_colors = {AxisType.DEVICE: "green", AxisType.GLOBAL: "blue", AxisType.THREA
 axis_to_pos = {AxisType.DEVICE: -2, AxisType.WEAK: -1, AxisType.LOOP: -1, AxisType.THREAD: 0, AxisType.GLOBAL: 0, AxisType.WARP: 1,
                AxisType.LOCAL: 2, AxisType.UPCAST: 3, AxisType.GROUP_REDUCE: 2, AxisType.REDUCE: 4, AxisType.UNROLL: 5}
 
-range_start = {Ops.STAGE: 1, Ops.REDUCE: 1, Ops.WMMA: 3, Ops.END: 1, Ops.CALL: 1, Ops.FUNCTION: 1,
-               Ops.SLICE: 2, Ops.LINEAR: 0}
+range_start = {Ops.STAGE: 1, Ops.REDUCE: 1, Ops.WMMA: 3, Ops.END: 1, Ops.CALL: 1, Ops.FUNCTION: 1, Ops.LINEAR: 0}
 
 # https://en.wikipedia.org/wiki/Identity_element
 def identity_element(op:Ops, dt:DType) -> PyConst: return dt.const({Ops.ADD:0, Ops.MUL:1, Ops.MAX:dt.min}[op])
@@ -151,9 +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
@@ -171,9 +167,6 @@ def dtype_from_uop(op:Ops, src:tuple[UOp,...], arg:Any) -> DType|None:
       return arg.dtype
     case Ops.BINARY:
       return dtypes.uint8
-    case Ops.SLICE:
-      # TODO: slice just shouldn't exist
-      return None
     case Ops.CAST | Ops.BITCAST:
       assert isinstance(arg, DType), f"CAST/BITCAST arg must be DType, got {arg}"
       return arg
@@ -221,7 +214,7 @@ class UOpMetaClass(type):
     return created
 
 # some uops map to other stuff
-buffers:weakref.WeakKeyDictionary[UOp, Buffer|MultiBuffer] = weakref.WeakKeyDictionary() # this maps BUFFER/SLICE uops to their device Buffers
+buffers:weakref.WeakKeyDictionary[UOp, Buffer|MultiBuffer] = weakref.WeakKeyDictionary() # this maps BUFFER/view uops to their device Buffers
 all_metadata:weakref.WeakKeyDictionary[UOp, tuple[Metadata, ...]] = weakref.WeakKeyDictionary() # TODO: should this be here?
 
 # recursive_property replaces functools.cached_property in recursive UOp functions to prevent RecursionError
@@ -381,15 +374,11 @@ 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
         return ()
-      case Ops.SLICE:
-        # HACK: SLICE is used inside kernels, so we set the shape to () if it's on an INDEX
-        if self.src[0].op is Ops.INDEX: return ()
-        return (self.arg,)
       case Ops.CUSTOM | Ops.CUSTOMI:
         if self.dtype is dtypes.void: return None
         input_shapes = [x._shape for x in self.src if x._shape is not None]
@@ -528,10 +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
@@ -568,9 +559,9 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
     in_tuple = self.src[0] if self.op is Ops.FUNCTION else self
     assert in_tuple.op is Ops.TUPLE, f"gettuple requires FUNCTION or TUPLE source, got {self.op}"
     return UOp(Ops.GETTUPLE, src=(self,), arg=idx)
-  def group(*srcs:UOp|None):  # pylint: disable=no-self-argument
+  def group(*srcs:UOp|None, **kwargs):  # pylint: disable=no-self-argument
     if len(srcs) == 1 and isinstance(srcs[0], UOp): return srcs[0]
-    return UOp(Ops.GROUP, src=tuple([x for x in srcs if x is not None]))
+    return UOp(Ops.GROUP, src=tuple([x for x in srcs if x is not None]), **kwargs)
   def index(self, *srcs:UOp|int|None, **kwargs):
     new_srcs: list[UOp] = [UOp.const(x) if isinstance(x, int) else x for x in srcs if x is not None]
     if len(new_srcs) == 1 and new_srcs[0].op is Ops.CONST and self.op is Ops.STACK: return self.src[new_srcs[0].val]
@@ -807,11 +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)
@@ -822,7 +811,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
   unique_num = itertools.count(0)
 
   def getaddr(self, device=None) -> UOp:
-    if self.without_after.op not in {Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT, Ops.PARAM}: return self
+    if self.without_after.op not in {Ops.BUFFER, Ops.SHRINK, Ops.BITCAST, Ops.BINARY, Ops.MSTACK, Ops.MSELECT, Ops.PARAM}: return self
     return UOp(Ops.GETADDR, src=(self,), arg=device or to_tuple(self.device)[0])
   @staticmethod
   def new_buffer(device:str|tuple[str, ...], size:int, dtype:DType, num=None):
@@ -901,8 +890,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
     while len(s.src) and s.op not in {Ops.BUFFER, Ops.PARAM, Ops.STAGE, Ops.MSTACK}: s = s.src[0]
     return s
 
-  def contiguous_view_offset(self) -> int|None:
-    """If movement ops on a BUFFER collapse to a contiguous range, return `offset` in elements. Otherwise None."""
+  def contiguous_view(self) -> tuple[UOp, int]|None:
     from tinygrad.schedule.rangeify import pm_mops
     from tinygrad.uop.symbolic import symbolic
 
@@ -915,14 +903,17 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
 
     idx = self.flatten().index(UOp.range(self.numel(), 0))
     out = graph_rewrite(idx, pm_mops+symbolic+pm_contiguous_view_offset, ctx=self, name="contiguous_view_offset")
-    return out.val if out.op is Ops.CONST and isinstance(out.val, int) else None
+    if out.op is not Ops.INDEX or not (b:=out.src[0]).tag or (c:=out.src[1]).op is not Ops.CONST or not isinstance(c.val, int): return None
+    return b.rtag(None), c.val
+
+  def contiguous_view_offset(self) -> int|None: return None if (view := self.contiguous_view()) is None else view[1]
 
   def has_buffer_identity(self, after_ok=False):
     """Check if this UOp has a concrete buffer identity in the graph (RESHAPE/UNSHARD -> BUFFER chain)."""
     # TODO: this is confusing because UOp.variable('v', 0, 1, dtypes.weakfloat) is True for jit to work, but it doesn't have a buffer
     if self.op in {Ops.RESHAPE, Ops.UNSHARD, Ops.MSELECT}: return self.src[0].has_buffer_identity(after_ok)
     if after_ok and self.op == Ops.AFTER: return self.src[0].has_buffer_identity(after_ok)
-    return self.op in {Ops.BUFFER, Ops.SLICE, Ops.PARAM}
+    return self.op in {Ops.BUFFER, Ops.PARAM}
 
   def _base_buffer_is_realized(self) -> bool:
     """Walk through AFTER chain to find if the underlying buffer is realized (has allocated memory)."""
@@ -932,30 +923,19 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
 
   @property
   def buffer(self) -> Buffer|MultiBuffer:
-    if self.op in {Ops.CONTIGUOUS, Ops.RESHAPE, Ops.UNSHARD, Ops.DETACH, Ops.AFTER}: return self.src[0].buffer
+    if self.op in {Ops.CONTIGUOUS, Ops.CONTIGUOUS_BACKWARD, Ops.RESHAPE, Ops.UNSHARD, Ops.DETACH, Ops.AFTER}: return self.src[0].buffer
     # this buffer can process disk tensors and simple movement ops
-    if self is not self.base:
-      buf = self.base.buffer
-      assert isinstance(buf, Buffer), "must be a Buffer for movement ops"
-      offset = self.contiguous_view_offset()
-      if offset is None: raise RuntimeError(f"non-contiguous view is not supported for {buf.device} buffer")
-      return buf.view(prod(self.max_shape), self.dtype, offset*self.dtype.itemsize)
-    if self.op is Ops.BITCAST:
-      buf = self.src[0].buffer
-      assert isinstance(buf, Buffer), "must be a Buffer for BITCAST"
-      return buf.view(prod(self.max_shape), self.dtype, 0)
-    if self.op is Ops.SLICE:
+    if self is not self.base or self.op is Ops.BITCAST:
       if (cret:=buffers.get(self)) is not None: return cret
-      buf = self.src[0].buffer
-      offset = self.src[1].val
+      if (cv := self.contiguous_view()) is None: raise RuntimeError(f"non-contiguous view is not supported for {self.device} buffer")
+      buf, offset = (b:=cv[0]).base.buffer, cv[1]
       if isinstance(buf, MultiBuffer):
         mbuf = MultiBuffer.__new__(MultiBuffer)
-        mbuf.bufs = [b.view(self.arg, self.dtype, offset * self.src[0].dtype.itemsize) for b in buf.bufs]
+        mbuf.bufs = [x.view(prod(self.max_shape), self.dtype, offset*b.dtype.itemsize) for x in buf.bufs]
         buffers[self] = mbuf
         return mbuf
-      assert isinstance(buf, Buffer), "must be a Buffer for SLICE"
-      buffers[self] = bv = buf.view(self.arg, self.dtype, offset * self.src[0].dtype.itemsize)
-      return bv
+      buffers[self] = buf.view(prod(self.max_shape), self.dtype, offset*b.dtype.itemsize)
+      return buffers[self]
     if self.op is Ops.MSELECT:
       ret = self.src[0].buffer
       assert isinstance(ret, MultiBuffer)
@@ -977,8 +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
@@ -989,29 +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 ***
 
@@ -1022,7 +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
@@ -1034,7 +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))
@@ -1099,9 +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
@@ -1118,7 +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']}"]
@@ -1173,7 +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)
 
@@ -1181,7 +1177,7 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
   def custom_function(name:str, *src:UOp) -> UOp: return UOp(Ops.CUSTOM_FUNCTION, src=src, arg=name)
 
   # opaque bodies stay as Ops.CALL; value-producing bodies become Ops.FUNCTION (wrapped in TUPLE)
-  _OPAQUE_CALL_BODIES = {Ops.SINK, Ops.PROGRAM, Ops.LINEAR, Ops.COPY, Ops.SLICE, Ops.CUSTOM_FUNCTION}
+  _OPAQUE_CALL_BODIES = {Ops.SINK, Ops.PROGRAM, Ops.LINEAR, Ops.COPY, Ops.CUSTOM_FUNCTION}
   def call(self, *srcs:UOp, ret_dtype:DType|None=None, grad_fxn:Callable|None=None,
            name:str|None=None, precompile:bool=False, precompile_backward:bool=False, aux:Any=None) -> UOp:
     if ret_dtype is not None: return UOp(Ops.CALL, ret_dtype, src=(self,)+srcs)
@@ -1427,9 +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__
@@ -1771,14 +1767,18 @@ 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([
-  (UPat(Ops.INDEX, src=(UPat(),)), lambda: UOp.const(0)),
-  (UPat(Ops.INDEX, src=(UPat(), UPat(Ops.RANGE))), lambda: UOp.const(0)),
-  (UPat(Ops.INDEX, src=(UPat(), UPat(Ops.RANGE)+UPat.cvar('c'))), lambda c: c),
-  (UPat(Ops.INDEX, src=(UPat(), UPat.cvar('c'))), lambda ctx, c: c if resolve(ctx.numel() == 1, False) else None),
+  # normalize to 1d bitcasts
+  (UPat(Ops.BITCAST, name="b"), lambda b: b.src[0].flatten().bitcast(b.dtype).reshape(b.shape) if len(b.shape) != 1 else None),
+  (UPat(Ops.BITCAST, name="b").index(UPat.cvar("c")), lambda ctx, b, c:
+   b.src[0].flatten().index(UOp.range(ctx.numel() * (osz:=b.element_size())//(isz:=b.src[0].element_size()), 0) + (c * osz//isz)) if b.tag else None),
+  (UPat(Ops.INDEX, src=(UPat.var("b"),)), lambda b: b.rtag().index(0)),
+  (UPat(Ops.INDEX, src=(UPat.var("b"), UPat(Ops.RANGE))), lambda b: b.rtag().index(0)),
+  (UPat(Ops.INDEX, src=(UPat.var("b"), UPat(Ops.RANGE)+UPat.cvar('c'))), lambda ctx, b, c: b.rtag().index(c)),
+  (UPat(Ops.INDEX, src=(UPat.var("b"), UPat.cvar('c'))), lambda ctx, b, c: b.rtag().index(c) if resolve(ctx.numel() == 1, False) else None),
 ])
 
 # *** what was symbolic.py ***

tinygrad/uop/render.py

@@ -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]]})"),

tinygrad/uop/spec.py

@@ -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)),
@@ -233,13 +232,6 @@ spec_hcq = PatternMatcher([
 spec_full = PatternMatcher([
   (UPat(Ops.REWRITE_ERROR, dtypes.void, name="x"), lambda x: isinstance(x.arg, str)),
 
-  # SLICE on BUFFER is allowed if BUFFER is
-  (UPat(Ops.SLICE, src=(UPat(GroupOp.Movement.union({Ops.BUFFER, Ops.PARAM, Ops.STAGE, Ops.AFTER})),
-                        UPat(Ops.CONST, dtype=dtypes.weakint)), allow_any_len=True, name="bv"),
-   lambda bv: isinstance(bv.arg, int)),
-
-  (UPat(Ops.CALL, dtypes.void, src=(UPat((Ops.SLICE,)),), allow_any_len=True), lambda: True),
-
   # codegen may end ranges after gpudims has replaced RANGE with SPECIAL.
   (UPat(Ops.END, src=(UPat(), UPat()), allow_any_len=True), lambda: True),
 
@@ -248,9 +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 *****
@@ -258,17 +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

tinygrad/uop/symbolic.py

@@ -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
 
@@ -136,10 +140,10 @@ 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),
@@ -245,10 +249,12 @@ 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
+  ((UPat.cvar("a") < UPat.var("b")).where(UPat.var("b"), UPat.cvar("c")), lambda a,b,c: UOp.maximum(a,b) if a.val == c.val else None),
+  ((UPat.var("a") < UPat.cvar("b")).where(UPat.cvar("c"), UPat.var("a")), lambda a,b,c: UOp.maximum(a,b) if b.val == c.val else None),
   (UPat.maximum(UPat.var("x"), UPat.var("y")), lambda x,y: x if x.vmin >= y.vmax else y if x.vmax <= y.vmin else None),
   # TODO: why does this rule break beautiful_mnist?
   #((UPat.var("x")+UPat.var("z")).maximum(UPat.var("y")+UPat.var("z")), lambda x,y,z: x.maximum(y) + z),
@@ -328,14 +334,14 @@ 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
@@ -405,7 +411,8 @@ 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)
@@ -430,9 +437,6 @@ 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 **

tinygrad/uop/validate.py

@@ -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:

tinygrad/uop/weak.py

@@ -12,7 +12,7 @@ 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)),
@@ -23,7 +23,7 @@ pm_lower_weak = PatternMatcher([
   # 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:
@@ -65,6 +65,7 @@ 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([

tinygrad/viz/serve.py

@@ -50,7 +50,7 @@ uops_colors = {Ops.LOAD: "#ffc0c0", Ops.STORE: "#87CEEB", Ops.CONST: "#e0e0e0",
                Ops.INDEX: "#CEF9B7", Ops.STACK: "#D8F9E4",
                Ops.WMMA: "#efefc0", Ops.UNSHARD: "#f6ccff", Ops.INS: "#eec4ff",
                **{x:"#D8F9E4" for x in GroupOp.Movement}, **{x:"#ffffc0" for x in GroupOp.ALU}, Ops.THREEFRY:"#ffff80",
-               Ops.SLICE: "#E5EAFF", Ops.BUFFER: "#B0BDFF", Ops.GETADDR: "#9DB1F0", Ops.COPY: "#a040a0", Ops.CUSTOM_FUNCTION: "#bf71b6",
+               Ops.BUFFER: "#B0BDFF", Ops.GETADDR: "#9DB1F0", Ops.COPY: "#a040a0", Ops.CUSTOM_FUNCTION: "#bf71b6",
                Ops.CALL: "#00B7C8", Ops.FUNCTION: "#C07788", Ops.PARAM: "#14686F", Ops.SOURCE: "#c0c0c0", Ops.BINARY: "#404040",
                Ops.LINEAR: "#7DF4FF",
                Ops.ALLREDUCE: "#ff40a0", Ops.MSELECT: "#d040a0", Ops.MSTACK: "#d040a0", Ops.CONTIGUOUS: "#FFC14D",