Merge branch 'master' into clean_load

okay w/e
2026-06-19 16:56:57 -07:00 · 2026-06-18 21:00:36 -07:00 · 2026-06-18 20:54:45 -07:00 · 2026-06-18 20:53:30 -07:00 · 2026-06-18 18:24:59 -07:00
165 changed files with 3233 additions and 3118 deletions
@@ -81,178 +81,82 @@ jobs:
  #      source /tmp/tinygrad_pytest_ci/bin/activate
  #      pytest -nauto --durations=20

-  sharedbenchmarks:
-    name: Benchmark (DEV=${{ matrix.dev }})
-    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
-    strategy:
-      fail-fast: false
-      matrix:
-        dev: ['METAL', 'AMD', 'NV']
+  testmacbenchmark:
+    name: Mac Benchmark
+    runs-on: [self-hosted, macOS]
    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
-    env:
-      DEV: ${{ matrix.dev }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
-    - name: Setup (AMD)
-      if: ${{ matrix.dev == 'AMD' }}
-      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py amd rmmod
-        ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: reset process replay
-      run: python3 test/external/process_replay/reset.py
-    - name: Test tensor cores
-      run: |
-        if [[ "${{ matrix.dev }}" == "METAL" ]]; then
-          python3 test/opt/test_tensor_cores.py
-          DEBUG=2 SHOULD_USE_TC=1 python3 extra/gemm/simple_matmul.py
-          DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3 extra/gemm/simple_matmul.py
-          DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3 extra/gemm/simple_matmul.py
-          M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
-        elif [[ "${{ matrix.dev }}" == "NV" ]]; then
-          ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-          DEV=NV:PTX ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-          DEV=CUDA SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-          DEV=CUDA SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-          DEV=CUDA SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-          DEV=CUDA SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-          DEV=NV:PTX SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-          SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-          # TODO: too slow
-          # M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
-          # DEV=NV:PTX M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
-        else
-          python3 test/opt/test_tensor_cores.py
-          # TODO: this is flaky
-          # DEV=AMD:LLVM python3 test/opt/test_tensor_cores.py
-          SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-          SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-          # TODO: AMD compiler bug causes this to fail
-          # HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
-        fi
-    - name: Run model inference benchmark
-      # TODO: unstable on AMD
-      if: ${{ matrix.dev != 'AMD' }}
-      run: CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
-    - name: Test speed vs torch
-      # TODO: unstable on AMD
-      if: ${{ matrix.dev != 'AMD' }}
-      env:
-        HALF: ${{ matrix.dev == 'NV' && '1' || '0' }}
-      run: CAPTURE_PROCESS_REPLAY=0 BIG=2 ${{ matrix.dev == 'METAL' && 'MPS=1' || 'TORCHCUDA=1' }} python3 test/speed/external_test_speed_v_torch.py
-    - name: Run Stable Diffusion
-      env:
-        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'METAL' && '720' || matrix.dev == 'AMD' && '550' || '0' }}
-      run: BENCHMARK_LOG=stable_diffusion python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
-    - name: Run SDXL
-      if: ${{ matrix.dev != 'NV' }}
-      env:
-        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'METAL' && '5000' || matrix.dev == 'AMD' && '3200' || '2000' }}
-      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
-    - name: Run llama3.2
-      run: BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
-    - name: Run qwen3.5
-      # qwen3.5:35b-a3b doesn't fit on mac
-      if: ${{ matrix.dev != 'METAL' }}
-      run: BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
-    - name: Run olmoe
-      # just metal for now
-      if: ${{ matrix.dev == 'METAL' }}
-      run: BENCHMARK_LOG=olmoe JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m olmoe --benchmark --warmup
-    - name: Train MNIST
-      run: time TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
-    - name: Run 10 CIFAR training steps
-      env:
-        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'NV' && '130' || matrix.dev == 'AMD' && '200' || '3000' }}
-      run: BENCHMARK_LOG=cifar_10steps STEPS=10 python3 examples/hlb_cifar10.py
-    - name: Run 10 CIFAR training steps w HALF
-      env:
-        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'NV' && '120' || matrix.dev == 'AMD' && '230' || '3000' }}
-      run: BENCHMARK_LOG=cifar_10steps_half STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    - uses: actions/upload-artifact@v7
-      if: ${{ matrix.dev != 'AMD' }}
-      with:
-        name: Speed (${{ matrix.dev }})
-        path: |
-          onnx_inference_speed.csv
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  tinyboxbenchmark:
-    name: Tinybox Benchmark (${{ matrix.dev }})
-    runs-on: [self-hosted, Linux, "${{ matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
-    strategy:
-      fail-fast: false
-      matrix:
-        dev: ['AMD', 'NV']
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    env:
-      DEV: ${{ matrix.dev }}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup (AMD)
-      if: ${{ matrix.dev == 'AMD' }}
-      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py amd rmmod
-        ./extra/hcq/hcq_smi.py amd kill_pids
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
-        mkdir -p extra/datasets
-        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+        mkdir -p extra/disassemblers
+        ln -s ~/tinygrad/extra/disassemblers/applegpu extra/disassemblers/applegpu
+        ln -s ~/tinygrad/weights/sd-v1-4.ckpt weights/sd-v1-4.ckpt
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
-      run: test/external/process_replay/reset.py
-    - name: Test tiny
-      run: |
-        DEBUG=2 python -m pytest -rA test/test_tiny.py
-        if [[ "${{ matrix.dev }}" == "NV" ]]; then
-          DEBUG=2 DEV=CUDA python -m pytest -rA test/test_tiny.py
-        fi
-    - name: Test speed vs theoretical
-      run: IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test benchmark allreduce
-      if: ${{ matrix.dev == 'NV' }}
-      run: python test/external/external_benchmark_multitensor_allreduce.py
-    - name: HEVC Decode Benchmark
-      if: ${{ matrix.dev == 'NV' }}
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 PYTHONPATH=. python3 extra/hevc/decode.py
-    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run MLPerf resnet eval on training data
-      run: time BENCHMARK_LOG=resnet_eval MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf Bert training steps (6 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: python3.11 test/external/process_replay/reset.py
+    - name: Print macOS version
+      run: sw_vers
+    - name: Run Stable Diffusion
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+    - name: Run Stable Diffusion without fp16
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
+    - name: Run Stable Diffusion v2
+      # TODO: very slow step time
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing
+    # process replay can't capture this, the graph is too large
+    - name: Run SDXL
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing
+    - name: Run model inference benchmark
+      run: DEV=METAL NOCLANG=1 python3.11 test/external/external_model_benchmark.py
+    - name: Test speed vs torch
+      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
+    - name: Test tensor cores
+      run: DEV=METAL python3.11 test/opt/test_tensor_cores.py
+    - name: Run Tensor Core GEMM (float)
+      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
+    - name: Run Tensor Core GEMM (half)
+      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
+    - name: Run Tensor Core GEMM (bfloat16)
+      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
+    - name: Fuzz Padded Tensor Core GEMM
+      run: DEV=METAL M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
+    - name: Run llama3.2
+      run: BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
+    - name: Run olmoe
+      run: BENCHMARK_LOG=olmoe JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 -m tinygrad.llm -m olmoe --benchmark --warmup
+    - name: Train MNIST
+      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py
+
+    # NOTE: this is failing in CI. it is not failing on my machine and I don't really have a way to debug it
+    # the error is "RuntimeError: Internal Error (0000000e:Internal Error)"
+    #- name: Run 10 CIFAR training steps
+    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py
+    #- name: Run 10 CIFAR training steps w HALF
+    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py
+
+    #- name: Run 10 CIFAR training steps w BF16
+    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
+    # TODO: too slow
+    # - name: Run 10 CIFAR training steps w winograd
+    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
+    - uses: actions/upload-artifact@v7
+      with:
+        name: Speed (Mac)
+        path: |
+          onnx_inference_speed.csv
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -292,6 +196,337 @@ jobs:
    - name: UsbGPU (USB4/TB) tiny tests
      run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py

+  testnvidiabenchmark:
+    name: tinybox green Benchmark
+    runs-on: [self-hosted, Linux, tinyboxgreen]
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Print nvidia-smi
+      run: nvidia-smi
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: Run model inference benchmark
+      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
+    - name: Test speed vs torch
+      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
+    - name: Test speed vs theoretical
+      run: DEV=NV IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+    - name: Test benchmark allreduce
+      run: DEV=NV python test/external/external_benchmark_multitensor_allreduce.py
+    - name: Test tensor cores
+      run: |
+        DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+        DEV=NV:PTX ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+    - name: Run Tensor Core GEMM (CUDA)
+      run: |
+        DEV=CUDA SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        DEV=CUDA SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        DEV=CUDA SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+        DEV=CUDA SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+    - name: Run Tensor Core GEMM (PTX)
+      run: DEV=NV:PTX SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+    - name: Run Tensor Core GEMM (NV)
+      run: DEV=NV SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+    - name: Test DEV=NV
+      run: DEBUG=2 DEV=NV python -m pytest -rA test/test_tiny.py
+    - name: Test DEV=CUDA
+      run: DEBUG=2 DEV=CUDA python -m pytest -rA test/test_tiny.py
+    - name: Run Stable Diffusion
+      run: BENCHMARK_LOG=stable_diffusion DEV=NV python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+    # TODO: too slow
+    # - name: Run SDXL
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
+    - name: Run llama3.2
+      run: DEV=NV BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
+    - name: Run qwen3.5
+      run: DEV=NV BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
+    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
+      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    # - name: Run LLaMA-3 8B on 6 GPUs
+    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    # - name: Run LLaMA-2 70B
+    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    - uses: actions/upload-artifact@v7
+      with:
+        name: Speed (NVIDIA)
+        path: |
+          onnx_inference_speed.csv
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testmorenvidiabenchmark:
+    name: tinybox green Training Benchmark
+    runs-on: [self-hosted, Linux, tinyboxgreen]
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
+        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    # TODO: too slow
+    # - name: Fuzz Padded Tensor Core GEMM (NV)
+    #   run: DEV=NV M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    # TODO: too slow
+    # - name: Fuzz Padded Tensor Core GEMM (PTX)
+    #   run: DEV=NV:PTX M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    - name: HEVC Decode Benchmark
+      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 DEV=NV PYTHONPATH=. python3 extra/hevc/decode.py
+    - name: Train MNIST
+      run: time PYTHONPATH=. DEV=NV TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+    - name: Run 10 CIFAR training steps
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=130 DEV=NV STEPS=10 python3 examples/hlb_cifar10.py
+    - name: Run 10 CIFAR training steps w HALF
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    - name: Run 10 CIFAR training steps w BF16
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
+    # - name: Run 10 CIFAR training steps w winograd
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 DEV=NV WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar DEV=NV DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run full CIFAR training steps w 6 GPUS
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEV=NV DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run MLPerf resnet eval on training data
+      run: time BENCHMARK_LOG=resnet_eval DEV=NV MODEL=resnet python3 examples/mlperf/model_eval.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      run: BENCHMARK_LOG=resnet_10steps DEV=NV DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+      run: BENCHMARK_LOG=resnet_10steps_6gpu DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf Bert training steps (6 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps_6gpu DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testamdbenchmark:
+    name: tinybox red Benchmark
+    runs-on: [self-hosted, Linux, tinybox]
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    #- name: Insert amdgpu
+    #  run: sudo modprobe amdgpu
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    #- name: setup perflevel
+    #  run: |
+    #    examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
+    #    rocm-smi
+    #- name: Show off tinybox
+    #  run: /opt/rocm/bin/rocm-bandwidth-test
+    # TODO: unstable on AMD
+    #- name: Run model inference benchmark
+    #  run: LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 NOCLANG=1 python3 test/external/external_model_benchmark.py
+    # TODO: unstable on AMD
+    #- name: Test speed vs torch
+    #  run: |
+    #    python3 -c "import torch; print(torch.__version__)"
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
+    - name: Test speed vs theoretical
+      run: DEV=AMD IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+    - name: Test tensor cores (no LLVM)
+      run: DEV=AMD python3 test/opt/test_tensor_cores.py
+    # TODO: this is flaky
+    # - name: Test tensor cores AMD:LLVM
+    #   run: DEV=AMD:LLVM python3 test/opt/test_tensor_cores.py
+    - name: Run Tensor Core GEMM (AMD)
+      run: |
+        DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+    - name: Test DEV=AMD
+      run: DEBUG=2 DEV=AMD python -m pytest -rA test/test_tiny.py
+    #- name: Test HIP=1
+    #  run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
+    # TODO: AMD compiler bug causes this to fail
+    #- name: Fuzz Padded Tensor Core GEMM
+    #  run: HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
+    #- name: Remove amdgpu
+    #  run: sleep 10 && sudo rmmod amdgpu # sleep a bit to let the driver unload the prev pid.
+    - name: Test AM cold start time
+      run: time DEV=AMD AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
+    - name: Test AM warm start time
+      run: time DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
+    - name: Run Stable Diffusion
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 DEV=AMD python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+    - name: Run SDXL
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 DEV=AMD python3 examples/sdxl.py --seed 0 --noshow --timing
+    - name: Run llama3.2
+      run: DEV=AMD BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
+    - name: Run qwen3.5
+      run: DEV=AMD BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
+    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
+      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    # - name: Run LLaMA-3 8B on 6 GPUs
+    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    #- name: Restore amdgpu
+    #  run: sudo modprobe amdgpu
+    # - name: Run LLaMA-2 70B
+    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testmoreamdbenchmark:
+    name: tinybox red Training Benchmark
+    runs-on: [self-hosted, Linux, tinybox]
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
+        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: Test GPU crash recovery
+      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
+    - name: Train MNIST
+      run: time PYTHONPATH=. DEV=AMD TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+    - name: Run 10 CIFAR training steps
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 DEV=AMD STEPS=10 python3 examples/hlb_cifar10.py
+    - name: Run 10 CIFAR training steps w HALF
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 DEV=AMD STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    # - name: Run 10 CIFAR training steps w BF16
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 DEV=AMD STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
+    # TODO: too slow
+    # - name: Run 10 CIFAR training steps w winograd
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 DEV=AMD WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar DEV=AMD DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run full CIFAR training steps w 6 GPUS
+      run: time BENCHMARK_LOG=cifar_6gpu DEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    # TODO: broken on some of the machines
+    #- name: Test full tinyfs load
+    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testmlperfamdbenchmark:
+    name: tinybox red MLPerf Benchmark
+    runs-on: [self-hosted, Linux, tinybox]
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
+        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: Run MLPerf resnet eval
+      run: time BENCHMARK_LOG=resnet_eval DEV=AMD MODEL=resnet python3 examples/mlperf/model_eval.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      run: BENCHMARK_LOG=resnet_10steps DEV=AMD DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+      run: BENCHMARK_LOG=resnet_10steps_6gpu DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf Bert training steps (6 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps_6gpu DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
  testcommalatest:
    name: comma Benchmark (0.11.0)
    runs-on: [self-hosted, Linux, comma]
@@ -315,7 +550,7 @@ jobs:
    - name: openpilot compile3 0.11.0 driving_vision (from pickle)
      run: BENCHMARK_LOG=openpilot_0_11_0_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM taskset -c 4-7 python3 examples/openpilot/compile3.py
    - name: IR3 openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=18 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
+      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
    - name: openpilot compile3 0.11.0 driving_policy
      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3.2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
    - name: openpilot compile3 0.11.0 dmonitoring
@@ -413,28 +648,31 @@ jobs:
    - name: openpilot run_pickle 0.10.1 driving_vision
      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py

-  driverbenchmarks:
-    name: PCI Driver Benchmark (DEV=${{ matrix.dev }})
+  testreddriverbenchmark:
+    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    strategy:
-      fail-fast: false
-      matrix:
-        dev: ['AMD', 'NV']
    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
-    env:
-      DEV: ${{ matrix.dev }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
-    - name: Setup
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove amd modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Symlink models and datasets
      run: |
-        ./extra/amdpci/setup_python_cap.sh
-        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} rmmod
-        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} kill_pids
+        mkdir -p weights
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
+        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
        mkdir -p extra/datasets
        ln -s /raid/datasets/imagenet extra/datasets/imagenet
    - name: setup staging db
@@ -445,47 +683,104 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Test driver cold start time
-      run: time DEBUG=3 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: time DEBUG=3 DEV=AMD AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Test driver warm start time
-      if: ${{ matrix.dev == 'AMD' }}
-      run: time DEBUG=3 python3 test/test_tiny.py TestTiny.test_plus
+      run: time DEBUG=3 DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
    - name: Test GPU crash recovery
-      if: ${{ matrix.dev == 'AMD' }}
-      run: python3 -m pytest -rA test/external/external_test_gpu_crash.py
-    - name: Test tensor cores
-      run: |
-        if [[ "${{ matrix.dev }}" == "AMD" ]]; then
-          # Fails on 9070
-          # python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-          # DEV=AMD:LLVM python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-          # SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-          SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-        else
-          ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-        fi
+      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
+    # Fails on 9070
+    # - name: Test tensor cores
+    #   run: |
+    #     DEV=AMD python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+    #     DEV=AMD:LLVM python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+    #     DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+    - name: Run Tensor Core GEMM (AMD)
+      run: DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+    - name: Test DEV=AMD
+      run: DEBUG=2 DEV=AMD python -m pytest -rA test/test_tiny.py
    - name: Test DISK copy time
-      run: TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+      run: DEV=AMD TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
    - name: Test CPU copy time
      run: |
-        GRAPH_ONE_KERNEL=1 NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        GRAPH_ONE_KERNEL=1 NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
+        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
+        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      if: ${{ matrix.dev == 'NV' }}
-      run: BENCHMARK_LOG=resnet_10steps MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: time BENCHMARK_LOG=cifar DEV=AMD DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #   run: BENCHMARK_LOG=resnet_10steps DEV=AMD MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    - name: Remote
      run: |
        pkill -f 'extra/remote/serve.py' || true
        PYTHONPATH=. python3 extra/remote/serve.py 6482 &
        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 python3 test/test_tiny.py
-        if [[ "${{ matrix.dev }}" == "AMD" ]]; then
-          DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD_AQL=1 python3 test/test_tiny.py
-        fi
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD python3 test/test_tiny.py
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD AMD_AQL=1 python3 test/test_tiny.py
+        pkill -f 'extra/remote/serve.py' || true
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testgreendriverbenchmark:
+    name: NV Benchmark
+    runs-on: [self-hosted, Linux, tinyboxrandom]
+    timeout-minutes: 20
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setcap to python
+      run: ./extra/amdpci/setup_python_cap.sh
+    - name: Remove nv modules
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
+    - name: Kill stale pids
+      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
+        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: Test driver start time
+      run: time DEBUG=3 DEV=NV python3 test/test_tiny.py TestTiny.test_plus
+    - name: Test tensor cores
+      run: DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+    - name: Test DISK copy time
+      run: DEV=NV TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+    - name: Test CPU copy time
+      run: |
+        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
+        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
+    - name: Test LLAMA-3
+      run: BENCHMARK_LOG=llama3_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar DEV=NV DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      run: BENCHMARK_LOG=resnet_10steps DEV=NV MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf Bert training steps (1 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Remote
+      run: |
+        pkill -f 'extra/remote/serve.py' || true
+        PYTHONPATH=. python3 extra/remote/serve.py 6483 &
+        sleep 1
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6483 DEV=NV python3 test/test_tiny.py
        pkill -f 'extra/remote/serve.py' || true
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
@@ -1,8 +1,8 @@
 name: Run MLPerf Training

 on:
-  #schedule:
-  #  - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
+  schedule:
+    - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
  push:
    branches:
      - update_mlperf
@@ -77,14 +77,45 @@ jobs:
      run: |
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
+    - name: Test one op
+      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Test ResNet-18
      run: DEBUG=2 python3 extra/torch_backend/example.py
+    - name: custom tests
+      run: python3 -m pytest -n auto extra/torch_backend/test.py --durations=20
    - name: Test one op in torch tests
      run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
    - name: Test Ops with TINY_BACKEND
      run: DEV=CPU:LLVM LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/backend/test_ops.py --durations=20
-    - name: Custom tests
-      run: DEV=CPU:LLVM GPUS=4 TINY_BACKEND=1 python3 -m pytest -nauto extra/torch_backend/test.py extra/torch_backend/test_inplace.py extra/torch_backend/test_multigpu.py extra/torch_backend/test_kernel_fusion.py --durations=20
+    - name: Test in-place operations on views
+      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
+    - name: Test multi-gpu
+      run: DEV=CPU:LLVM GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
+    - name: Test kernel fusion
+      run: python3 extra/torch_backend/test_kernel_fusion.py
+
+
+  torchbackendmore:
+    name: Torch Backend Tests More
+    runs-on: *linux
+    timeout-minutes: 15
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: torch-backend-pillow-torchvision-et-pt
+        deps: testing_unit
+        llvm: 'true'
+    - name: Install ninja
+      run: |
+        sudo apt update || true
+        sudo apt install -y --no-install-recommends ninja-build
+    - name: Test beautiful_mnist in torch with TINY_BACKEND
+      run: STEPS=20 DEV=CPU TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+    - name: Test some torch tests (expect failure)
+      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true

  bepython:
    name: Python Backend
@@ -102,26 +133,46 @@ jobs:
      run: SKIP_SLOW_TEST=1 DEV=PYTHON python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_ops.py test/backend/test_uops.py test/backend/test_symbolic_ops.py test/backend/test_renderer_failures.py::TestRendererFailures --durations=20
    - name: Test IMAGE support
      run: IMAGE=1 DEV=PYTHON python3 test/backend/test_ops.py TestOps.test_gemm TestOps.test_simple_conv2d
-    - name: Test emulated tensor cores
+    - name: Test emulated METAL tensor cores
      env:
-        DEBUG: 2
-        N: 64
-        CNT: 1
-        SHOULD_USE_TC: 1
+        DEV: 'PYTHON::METAL'
      run: |
-        parallel -k --link --tagstring '[{1}]' '{2} python3 ./extra/gemm/simple_matmul.py' \
-          ::: metal gfx950 gfx1100 gfx1100_acchalf gfx1201 gfx1201_acchalf sm_75 sm_80_half sm_80_tf32 \
-          ::: 'DEV=PYTHON::METAL' 'DEV=PYTHON::gfx950 HALF=1 ACC_HALF=0' \
-              'DEV=PYTHON::gfx1100 HALF=1 ACC_HALF=0' 'DEV=PYTHON::gfx1100 HALF=1 ACC_HALF=1 ATOL=1e-3' \
-              'DEV=PYTHON::gfx1201 HALF=1 ACC_HALF=0' 'DEV=PYTHON::gfx1201 HALF=1 ACC_HALF=1 ATOL=1e-3' \
-              'DEV=PYTHON::sm_75 HALF=1' 'DEV=PYTHON::sm_80 HALF=1' 'DEV=PYTHON::sm_80 ALLOW_TF32=1'
-    - name: Run additional tensor core tests
+        DEBUG=2 python3 test/backend/test_ops.py TestOps.test_big_gemm
+        python3 -m pytest -nauto test/opt/test_tensor_cores.py
+    - name: Test emulated AMD tensor cores
+      env:
+        DEV: 'PYTHON::gfx1100'
      run: |
-        DEV=PYTHON::METAL python3 -m pytest -nauto test/opt/test_tensor_cores.py test/null/test_uops_stats.py::TestUOpsStatsMatmulHalf
-        DEV=PYTHON::gfx1100 python3 -m pytest -nauto test/opt/test_tensor_cores.py test/null/test_uops_stats.py::TestUOpsStatsMatmulHalf
-        DEV=PYTHON::gfx950 python3 -m pytest -nauto test/opt/test_tensor_cores.py
-        DEV=PYTHON::gfx1201 python3 -m pytest -nauto test/opt/test_tensor_cores.py
+        DEBUG=2 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        python3 -m pytest -nauto test/opt/test_tensor_cores.py
+    - name: Test emulated AMD MFMA tensor cores
+      env:
+        DEV: 'PYTHON::gfx950'
+      run: |
+        DEBUG=2 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        python3 -m pytest -nauto test/opt/test_tensor_cores.py
+    - name: Test emulated AMD RDNA4 tensor cores
+      env:
+        DEV: 'PYTHON::gfx1201'
+      run: |
+        DEBUG=2 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        python3 -m pytest -nauto test/opt/test_tensor_cores.py
+    - name: Test emulated CUDA tensor cores
+      run: |
+        DEBUG=2 DEV=PYTHON::sm_80 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 ALLOW_TF32=1 DEV=PYTHON::sm_80 python3 test/backend/test_ops.py TestOps.test_gemm
+        DEBUG=2 DEV=PYTHON::sm_75 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
        ALLOW_TF32=1 DEV=PYTHON::sm_89 python3 -m pytest -nauto test/opt/test_tensor_cores.py
+    - name: Test device flop counts
+      run: |
+        DEBUG=2 DEV=PYTHON::METAL python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 DEV=PYTHON::gfx1100 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
        DEBUG=2 DEV=PYTHON::sm_80 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf

  linter:
@@ -216,6 +267,13 @@ jobs:
      run: python3 test/external/external_benchmark_schedule.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
+    - name: Regen dataset on test_tiny
+      run: |
+        test/external/process_replay/reset.py
+        CAPTURE_PROCESS_REPLAY=1 python test/test_tiny.py TestTiny.test_plus
+        python extra/optimization/extract_dataset.py
+        gzip -c /tmp/sops > extra/datasets/sops.gz
+        #DEBUG=1 MIN_ASTS=1 python extra/optimization/get_action_space.py
    - name: Repo line count < 25000 lines
      run: MAX_LINE_COUNT=25000 python sz.py

@@ -280,6 +338,31 @@ jobs:
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

+  testgpumisc:
+    name: CL Misc tests
+    runs-on: *linux
+    timeout-minutes: 10
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v6
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: gen-dataset
+          deps: testing
+          opencl: 'true'
+      - name: Generate Dataset
+        run: DEV=CL extra/optimization/generate_dataset.sh
+      - name: Run Kernel Count Test
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_opt.py
+      - name: Run fused optimizer tests
+        run: DEV=CL FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/backend/test_optim.py -k "not muon"
+      - name: Upload artifact
+        uses: actions/upload-artifact@v7
+        with:
+          name: sops.gz
+          path: /tmp/sops.gz
+
  testopenpilot:
    name: openpilot Compile Tests
    runs-on: *linux
@@ -296,7 +379,7 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=55 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1468 ALLOWED_GATED_READ_IMAGE=10 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp32 (test correctness)
        run: |
          DEV=CL IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
@@ -339,6 +422,7 @@ jobs:
        with:
          key: optim
          deps: testing
+          pydeps: "tensorflow==2.19"
          opencl: 'true'
      #- name: Test Optimization Helpers
      #  run: DEBUG=1 python3 extra/optimization/test_helpers.py
@@ -347,7 +431,7 @@ jobs:
      - name: Test Beam Search
        run: DEV=CL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
-        run: DEV=CL python -m pytest -n=auto test/external/external_test_lr_schedule.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
      - name: DEV=NULL beautiful_mnist_multigpu
        run: DEV=NULL NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
      - name: Test Bert training
@@ -653,33 +737,6 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

-  testmetal:
-    strategy:
-      fail-fast: false
-      matrix:
-        group: [1, 2]
-    name: MacOS (DEV=METAL) (${{ matrix.group }})
-    runs-on: *macos
-    timeout-minutes: 20
-    env:
-      DEV: METAL
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: macos-metal
-        deps: testing_unit
-    - name: Check Device.DEFAULT and print some source
-      run: |
-        python -c "from tinygrad import Device; assert Device.DEFAULT == 'METAL'"
-        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run backend tests
-      run: python -m pytest -n=auto test/backend --durations=20 --splits 2 --group ${{ matrix.group }}
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
  testmacos:
    strategy:
      fail-fast: false
@@ -688,6 +745,7 @@ jobs:
          - 'CPU:CLANG'
          - 'CPU:LLVM'
          - 'CPU:LVP'
+          - 'METAL'
          - 'WEBGPU'

    name: MacOS (DEV=${{ matrix.dev }})
@@ -710,8 +768,8 @@ jobs:
      run: |
        python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run test_tiny
-      run: python -m pytest -n=auto test/test_tiny.py --durations=20
+    - name: Run backend tests
+      run: python -m pytest -n=auto test/backend --durations=20
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -72,7 +72,7 @@ As it turns out, 90% of what you need for neural networks are a decent autograd/
 Throw in an optimizer, a data loader, and some compute, and you have all you need.

 ```python
-from tinygrad import Tensor, nn, Context
+from tinygrad import Tensor, nn

 class LinearNet:
  def __init__(self):
@@ -86,7 +86,7 @@ optim = nn.optim.Adam([model.l1, model.l2], lr=0.001)

 x, y = Tensor.rand(4, 1, 28, 28), Tensor([2,4,3,7])  # replace with real mnist dataloader

-with Context(TRAINING=1):
+with Tensor.train():
  for i in range(10):
    optim.zero_grad()
    loss = model(x).sparse_categorical_crossentropy(y).backward()
@@ -11,7 +11,7 @@ X_train -= X_train.mean()
 # *****
 # 1. Define an MNIST model.

-from tinygrad import Tensor, Context
+from tinygrad import Tensor

 l1 = Tensor.kaiming_uniform(128, 784)
 l2 = Tensor.kaiming_uniform(10, 128)
@@ -24,11 +24,11 @@ l1n, l2n = l1.numpy(), l2.numpy()
 from tinygrad.nn.optim import SGD
 optim = SGD([l1, l2])

-with Context(TRAINING=1):
-  X, Y = X_train[(samples:=Tensor.randint(128, high=X_train.shape[0]))], Y_train[samples]
-  optim.zero_grad()
-  model(X).sparse_categorical_crossentropy(Y).backward()
-  optim.schedule_step()   # this will step the optimizer without running realize
+Tensor.training = True
+X, Y = X_train[(samples:=Tensor.randint(128, high=X_train.shape[0]))], Y_train[samples]
+optim.zero_grad()
+model(X).sparse_categorical_crossentropy(Y).backward()
+optim.schedule_step()   # this will step the optimizer without running realize

 # *****
 # 3. Create a schedule (linear uop).
@@ -67,7 +67,8 @@ def example_2_hip(a:Tensor, correct):
    # the sink specifies the GLOBAL and LOCAL sizes, along with the input buffers and name
    sink = UOp.sink(UOp.special(GLOBALS, 'gidx0'), UOp.special(THREADS, 'lidx0'), out, buf,
                    arg=KernelInfo(name="hip_reduce_sum_kernel"))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
+                UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
  eval_harness("HIP kernel", a, lambda x: Tensor.empty(GLOBALS).custom_kernel(x, fxn=hip_reduce_sum)[0].sum(), check=correct)

 def example_3_custom_uop(a:Tensor, correct):
@@ -122,7 +123,8 @@ def example_5_custom_assembly(a:Tensor, correct):
        offset_dwords = (self.labels[inst._target] - inst._pos - inst.size()) // 4
        if not -32768 <= offset_dwords <= 32767: raise ValueError(f"branch to '{inst._target}' offset {offset_dwords} exceeds simm16 range")
        inst.simm16 = offset_dwords
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in self.instructions]))))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
+                                   UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in self.instructions]))))

  CU_COUNT = 32
  LANES = 64
@@ -24,7 +24,7 @@ You will see `CUDA` here on a GPU instance, or `CPU` here on a CPU instance.
 We'll use the model from [the Keras tutorial](https://keras.io/examples/vision/mnist_convnet/).

 ```python
-from tinygrad import Tensor, nn, Context
+from tinygrad import Tensor, nn

 class Model:
  def __init__(self):
@@ -74,8 +74,8 @@ We'll use the Adam optimizer. The `nn.state.get_parameters` will walk the model
 ```python
 optim = nn.optim.Adam(nn.state.get_parameters(model))
 batch_size = 128
-@Context(TRAINING=1)
 def step():
+  Tensor.training = True  # makes dropout work
  samples = Tensor.randint(batch_size, high=X_train.shape[0])
  X, Y = X_train[samples], Y_train[samples]
  optim.zero_grad()
@@ -143,6 +143,7 @@ Since we are just randomly sampling from the dataset, there's no real concept of
 for step in range(7000):
  loss = jit_step()
  if step%100 == 0:
+    Tensor.training = False
    acc = (model(X_test).argmax(axis=1) == Y_test).mean().item()
    print(f"step {step:4d}, loss {loss.item():.2f}, acc {acc*100.:.2f}%")
 ```
@@ -165,14 +165,13 @@ from extra.datasets import fetch_mnist
 Now we have everything we need to start training our neural network.
 We will be training for 1000 steps with a batch size of 64.

-We use `with Context(TRAINING=1)` to enable training mode.
+We use `with Tensor.train()` to set the internal flag `Tensor.training` to `True` during training.
 Upon exit, the flag is restored to its previous value by the context manager.

 ```python
-from tinygrad import Context
 X_train, Y_train, X_test, Y_test = fetch_mnist()

-with Context(TRAINING=1):
+with Tensor.train():
  for step in range(1000):
    # random sample a batch
    samp = np.random.randint(0, X_train.shape[0], size=(64))
@@ -182,7 +182,7 @@ if __name__ == "__main__":
  # *** run on Machine and compare ***

  # NOTE: the scratch size needs to be reduced to 1536 when you have a register allocator
-  src = eval(prg.src[2].arg)
+  src = eval(prg.src[3].arg)
  max_regs = max(t[1] for instr in src for v in instr.values() for t in v if len(t) > 1) + 8
  print(f"{max_regs:5d} regs used" + ("" if max_regs <= 1536 else "       <-- WARNING: TOO MANY REGISTERS, MUST BE <= 1536"))
  machine = problem.Machine(mem, src, problem.DebugInfo(scratch_map={}), n_cores=1, trace=False, scratch_size=max_regs)
@@ -1,6 +1,6 @@
 from typing import Tuple
 import time
-from tinygrad import Tensor, TinyJit, nn, Context
+from tinygrad import Tensor, TinyJit, nn
 import gymnasium as gym
 from tinygrad.helpers import trange
 import numpy as np  # TODO: remove numpy import
@@ -55,7 +55,7 @@ if __name__ == "__main__":

  @TinyJit
  def train_step(x:Tensor, selected_action:Tensor, reward:Tensor, old_log_dist:Tensor) -> Tuple[Tensor, Tensor, Tensor]:
-    with Context(TRAINING=1):
+    with Tensor.train():
      log_dist, value = model(x)
      action_mask = (selected_action.reshape(-1, 1) == Tensor.arange(log_dist.shape[1]).reshape(1, -1).expand(selected_action.shape[0], -1)).float()

@@ -122,7 +122,7 @@ if __name__ == "__main__":
    return ret.mul(hyp['opt']['loss_scale_scaler']*loss_batchsize_scaler).sum().div(hyp['opt']['loss_scale_scaler'])

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def train_step(idxs:Tensor) -> Tensor:
    X, Y = X_train[idxs], Y_train[idxs]
    if len(GPUS) > 1:
@@ -1,6 +1,6 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function, Context
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

@@ -19,7 +19,7 @@ class Model:
  def __call__(self, x:Tensor) -> Tensor: return x.sequential(self.layers)

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def train_step(self, X_train:Tensor, Y_train:Tensor) -> Tensor:
    opt.zero_grad()
    samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
@@ -1,6 +1,6 @@
 # model based off https://towardsdatascience.com/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import List, Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, Device, Context
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, Device
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

@@ -31,7 +31,7 @@ if __name__ == "__main__":

  @TinyJit
  def train_step() -> Tensor:
-    with Context(TRAINING=1):
+    with Tensor.train():
      opt.zero_grad()
      samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
      Xt, Yt = X_train[samples].shard_(GPUS, axis=0), Y_train[samples].shard_(GPUS, axis=0)  # we shard the data on axis 0
@@ -1,6 +1,6 @@
 import itertools
 from typing import Callable
-from tinygrad import nn, Tensor, dtypes, Device, TinyJit, Context
+from tinygrad import nn, Tensor, dtypes, Device, TinyJit
 from tinygrad.helpers import getenv, trange, partition

 class Model:
@@ -59,7 +59,7 @@ if __name__ == "__main__":
  Tensor.realize(*params, *buffers, *adam_params, loss, grads)

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def microbatch():
    samples = Tensor.randint(BS // ACC_STEPS, high=X_train.shape[0])
    for t in params: t.grad = None
@@ -10,7 +10,7 @@ from extra.lr_scheduler import OneCycleLR
 from tinygrad import nn, dtypes, Tensor, Device, GlobalCounters, TinyJit, Variable
 from tinygrad.nn.state import get_state_dict
 from tinygrad.nn import optim
-from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod, TRAINING
+from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod
 from extra.bench_log import BenchEvent, WallTimeEvent

 cifar_mean = [0.4913997551666284, 0.48215855929893703, 0.4465309133731618]
@@ -44,7 +44,7 @@ class UnsyncedBatchNorm:
    return ret.reshape(x.shape).cast(x.dtype)

  def calc_stats(self, x:Tensor):
-    if TRAINING:
+    if Tensor.training:
      # This requires two full memory accesses to x
      # https://github.com/pytorch/pytorch/blob/c618dc13d2aa23625cb0d7ada694137532a4fa33/aten/src/ATen/native/cuda/Normalization.cuh
      # There's "online" algorithms that fix this, like https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Welford's_Online_algorithm
@@ -359,11 +359,11 @@ def train_cifar():
  i = 0
  eval_acc_pct = 0.0
  batcher = fetch_batches(X_train, Y_train, BS=BS, is_train=True)
-  with Context(TRAINING=1):
+  with Tensor.train():
    st = time.monotonic()
    while i <= STEPS:
      if i % getenv("EVAL_STEPS", STEPS) == 0 and i > 1 and not getenv("DISABLE_BACKWARD"):
-        # Using Context(TRAINING=0) here actually bricks batchnorm, even with track_running_stats=True
+        # Use Tensor.training = False here actually bricks batchnorm, even with track_running_stats=True
        corrects = []
        corrects_ema = []
        losses = []
@@ -3,7 +3,7 @@ import os
 if "NOOPT" not in os.environ: os.environ["NOOPT"] = "1"
 from tinygrad import Device, nn, Tensor, dtypes
 from train_gpt2 import GPT, GPTConfig
-from tinygrad.helpers import DEV, dedup, flatten, getenv, GlobalCounters, to_function_name, Context
+from tinygrad.helpers import DEV, dedup, flatten, getenv, GlobalCounters, to_function_name
 from tinygrad.engine.realize import get_kernel
 from tinygrad.schedule.memory import memory_planner
 from tinygrad.uop.ops import Ops
@@ -23,23 +23,23 @@ if __name__ == "__main__":
  #B, T = Variable("B", 1, 128).bind(4), 64 #Variable("T", 1, 1024).bind(64)
  B, T = 4, 64

+  Tensor.training = True
  optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=1e-4)
  warmup_count = getenv("WARMUP", 3)
-  with Context(TRAINING=1):
-    for i in range(warmup_count):  # TODO: why does it take three and not two to stabilize
-      GlobalCounters.reset()
-      X = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
-      Y = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
-      _, loss = model(X, Y)
-      optimizer.zero_grad()
-      if getenv("BACKWARD", 1):
-        loss.backward()
-        tensors = optimizer.schedule_step()
-      else:
-        tensors = []
-      sched = loss.schedule(*tensors)
-      print(f"calls {i}:", len(sched))
-      #run_schedule(sched[:])
+  for i in range(warmup_count):  # TODO: why does it take three and not two to stabilize
+    GlobalCounters.reset()
+    X = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
+    Y = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
+    _, loss = model(X, Y)
+    optimizer.zero_grad()
+    if getenv("BACKWARD", 1):
+      loss.backward()
+      tensors = optimizer.schedule_step()
+    else:
+      tensors = []
+    sched = loss.schedule(*tensors)
+    print(f"calls {i}:", len(sched))
+    #run_schedule(sched[:])
  sched = memory_planner(sched)
  ast_dedup = dedup([si.ast for si in sched if si.ast.op is Ops.SINK])
  srcs = {}
@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 import os, math, time
 import numpy as np
-from tinygrad import Tensor, nn, fetch, Device, TinyJit, GlobalCounters, Context
+from tinygrad import Tensor, nn, fetch, Device, TinyJit, GlobalCounters
 from dataclasses import dataclass

@dataclass
@@ -177,7 +177,7 @@ if __name__ == "__main__":
  if args.gpus > 1: x, y = x.shard(GPUS, axis=0), y.shard(GPUS, axis=0)

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def step(x:Tensor, y:Tensor) -> Tensor:
    _, loss = model(x, y)
    optimizer.zero_grad()
@@ -204,3 +204,4 @@ if __name__ == "__main__":
    top_k = 40
    y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
    print(decode(y[0].tolist()))
+
@@ -1,5 +1,5 @@
 # much taken from https://github.com/cloneofsimo/minRF
-from tinygrad import Tensor, nn, GlobalCounters, TinyJit, Context
+from tinygrad import Tensor, nn, GlobalCounters, TinyJit
 from tinygrad.helpers import getenv, trange
 from extra.models.llama import Attention, FeedForward, precompute_freqs_cis

@@ -135,7 +135,7 @@ if __name__ == "__main__":
  optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=5e-4)

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def train_step():
    if getenv("OVERFIT"): samples = Tensor.zeros(getenv("BS", 256), dtype='int')
    else: samples = Tensor.randint(getenv("BS", 256), high=X_train.shape[0])
@@ -2,7 +2,7 @@ import math
 from typing import Union

 from tinygrad import Tensor, nn, dtypes
-from tinygrad.helpers import prod, argfix, Context, TRAINING
+from tinygrad.helpers import prod, argfix, Context
 from tinygrad.nn.state import get_parameters
 from extra.models.unet import UNetModel

@@ -85,7 +85,7 @@ class FrozenBatchNorm2dRetinaNet(nn.BatchNorm2d):

  def __call__(self, x:Tensor) -> Tensor:
    batch_mean, batch_var = super().calc_stats(x.cast(dtypes.float32))
-    if self.track_running_stats and TRAINING:
+    if self.track_running_stats and Tensor.training:
      self.running_mean.assign((1-self.momentum) * self.running_mean + self.momentum * batch_mean.detach().cast(self.running_mean.dtype))
      self.running_var.assign((1-self.momentum) * self.running_var + self.momentum * x.numel()/(x.numel()-x.shape[1]) * batch_var.detach().cast(self.running_var.dtype))
      self.num_batches_tracked += 1
@@ -358,7 +358,7 @@ def eval_stable_diffusion():
      batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
    return batch, unpadded_bs 

-  @Context(TRAINING=0)
+  @Tensor.train(mode=False)
  def eval_unet(eval_inputs:list[dict], unet:UNetModel, cond_stage:FrozenOpenClipEmbedder, first_stage:AutoencoderKL,
                inception:FidInceptionV3, clip:OpenClipEncoder) -> tuple[float, float]:
    # Eval is divided into 5 jits, one per model
@@ -498,10 +498,11 @@ def eval_stable_diffusion():

 if __name__ == "__main__":
  # inference only
+  Tensor.training = False
+
  models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert").split(",")
-  with Context(TRAINING=0):
-    for m in models:
-      nm = f"eval_{m}"
-      if nm in globals():
-        print(f"eval {m}")
-        globals()[nm]()
+  for m in models:
+    nm = f"eval_{m}"
+    if nm in globals():
+      print(f"eval {m}")
+      globals()[nm]()
@@ -1,6 +1,6 @@
 # load each model here, quick benchmark
 from tinygrad import Tensor, GlobalCounters
-from tinygrad.helpers import getenv, Context
+from tinygrad.helpers import getenv
 import numpy as np

 def test_model(model, *inputs):
@@ -59,10 +59,11 @@ def spec_mrcnn():

 if __name__ == "__main__":
  # inference only for now
-  with Context(TRAINING=0):
-    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
-      nm = f"spec_{m}"
-      if nm in globals():
-        print(f"testing {m}")
-        globals()[nm]()
+  Tensor.training = False
+
+  for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
+    nm = f"spec_{m}"
+    if nm in globals():
+      print(f"testing {m}")
+      globals()[nm]()

@@ -2,7 +2,7 @@ import os, time, math, functools, random, contextlib
 from pathlib import Path
 import multiprocessing

-from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes, Context
+from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
 from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling, profile_marker, DEBUG
 from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW
@@ -157,7 +157,6 @@ def train_resnet():
  # input_std = Tensor([0.229, 0.224, 0.225], device=GPUS, dtype=dtypes.float32).reshape(1, -1, 1, 1)
  def normalize(x): return (x.permute([0, 3, 1, 2]) - input_mean).cast(dtypes.default_float)
  @TinyJit
-  @Context(TRAINING=1)
  def train_step(X, Y):
    optimizer_group.zero_grad()
    X = normalize(X)
@@ -171,7 +170,6 @@ def train_resnet():
    return loss.realize(), top_1.realize()

  @TinyJit
-  @Context(TRAINING=0)
  def eval_step(X, Y):
    X = normalize(X)
    out = model.forward(X)
@@ -194,6 +192,7 @@ def train_resnet():
    # ** train loop **
    if MLLOGGER and RUNMLPERF:
      MLLOGGER.start(key=mllog_constants.EPOCH_START, value=e+1, metadata=dict(epoch_num=e+1))
+    Tensor.training = True
    BEAM.value = TRAIN_BEAM

    if INITMLPERF:
@@ -272,6 +271,7 @@ def train_resnet():
      eval_loss = 0.0
      eval_top_1 = 0
      eval_num_samples = 0
+      Tensor.training = False
      BEAM.value = EVAL_BEAM

      if INITMLPERF:
@@ -614,7 +614,7 @@ def train_retinanet():

      if getenv("RESET_STEP", 1): _train_step.reset()

-      with Context(TRAINING=0):
+      with Tensor.train(mode=False):
        if not RUNMLPERF:
          i, proc = 0, _fake_data_get(EVAL_BS, val=(val:=True))
        else:
@@ -784,7 +784,7 @@ def train_unet3d():
    return x.shard(GPUS, axis=0).realize(), y.shard(GPUS, axis=0), cookie

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def train_step(model, x, y):
    optim.zero_grad()

@@ -795,7 +795,7 @@ def train_unet3d():
    optim.step()
    return loss.realize()

-  @Context(TRAINING=0)
+  @Tensor.train(mode=False)
  def eval_step(model, x, y):
    y_hat, y = sliding_window_inference(model, x, y, gpus=GPUS)
    y_hat, y = Tensor(y_hat), Tensor(y)
@@ -919,7 +919,6 @@ def train_rnnt():
  pass

@TinyJit
-@Context(TRAINING=0)
 def eval_step_bert(model, input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor, masked_positions:Tensor, masked_lm_ids:Tensor,
                   masked_lm_weights:Tensor, next_sentence_labels:Tensor, GPUS):
  for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
@@ -1107,7 +1106,6 @@ def train_bert():
      MLLOGGER.start(key=mllog_constants.EPOCH_START, value=i*GBS, metadata={"epoch_num": i*GBS})

  @TinyJit
-  @Context(TRAINING=1)
  def train_step_bert(input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor,
                      masked_positions:Tensor, masked_lm_ids:Tensor, masked_lm_weights:Tensor, next_sentence_labels:Tensor):
    for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
@@ -1135,6 +1133,7 @@ def train_bert():

  while train_data is not None and i < train_steps and not achieved:
    if getenv("TRAIN", 1):
+      Tensor.training = True
      BEAM.value = TRAIN_BEAM
      st = time.perf_counter()
      GlobalCounters.reset()
@@ -1187,6 +1186,7 @@ def train_bert():
      eval_lm_accs = []
      eval_clsf_accs = []
      eval_times = []
+      Tensor.training = False
      BEAM.value = EVAL_BEAM

      for j in tqdm(range(max_eval_steps), desc="Evaluating", total=max_eval_steps, disable=BENCHMARK):
@@ -1490,7 +1490,7 @@ def train_llama3():
    return lr_cpu, grad_norm_cpu

  @TinyJit
-  @Context(TRAINING=0)
+  @Tensor.train(False)
  def eval_step(tokens:Tensor):
    if is_dp: tokens = tokens.to(None).shard(device, 0)
    if is_mp: tokens = tokens.shard(device)
@@ -1803,7 +1803,7 @@ if __name__ == "__main__":
  elif getenv("RUNMLPERF"): bench_log_manager = WallTimeEvent(BenchEvent.MLPERF_RUN)
  else: bench_log_manager = contextlib.nullcontext()

-  with Context(TRAINING=1):
+  with Tensor.train():
    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
      nm = f"train_{m}"
      if nm in globals():
@@ -38,7 +38,7 @@ def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):

 def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
           x_fp8:Tensor|None=None, x_new_amax:Tensor|None=None,
-           grad_amax_state:Tensor|None=None, x_prequant_mx:tuple|None=None) -> tuple[Tensor,...]:
+           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
  if not fp8:
    if ASM_GEMM:
      from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
@@ -47,14 +47,12 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
  assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
  if MXFP8:
    from extra.gemm.cdna_asm_gemm import asm_gemm, quantize_mxfp8, mx_pack, can_use_asm_gemm, _mx_block_scale
-    if x_prequant_mx is not None: x_q, x_e8, x_si = x_prequant_mx       # fused producer already quantized (2d)
-    else: x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
-    l_shape = x.shape[:-1] if x is not None else x_q.shape[:-1]
+    x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
    if can_use_asm_gemm(x_q, w.T):
      out = asm_gemm(x_q, w.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(w_inv_scale), w_inv_scale),
-                     mx_w_stored=True).reshape(*l_shape, w.shape[0])
+                     mx_w_stored=True).reshape(*x.shape[:-1], w.shape[0])
    else:
-      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*l_shape, x_q.shape[-1])
+      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*x.shape[:-1], x.shape[-1])
      out = x_phys @ (w.cast(dtypes.bfloat16) * _mx_block_scale(w_inv_scale)).T
    return out, (amax_x.detach() if amax_x is not None else None), x_q
  if x_fp8 is None:
@@ -128,8 +126,10 @@ class FlatTransformer:

    # FeedForward
    if SPLIT_W13:
-      self.w1, s_1 = self.lin_per_layer(dim, hidden_dim)
-      self.w3, s_3 = self.lin_per_layer(dim, hidden_dim)
+      if getenv("ZEROS"): w13_raw = Tensor.zeros(2, self.n_layers, hidden_dim, dim)
+      else: w13_raw = Tensor.normal(2, self.n_layers, hidden_dim, dim, mean=0.0, std=0.02)
+      self.w1, s_1 = self.lin_per_layer(dim, hidden_dim, w=w13_raw[0])
+      self.w3, s_3 = self.lin_per_layer(dim, hidden_dim, w=w13_raw[1])
    else:
      self.w13, s_13 = self.lin_per_layer(dim, hidden_dim * 2)
    self.w2, s_2 = self.lin_per_layer(hidden_dim, dim, std=scaled_std)
@@ -160,7 +160,7 @@ class FlatTransformer:
  def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02, w:Tensor|None=None):
    if w is None:
      if getenv("ZEROS"): w = Tensor.zeros(self.n_layers, out_features, in_features)
-      else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std)
+      else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std).realize()
    if MXFP8:
      from extra.gemm.cdna_asm_gemm import quantize_mxfp8
      w_q, w_e8, _ = quantize_mxfp8(w.reshape(self.n_layers * out_features, in_features))
@@ -216,15 +216,8 @@ class FlatTransformer:
      x_w3, new_amax, *s = matmul(inp, kwargs["w3"], amax_x=kwargs["amax_x3"], w_inv_scale=kwargs["s_3"], grad_amax_state=kwargs["grad_amax_xw3"])
      amaxs.append(new_amax)
      saves.extend([*s, x_w3])
-      if FUSED_SILU_W13 and MXFP8:
-        from extra.llama_kernels.fused_silu_mul_quantize_mxfp8 import fused_silu_mul_quantize_mxfp8
-        aq, ae8, asi = fused_silu_mul_quantize_mxfp8(x_w1.reshape(-1, x_w1.shape[-1]), x_w3.reshape(-1, x_w3.shape[-1]))
-        out, new_amax, *s = matmul(None, kwargs["w2"], x_prequant_mx=(aq, ae8, asi), amax_x=kwargs["amax_x2"],
-                                   w_inv_scale=kwargs["s_2"], grad_amax_state=kwargs["grad_amax_xout"])
-        out = out.reshape(*x_w1.shape[:-1], kwargs["w2"].shape[0])
-      else:
-        out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
-                                   grad_amax_state=kwargs["grad_amax_xout"])
+      out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
+                                 grad_amax_state=kwargs["grad_amax_xout"])
      amaxs.append(new_amax)
      saves.extend([*s, out])
    else:
@@ -254,30 +247,20 @@ class FlatTransformer:
      for v in get_parameters(self): v.shard_(device, axis=None)
    else:
      # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
-      def _shard_fp8(name:str, axis:int, std:float=0.02):
-        w = getattr(self, name)
-        if MXFP8:
-          from extra.gemm.cdna_asm_gemm import quantize_mxfp8
-          w_bf16 = Tensor.empty(self.n_layers, w.shape[1], w.shape[2], dtype=dtypes.bfloat16).shard(device, axis=axis).randn_like() * std
-          w_q, w_e8, _ = quantize_mxfp8(w_bf16)
-          w.replace(w_q)
-          self._fp8_inv_scale[name].replace(w_e8.contiguous()).is_param_(False)
-          self._fp8_next_inv_scale[name].replace(w_e8.contiguous()).is_param_(False)
-        else:
-          w.shard_(device, axis=axis)
-          scale_axis = (1 if axis == 1 else None) if COLUMNWISE_WEIGHT_SCALE else None
-          self._fp8_inv_scale[name] = self._fp8_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
-          self._fp8_next_inv_scale[name] = self._fp8_next_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
-        Tensor.realize(w, self._fp8_inv_scale[name], self._fp8_next_inv_scale[name])
-      sstd = 0.02 / math.sqrt(2 * self.n_layers)
+      def _shard_fp8(name:str, axis:int):
+        getattr(self, name).shard_(device, axis=axis)
+        scale_axis = axis if MXFP8 else (1 if axis == 1 else None) if COLUMNWISE_WEIGHT_SCALE else None
+        self._fp8_inv_scale[name] = self._fp8_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
+        self._fp8_next_inv_scale[name] = self._fp8_next_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
+        Tensor.realize(getattr(self, name), self._fp8_inv_scale[name], self._fp8_next_inv_scale[name])
      _shard_fp8("wqkv", 1)          # (n_layers, out, dim) shard out
-      _shard_fp8("wo", 2, sstd)      # (n_layers, dim, in) shard in
+      _shard_fp8("wo", 2)            # (n_layers, dim, in) shard in
      if SPLIT_W13:
        _shard_fp8("w1", 1)
        _shard_fp8("w3", 1)
      else:
        _shard_fp8("w13", 1)         # (n_layers, hidden*2, dim) shard out
-      _shard_fp8("w2", 2, sstd)      # (n_layers, dim, hidden) shard in
+      _shard_fp8("w2", 2)            # (n_layers, dim, hidden) shard in
      self.attention_norm.shard_(device, axis=None).realize()
      self.ffn_norm.shard_(device, axis=None).realize()
      self.norm.weight.shard_(device, axis=None).realize()
@@ -1,6 +1,6 @@
 #!/bin/bash
-export BENCHMARK=${BENCHMARK:-5}
+export BENCHMARK=5
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
 SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
-[ "$BENCHMARK" -le 3 ] || python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
+python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
@@ -3,7 +3,7 @@ import torch
 from torchvision.utils import make_grid, save_image
 from tinygrad.nn.state import get_parameters
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import trange, Context
+from tinygrad.helpers import trange
 from tinygrad.nn import optim
 from tinygrad.nn.datasets import mnist

@@ -86,7 +86,7 @@ if __name__ == "__main__":
  optim_g = optim.Adam(get_parameters(generator), lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
  optim_d = optim.Adam(get_parameters(discriminator), lr=0.0002, b1=0.5)
  # training loop
-  with Context(TRAINING=1):
+  with Tensor.train():
    for epoch in (t := trange(epochs)):
      loss_g, loss_d = 0.0, 0.0
      for _ in range(n_steps):
@@ -42,7 +42,7 @@ def compile(onnx_file):
  kernel_calls = [u for u in run_onnx_jit.captured.linear.toposort(gate=lambda x: x.op not in kernel_asts)
                  if u.op is Ops.CALL and u.src[0].op in kernel_asts]
  print(f"captured {len(kernel_calls)} kernels")
-  if getenv("TEST", 1): np.testing.assert_equal(test_val, ret, "JIT run failed")
+  np.testing.assert_equal(test_val, ret, "JIT run failed")
  print("jit run validated")

  # check gated read_image usage
@@ -50,7 +50,7 @@ def compile(onnx_file):
  read_image_count = 0
  gated_read_image_count = 0
  for call in kernel_calls:
-    _, _, source, _ = call.src[0].src
+    _, _, _, source, _ = call.src[0].src
    src = source.arg
    kernel_count += 1
    read_image_count += src.count("read_image")
@@ -5,7 +5,7 @@
 #  - symbolic removal

 from examples.beautiful_mnist import Model
-from tinygrad import Tensor, nn, getenv, GlobalCounters, Variable, Context
+from tinygrad import Tensor, nn, getenv, GlobalCounters, Variable
 from tinygrad.nn.datasets import mnist
 from tinygrad.helpers import trange

@@ -26,7 +26,7 @@ if __name__ == "__main__":
  X_samp, Y_samp = X_train[samples], Y_train[samples]
  print("*** got samples")

-  with Context(TRAINING=1):
+  with Tensor.train():
    """
    i = UOp.range(samples.shape[0])  # TODO: fix range function on UOp
    losses = model(X_samp[i]).sparse_categorical_crossentropy(Y_samp[i]).backward().contract(i)
@@ -38,7 +38,7 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li
    arg_uops = [b for b in call.src[1:] if b.op is not Ops.BIND]
    prg = to_program(call.src[0], Device[arg_uops[0].device].renderer)
    info = prg.arg
-    functions[info.function_name] = prg.src[2].arg
+    functions[info.function_name] = prg.src[3].arg
    cargs = [name_of(bu, i == 0) for i, bu in enumerate(arg_uops)] + list(info.vars)
    statements.append((info.function_name, cargs, info.global_size, info.local_size))

@@ -462,7 +462,7 @@ def test_matmul():
    lds = UOp.placeholder((lds_size,), dtypes.uint8, 0, AddrSpace.LOCAL)
    sink = UOp.sink(A.base, B.base, C.base, lds, *gidxs, *lidxs, arg=KernelInfo(name=colored("kernel", "cyan"),
                                                                                  estimates=Estimates(ops=N*N*N*2, mem=N*N*4*3)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
  c = Tensor.custom_kernel(a, b, c, fxn=asm_kernel)[2]
  linear = c.schedule_linear()

@@ -66,7 +66,7 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:

  # accumulator (unified: both paths use (TM, TN) with scalar dtypes.float)
  acc = UOp.placeholder((TM, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(acc.zeros_like(buffer=False)))
+  acc = acc.after(acc.store(acc.zeros_like()))

  if use_wmma:
    k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)
@@ -17,7 +17,7 @@ def make_matmul_kernel(name:str, src:str, local_size:int):
    wg_y = UOp.special(N//128, "gidx1")
    sink = UOp.sink(a.base, b.base, c.base, threads, wg_x, wg_y, arg=KernelInfo(name, estimates=Estimates(ops=2*N**3, mem=3*N*N*4)))
    lib = Device[Device.DEFAULT].compiler.compile_cached(src)
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                 UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
  return fxn

@@ -2622,7 +2622,8 @@ def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
  lds = UOp.placeholder((133_120,), dtypes.uint8, 0, AddrSpace.LOCAL)
  sink = UOp.sink(C.base, A.base, B.base, lds, lidx, gidx,
                  arg=KernelInfo(name=f"gemm_{batch}_{M}_{N}_{K}", estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname),
+                                UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

 # ** FP8 GEMM custom kernel

@@ -2645,7 +2646,7 @@ def custom_hk_fp8_gemm(C:UOp, A:UOp, B:UOp, *args:UOp, dname:str, scale_mode:int
  lib = HIPCCCompiler("gfx950", [f"-I{(kittens_path/'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-ffast-math",
                                 "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DGEMM_M={M}", f"-DGEMM_N={N}", f"-DGEMM_K={K}",
                                 f"-DSCALE_MODE={scale_mode}"]).compile_cached(src)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
                               UOp(Ops.BINARY, arg=lib)))

 # ** MXFP8 GEMM custom kernel
@@ -2668,19 +2669,19 @@ def custom_hk_mxfp8_gemm(C:UOp, A:UOp, B:UOp, scale_A:UOp, scale_B:UOp, *extra:U
  src = (kittens_path/"gemm_mxfp8.cpp").read_text()
  lib = HIPCCCompiler("gfx950", [f"-I{(kittens_path/'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-ffast-math",
                                 "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DGEMM_M={M}", f"-DGEMM_N={N}", f"-DGEMM_K={K}"]).compile_cached(src)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
                               UOp(Ops.BINARY, arg=lib)))

 def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
  # 1x32 block scaling along the last axis
-  *batch, K = x.shape
-  scale_K = K // 32
-  amax = x.detach().float().reshape(*batch, scale_K, 32).abs().max(axis=-1)
+  rows, K = x.shape
+  scale_K, k_iters = K // 32, K // 128
+  amax = x.detach().float().reshape(rows, scale_K, 32).abs().max(axis=-1)
  e8 = (amax.maximum(1e-38).log2().floor() + 127).clamp(0, 254).cast(dtypes.uint8)
-  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(*batch, scale_K, 1).expand(*batch, scale_K, 32).reshape(*batch, K)
+  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(rows, scale_K, 1).expand(rows, scale_K, 32).reshape(rows, K)
  x_scaled = x.float() * qscale
  x_clamped = x_scaled + (x_scaled.detach().clamp(-448.0, 448.0) - x_scaled.detach())  # STE
-  return x_clamped.cast(FP8_DTYPE), e8, (mx_pack(e8) if len(batch) == 1 else None)
+  return x_clamped.cast(FP8_DTYPE), e8, mx_pack(e8)

 def mx_pack(e8:Tensor) -> Tensor:
  rows, scale_K = e8.shape
@@ -2758,7 +2759,7 @@ def custom_hk_bf16_gemm(C:UOp, A:UOp, B:UOp, *args:UOp, dname:str) -> UOp:
  src = (kittens_path/"gemm_bf16.cpp").read_text()
  lib = HIPCCCompiler("gfx950", [f"-I{(kittens_path/'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-ffast-math",
                                 "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DGEMM_M={M}", f"-DGEMM_N={N}", f"-DGEMM_K={K}"]).compile_cached(src)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
                                UOp(Ops.BINARY, arg=lib)))

@functools.cache
@@ -2776,7 +2777,7 @@ def custom_hk_bf16_atb_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
  src = (kittens_path/"gemm_bf16_atb.cpp").read_text()
  lib = HIPCCCompiler("gfx950", [f"-I{(kittens_path/'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-ffast-math",
                                 "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DGEMM_M={M}", f"-DGEMM_N={N}", f"-DGEMM_K={K}"]).compile_cached(src)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
                                UOp(Ops.BINARY, arg=lib)))

 def hk_bf16_atb_gemm(a:Tensor, b:Tensor) -> Tensor:
@@ -2850,7 +2851,11 @@ def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=
    grad_a = asm_gemm(g_fp8, b_t, x_scale=s_x_t, w_scale=s_w_t, g_scale=g_scale) if has_w else asm_gemm(g_fp8, b_t, x_scale=s_x_t, w_scale=g_scale)
    # wgrad: no w_scale
    g_fp8_2d = g_fp8.reshape(-1, g_fp8.shape[-1])
-    g_fp8_T = g_fp8.permute(2, 0, 1).reshape(g_t.shape[-1], -1)
+    if getenv("FAST_FP8_TRANSPOSE", 0) and g_fp8_2d.shape[0] % 64 == 0 and g_fp8_2d.shape[1] % 64 == 0:
+      from extra.llama_kernels.fp8_transpose import fast_fp8_transpose
+      g_fp8_T = fast_fp8_transpose(g_fp8_2d)
+    else:
+      g_fp8_T = g_fp8.permute(2, 0, 1).reshape(g_t.shape[-1], -1)
    grad_b = asm_gemm(g_fp8_T, a_t.reshape(-1, a_t.shape[-1]), x_scale=s_x_t, w_scale=g_scale)
    # wgrad: rescale if not scalar
    if w_post_t is not None:
@@ -2894,7 +2899,7 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F

  g = Tensor(gradient, device=aq.device)[:aq.shape[0]].reshape(aq.shape[0]*aq.shape[1], bq.shape[0]).cast(dtypes.bfloat16)
  grad_a = asm_gemm(g, b_phys, mx=True)
-  grad_b = asm_gemm(g.T, a_phys, mx=True, a_pretranspose=g)
+  grad_b = asm_gemm(g.T, a_phys, mx=True)

  grad_a = (grad_a * _mx_block_scale(ae8)).reshape(aq.shape)
  if not w_stored: grad_b = grad_b * _mx_block_scale(be8)
@@ -2904,8 +2909,7 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F
 # ** main gemm function

 def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=None, grad_amax_state:Tensor|None=None,
-             w_post_scale:Tensor|None=None, mx:bool=False, mx_scales:tuple|None=None, mx_w_stored:bool=False, g_scale:Tensor|None=None,
-             a_pretranspose:Tensor|None=None) -> Tensor:
+             w_post_scale:Tensor|None=None, mx:bool=False, mx_scales:tuple|None=None, mx_w_stored:bool=False, g_scale:Tensor|None=None) -> Tensor:
  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
  counters["used"] += 1
  unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
@@ -2942,11 +2946,6 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
      if mx_scales is not None:
        a_si, a_e8, b_si, b_e8 = mx_scales
        a_q, b_q = a.reshape(-1, a.shape[-1]), b.T
-      elif (a_pretranspose is not None and getenv("FUSED_GRAD_QUANTIZE", 0) and a_pretranspose.dtype == dtypes.bfloat16
-            and a_pretranspose.shape[0] % 32 == 0 and a_pretranspose.shape[1] % 256 == 0):
-        from extra.llama_kernels.transpose_quantize_mxfp8 import transpose_quantize_mxfp8
-        a_q, a_e8, a_si = transpose_quantize_mxfp8(a_pretranspose)
-        b_q, b_e8, b_si = quantize_mxfp8(b.T)
      else:
        a_q, a_e8, a_si = quantize_mxfp8(a.reshape(-1, a.shape[-1]))
        b_q, b_e8, b_si = quantize_mxfp8(b.T)
@@ -223,7 +223,7 @@ def test_matmul():
    lds = UOp.placeholder((lds_size,), dtypes.uint8, 0, AddrSpace.LOCAL)
    sink = UOp.sink(A.base, B.base, C.base, lds, *gidxs, *lidxs,
                    arg=KernelInfo(name=colored("kernel","cyan"), estimates=Estimates(ops=N*N*N*2, mem=N*N*2*3)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

  c = Tensor.custom_kernel(a, b, c, fxn=asm_kernel)[2]
  linear = c.schedule_linear()
@@ -93,7 +93,7 @@ if __name__ == "__main__":
  info = ProgramInfo(name="matmul_kernel",
                     global_size=(M//BLOCK_SIZE_M, N//BLOCK_SIZE_N, 1), local_size=(32*compiled.metadata.num_warps, 1, 1))
  sink = UOp.sink(arg=KernelInfo(name="matmul_kernel"))
-  prg_uop = to_program(UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR), UOp(Ops.SOURCE, arg=src)), arg=info),
+  prg_uop = to_program(UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR), UOp(Ops.SOURCE, arg=src)), arg=info),
                       Device.default.renderer)
  rt = get_runtime(Device.DEFAULT, prg_uop)
  all_bufs = [x.ensure_allocated() for x in bufs]
@@ -27,7 +27,7 @@ class HCQ2Compiled(Compiled):
      (UPat(Ops.BUFFER, tag="timeline_value"), lambda ctx: ctx.timeline_value()),
      (UPat(Ops.BUFFER, tag="sentinel_signal"), lambda ctx: ctx.timeline_signal("sentinel", (1 << 64) - 1)),
      (UPat(Ops.BUFFER, name="b"), lambda ctx, b:
-        Buffer(ctx.device, b.max_numel(), b.dtype, options=BufferSpec(host=False, uncached=True, cpu_access=True, nolru=True))), # TODO: remove nolru
+        Buffer(ctx.device, b.arg, b.dtype, options=BufferSpec(host=False, uncached=True, cpu_access=True, nolru=True))), # TODO: remove nolru
    ])

    super().__init__(device, allocator, compilers, lambda *a, **kw: None, None, arch=arch)
@@ -137,23 +137,20 @@ def unwrap_after(uop):
  return uop

 def make_getaddr(u, device=None):
-  if unwrap_after(u).op not in (Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT, Ops.PARAM): return u
+  if unwrap_after(u).op not in (Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT): return u
  return UOp(Ops.GETADDR, dtypes.uint64, src=(u, UOp(Ops.DEVICE, arg=device or to_tuple(u.device)[0])))

 def make_ins(op, *srcs):
  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)

-def make_patch(buf:UOp, off:sint, val:UOp, dtype=None) -> UOp:
-  dt = dtype or val.dtype
-  return UOp(Ops.SHRINK, buf.dtype.base, (buf, UOp.const(dtypes.int, off), UOp.const(dtypes.int, dt.itemsize))).bitcast(dt).store(val.cast(dt))
-
 def make_cmdbuf(lin, devs, tag):
  blob, patches = b'', []
  for s in (s for ins in lin.src for s in ins.src):
    if s.op is not Ops.CONST: patches.append((len(blob), s))
    blob += struct.pack(f'<{s.dtype.fmt}', s.arg if s.op is Ops.CONST else 0x0)
  buf = UOp.new_buffer(devs, len(blob), dtypes.uint8).rtag(tag)
-  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *[make_patch(buf, off, s) for off, s in patches])
+  stores = [buf.index(UOp.const(dtypes.int, off), dtype=buf.dtype.ptr()).cast(s.dtype.ptr()).store(s) for off, s in patches]
+  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *stores)

 def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))

@@ -183,19 +180,19 @@ class HCQInfo:
  def from_call(call:UOp) -> HCQInfo: return HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), get_call_outs_ins(call)[0])

 # *****************
-# 1.1. prep: staging copies
+# 1.1. prep runtimes: 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 stage_copy(dst:UOp, src:UOp) -> UOp|None:
  if not (_need_staging(src, dst) or _need_staging(dst, src)): return None

-  stage = UOp.new_buffer("CPU", src.nbytes(), dtypes.uint8)
+  stage = UOp.new_buffer("CPU", src.buffer.nbytes, dtypes.uint8)
  return UOp(Ops.LINEAR, dtypes.void, (src.copy_to_device("CPU").call(stage, src), stage.copy_to_device(dst.device).call(dst, stage)))
 pm_insert_copy_staging = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src"))), stage_copy)])

 # *****************
-# 2.1. hcq lowering: programs/kernargs
+# 1.2. prep runtimes: programs/kernargs

@functools.cache
 def get_pm_prep_program(name:str) -> PatternMatcher|None:
@@ -214,15 +211,19 @@ def prep_program(call:UOp, prg:UOp) -> UOp|None:
  return prg.replace(src=(buf.after(buf.store(blob)),), arg=(data, prg.arg)).call(*call.src[1:], aux=HCQInfo.from_call(call))

 def prep_kernargs(call:UOp, prg:UOp) -> UOp:
-  (data, info), dev_uop = prg.arg, UOp(Ops.DEVICE, arg=call.src[1].device)
-  buf = UOp.new_buffer(dev_uop.arg, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
-  patches = [make_patch(buf, i*8, make_getaddr(call.src[1+gi], dev_uop.arg)) for i,gi in enumerate(info.globals)] \
-          + [make_patch(buf, len(info.globals)*8 + i*4, v, dtypes.uint32) for i,v in enumerate(info.vars)]
-  return call.replace(src=(prg.replace(src=prg.src + (buf.after(*patches),), arg=(data, info)),) + call.src[1:])
+  data, info = prg.arg
+  patches = [(i*dtypes.uint64.itemsize, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], UOp(Ops.DEVICE, arg=call.src[1+gi].device))),
+              dtypes.uint64) for i,gi in enumerate(info.globals)] \
+          + [(len(info.globals)*dtypes.uint64.itemsize + i*dtypes.uint32.itemsize, v, dtypes.uint32) for i,v in enumerate(info.vars)]
+
+  buf = UOp.new_buffer(call.src[1].device, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
+  kernargs = buf.after(*tuple(buf.index(UOp.const(dtypes.int, o), dtype=buf.dtype.ptr()).cast(dt.ptr()).store(val.cast(dt)) for o, val, dt in patches))
+
+  return call.replace(src=(prg.replace(src=prg.src + (kernargs,), arg=(data, info)),) + call.src[1:])

 pm_prep_runtime = PatternMatcher([
  # bind generic PROGRAM device to the call's actual dev(s), then run device-specific lowering
-  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"),),
+  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(), UPat(), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"),),
    name="call", allow_any_len=True), prep_program),

  # lower kernargs (PROGRAM.src[0] is now AFTER(BUFFER, COPY) — the lowered program image)
@@ -230,7 +231,7 @@ pm_prep_runtime = PatternMatcher([
 ])

 # *****************
-# 2.2. hcq lowering: ops to ir
+# 2. lowering to hcq ir

 def make_submit(*cmds, devs:str|tuple[str, ...], queue:str) -> UOp:
  devs:tuple[str, ...] = to_tuple(devs)
@@ -306,7 +307,7 @@ def schedule_inner_sync(ctx:DepsCtx, linear:UOp) -> UOp:
    for (_, lane), dep in latest.items(): deps[dep] += (lane,)

    if deps: new_q = new_q.after(*deps, arg=tuple(deps.values())).rtag("deps")
-    new_src.append(call.replace(src=(call.src[0].substitute({q:new_q}),)))
+    new_src.append(call.replace(src=(call.src[0].substitute({q:new_q}), *call.src[1:])))
  return linear.replace(src=tuple(new_src))
 pm_schedule_inner_sync = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), schedule_inner_sync)])

@@ -470,16 +471,15 @@ def pack_hcq_placeholders(call:UOp) -> UOp|None:
  off_per_buf:dict[UOp, int] = {}
  size_per_tag:dict[str, int] = {}
  for b in bufs:
-    bsz = b.max_numel()
-    if b.tag in maxtags: size_per_tag[b.tag] = max(size_per_tag.get(b.tag, 0), bsz)
+    if b.tag in maxtags: size_per_tag[b.tag] = max(size_per_tag.get(b.tag, 0), b.arg)
    elif b.tag in sumtags:
      off_per_buf[b] = round_up(size_per_tag.get(b.tag, 0), {"program": 0x1000}.get(b.tag, 128))
-      size_per_tag[b.tag] = off_per_buf[b] + bsz
+      size_per_tag[b.tag] = off_per_buf[b] + b.arg

  count_per_tag = collections.Counter(b.tag for b in bufs)
  ref_bufs = {b.tag:b for b in bufs if count_per_tag[b.tag] > 1}
-  bases = {tag:UOp.new_buffer(b.device, size_per_tag[tag], b.dtype).rtag(tag) for tag,b in ref_bufs.items()}
-  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.weakint, off_per_buf.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
+  bases = {tag:UOp.new_buffer(b.src[1].arg, size_per_tag[tag], b.dtype).rtag(tag) for tag,b in ref_bufs.items()}
+  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.weakint, off_per_buf.get(b, 0))), b.arg) 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, tag="hcq", name="call"), pack_hcq_placeholders)])

@@ -496,7 +496,7 @@ pm_hold_call_buffers = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"),

 def bufferize_buf(buf:UOp) -> UOp|None:
  if buf.tag is None: return None
-  uops = tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=dv), "CPU") for dev in to_tuple(buf.device))
+  uops = tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=dv), "CPU") for dev in to_tuple(buf.src[1].arg))
  return make_mstack(uops)
 pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, name="buf"), bufferize_buf)])

@@ -532,9 +532,9 @@ pm_resolve_patches = PatternMatcher([
  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
  (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),

-  # shrink on slice is shrink on base at offset
-  (UPat(Ops.SHRINK, src=(UPat(Ops.SLICE, name="bv"), UPat(), UPat()), name="shr"),
-    lambda shr, bv: shr.replace(src=(bv.src[0], shr.src[1] + bv.src[1].cast(shr.src[1].dtype), shr.src[2]))),
+  # index on slice is index
+  (UPat(Ops.INDEX, src=(UPat(Ops.SLICE, name="bv"), UPat()), name="idx", allow_any_len=True),
+    lambda idx, bv: idx.replace(src=(bv.src[0], idx.src[1] + bv.src[1].cast(idx.src[1].dtype), *idx.src[2:]))),

  # getaddr
  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"), UPat(Ops.DEVICE, name="dev"))), resolve_getaddr_slice), # getaddr(slice(x)) -> offset+getaddr(x)
@@ -542,8 +542,8 @@ pm_resolve_patches = PatternMatcher([

  # folders
  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
-  (UPat(Ops.SHRINK, src=(UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf"), UPat.cvar("off"), UPat(Ops.CONST))).bitcast()
-    .store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))), fold_const_store),
+  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").index(UPat.cvar("off")).or_casted().store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))),
+    fold_const_store),
 ]) + symbolic_simple

 # *****************
@@ -570,8 +570,8 @@ def callify_hcq(call:UOp) -> UOp:
  return UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(prg,), arg="hcq").call(*call.src[1:], aux=call.arg.aux)
 pm_callify_hcq = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), callify_hcq)])

-@track_rewrites(lambda _,ret: f"HCQ Compile {pluralize('Kernel', len(ret.src))}")
-def hcq_compile(linear:UOp) -> UOp:
+@track_rewrites(lambda _,ret: f"HCQ Schedule {pluralize('Kernel', len(ret.src))}")
+def hcq_schedule(linear:UOp) -> UOp:
  linear = graph_rewrite(linear, pm_insert_copy_staging + pm_flatten_linear, name="insert copy staging")
  linear = graph_rewrite(linear, pm_prep_runtime, name="prepare runtime")

@@ -587,11 +587,12 @@ def hcq_compile(linear:UOp) -> UOp:
  linear = graph_rewrite(linear, pm_encode_cmdbufs, walk=True, name="encode cmdbufs", enter_calls=True)
  linear = graph_rewrite(linear, pm_lift_patches_to_cmdbuf, name="lift patches to cmdbuf", enter_calls=True)
  linear = graph_rewrite(linear, pm_pack_placeholders, walk=True, name="pack placeholders")
-  return graph_rewrite(linear, pm_hold_call_buffers, walk=True, name="hold call buffers")
+  linear = graph_rewrite(linear, pm_hold_call_buffers, walk=True, name="hold call buffers")

-@track_rewrites(lambda _,ret: f"HCQ Link {pluralize('Kernel', len(ret.src))}")
-def hcq_link(linear:UOp) -> UOp:
+  # realize starts from here
  linear = graph_rewrite(linear, pm_bufferize, bottom_up=True, walk=True, name="bufferize placeholders", enter_calls=True)
  linear = graph_rewrite(linear, pm_resolve_patches, bottom_up=False, name="simplify patches", enter_calls=True)
  linear = graph_rewrite(linear, pm_parametrize_host_buffers, walk=True, name="parametrize host buffers")
-  return graph_rewrite(linear, pm_callify_hcq, name="callify hcq")
+  linear = graph_rewrite(linear, pm_callify_hcq, name="callify hcq")
+
+  return linear
@@ -142,7 +142,7 @@ pm_pm4_opsel = PatternMatcher([
 ])

 def pm4_submit(cmdbuf, devs):
-  size, zero = UOp.const(dtypes.uint32, cmdbuf.nbytes() // dtypes.uint32.itemsize), UOp.const(dtypes.int, 0)
+  size, zero = UOp.const(dtypes.uint32, cmdbuf.src[0].arg // dtypes.uint32.itemsize), UOp.const(dtypes.int, 0)

  # the compute queue's ring and its host-side ring/write/put pointers (placeholders, resolved in pm_bufferize)
  for d in devs: q = Device[d].compute_queue
@@ -205,7 +205,7 @@ pm_sdma_opsel = PatternMatcher([

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

  # the sdma queue's ring and its host-side ring/write/put pointers
  for d in devs: q = Device[d].sdma_queue(0)
@@ -247,8 +247,8 @@ class AMDProgramData:
 _amd_program_cache:dict[tuple[bytes,str], tuple[AMDProgramData,bytes]] = {}

 def amd_build_program(prg:UOp) -> UOp:
-  dev = Device[to_tuple(prg.device)[0]] # TODO: rm this
-  if (cached:=_amd_program_cache.get(key:=(lib:=prg.src[3].arg, dev.device))) is None:
+  dev = Device[prg.src[1].arg] # TODO: rm this
+  if (cached:=_amd_program_cache.get(key:=(lib:=prg.src[4].arg, dev.device))) is None:
    image, sections, relocs = elf_loader(lib)
    rodata = next(sh.header.sh_addr for sh in sections if sh.name == ".rodata")
    for off, sym, typ, addent in relocs:
@@ -271,7 +271,7 @@ def amd_build_program(prg:UOp) -> UOp:
  return cached

 pm_prep_program = PatternMatcher([
-  (UPat(Ops.PROGRAM, src=(UPat(), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"), amd_build_program),
+  (UPat(Ops.PROGRAM, src=(UPat(), UPat(Ops.DEVICE, arg="AMD"), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"), amd_build_program),
 ])

 class AMDAllocator(HCQAllocator['AMDDevice']):
@@ -579,7 +579,7 @@ class AMDDevice(HCQ2Compiled):

    # Scratch setup
    self.max_private_segment_size = 0
-    self.pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, tag="scratch", name="b"), lambda ctx, b: ctx.scratch_buffer(b.max_numel()))]) + self.pm_bufferize
+    self.pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, tag="scratch", name="b"), lambda ctx, b: ctx.scratch_buffer(b.arg))]) + self.pm_bufferize

    self.pmc_enabled:bool = PROFILE > 0 and PMC > 0
    if self.pmc_enabled:
@@ -21,7 +21,7 @@ def _custom_fused_bwd_w13(grad_xw13_fp8:UOp, grad_amax_buf:UOp,
                  xw13.base, grad_x2.base, amax_state.base, grad_amax_state.base, threads, workgroups,
                  arg=KernelInfo(f"fused_silu_mul_bwd_w13_{n_elems}", estimates=Estimates(ops=10*n_elems, mem=mem)))
  src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_bwd_w13.cpp", n_elems, hidden)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))

@functools.cache
@@ -34,7 +34,7 @@ def _custom_fused_cast_amax_w13(fp8_out:UOp, amax_buf:UOp, xw13:UOp, amax_state:
  sink = UOp.sink(fp8_out.base, amax_buf.base, xw13.base, amax_state.base, threads, workgroups,
                  arg=KernelInfo(f"fused_silu_mul_cast_amax_w13_{n_elems}", estimates=Estimates(ops=5*n_elems, mem=mem)))
  src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_fwd_w13.cpp", n_elems, hidden)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))

 def _fused_quantize_bwd_w13(gradient:UOp, kernel:UOp):
@@ -0,0 +1,41 @@
+from __future__ import annotations
+import functools, pathlib
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import THREADS_PER_WG, alloc_like, dname_of, compile_hip
+
+TILE = 64
+
+@functools.cache
+def _custom_fp8_transpose(out:UOp, inp:UOp, dname:str) -> UOp:
+  M, N = inp.shape
+  num_wg = (M // TILE) * (N // TILE)
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(num_wg, "gidx0")
+  mem = M * N * 2  # one byte read + one byte write per element
+  sink = UOp.sink(out.base, inp.base, threads, workgroups,
+                  arg=KernelInfo(f"fp8_transpose_{M}_{N}",
+                                 estimates=Estimates(ops=M*N, mem=mem)))
+  src = (pathlib.Path(__file__).parent/"fp8_transpose.cpp").read_text()
+  defines = [f"-DM_DIM={M}", f"-DN_DIM={N}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def fast_fp8_transpose(t:Tensor) -> Tensor:
+  assert t.ndim == 2, f"fast_fp8_transpose needs 2D input, got shape {t.shape}"
+  assert t.dtype in dtypes.fp8s, f"fast_fp8_transpose needs fp8 dtype, got {t.dtype}"
+  M, N = t.shape
+  assert M % TILE == 0 and N % TILE == 0, f"M={M}, N={N} must be multiples of {TILE}"
+
+  device = t.device
+  axis = t.uop.axis if isinstance(device, tuple) else None
+  out_axis = None
+  if axis == 0: out_axis = 1
+  elif axis == 1: out_axis = 0
+  elif axis is not None:
+    raise ValueError(f"fast_fp8_transpose: unsupported axis {axis}")
+
+  out = alloc_like((N, M), t.dtype, device, out_axis)
+  fxn = functools.partial(_custom_fp8_transpose, dname=dname_of(device))
+  out, _ = Tensor.custom_kernel(out, t, fxn=fxn)
+  return out
@@ -0,0 +1,74 @@
+#include <hip/hip_runtime.h>
+
+// LDS-staged 64x64 fp8 transpose.
+//   in : (M_DIM, N_DIM) fp8 contiguous
+//   out: (N_DIM, M_DIM) fp8 contiguous, out[c][r] = in[r][c]
+//
+// One WG processes one 64x64 output tile.  Each thread reads one uint4 (16 fp8) coalesced
+// from input rows, stages into LDS, then writes one uint4 coalesced to the output (whose
+// 16 fp8 come from 16 different input rows via in-LDS gather).
+//
+// LDS layout: lds[64][LDS_STRIDE] with LDS_STRIDE=65 (1 byte pad) to mitigate bank conflicts
+// during the column-direction read of the write phase.
+
+#ifndef M_DIM
+#define M_DIM 16384
+#endif
+#ifndef N_DIM
+#define N_DIM 28672
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int TILE = 64;
+constexpr int VEC = 16;                    // fp8 per uint4 (128-bit) load/store
+constexpr int LDS_PAD = 1;
+constexpr int LDS_STRIDE = TILE + LDS_PAD; // 65 fp8 per row
+
+static_assert(THREADS_PER_WG * VEC == TILE * TILE, "256 threads * 16 fp8 = 64*64");
+static_assert(M_DIM % TILE == 0, "M_DIM must be a multiple of 64");
+static_assert(N_DIM % TILE == 0, "N_DIM must be a multiple of 64");
+
+constexpr int N_TILES_N = N_DIM / TILE;
+
+struct alignas(16) fp8x16 { uint8_t v[16]; };
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+fp8_transpose(uint8_t* __restrict__ out,         // (N_DIM, M_DIM)
+              const uint8_t* __restrict__ in)    // (M_DIM, N_DIM)
+{
+  __shared__ uint8_t lds[TILE * LDS_STRIDE];
+
+  const int tid     = threadIdx.x;
+  const int wg_id   = blockIdx.x;
+  const int tile_r  = wg_id / N_TILES_N;       // tile index along M dim of input
+  const int tile_c  = wg_id % N_TILES_N;       // tile index along N dim of input
+
+  const int a = tid / (TILE / VEC);            // 0..63 (row within tile during read; col within tile during write)
+  const int b = tid % (TILE / VEC);            // 0..3
+  const int b16 = b * VEC;                     // 0,16,32,48
+
+  // ---- Read phase: input rows -> LDS rows
+  {
+    const long long src = (long long)(tile_r * TILE + a) * (long long)N_DIM
+                        + (long long)(tile_c * TILE + b16);
+    fp8x16 v = *reinterpret_cast<const fp8x16*>(&in[src]);
+    *reinterpret_cast<fp8x16*>(&lds[a * LDS_STRIDE + b16]) = v;
+  }
+  __syncthreads();
+
+  // ---- Write phase: LDS columns (gathered) -> output rows
+  // out[(tile_c*TILE + a)][(tile_r*TILE + b16 + i)] = in[(tile_r*TILE + b16 + i)][(tile_c*TILE + a)]
+  //                                                  = lds[b16 + i][a]
+  {
+    fp8x16 v;
+    #pragma unroll
+    for (int i = 0; i < VEC; ++i) {
+      v.v[i] = lds[(b16 + i) * LDS_STRIDE + a];
+    }
+    const long long dst = (long long)(tile_c * TILE + a) * (long long)M_DIM
+                        + (long long)(tile_r * TILE + b16);
+    *reinterpret_cast<fp8x16*>(&out[dst]) = v;
+  }
+}
@@ -22,7 +22,7 @@ def _custom_fwd(fp8_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_buf:UOp,
  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
             f"-DEPS_LITERAL={eps_val}f"]
  src = _src()
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))

@functools.cache
@@ -39,7 +39,7 @@ def _custom_fwd_add(fp8_out:UOp, h_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax
  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
             f"-DEPS_LITERAL={eps_val}f", f"-DHAS_RESIDUAL=1"]
  src = _src()
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))

@functools.cache
@@ -55,7 +55,7 @@ def _custom_bwd(grad_x:UOp, grad_weight_partial:UOp,
                                 estimates=Estimates(ops=8*n_elems, mem=mem)))
  defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
  src = _src_bwd()
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))

 def _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel:UOp):
@@ -1,104 +0,0 @@
-import functools
-from tinygrad import Tensor, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
-from extra.llama_kernels import FP8_MAX, THREADS_PER_WG, alloc_like
-
-BLK = 32
-PACK = 4
-LOG2E = 1.4426950408889634
-
-@functools.cache
-def _custom_silu_mul_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x_w1:UOp, x_w3:UOp) -> UOp:
-  rows, K = x_w1.shape
-  scale_K = K // BLK
-  n_elems = rows * K
-  n_super = n_elems // (BLK * PACK)
-  sk4 = scale_K // PACK
-  assert n_super % THREADS_PER_WG == 0, f"{n_super=} must divide over {THREADS_PER_WG=}"
-  nwg = n_super // THREADS_PER_WG
-
-  x_w1, x_w3 = x_w1.reshape(n_elems), x_w3.reshape(n_elems)
-  fp8_out = fp8_out.reshape(n_elems)
-  e8_out = e8_out.reshape(rows * scale_K)
-  si_out = si_out.reshape(sk4 * rows)
-
-  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
-  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
-  sb = UOp.range(PACK, 2, AxisType.UNROLL)
-  lane = UOp.range(BLK, 3, AxisType.UNROLL)
-
-  super_idx = wg * THREADS_PER_WG + tid
-  idx = super_idx * (BLK * PACK) + sb * BLK + lane
-
-  w1 = x_w1[idx].cast(dtypes.float)
-  w3 = x_w3[idx].cast(dtypes.float)
-  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
-  act = w1 * sig * w3
-  abs_a = (act < 0.0).where(-act, act)
-  blk_max = abs_a.reduce(lane, arg=Ops.MAX)
-  e8f = (blk_max.maximum(1e-38).log2().floor() + 127.0).maximum(0.0).minimum(254.0)
-  qscale = (127.0 - e8f).exp2()
-  scaled = (act * qscale).maximum(-FP8_MAX).minimum(FP8_MAX)
-  e8u8 = e8f.cast(dtypes.uint8)
-
-  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
-  e8_store = e8_out.after(fp8_store)[super_idx * PACK + sb].store(e8u8)
-  packed = (e8u8.cast(dtypes.uint32) << (sb.cast(dtypes.uint32) * 8)).reduce(sb, arg=Ops.ADD)
-  row, col4 = super_idx // sk4, super_idx % sk4
-  si_store = si_out.after(e8_store.end(sb))[col4 * rows + row].store(packed)
-  return si_store.end(tid, wg).sink(arg=KernelInfo(f"silu_mul_quantize_mxfp8_{n_elems}", opts_to_apply=()))
-
-@functools.cache
-def _custom_silu_mul_bwd_mxfp8(gx1_out:UOp, gx3_out:UOp, x_w1:UOp, x_w3:UOp, grad_aq:UOp, e8:UOp) -> UOp:
-  rows, K = x_w1.shape
-  scale_K = K // BLK
-  n_elems = rows * K
-  VEC = 8
-  assert n_elems % (THREADS_PER_WG * VEC) == 0, f"{n_elems=} must divide {THREADS_PER_WG*VEC=}"
-  nwg = n_elems // (THREADS_PER_WG * VEC)
-  x_w1, x_w3, grad_aq = x_w1.reshape(n_elems), x_w3.reshape(n_elems), grad_aq.reshape(n_elems)
-  gx1_out, gx3_out, e8 = gx1_out.reshape(n_elems), gx3_out.reshape(n_elems), e8.reshape(rows * scale_K)
-
-  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
-  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
-  lane = UOp.range(VEC, 2, AxisType.UNROLL)
-  idx = (wg * THREADS_PER_WG + tid) * VEC + lane
-
-  e8v = e8[idx // BLK].cast(dtypes.float)
-  qscale = (127.0 - e8v).exp2()
-  ga = grad_aq[idx].cast(dtypes.float) * qscale
-  w1 = x_w1[idx].cast(dtypes.float)
-  w3 = x_w3[idx].cast(dtypes.float)
-  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
-  s = w1 * sig
-  sprime = sig * (1.0 + w1 * (1.0 - sig))
-  gx1 = gx1_out[idx].store((ga * sprime * w3).cast(gx1_out.dtype.base))
-  gx3 = gx3_out.after(gx1)[idx].store((ga * s).cast(gx3_out.dtype.base))
-  return gx3.end(lane, tid, wg).sink(arg=KernelInfo(f"silu_mul_bwd_mxfp8_{n_elems}", opts_to_apply=()))
-
-def _silu_mul_quantize_mxfp8_bwd(gradient:UOp, kernel:UOp):
-  _, e8_out, _, x_w1, x_w3 = kernel.src[1:]
-  device = x_w1.device
-  rows, K = x_w1.shape
-  axis = x_w1.axis if isinstance(device, tuple) else None
-  gx1 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
-  gx3 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
-  gx1, gx3, *_ = Tensor.custom_kernel(gx1, gx3, Tensor(x_w1, device=device), Tensor(x_w3, device=device),
-                                      Tensor(gradient, device=device).cast(dtypes.bfloat16), Tensor(e8_out.after(kernel), device=device),
-                                      fxn=_custom_silu_mul_bwd_mxfp8)
-  return (None, None, None, gx1.uop, gx3.uop)
-
-def fused_silu_mul_quantize_mxfp8(x_w1:Tensor, x_w3:Tensor) -> tuple[Tensor, Tensor, Tensor]:
-  assert x_w1.shape == x_w3.shape, f"{x_w1.shape} != {x_w3.shape}"
-  assert x_w1.dtype == dtypes.bfloat16 and x_w3.dtype == dtypes.bfloat16
-  assert x_w1.ndim == 2, f"expected 2d, got {x_w1.shape}"
-  from extra.gemm.cdna_asm_gemm import FP8_DTYPE
-  rows, K = x_w1.shape
-  scale_K = K // BLK
-  axis = x_w1.uop.axis if isinstance(x_w1.device, tuple) else None
-  fp8_out = alloc_like((rows, K), FP8_DTYPE, x_w1.device, axis)
-  e8_out = alloc_like((rows, scale_K), dtypes.uint8, x_w1.device, axis)
-  si_out = alloc_like((scale_K // PACK, rows), dtypes.uint32, x_w1.device, None if axis is None else (1 if axis == 0 else 0))
-  fp8_out, e8_out, si_out, *_ = Tensor.custom_kernel(fp8_out, e8_out, si_out, x_w1, x_w3,
-                                                     fxn=_custom_silu_mul_quantize_mxfp8, grad_fxn=_silu_mul_quantize_mxfp8_bwd)
-  return fp8_out, e8_out, si_out
@@ -42,8 +42,8 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_st
  step = THREADS_PER_WG // 2
  while step:
    active = tid < step
-    other = lds[(tid + step).valid(active)].load()
-    lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)).barrier())
+    other = lds[tid + step].load(UOp.const(dtypes.float, 0.0), active)
+    lds = lds.after(lds[tid].store(lds[tid].maximum(other), gate=active).barrier())
    step //= 2

  amax_store = amax_partial[tid.eq(0).where(wg, UOp.invalid())].store(lds[0])
@@ -1,37 +0,0 @@
-from __future__ import annotations
-import functools, pathlib
-from tinygrad import Tensor, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.renderer import Estimates
-from extra.llama_kernels import THREADS_PER_WG, alloc_like, dname_of, compile_hip
-
-TILE_N = THREADS_PER_WG   # 256
-BLK = 32
-
-@functools.cache
-def _custom_transpose_quantize_mxfp8(q:UOp, e8:UOp, g:UOp, dname:str) -> UOp:
-  M, N = g.shape
-  num_wg = (M // BLK) * (N // TILE_N)
-  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(num_wg, "gidx0")
-  mem = M * N * 2 + M * N + (M // BLK) * N   # read bf16, write fp8 + e8
-  sink = UOp.sink(q.base, e8.base, g.base, threads, workgroups,
-                  arg=KernelInfo(f"transpose_quantize_mxfp8_{M}_{N}", estimates=Estimates(ops=M*N, mem=mem)))
-  src = (pathlib.Path(__file__).parent/"transpose_quantize_mxfp8.cpp").read_text()
-  defines = [f"-DM_DIM={M}", f"-DN_DIM={N}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
-                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
-
-def transpose_quantize_mxfp8(g:Tensor) -> tuple[Tensor, Tensor, Tensor]:
-  # fused g.T quantize: returns (q, e8, si) == quantize_mxfp8(g.T) — q (N,M) fp8, e8 (N, M/32), si packed (M/128, N)
-  assert g.ndim == 2 and g.dtype == dtypes.bfloat16, f"{g.shape} {g.dtype}"
-  from extra.gemm.cdna_asm_gemm import FP8_DTYPE, mx_pack
-  M, N = g.shape
-  assert M % BLK == 0 and N % TILE_N == 0, f"M={M} must%{BLK}, N={N} must%{TILE_N}"
-  device = g.device
-  axis = g.uop.axis if isinstance(device, tuple) else None
-  out_axis = None if axis is None else (1 if axis == 0 else 0)
-  q = alloc_like((N, M), FP8_DTYPE, device, out_axis)
-  e8 = alloc_like((N, M // BLK), dtypes.uint8, device, out_axis)
-  fxn = functools.partial(_custom_transpose_quantize_mxfp8, dname=dname_of(device))
-  q, e8, *_ = Tensor.custom_kernel(q, e8, g, fxn=fxn)
-  return q, e8, mx_pack(e8)
@@ -1,62 +0,0 @@
-#include <hip/hip_runtime.h>
-#include <hip/hip_fp8.h>
-#include <hip/hip_bf16.h>
-
-#ifndef M_DIM
-#define M_DIM 8192
-#endif
-#ifndef N_DIM
-#define N_DIM 14336
-#endif
-#ifndef THREADS_PER_WG
-#define THREADS_PER_WG 256
-#endif
-
-constexpr int BLK = 32;
-constexpr int TILE_M = BLK;                 // one mxfp8 block along M per tile
-constexpr int TILE_N = THREADS_PER_WG;      // 256, one output column per thread
-constexpr int LDS_STRIDE = TILE_N + 1;      // +1 pad: stride 257 ≡ 1 (mod 32) -> conflict-free column reads
-constexpr int N_TILES_N = N_DIM / TILE_N;
-constexpr float FP8_MAX = 448.0f;
-
-static_assert(M_DIM % TILE_M == 0, "M_DIM must be a multiple of 32");
-static_assert(N_DIM % TILE_N == 0, "N_DIM must be a multiple of 256");
-
-extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
-transpose_quantize_mxfp8(__hip_fp8_storage_t* __restrict__ q,        // (N_DIM, M_DIM)
-                         uint8_t* __restrict__ e8_out,               // (N_DIM, M_DIM/32)
-                         const __hip_bfloat16* __restrict__ g)       // (M_DIM, N_DIM)
-{
-  __shared__ __hip_bfloat16 lds[TILE_M * LDS_STRIDE];
-  const int tid    = threadIdx.x;
-  const int tile_m = blockIdx.x / N_TILES_N;     // which 32-block along M
-  const int tile_n = blockIdx.x % N_TILES_N;
-
-  #pragma unroll
-  for (int mm = 0; mm < TILE_M; mm++)
-    lds[mm * LDS_STRIDE + tid] = g[(long long)(tile_m * TILE_M + mm) * N_DIM + (tile_n * TILE_N + tid)];
-  __syncthreads();
-
-  float vals[TILE_M];
-  float amax = 0.0f;
-  #pragma unroll
-  for (int mm = 0; mm < TILE_M; mm++) {
-    float v = (float)lds[mm * LDS_STRIDE + tid];
-    vals[mm] = v;
-    amax = fmaxf(amax, fabsf(v));
-  }
-  int e8 = (int)floorf(log2f(fmaxf(amax, 1e-38f))) + 127;
-  e8 = max(0, min(254, e8));
-  float qscale = exp2f((float)(127 - e8));
-
-  const long long n = tile_n * TILE_N + tid;
-  __hip_fp8_storage_t out[TILE_M];
-  #pragma unroll
-  for (int mm = 0; mm < TILE_M; mm++)
-    out[mm] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, vals[mm] * qscale)), __HIP_SATFINITE, __HIP_E4M3);
-  // 32 contiguous fp8 along M -> two 16-byte vector stores
-  long long obase = n * M_DIM + (long long)(tile_m * TILE_M);
-  *reinterpret_cast<uint4*>(&q[obase])      = *reinterpret_cast<uint4*>(&out[0]);
-  *reinterpret_cast<uint4*>(&q[obase + 16]) = *reinterpret_cast<uint4*>(&out[16]);
-  e8_out[n * (M_DIM / BLK) + tile_m] = (uint8_t)e8;
-}
@@ -37,7 +37,7 @@ def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, **kwargs)
    gidx = UOp.special(NUM_WORKGROUPS, "gidx0")
    FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
    sink = UOp.sink(A.base, threads, gidx, arg=KernelInfo(inst.op.name.lower(), estimates=Estimates(ops=FLOPs, mem=0)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
  dummy = Tensor.zeros(1).contiguous().realize()
  out = Tensor.custom_kernel(dummy, fxn=fxn)[0]
  linear = out.schedule_linear()
@@ -1,6 +1,6 @@
 from tinygrad.tensor import Tensor
 from tinygrad.nn import Conv2d, LayerNorm, LayerNorm2d, Linear
-from tinygrad.helpers import fetch, get_child, Context
+from tinygrad.helpers import fetch, get_child

 class Block:
  def __init__(self, dim):
@@ -58,6 +58,7 @@ if __name__ == "__main__":
  from test.models.test_efficientnet import chicken_img, preprocess, _LABELS
  img = Tensor(preprocess(chicken_img))

-  with Context(TRAINING=0):
-    out = model(img).numpy()
+  Tensor.training = False
+
+  out = model(img).numpy()
  print(_LABELS[out.argmax()])
@@ -5,7 +5,7 @@ import numpy as np
 from pathlib import Path
 from tinygrad import nn, Tensor, dtypes
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.helpers import get_child, fetch, TRAINING
+from tinygrad.helpers import get_child, fetch
 from tinygrad.nn.state import torch_load
 from examples.mlperf.helpers import BoxCoder
 from extra.models.resnet import ResNet
@@ -1069,7 +1069,7 @@ class RoIBoxHead:
  def __call__(self, features, proposals, targets=None):
    x = self.feature_extractor(features, proposals)
    class_logits, box_regression = self.predictor(x)
-    if not TRAINING:
+    if not Tensor.training:
      result = self.post_processor((class_logits, box_regression), proposals)
      return x, result, {}

@@ -1111,7 +1111,7 @@ class Mask:
    x = self.feature_extractor(features, proposals)
    if x is not None:
      mask_logits = self.predictor(x)
-      if not TRAINING:
+      if not Tensor.training:
        result = self.post_processor(mask_logits, proposals)
        return x, result, {}
    return x, [], {}
@@ -1,6 +1,6 @@
 import math
 from tinygrad import Tensor, dtypes
-from tinygrad.helpers import flatten, get_child, TRAINING
+from tinygrad.helpers import flatten, get_child
 from examples.mlperf.helpers import generate_anchors, BoxCoder
 from examples.mlperf.losses import sigmoid_focal_loss, l1_loss
 from extra.models.resnet import ResNet
@@ -141,7 +141,7 @@ class ClassificationHead:
    out = [self.cls_logits(feat.sequential(self.conv)).permute(0, 2, 3, 1).reshape(feat.shape[0], -1, self.num_classes) for feat in x]
    out = out[0].cat(*out[1:], dim=1)

-    if TRAINING:
+    if Tensor.training:
      assert labels is not None and matches is not None, "labels and matches should be passed in when training"
      return self._compute_loss(out.cast(dtypes.float32), labels, matches)

@@ -167,7 +167,7 @@ class RegressionHead:
    out = [self.bbox_reg(feat.sequential(self.conv)).permute(0, 2, 3, 1).reshape(feat.shape[0], -1, 4) for feat in x]
    out = out[0].cat(*out[1:], dim=1)

-    if TRAINING:
+    if Tensor.training:
      assert bboxes is not None and matches is not None and anchors is not None, "bboxes, matches, and anchors should be passed in when training"
      return self._compute_loss(out, bboxes, matches, anchors)

@@ -187,7 +187,7 @@ class RetinaHead:
    self.regression_head = RegressionHead(in_channels, num_anchors)

  def __call__(self, x:Tensor, **kwargs) -> Tensor|dict[str, Tensor]:
-    if TRAINING:
+    if Tensor.training:
      return {
        "classification_loss": self.classification_head(x, labels=kwargs["labels"], matches=kwargs["matches"]),
        "regression_loss": self.regression_head(x, bboxes=kwargs["bboxes"], matches=kwargs["matches"], anchors=kwargs["anchors"])
@@ -111,7 +111,7 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:st
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
 def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
@@ -141,7 +141,7 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             src=(sink, UOp(Ops.DEVICE, arg=device), 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):
@@ -171,7 +171,7 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
 def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
@@ -201,4 +201,4 @@ def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int,
  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
@@ -1,14 +1,10 @@
 import unittest
 import torch
 import tinygrad.nn.torch
+torch.set_default_device("tiny")
 import numpy as np

 class TestTorchBackendInplace(unittest.TestCase):
-  def setUp(self):
-    self._prev_device = torch.get_default_device()
-    torch.set_default_device("tiny")
-  def tearDown(self): torch.set_default_device(self._prev_device)
-
  def test_zero(self):
    a = torch.ones(4)
    a.zero_()
@@ -2,15 +2,11 @@ import unittest
 from tinygrad.helpers import getenv
 import torch
 import tinygrad.nn.torch
+torch.set_default_device("tiny")
 import numpy as np

@unittest.skipIf(getenv("GPUS",1)<=1, "only single GPU")
 class TestTorchBackendMultiGPU(unittest.TestCase):
-  def setUp(self):
-    self._prev_device = torch.get_default_device()
-    torch.set_default_device("tiny")
-  def tearDown(self): torch.set_default_device(self._prev_device)
-
  def test_transfer(self):
    a = torch.Tensor([[1,2],[3,4]]).to("tiny:0")
    b = torch.Tensor([[3,2],[1,0]]).to("tiny:1")
@@ -1,6 +1,6 @@
 import numpy as np
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import trange, Context
+from tinygrad.helpers import trange
 from tinygrad.engine.jit import TinyJit


@@ -22,7 +22,7 @@ def train(model, X_train, Y_train, optim, steps, BS=128, lossfn=lambda out,y: ou

  if allow_jit: train_step = TinyJit(train_step)

-  with Context(TRAINING=1):
+  with Tensor.train():
    losses, accuracies = [], []
    for i in (t := trange(steps, disable=None)):
      samp = np.random.randint(0, X_train.shape[0], size=(BS))
@@ -40,7 +40,7 @@ def train(model, X_train, Y_train, optim, steps, BS=128, lossfn=lambda out,y: ou

 def evaluate(model, X_test, Y_test, num_classes=None, BS=128, return_predict=False, transform=lambda x: x,
             target_transform=lambda y: y):
-  @Context(TRAINING=0)
+  Tensor.training = False
  def numpy_eval(Y_test, num_classes):
    Y_test_preds_out = np.zeros(list(Y_test.shape)+[num_classes])
    for i in trange((len(Y_test)-1)//BS+1, disable=None):
@@ -55,3 +55,4 @@ def evaluate(model, X_test, Y_test, num_classes=None, BS=128, return_predict=Fal
  acc, Y_test_pred = numpy_eval(Y_test, num_classes)
  print("test set accuracy is %f" % acc)
  return (acc, Y_test_pred) if return_predict else acc
+
@@ -20,7 +20,6 @@ include-package-data = true
 packages = [
  'tinygrad',
  'tinygrad.codegen',
-  'tinygrad.codegen.decomp',
  'tinygrad.codegen.opt',
  'tinygrad.codegen.late',
  'tinygrad.engine',
@@ -30,7 +30,7 @@ def custom_add_one(A:UOp) -> UOp:
    s_endpgm(),
  ]
  sink = UOp.sink(A.base, threads, arg=KernelInfo(f"custom_add_one_{A.numel()}", estimates=Estimates(ops=A.numel(), mem=A.numel()*4*2)))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

 def custom_add_var(A:UOp, B:UOp) -> UOp:
  A,B = A.flatten(), B.flatten()
@@ -49,7 +49,7 @@ def custom_add_var(A:UOp, B:UOp) -> UOp:
    s_endpgm(),
  ]
  sink = UOp.sink(A.base, B.base, var, threads, arg=KernelInfo(f"custom_add_var_{A.numel()}"))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

 def custom_wave_sync(A:UOp, arch:str) -> UOp:
  # 4 waves across 1024 WG — enough to saturate a SIMD with many concurrent WGs
@@ -63,7 +63,7 @@ def custom_wave_sync(A:UOp, arch:str) -> UOp:
    insts += [s_nop(0)]*4
  insts.append(s_endpgm())
  sink = UOp.sink(A.base, threads, wg, arg=KernelInfo("custom_wave_sync"))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

 def custom_lds_sync(A:UOp, arch:str) -> UOp:
  A = A.flatten()
@@ -97,7 +97,7 @@ def custom_lds_sync(A:UOp, arch:str) -> UOp:
    isa.s_endpgm(),
  ]
  sink = UOp.sink(A.base, lds, threads, wg, arg=KernelInfo("custom_lds_sync"))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

 def custom_handwritten(A:UOp) -> UOp:
  A = A.flatten()
@@ -143,7 +143,7 @@ def custom_handwritten(A:UOp) -> UOp:
  k.emit(r4.s_endpgm())
  insts = k.finalize()
  sink = UOp.sink(A.base, threads, wg, lds, arg=KernelInfo("custom_handwritten"))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

 def custom_data_deps(A:UOp) -> UOp:
  A = A.flatten()
@@ -159,7 +159,7 @@ def custom_data_deps(A:UOp) -> UOp:
  k.emit(s_endpgm())
  insts = k.finalize()
  sink = UOp.sink(A.base, threads, arg=KernelInfo("custom_data_deps"))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

@unittest.skipUnless(Device.DEFAULT == "AMD", "requires AMD device")
 class TestAsmKernel(unittest.TestCase):
@@ -89,7 +89,7 @@ class TestTinygradIntegration(unittest.TestCase):
    assert len(sink_items) > 0, "No SINK in schedule"
    renderer = AMDLLVMRenderer(Target("AMD", arch='gfx1100'))
    prg = to_program(sink_items[0].src[0], renderer)
-    lib = renderer.compiler.compile(prg.src[2].arg)
+    lib = renderer.compiler.compile(prg.src[3].arg)
    return next(s.content for s in elf_loader(lib)[1] if s.name == ".text")

  def test_simple_add_kernel(self):
@@ -15,7 +15,7 @@ dev = Device["AMD"]
 a = Tensor([1.0]).realize()
 b = a + 1
 linear = compile_linear(b.schedule_linear())
-lib = bytearray(linear.src[-1].src[0].src[3].arg)
+lib = bytearray(linear.src[-1].src[0].src[4].arg)

 # Find s_endpgm (0xBFB00000) and replace with V_MOVRELD_B32 (op=66) which has no pcode
 # VOP1 encoding: bits[31:25]=0x7E, op=bits[16:9], so op=66 -> 66<<9 = 0x8400
@@ -83,8 +83,8 @@ def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelSnapshot], dict[int, in
            buf_data[dst_id] = src_data
      elif ast.op is Ops.PROGRAM:
        info = ast.arg
-        if len(ast.src) > 3 and ast.src[3].op is Ops.BINARY:
-          lib = bytes(ast.src[3].arg)
+        if len(ast.src) > 4 and ast.src[4].op is Ops.BINARY:
+          lib = bytes(ast.src[4].arg)
          _, sections, _ = elf_loader(lib)
          for sec in sections:
            if sec.name == '.text':
@@ -98,7 +98,7 @@ def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelSnapshot], dict[int, in
                buf_sizes.append(b.nbytes)
              kernels.append(KernelSnapshot(
                code=bytes(sec.content),
-                src=ast.src[2].arg,
+                src=ast.src[3].arg,
                global_size=tuple(info.global_size),
                local_size=tuple(info.local_size),
                buf_idxs=buf_idxs,
@@ -430,7 +430,7 @@ class TestCustomKernel(unittest.TestCase):
    binary = Device[a.device].renderer.compiler.compile(src)
    def custom_src_kernel(A:UOp) -> UOp:
      sink = UOp.sink(A, arg=KernelInfo(name="test_src"))
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(sink.toposort())),
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="CPU"), UOp(Ops.LINEAR, src=tuple(sink.toposort())),
                                   UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=binary)))

    a = Tensor.custom_kernel(a, fxn=custom_src_kernel)[0]
@@ -25,7 +25,7 @@
 import unittest
 import numpy as np
 import torch
-from tinygrad import Tensor, dtypes, nn, Context
+from tinygrad import Tensor, dtypes, nn
 from tinygrad.device import Device
 from tinygrad.helpers import DEV
 from tinygrad.renderer.nir import NIRRenderer
@@ -101,7 +101,7 @@ class TestDropoutProbabilityEdgeCases(unittest.TestCase):
  # we don't need more of these

  def test_dropout_rate_one(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      out = Tensor.ones(100).dropout(1.0)
      np.testing.assert_allclose(out.numpy(), np.zeros(100))

@@ -109,7 +109,7 @@ class TestDropoutProbabilityEdgeCases(unittest.TestCase):
    with self.assertRaises(ValueError):
      torch.nn.functional.dropout(torch.ones(10), -0.1, True)
    with self.assertRaises(ValueError):
-      with Context(TRAINING=1):
+      with Tensor.train():
        Tensor.ones(10).dropout(-0.1)

 class TestInputValidation(unittest.TestCase):
@@ -47,7 +47,7 @@ class TestLinearizer(unittest.TestCase):
    tst = Tensor.ones(16, dtype=dtypes.int).contiguous().realize()
    out = tst.neg().cast(dtypes.char).cast(dtypes.int).cast(dtypes.char) * 2
    ast = helper_linearizer_opt(out)
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    self.assertEqual(len([x for x in uops if x.op is Ops.CAST]), 1)

  @unittest.expectedFailure
@@ -55,7 +55,7 @@ class TestLinearizer(unittest.TestCase):
    tst = Tensor.ones(16, dtype=dtypes.int).contiguous().realize()
    out = tst.neg().cast(dtypes.char).cast(dtypes.int) * 2
    ast = helper_linearizer_opt(out)
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    self.assertEqual(len([x for x in uops if x.op is Ops.CAST]), 0)

  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "broken on ptx")
@@ -65,7 +65,7 @@ class TestLinearizer(unittest.TestCase):
    b = Tensor.empty(16)
    out = img.conv2d(w, b)
    ast = helper_linearizer_opt(out)
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    # slice at the last loop end
    uslice = [i for i,u in enumerate(uops) if u.op == Ops.END][-1]
    # only valid test if outermost range is the reduce
@@ -86,7 +86,7 @@ class TestLinearizer(unittest.TestCase):
    a = Tensor.randn(2, ).realize()
    out = a.reshape(2, 1).expand(2, 3).sum()
    ast = helper_linearizer_opt(out, wanna_output=[np.broadcast_to(a.numpy().reshape(2, 1), (2, 3)).sum()])
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
    assert len(ranges) == 1 # NOTE: it collapses now

@@ -94,7 +94,7 @@ class TestLinearizer(unittest.TestCase):
    a = Tensor.randn(2, ).realize()
    out = a.reshape(2, 1).expand(2, 3).expand(2, 2, 3).sum()
    ast = helper_linearizer_opt(out, wanna_output=[np.broadcast_to(np.broadcast_to(a.numpy().reshape(2, 1), (2, 3)), (2, 2, 3)).sum()])
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
    assert len(ranges) == 1 # NOTE: it collapses now

@@ -102,7 +102,7 @@ class TestLinearizer(unittest.TestCase):
    a = Tensor([2, 2]).realize()
    out = a.reshape(2, 1).pad(((1, 1), (1, 1)), value=2).sum()
    ast = helper_linearizer_opt(out, wanna_output=[24])
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
    # RANGE -> ALU -> RANGE -> ALU + LOAD -> STORE
    assert any(x.op in GroupOp.ALU for x in uops[ranges[0]:ranges[1]])
@@ -115,7 +115,7 @@ class TestLinearizer(unittest.TestCase):
    b = Tensor.randn(1, 1).realize()
    out = (a + b[0]).sum() + b[0]
    ast = helper_linearizer_opt(out, wanna_output=[(a.numpy()+b.numpy()[0]).sum()+b.numpy()])
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
    # LOAD -> RANGE -> LOAD -> STORE
    assert len([x for x in uops[:ranges[0]] if x.op is Ops.LOAD]) == 1
@@ -125,7 +125,7 @@ class TestLinearizer(unittest.TestCase):
    b = Tensor.randn(1, 1).realize()
    out = (a.reshape(2, 1).expand(2, 3) + b[0]).sum() + b[0]
    ast = helper_linearizer_opt(out, wanna_output=[(np.broadcast_to(a.numpy().reshape(2, 1), (2, 3)) + b.numpy()[0]).sum() + b.numpy()])
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    ranges = [i for i,u in enumerate(uops) if u.op is Ops.RANGE]
    assert len(ranges) == 1 # NOTE: it collapses now

@@ -137,10 +137,10 @@ class TestLinearizer(unittest.TestCase):
    r = a[:-1] + a[1:]

    uops = tuple(to_program(replace_opts(r.schedule_linear().src[-1].src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=0)]),
-                       renderer=Device[Device.DEFAULT].renderer).src[1].src)
+                       renderer=Device[Device.DEFAULT].renderer).src[2].src)
    num_loads = len([uop for uop in uops if uop.op is Ops.LOAD])
    assert num_loads <= 4, "more load uops than needed"
-    assert num_loads >= 1, "expected at least one load uop"
+    assert num_loads >= 4, "unexpected number of uops, maybe this test needs updating?"

  @unittest.skip("this is handled at higher level now")
  def test_upcast_cse(self):
@@ -150,7 +150,7 @@ class TestLinearizer(unittest.TestCase):
    r = a.expand([2]) + b.expand([2])

    uops = tuple(to_program(replace_opts(r.schedule_linear().src[-1].src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=0)]),
-                       renderer=Device[Device.DEFAULT].renderer).src[1].src)
+                       renderer=Device[Device.DEFAULT].renderer).src[2].src)
    num_ops = len([uop for uop in uops if uop.op in GroupOp.ALU])
    assert num_ops <= 1, "more alu uops than needed"

@@ -160,7 +160,7 @@ class TestLinearizer(unittest.TestCase):
    r = Tensor.conv2d(x,w,padding=1).relu()

    uops = tuple(to_program(replace_opts(r.schedule_linear().src[-1].src[0],
-      [Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=0)]), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+      [Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=0)]), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    accs = [u for u in uops if u.op is Ops.BUFFER and u.addrspace is AddrSpace.REG]
    stores = [u for u in uops if u.op is Ops.STORE]
    assert len(accs) == 0  # it's removed now
@@ -173,7 +173,7 @@ class TestLinearizer(unittest.TestCase):
    out = Tensor.ones(64,64).contiguous() @ Tensor.ones(64,64).contiguous()
    prg = to_program(replace_opts(out.schedule_linear().src[-1].src[0], [Opt(OptOps.LOCAL, axis=0, arg=4)]),
                      renderer=Device[Device.DEFAULT].renderer)
-    self.assertEqual(len(prg.src[2].arg.split("for")), 5)
+    self.assertEqual(len(prg.src[3].arg.split("for")), 5)

  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
@@ -185,7 +185,7 @@ class TestLinearizer(unittest.TestCase):
    opts_to_apply = [Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4)]
    program = to_program(replace_opts(r.schedule_linear().src[-1].src[0], opts_to_apply), renderer=Device[Device.DEFAULT].renderer)

-    stores = [u for u in tuple(program.src[1].src) if u.op is Ops.STORE and u.src[0].addrspace != AddrSpace.REG]
+    stores = [u for u in tuple(program.src[2].src) if u.op is Ops.STORE and u.src[0].addrspace != AddrSpace.REG]

    # the first store is to lds and can be upcasted
    assert stores[0].src[1].max_numel() == 4
@@ -199,7 +199,7 @@ class TestLinearizer(unittest.TestCase):
    a, b = Tensor.randn(1).realize(), Tensor.randn(1).realize()
    r = Tensor.stack(a, b)
    uops = tuple(to_program(replace_opts(r.schedule_linear().src[-1].src[0], [Opt(op=OptOps.UPCAST, axis=0, arg=0)]),
-                       renderer=Device[Device.DEFAULT].renderer).src[1].src)
+                       renderer=Device[Device.DEFAULT].renderer).src[2].src)
    num_ops = len([uop for uop in uops if uop.op in GroupOp.ALU])
    assert num_ops == 0, "more alu uops than needed"

@@ -210,14 +210,14 @@ class TestLinearizer(unittest.TestCase):
        a = Tensor([1, 2, 3], dtype=tensor_dtype).sum()
        realized_ast = a.schedule_linear().src[-1].src[0]
        program = to_program(replace_opts(realized_ast, []), renderer=Device[Device.DEFAULT].renderer)
-        local = [uop for uop in tuple(program.src[1].src) if uop.op is Ops.BUFFER and uop.addrspace in (AddrSpace.LOCAL, AddrSpace.REG)]
+        local = [uop for uop in tuple(program.src[2].src) if uop.op is Ops.BUFFER and uop.addrspace in (AddrSpace.LOCAL, AddrSpace.REG)]
        assert local[0].dtype.base == acc_dtype

  def test_arg_acc_dtype(self):
    def helper_arg_acc_dtype(c: Tensor, expected_dtype:DType):
      realized_ast = c.schedule_linear().src[-1].src[0]
      program = to_program(replace_opts(realized_ast, []), renderer=Device[Device.DEFAULT].renderer)
-      local = [uop for uop in tuple(program.src[1].src) if uop.op is Ops.BUFFER and uop.addrspace in (AddrSpace.LOCAL, AddrSpace.REG)]
+      local = [uop for uop in tuple(program.src[2].src) if uop.op is Ops.BUFFER and uop.addrspace in (AddrSpace.LOCAL, AddrSpace.REG)]
      self.assertEqual(local[0].dtype.base, expected_dtype)

    tests = (
@@ -244,7 +244,7 @@ class TestLinearizer(unittest.TestCase):
    opt = [Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4)]
    ast = helper_linearizer_opt(r, [opt])
    # the uops graph is reg BUFFER -> 4x STORE 0.0 -> RANGE -> 4x ALU -> 4x STORE -> ENDRANGE
-    uops = tuple(to_program(replace_opts(ast, opt), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, opt), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    begin_range = [i for i, x in enumerate(uops) if x.op is Ops.RANGE][-1]
    end_range = [i for i, x in enumerate(uops) if x.op is Ops.END][0]
    for i,u in enumerate(uops): print(i, u.op, [uops.index(s) for s in u.src], u.arg, u.dtype)
@@ -264,7 +264,7 @@ class TestLinearizer(unittest.TestCase):
    # shrink so that the dims do not collapse
    t = Tensor.ones(5, 6, 7).contiguous().realize().shrink(((0, 4), (0, 5), (0, 6)))
    ast = helper_linearizer_opt(t+1)
-    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[2].src)
    idxs = dedup([uop for uop in uops if uop.op is Ops.SPECIAL])
    idxs = sorted(idxs, key=lambda uop: uop.arg)
    assert (idxs[0].arg, idxs[0].src[0].arg) == ('gidx0', 6), idxs[0]
@@ -297,12 +297,12 @@ class TestLinearizer(unittest.TestCase):
    run_linear(linear, var_vals)
    np.testing.assert_equal(a.flatten().numpy(), [1.,1.,1.,1.,2.,2.,2.,2.,1.,1.,1.,1.,1.,1.,1.,1.])
    program = to_program(replace_opts(linear.src[-1].src[0], []), renderer=Device[Device.DEFAULT].renderer)
-    assert not any(u.op == Ops.WHERE for u in tuple(program.src[1].src)), "found where where where should be folded"
+    assert not any(u.op == Ops.WHERE for u in tuple(program.src[2].src)), "found where where where should be folded"

  def test_phi_simplification(self):
    def helper(t, max_ops=0):
      ast = helper_linearizer_opt(t)
-      uops = tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[1].src)
+      uops = tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[2].src)
      # ignore kernel optimized IF statements for now
      if if_op:=next((u for u in uops if u.op is Ops.IF), None):
        uops = uops[:uops.index(if_op)]
@@ -335,7 +335,7 @@ class TestLinearizer(unittest.TestCase):
    out = x.matmul(y)
    with Context(TC=0):
      ast = helper_linearizer_opt(out)
-      uops = tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[1].src)
+      uops = tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[2].src)
    # check that the float4 cast collapses
    store_vals = [u.src[1] for u in uops if u.op is Ops.STORE and u.src[0].dtype.addrspace != AddrSpace.REG]
    for val in store_vals:
@@ -347,7 +347,7 @@ class TestLinearizer(unittest.TestCase):
    x = Tensor.randn((4,3,6,6)).realize()
    out = x.flip((0,1)).contiguous()
    ast = helper_linearizer_opt(out)
-    store_val = [u.src[1] for u in tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[1].src) if u.op is Ops.STORE][0]
+    store_val = [u.src[1] for u in tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[2].src) if u.op is Ops.STORE][0]
    assert store_val.dtype == dtypes.float.vec(4) and store_val.op is not Ops.STACK

  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
@@ -360,14 +360,15 @@ class TestLinearizer(unittest.TestCase):
            Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 2)] # upcast accs in both reduces
    ast = helper_linearizer_opt(out, opts=[opt])
    def get_recursive(uop): return set.union(set(uop.src), [uop], *[get_recursive(v) for v in uop.src])
-    uops = tuple(to_program(replace_opts(ast, opt), renderer=Device[Device.DEFAULT].renderer).src[1].src)
-    local_stores = [u for u in uops if u.op is Ops.STORE and u.addrspace == AddrSpace.LOCAL]
-    global_stores = [u for u in uops if u.op is Ops.STORE and u.addrspace == AddrSpace.GLOBAL]
+    uops = tuple(to_program(replace_opts(ast, opt), renderer=Device[Device.DEFAULT].renderer).src[2].src)
+    local_stores = [u for u in uops if u.op is Ops.STORE and any(
+      x.op is Ops.BUFFER and x.addrspace is AddrSpace.LOCAL for x in get_recursive(u.src[0]))]
+    global_stores = [u for u in uops if u.op is Ops.STORE and any(x.op is Ops.PARAM for x in get_recursive(u.src[0]))]
    barrier = [u for u in uops if u.op is Ops.BARRIER]
    assert len(barrier) == 1
    # check that the float4 cast collapses for all stores
    for store in local_stores+global_stores:
-      assert store.src[1].max_numel() > 1, f"store shape {store.src[1].shape} on {store.addrspace}"
+      assert store.src[1].max_numel() > 1 # and store.src[2].op is not Ops.VECTORIZE
    # # check the children's vins
    # TODO: src ALU are not the same, should it?
    # assert barrier.src == tuple(local_stores)
@@ -382,7 +383,7 @@ class TestLinearizer(unittest.TestCase):
    x, y = Tensor.rand(1,128), Tensor.rand(128, 128)
    r = (x@y).relu()
    ast = helper_linearizer_opt(r)
-    uops = tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[1].src)
+    uops = tuple(to_program(ast, renderer=Device[Device.DEFAULT].renderer).src[2].src)
    stores = [u for u in uops if u.op is Ops.STORE and u.src[0].dtype.addrspace != AddrSpace.REG]

    # the float4 value stores directly in lds and we skip upcast
@@ -14,7 +14,7 @@ from test.helpers import not_support_multi_device, needs_second_gpu, slow
@slow
 class TestNN(unittest.TestCase):
  def test_batchnorm2d(self, training=False, threed=False, track_running_stats=True):
-    with Context(TRAINING=training):
+    with Tensor.train(training):
      szs = [4, 8, 16, 32]
      for sz in szs:
        # create in tinygrad
@@ -3,7 +3,6 @@ import torch
 import unittest
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.nn.optim import Adam, SGD, AdamW, Muon, LAMB
-from tinygrad.helpers import Context
 from test.helpers import needs_second_gpu, slow

 np.random.seed(1337)
@@ -46,7 +45,11 @@ def step(tensor, optim, steps=1, teeny=False, **kwargs):

@slow
 class TestOptim(unittest.TestCase):
-  def setUp(self): self.enterContext(Context(TRAINING=1))
+  def setUp(self):
+    self.old_training = Tensor.training
+    Tensor.training = True
+  def tearDown(self):
+    Tensor.training = self.old_training

  def _test_optim(self, tinygrad_optim, torch_optim, steps, opts, atol, rtol):
    for x,y in zip(step(Tensor, tinygrad_optim, steps, **opts),
@@ -154,11 +157,13 @@ class TestOptim(unittest.TestCase):
    t = Tensor.ones((1,1))
    optimizer = Adam([t])
    optimizer.zero_grad()
+    old_state = Tensor.training
    t.sum().backward()
-    with Context(TRAINING=0):
-      self.assertRaises(RuntimeError, optimizer.step)
-    with Context(TRAINING=1):
-      optimizer.step()
+    Tensor.training = False
+    self.assertRaises(RuntimeError, optimizer.step)
+    Tensor.training = True
+    optimizer.step()
+    Tensor.training = old_state

  def test_lamb_cpu_offload(self):
    # test that LAMB works when optimizer params (m, v, b1_t, b2_t) are moved to CPU
@@ -43,10 +43,10 @@ def sexec(out:Tensor, opts:list[Opt], replace_src=None, run_count=3):
  call = linear.src[-1]
  prg = to_program(replace_opts(call.src[0], opts), renderer=Device[Device.DEFAULT].renderer)
  if replace_src is not None:
-    old_name = prg.src[2].arg.split("__attribute__((noinline)) void ")[1].split("(")[0]
-    new_src = replace_src + "/* DSP boilerplate */" + prg.src[2].arg.split("/* DSP boilerplate */")[1].replace(old_name, "fxn")
+    old_name = prg.src[3].arg.split("__attribute__((noinline)) void ")[1].split("(")[0]
+    new_src = replace_src + "/* DSP boilerplate */" + prg.src[3].arg.split("/* DSP boilerplate */")[1].replace(old_name, "fxn")
    # drop BINARY and replace SOURCE so run_linear recompiles
-    prg = prg.replace(src=prg.src[:2] + (UOp(Ops.SOURCE, arg=new_src),))
+    prg = prg.replace(src=prg.src[:3] + (UOp(Ops.SOURCE, arg=new_src),))
  linear = linear.replace(src=linear.src[:-1] + (call.replace(src=(prg, *call.src[1:])),))
  for _ in range(run_count): run_linear(linear)

@@ -81,7 +81,7 @@ class TestQuantizeOnnxCPU(unittest.TestCase):
    with Context(QUANTIZE=1):
      linear = run_onnx({"input":inp})["output"].schedule_linear()
      prg = to_program(linear.src[-2].src[0], renderer=Device[Device.DEFAULT].renderer)
-      daccs = [u for u in tuple(prg.src[1].src) if u.op is Ops.BUFFER and u.addrspace is AddrSpace.REG]
+      daccs = [u for u in tuple(prg.src[2].src) if u.op is Ops.BUFFER and u.addrspace is AddrSpace.REG]
      assert all(u.dtype.scalar() is dtypes.int for u in daccs)

@unittest.skipIf(Device.DEFAULT != "DSP", "only tests for DSP")
@@ -77,7 +77,7 @@ class TestRandomness(unittest.TestCase):
      ast = call.src[0]
      if ast.op is Ops.SINK:
        prg = to_program(ast, renderer=Device[Device.DEFAULT].renderer)
-        for u in tuple(prg.src[1].src):
+        for u in tuple(prg.src[2].src):
          self.assertNotIn(u.dtype, {dtypes.long, dtypes.ulong}, msg=f"long found in {prg.arg.name}")

  def test_threefry_against_reference_full(self):
@@ -63,7 +63,7 @@ class TestCStyleFailures(unittest.TestCase):
    for _ in range(5): ret = python_alu[op](ret, Tensor.empty(1, dtype=dtype))
    linear = ret.schedule_linear()
    assert len(linear.src) == 1
-    src = to_program(linear.src[0].src[0], Device[Device.DEFAULT].renderer).src[2].arg
+    src = to_program(linear.src[0].src[0], Device[Device.DEFAULT].renderer).src[3].arg
    self.assertEqual("("*5 not in src, should_strip_paren)

  def test_repeat_add(self): self._test_src_strip_paren(Ops.ADD)
@@ -166,13 +166,6 @@ class TestSchedule(unittest.TestCase):
    self.assertEqual(t.device, dev)
    np.testing.assert_equal(t[:64].numpy(), np.arange(64).cumsum())

-  def test_copy_multi_scalar(self):
-    devs = ("CPU:0", "CPU:1")
-    x = Tensor.ones(2, device="CPU").shard(devs, axis=0).realize()
-    out = (x.sum()*2).reshape(1).to("CPU")
-    run_linear(*check_schedule(out, 5))
-    np.testing.assert_equal(out.numpy(), [4.])
-
 class TestLimitBufs(unittest.TestCase):
  @unittest.skipIf(DEV.interface.startswith("MOCK") and Device.DEFAULT == "NV", "crashes in ocelot")
  def test_limit_bufs_with_var(self):
@@ -455,14 +448,6 @@ class TestCopyFolding(unittest.TestCase):
    b.realize()
    self.assertListEqual(b.tolist(), [[0, 2], [1, 3]])

-  def test_permute_copy_to_device(self):
-    b = Tensor([[0, 1, 2, 3], [4, 5, 6, 7]], device="CPU").permute(1, 0).to("PYTHON")
-    self.assertListEqual(b.tolist(), [[0, 4], [1, 5], [2, 6], [3, 7]])
-
-  def test_flip_copy_to_device(self):
-    b = Tensor([0, 1, 2, 3], device="CPU").flip(0).to("PYTHON")
-    self.assertListEqual(b.tolist(), [3, 2, 1, 0])
-
 class TestUOpBecome(unittest.TestCase):
  def test_setitem_offset(self):
    a = Tensor.full((16,), 0.).contiguous().realize()
@@ -41,7 +41,7 @@ class TestStunning(unittest.TestCase):
    X_samp, Y_samp = X_train[samples], Y_train[samples]
    vi = Variable('i', 0, samples.shape[0]-1)
    with Context(SPLIT_REDUCEOP=0):
-      with Context(TRAINING=1):
+      with Tensor.train():
        losses = []
        for i in range(samples.shape[0]):
          vib = vi.bind(i)
@@ -1,5 +1,5 @@
 import unittest
-from tinygrad import Tensor, Variable, GlobalCounters, Context
+from tinygrad import Tensor, Variable, GlobalCounters
 from tinygrad.uop.ops import sym_infer
 from tinygrad.dtype import dtypes
 from examples.gpt2 import Attention
@@ -63,7 +63,7 @@ class TestSymbolicOps(unittest.TestCase):
    self.test_attention(imin=4, imax=5, use_symbolic=True)

  def test_attention_training(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      self.test_attention(dropout_p=0.0)
      with self.assertRaises(ValueError):
        # symbolic shape dropout is not supported
@@ -1,7 +1,7 @@
 import numpy as np
 import torch
 import unittest, copy, mmap, random, math, array
-from tinygrad import Tensor, Device, dtypes, nn, Context
+from tinygrad import Tensor, Device, dtypes, nn
 from tinygrad.helpers import getenv, temp, mv_address
 from extra.gradcheck import numerical_jacobian, jacobian, gradcheck
 from hypothesis import given, settings, strategies as strat
@@ -203,7 +203,7 @@ class TestTinygrad(unittest.TestCase):
      np.testing.assert_allclose(x, y, atol=1e-5)

  def test_dropout(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      n, rate = 1_000_000, 0.1
      w = Tensor.ones(n).dropout(rate)
      non_zeros = np.count_nonzero(w.numpy())
@@ -249,7 +249,7 @@ class TestAssembly(unittest.TestCase):
    opts_to_apply = [Opt(OptOps.UNROLL, 0, 4)]
    ast = ast.replace(arg=KernelInfo(opts_to_apply=tuple(opts_to_apply)))
    program = to_program(ast, Device[Device.DEFAULT].renderer)
-    uops = tuple(program.src[1].src)
+    uops = tuple(program.src[2].src)
    self.assertGreaterEqual(len([x.op for x in uops if x.op is Ops.MULACC]), 4)

  def test_mulacc_shl(self):
@@ -1,68 +0,0 @@
-from tinygrad import Tensor, dtypes, Device, TinyJit
-from tinygrad.helpers import getenv, Timing, round_up
-
-# llama 70b shapes
-DIM, HIDDEN, N_HEADS, N_KV_HEADS, HEAD_DIM = 8192, 28672, 64, 8, 128
-QKV_OUT = N_HEADS*HEAD_DIM + 2*N_KV_HEADS*HEAD_DIM
-LAYER_SHAPES = {
-  "wqkv": ((QKV_OUT, DIM), 0),
-  "wo": ((DIM, N_HEADS*HEAD_DIM), 1),
-  "w1": ((HIDDEN, DIM), 0),
-  "w3": ((HIDDEN, DIM), 0),
-  "w2": ((DIM, HIDDEN), 1),
-}
-
-def build_params(layers:int, devs:tuple[str,...], dtype):
-  ts, nbytes = [], 0
-  ndev = len(devs)
-  for _ in range(layers):
-    for (a,b), ax in LAYER_SHAPES.values():
-      shape = (round_up(a, ndev) if ax==0 else a, round_up(b, ndev) if ax==1 else b)
-      t = Tensor.empty(*shape, dtype=dtype).shard(devs, axis=ax).realize()
-      ts.append(t)
-      nbytes += t.numel()*dtype.itemsize
-  return ts, nbytes
-
-def sync():
-  Device["CPU"].synchronize()
-  Device[Device.DEFAULT].synchronize()
-
-if __name__ == "__main__":
-  ngpus = getenv("GPUS", 8)
-  layers = getenv("LAYERS", 8)
-  devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(ngpus))
-
-  # to cpu
-  grads, nbytes = build_params(layers, devs, dtypes.bfloat16)
-  cpu_dst = [Tensor.rand(*g.shape, dtype=dtypes.bfloat16, device="CPU").contiguous().realize() for g in grads]
-  sync()
-  def to_cpu():
-    for d,g in zip(cpu_dst, grads):
-      d.assign(g.to("CPU"))
-    return Tensor.realize(*cpu_dst)
-  to_cpu_jit = TinyJit(to_cpu)
-  for _ in range(3):
-    to_cpu_jit()
-    sync()
-  for _ in range(5):
-    with Timing("to cpu ", on_exit=lambda ns: f" @ {nbytes/ns:.2f} GB/s"):
-      to_cpu_jit()
-      sync()
-
-  # to gpu
-  from extra.gemm.cdna_asm_gemm import FP8_DTYPE
-  w_gpu, nbytes = build_params(layers, devs, FP8_DTYPE)
-  cpu_src = [Tensor.rand(*w.shape, dtype=FP8_DTYPE, device="CPU").contiguous().realize() for w in w_gpu]
-  sync()
-  def to_gpu():
-    for w,c in zip(w_gpu, cpu_src):
-      w.assign(c.shard_like(w))
-    return Tensor.realize(*w_gpu)
-  to_gpu_jit = TinyJit(to_gpu)
-  for _ in range(3):
-    to_gpu_jit()
-    sync()
-  for _ in range(5):
-    with Timing("to gpu ", on_exit=lambda ns: f" @ {nbytes/ns:.2f} GB/s"):
-      to_gpu_jit()
-      sync()
@@ -1,15 +1,16 @@
 from tinygrad import nn, Tensor, dtypes
-from tinygrad.helpers import DEV, Timing, Context
+from tinygrad.helpers import DEV, Timing

 from extra.models.llama import Transformer
 from examples.llama3 import MODEL_PARAMS

 if __name__ == "__main__":
  DEV.value = "NULL"
+  Tensor.training = True
  #model_size = "8B"
  model_size = "405B"

-  with Context(TRAINING=1), Timing("total "):
+  with Timing("total "):
    with Timing("***** create model in    "):
      model = Transformer(**MODEL_PARAMS[model_size]["args"], linear=nn.Linear, embedding=nn.Embedding,
                          max_context=1024, jit=True, disable_kv_cache=True)
@@ -63,8 +63,8 @@ class BenchmarkResnetTrain(unittest.TestCase):
    with Context(TRACK_MATCH_STATS=0): Tensor.realize(*[t.assign(t.detach().contiguous()) for t in get_parameters(optim)])

    JITCNT = getenv("JITCNT", 1)
+    Tensor.training = True
    @TinyJit
-    @Context(TRAINING=1)
    def step(x):
      optim.zero_grad()
      x.grad = None
@@ -1,67 +0,0 @@
-import unittest, math
-import numpy as np
-from tinygrad.tensor import Tensor
-from tinygrad.nn.optim import AdamW
-from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup, LambdaLR, LambdaLinearScheduler
-
-np.random.seed(1337)
-x_init = np.random.randn(1,4).astype(np.float32)
-W_init = np.random.randn(4,4).astype(np.float32)
-m_init = np.random.randn(1,4).astype(np.float32)
-
-class TinyNet:
-  def __init__(self):
-    self.x = Tensor(x_init.copy())
-    self.W = Tensor(W_init.copy())
-    self.m = Tensor(m_init.copy())
-
-  def forward(self):
-    out = self.x.matmul(self.W).relu()
-    out = out.log_softmax(1)
-    out = out.mul(self.m).add(self.m).sum()
-    return out
-
-class TestCosineAnnealingLRWithWarmup(unittest.TestCase):
-  # only tests the lr
-  def _test_lr(self, base_lr, end_lr, warmup_steps, decay_steps):
-    net = TinyNet()
-    optim = AdamW([net.W], lr=0.0)
-    tiny_lr = CosineAnnealingLRWithWarmup(optim, base_lr, end_lr, warmup_steps, decay_steps)
-    lr = []
-    for _ in range(warmup_steps+decay_steps):
-      lr.append(optim.lr.item())
-      tiny_lr.step()
-    # reimplemented in python
-    expected = []
-    for i in range(warmup_steps): expected.append((i+1)/warmup_steps*base_lr)
-    for i in range(decay_steps): expected.append(end_lr+(base_lr-end_lr)*(1+math.cos((i+1)/decay_steps*math.pi))/2)
-    np.testing.assert_allclose(lr, expected, rtol=1e-5)
-
-  def test_lr_0(self): self._test_lr(3e-4, 8e-5, 3, 5)
-  def test_lr_1(self): self._test_lr(3e-4, 8e-5, 10, 20)
-  def test_lr_llama3(self): self._test_lr(8e-5, 8e-7, 20, 100)
-
-class TestLambdaLRLinearWarmup(unittest.TestCase):
-  def test_linear_lr_warmup(self):
-    BS, BASE_LR = 304, 2.5e-7
-    lr = BS * BASE_LR
-    # Use a dummy Tensor parameter for optimizer because the lr_scheduler only needs the optimizer's device and lr, the params aren't touched.
-    optimizer = AdamW([Tensor([1.])])
-    lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
-    lr_scheduler = LambdaLR(optimizer, Tensor(lr, device=optimizer.device), lambda_lr_callback)
-    lrs = {}
-
-    # with above settings, optimizer.lr should warm up to lr over 1000 steps linearly
-    for i in range(1200):
-      lr_scheduler.step()
-      if i in {0, 499, 998, 999, 1000, 1199}:
-        lrs[i] = optimizer.lr.item()
-
-    np.testing.assert_allclose(lr, lrs[999], rtol=0, atol=1e-11)
-    np.testing.assert_equal(lrs[999], lrs[1000])
-    np.testing.assert_equal(lrs[999], lrs[1199])
-    np.testing.assert_allclose(lrs[999] / lrs[0], 1000, rtol=0, atol=1)
-    np.testing.assert_allclose(lrs[999] / lrs[499], 2, rtol=0, atol=1e-5)
-
-if __name__ == '__main__':
-  unittest.main()
@@ -4,7 +4,7 @@ import numpy as np
 import torch

 from tinygrad import GlobalCounters, Tensor, Device
-from tinygrad.helpers import getenv, Context
+from tinygrad.helpers import getenv
 from tinygrad.nn.state import get_parameters
 from tinygrad.engine.realize import capturing, run_linear
 from tinygrad.tensor import _to_np_dtype
@@ -38,7 +38,11 @@ from extra.models.vit import ViT

@unittest.skipUnless(Device.DEFAULT == "CL", "Not Implemented")
 class TestInferenceMinKernels(unittest.TestCase):
-  def setUp(self): self.enterContext(Context(TRAINING=0))
+  def setUp(self):
+    self.training_old = Tensor.training
+    Tensor.training = False
+  def tearDown(self):
+    Tensor.training = self.training_old

  def test_convnext(self):
    model = ConvNeXt()
@@ -1,5 +1,5 @@
 #!/usr/bin/env python
-import unittest
+import unittest, math
 import numpy as np
 import tensorflow as tf
 from tensorflow.keras.optimizers import Lamb
@@ -7,12 +7,11 @@ from tensorflow.python.ops import math_ops
 from extra.lr_scheduler import LRSchedulerGroup

 from tinygrad.tensor import Tensor
-from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup
-from tinygrad.helpers import Context
+from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, AdamW

 from test.external.mlperf_resnet.lars_optimizer import LARSOptimizer

-from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup
+from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup, CosineAnnealingLRWithWarmup, LambdaLR, LambdaLinearScheduler
 from test.external.mlperf_resnet.lars_util import PolynomialDecayWithWarmup as PolynomialDecayWithWarmup_tf

 np.random.seed(1337)
@@ -104,7 +103,11 @@ def create_tf_polylr(initial_lr, end_lr, train_steps, warmup, power=2, skip_list
                                      initial_learning_rate=initial_lr, end_learning_rate=end_lr, warmup_epochs=warmup)

 class ExternalTestOptim(unittest.TestCase):
-  def setUp(self): self.enterContext(Context(TRAINING=1))
+  def setUp(self):
+    self.old_training = Tensor.training
+    Tensor.training = True
+  def tearDown(self):
+    Tensor.training = self.old_training

  def _test_optim(self, tinygrad_optim, tensorflow_optim, steps, opts, atol, rtol, tiny_sched=None, tf_sched=None, schedopts=None, do_optim=True):
    for x,y in zip(step(tinygrad_optim, steps=steps, kwargs=opts, scheduler=tiny_sched, schedopts=schedopts, do_optim=do_optim),
@@ -170,5 +173,48 @@ class ExternalTestOptim(unittest.TestCase):
      'warmup': steps_per_epoch * warmup_epochs,
    }, 1e-5, 1e-5, do_optim=False)

+
+class TestCosineAnnealingLRWithWarmup(unittest.TestCase):
+  # only tests the lr
+  def _test_lr(self, base_lr, end_lr, warmup_steps, decay_steps):
+    net = TinyNet()
+    optim = AdamW([net.W], lr=0.0)
+    tiny_lr = CosineAnnealingLRWithWarmup(optim, base_lr, end_lr, warmup_steps, decay_steps)
+    lr = []
+    for _ in range(warmup_steps+decay_steps):
+      lr.append(optim.lr.item())
+      tiny_lr.step()
+    # reimplemented in python
+    expected = []
+    for i in range(warmup_steps): expected.append((i+1)/warmup_steps*base_lr)
+    for i in range(decay_steps): expected.append(end_lr+(base_lr-end_lr)*(1+math.cos((i+1)/decay_steps*math.pi))/2)
+    np.testing.assert_allclose(lr, expected, rtol=1e-5)
+
+  def test_lr_0(self): self._test_lr(3e-4, 8e-5, 3, 5)
+  def test_lr_1(self): self._test_lr(3e-4, 8e-5, 10, 20)
+  def test_lr_llama3(self): self._test_lr(8e-5, 8e-7, 20, 100)
+
+class TestLambdaLRLinearWarmup(unittest.TestCase):
+  def test_linear_lr_warmup(self):
+    BS, BASE_LR = 304, 2.5e-7
+    lr = BS * BASE_LR
+    # Use a dummy Tensor parameter for optimizer because the lr_scheduler only needs the optimizer's device and lr, the params aren't touched.
+    optimizer = AdamW([Tensor([1.])])
+    lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
+    lr_scheduler = LambdaLR(optimizer, Tensor(lr, device=optimizer.device), lambda_lr_callback)
+    lrs = {}
+
+    # with above settings, optimizer.lr should warm up to lr over 1000 steps linearly
+    for i in range(1200):
+      lr_scheduler.step()
+      if i in {0, 499, 998, 999, 1000, 1199}:
+        lrs[i] = optimizer.lr.item()
+
+    np.testing.assert_allclose(lr, lrs[999], rtol=0, atol=1e-11)
+    np.testing.assert_equal(lrs[999], lrs[1000])
+    np.testing.assert_equal(lrs[999], lrs[1199])
+    np.testing.assert_allclose(lrs[999] / lrs[0], 1000, rtol=0, atol=1)
+    np.testing.assert_allclose(lrs[999] / lrs[499], 2, rtol=0, atol=1e-5)
+
 if __name__ == '__main__':
  unittest.main()
@@ -3,7 +3,7 @@ from tinygrad import Tensor, UOp, Device, nn
 from tinygrad.schedule import schedule_cache
 from tinygrad.codegen import to_program, to_program_cache
 from tinygrad.schedule.indexing import apply_movement_op, _apply_reshape
-from tinygrad.codegen.decomp.divandmod import fold_divmod_general
+from tinygrad.uop.divandmod import fold_divmod_general
 from test.test_tiny import TestTiny

 def uops_allocated(): return sum([isinstance(x, UOp) for x in gc.get_objects()])
@@ -3,7 +3,7 @@ import z3
 from tinygrad import dtypes, Device
 from tinygrad.uop.validate import uops_to_z3, z3_cdiv
 from tinygrad.uop.ops import UOp
-from tinygrad.codegen.decomp.op import fast_idiv
+from tinygrad.uop.decompositions import fast_idiv
 random.seed(42)

 powers_of_two = [2**i for i in range(64)]
@@ -11,7 +11,7 @@ if __name__ == "__main__":
  for i in range(10_000):
    if i % 1000 == 0:
      print(f"Progress: {i}")
-    dt = random.choice(dtypes.ints)
+    dt = random.choice(dtypes.ints + tuple(dt.vec(4) for dt in dtypes.ints))
    u = UOp.variable('x', random.randint(dt.min, 0), random.randint(1, dt.max), dtype=dt)
    d = random.randint(1, max(1, u.vmax)*2)
    if d in powers_of_two: continue
@@ -41,8 +41,8 @@ class BenchmarkBertTrain(unittest.TestCase):
    with Context(TRACK_MATCH_STATS=0): Tensor.realize(*[t.assign(t.detach().contiguous()) for t in get_parameters(optim)])

    JITCNT = getenv("JITCNT", 1)
+    Tensor.training = True
    @TinyJit
-    @Context(TRAINING=1)
    def step(inputs):
      optim.zero_grad()
      for i in inputs: i.grad = None
@@ -31,6 +31,8 @@ if __name__ == "__main__":
  assert all(os.path.exists(os.path.join(INIT_CKPT_DIR, f)) for f in required_files), \
    f"Missing checkpoint files in INIT_CKPT_DIR: {required_files}"

+  Tensor.training = False
+
  model = get_mlperf_bert_model(INIT_CKPT_DIR)

  for _, x in get_state_dict(model).items():
@@ -1,6 +1,6 @@
 import unittest, os
 from tempfile import TemporaryDirectory
-from tinygrad import Context
+from tinygrad import Tensor
 from tinygrad.helpers import getenv
 from examples.mlperf.model_train import train_stable_diffusion

@@ -14,10 +14,10 @@ class TestTrain(unittest.TestCase):
    if not getenv("CKPTDIR", ""): os.environ["CKPTDIR"] = "/raid/weights/stable_diffusion"
    with TemporaryDirectory(prefix="test-train") as tmp:
      os.environ["UNET_CKPTDIR"] = tmp
-      with Context(TRAINING=1):
+      with Tensor.train():
        saved_ckpts = train_stable_diffusion()
      expected_ckpt = f"{tmp}/{num_steps}.safetensors"
      assert len(saved_ckpts) == 1 and saved_ckpts[0] == expected_ckpt

 if __name__=="__main__":
-  unittest.main()
+  unittest.main()
@@ -48,9 +48,9 @@ def replay_to_program(p:UOp, ast:UOp, renderer:Renderer) -> tuple[str, str, tupl
    if sink_arg.beam: sink_arg = replace(sink_arg, opts_to_apply=p.src[0].arg.applied_opts)
    input_ast = ast.replace(arg=replace(sink_arg, name=p.src[0].arg.name))
  p2 = to_program(input_ast, renderer=renderer)
-  device = renderer.target.device
+  device = p.src[1].arg
  def to_str(ret:UOp) -> str:
-    src = next(x.arg for x in ret.src if x.op is Ops.SOURCE)
+    src = ret.src[3].arg
    # PYTHON renderer pickles UOps, first unpickle and decode here
    if device.startswith("PYTHON"): return "\n".join([str(x) for x in pickle.loads(base64.b64decode(src))])
    return src
@@ -84,7 +84,7 @@ def eval_uop(uop:UOp, inputs:list[tuple[DType, list[Any]]]|None=None, vals:tuple
    allocator._copyin(buf, memoryview(struct.pack(str(len(data)) + (buf_dt.fmt or ""), *data)))
  g = UOp.param(0, uop.dtype.ptr(1))
  prg = to_program(UOp.store(g.index(UOp.const(dtypes.int, 0), ptr=True), uop).sink(arg=KernelInfo()), PythonRenderer(Target("PYTHON")))
-  prog = PythonProgram("run", PythonCompiler().compile(prg.src[2].arg))
+  prog = PythonProgram("run", PythonCompiler().compile(prg.src[3].arg))
  prog(out_buf:=allocator.alloc(uop.dtype.itemsize), *bufs, vals=vals)
  return out_buf.cast(uop.dtype.fmt or "").tolist()[0]

@@ -375,7 +375,7 @@ def _mem_store(mem: UOp, addr: UOp, val: UOp, active: UOp, addr_bits: int = 32,
  """Conditional memory store with sub-word support. Returns list of store UOps."""
  adt = dtypes.uint64 if addr_bits == 64 else dtypes.uint32
  word_addr = addr >> UOp.const(adt, 2)
-  idx = mem.index(word_addr.valid(active))
+  idx = mem.index(word_addr.cast(dtypes.int).valid(active))
  if data_bits == 32: return [idx.store(active.where(_to_u32(val), idx))]
  # Sub-word store: read-modify-write with mask
  byte_pos = addr.cast(dtypes.uint32) & _c(3)
@@ -388,7 +388,7 @@ def _mem_store(mem: UOp, addr: UOp, val: UOp, active: UOp, addr_bits: int = 32,
  is_cross = byte_pos.eq(_c(3))
  cross_word0 = (idx & _c(0x00FFFFFF)) | ((val_u32 & _c(0xFF)) << _c(24))
  store0 = idx.store(active.where(is_cross.where(cross_word0, new_word), idx))
-  next_idx = mem.index((word_addr + UOp.const(adt, 1)).valid(active & is_cross))
+  next_idx = mem.index((word_addr + UOp.const(adt, 1)).cast(dtypes.int).valid(active & is_cross))
  cross_word1 = (next_idx & _c(0xFFFFFF00)) | ((val_u32 >> _c(8)) & _c(0xFF))
  return [store0, next_idx.store((active & is_cross).where(cross_word1, next_idx))]

@@ -398,7 +398,7 @@ def _mem_store_bytes(mem: UOp, addr: UOp, val: UOp, active: UOp, data_bits: int
  val_u32 = val.cast(dtypes.uint32) if val.dtype != dtypes.uint32 else val
  for i in range(data_bits // 8):
    byte_val = (val_u32 >> UOp.const(dtypes.uint32, i * 8)) & UOp.const(dtypes.uint32, 0xFF)
-    stores.append(mem.index((addr + UOp.const(dtypes.uint64, i)).valid(active)).store(byte_val.cast(dtypes.uint8)))
+    stores.append(mem.index((addr + UOp.const(dtypes.uint64, i)).cast(dtypes.int).valid(active)).store(byte_val.cast(dtypes.uint8)))
  return stores

 def _collect_data_slices(assigns: list[tuple[str, UOp]], data_prefix: str, pcode_vars: dict | None = None, op_name: str = "") -> dict[int, UOp]:
@@ -463,7 +463,7 @@ class _Ctx:
    """Read instruction dword from vmem at PC + dword_idx*4."""
    pc = self.rpc()
    addr = pc if dword_idx == 0 else pc + UOp.const(dtypes.uint64, dword_idx * 4)
-    return self.vmem.index(addr >> UOp.const(dtypes.uint64, 2), ptr=True).load()
+    return self.vmem.index((addr >> UOp.const(dtypes.uint64, 2)).cast(dtypes.int), ptr=True).load()

  def inst_field(self, field) -> UOp:
    """Extract field bits from instruction encoding. Tracks field for canonical key computation."""
@@ -516,14 +516,14 @@ class _Ctx:
  # Dynamic register access (takes UOp index instead of int)
  def rsgpr_dyn(self, reg: UOp, valid: UOp | None = None) -> UOp:
    """Read SGPR with dynamic register index."""
-    if valid is not None: return self.sgpr.index(reg.valid(valid), ptr=True).load()
-    return self.sgpr.index(reg, ptr=True).load()
+    if valid is not None: return self.sgpr.index(reg.cast(dtypes.int).valid(valid), ptr=True).load()
+    return self.sgpr.index(reg.cast(dtypes.int), ptr=True).load()

  def wsgpr_dyn(self, reg: UOp, val: UOp) -> UOp:
    """Write SGPR with dynamic register index. On RDNA, index 124 = NULL (writes discarded). On CDNA, index 124 = M0 (read/write)."""
    # RDNA: NULL (124) discards writes. CDNA: M0 (124) is writable.
    valid = None if self.wave_size == 64 else reg.ne(_c(124))
-    return self.sgpr.index(reg.valid(valid) if valid is not None else reg, ptr=True).store(val.cast(dtypes.uint32))
+    return self.sgpr.index(reg.cast(dtypes.int).valid(valid) if valid is not None else reg.cast(dtypes.int), ptr=True).store(val.cast(dtypes.uint32))

  def wmask(self, reg: UOp, val: UOp) -> list[UOp]:
    """Write a lane mask (VCC/EXEC). Splits into lo/hi for wave64."""
@@ -656,7 +656,7 @@ class _Ctx:
    stores = []
    for dest, val in assigns:
      if dest.startswith('D0'): stores.append(self.wsgpr_dyn(vdst_off, val.cast(dtypes.uint32)))
-      elif dest.startswith('VGPR['): stores.append(self.vgpr.index(val[0]).store(val[1].cast(dtypes.uint32)))
+      elif dest.startswith('VGPR['): stores.append(self.vgpr.index(val[0].cast(dtypes.int)).store(val[1].cast(dtypes.uint32)))
    return UOp.sink(*stores, *self.inc_pc())

  def compile_vop_pcode(self, op, srcs: dict[str, UOp | int], lane: UOp, vdst_reg: UOp, exec_mask: UOp,
@@ -711,11 +711,11 @@ class _Ctx:
        # VGPR bit-slice: (vgpr_idx, rhs_val, hi_bit, lo_bit) - hi/lo are UOp constants
        hi_bit, lo_bit = int(val[2].arg), int(val[3].arg)
        width = hi_bit - lo_bit + 1
-        old = self.vgpr.index(val[0], ptr=True).load()
+        old = self.vgpr.index(val[0].cast(dtypes.int), ptr=True).load()
        new_val = _set_bits(old, _val_to_bits(val[1]), width, lo_bit).cast(dtypes.uint32)
        active = _lane_active(exec_mask, lane)
        if len(val) > 4: active = active & _to_bool(val[4])
-        raw_stores.append(('vgpr_direct', self.vgpr.index(val[0].valid(active)).store(new_val)))
+        raw_stores.append(('vgpr_direct', self.vgpr.index(val[0].cast(dtypes.int).valid(active)).store(new_val)))
        continue
      if 'D0' in dest and '[laneId]' in dest:
        old_vcc = self.rmask(_c(VCC_LO.offset))
@@ -834,7 +834,7 @@ def _compile_smem(inst: ir3.SMEM | ir4.SMEM, ctx: _Ctx) -> UOp:
  nval = int(part.removeprefix('DWORD').removeprefix('X') or '1') if 'DWORD' in part else int(part[1:]) / 32 * (-1 if part[0] == 'I' else 1)
  ndwords = max(1, int(abs(nval)))
  dword_base = addr >> UOp.const(dtypes.uint64, 2)
-  vals = [ctx.vmem.index(dword_base + UOp.const(dtypes.uint64, i)) for i in range(ndwords)]
+  vals = [ctx.vmem.index((dword_base + UOp.const(dtypes.uint64, i)).cast(dtypes.int)) for i in range(ndwords)]
  if abs(nval) < 1:
    nbits = int(abs(nval) * 32)
    byte_off = (addr & UOp.const(dtypes.uint64, 3)).cast(dtypes.uint32) * UOp.const(dtypes.uint32, 8)
@@ -1819,7 +1819,7 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
    srcs = {'ADDR': addr_reg, 'DATA0': vdata_reg, 'VDST': vdst_reg, 'OFFSET': offset,
            'EXEC': exec_mask.cast(dtypes.uint64), '_vgpr': ctx.vgpr, '_wave_size': ctx.wave_size}
    _, assigns = parse_pcode(pcode, srcs)
-    stores = [ctx.vgpr.index(val[0]).store(val[1].cast(dtypes.uint32)) for dest, val in assigns if dest.startswith('VGPR[')]
+    stores = [ctx.vgpr.index(val[0].cast(dtypes.int)).store(val[1].cast(dtypes.uint32)) for dest, val in assigns if dest.startswith('VGPR[')]
    return UOp.sink(*stores, *ctx.inc_pc())

  def make_addr(lane: UOp) -> UOp:
@@ -1853,7 +1853,7 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
  def wmem(addr: UOp, val: UOp, active: UOp, data_bits: int = 32) -> UOp:
    if data_bits < 32:
      # Sub-dword LDS write: read-modify-write within the uint32 slot
-      word_addr = addr >> addr_shift
+      word_addr = (addr >> addr_shift).cast(dtypes.int)
      idx = mem.index(word_addr.valid(active))
      byte_pos = addr.cast(dtypes.uint32) & _c(3)
      byte_shift = byte_pos * _c(8)
@@ -1861,7 +1861,7 @@ def _compile_mem_op(inst: ir3.DS|ir3.FLAT|ir3.GLOBAL|ir3.SCRATCH|ir4.DS|ir4.VFLA
      mask = size_mask << byte_shift
      new_word = (idx & (mask ^ _c(0xFFFFFFFF))) | ((val.cast(dtypes.uint32) & size_mask) << byte_shift)
      return idx.store(active.where(new_word, idx))
-    idx = mem.index(addr >> addr_shift)
+    idx = mem.index((addr >> addr_shift).cast(dtypes.int))
    return idx.store(active.where(val, idx.load()))

  def make_srcs(lane: UOp) -> dict:
@@ -2011,7 +2011,7 @@ def _compile_mubuf(inst: irc.MUBUF, ctx: _Ctx) -> UOp:
    for i in range(n_dwords):
      word_addr = (addr + UOp.const(dtypes.uint64, i * 4)) >> UOp.const(dtypes.uint64, 2)
      val = in_bounds.where(mem.index(word_addr.cast(dtypes.int64), ptr=True).load(), _c(0))
-      lds_idx = (lds_addr + _c(i * 4)) >> _c(2)
+      lds_idx = ((lds_addr + _c(i * 4)) >> _c(2)).cast(dtypes.int)
      lds_slot = ctx.lds.index(lds_idx.valid(active))
      stores.append(lds_slot.store(active.where(val, lds_slot)))
  elif is_store:
@@ -2,7 +2,7 @@
 from typing import Any, Callable
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import Ops, UOp
-from tinygrad.codegen.decomp.dtype import f2f
+from tinygrad.uop.decompositions import f2f

 # Type alias for vars dict: stores UOps and tuples for lambda definitions
 VarVal = UOp | tuple[str, list[str], str]
@@ -831,7 +831,7 @@ class Parser:
    def mindex(idx:UOp, ptr=False): return mem.index(idx.valid(active) if active is not None else idx, ptr=ptr)
    byte_mem = mem.dtype.base == dtypes.uint8
    if byte_mem:
-      idx = addr
+      idx = addr.cast(dtypes.int)
      if dt in (dtypes.uint64, dtypes.int64, dtypes.float64):
        val = _u32(0).cast(dtypes.uint64)
        for i in range(8): val = val | (mindex(idx + _const(dtypes.int, i), ptr=True).load().cast(dtypes.uint64) << _u64(i * 8))
@@ -845,10 +845,10 @@ class Parser:
        val = _u32(0)
        for i in range(4): val = val | (mindex(idx + _const(dtypes.int, i), ptr=True).load().cast(dtypes.uint32) << _u32(i * 8))
    else:
-      idx = addr >> _const(addr.dtype, 2)
+      idx = (addr >> _const(addr.dtype, 2)).cast(dtypes.int)
      val = mindex(idx)
      if dt in (dtypes.uint64, dtypes.int64, dtypes.float64):
-        idx2 = (addr + _const(adt, 4)) >> _const(adt, 2)
+        idx2 = ((addr + _const(adt, 4)) >> _const(adt, 2)).cast(dtypes.int)
        val = val.cast(dtypes.uint64) | (mindex(idx2).cast(dtypes.uint64) << _u64(32))
      elif dt in (dtypes.uint8, dtypes.int8): val = (val >> ((addr & _const(adt, 3)).cast(dtypes.uint32) * _u32(8))) & _u32(0xFF)
      elif dt in (dtypes.uint16, dtypes.int16):
@@ -3,7 +3,7 @@ import ast, pathlib, unittest
 import numpy as np
 from PIL import Image

-from tinygrad import Tensor, Context
+from tinygrad import Tensor
 from tinygrad.helpers import getenv
 from test.helpers import slow
 from extra.models.efficientnet import EfficientNet
@@ -40,7 +40,7 @@ def preprocess(img, new=False):
  return img

 def _infer(model: EfficientNet, img):
-  with Context(TRAINING=0):
+  with Tensor.train(False):
    out = model.forward(Tensor(img)).argmax(axis=-1)
  return out.tolist()

@@ -5,11 +5,10 @@ import numpy as np
 from tinygrad.nn.state import get_parameters, get_state_dict
 from tinygrad.nn import optim, Linear, Conv2d, BatchNorm2d
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import Context
 from extra.datasets import fetch_mnist

 def compare_tiny_torch(model, model_torch, X, Y):
-  with Context(TRAINING=1):
+  with Tensor.train():
    model_torch.train()
    model_state_dict = get_state_dict(model)
    for k,v in model_torch.named_parameters():
@@ -21,7 +21,7 @@ class TestDisassembly(unittest.TestCase):
    c = Tensor([1], dtype=dtypes.float16) + Tensor([1], dtype=dtypes.float16)
    s = c.schedule_linear().src[-1]
    p = to_program(s.src[0], Device[Device.DEFAULT].renderer)
-    lib = Device[Device.DEFAULT].compiler.compile(p.src[2].arg)
+    lib = Device[Device.DEFAULT].compiler.compile(p.src[3].arg)
    out = io.StringIO()
    with redirect_stdout(out): Device[Device.DEFAULT].compiler.disassemble(lib)
    assert "fcvt" not in out.getvalue()
@@ -15,7 +15,7 @@ class TestLinearizerRewrite(unittest.TestCase):
      opts_to_apply.append(Opt(OptOps.UNROLL, 0, 4))
      ast = si.src[0].replace(arg=KernelInfo(opts_to_apply=tuple(opts_to_apply)))
      prg = to_program(ast, Device["CPU"].renderer)
-      print(prg.src[2].arg)
+      print(prg.src[3].arg)

  def test_arange(self):
    out = Tensor.arange(32).clone("NULL")
@@ -25,7 +25,7 @@ class TestLinearizerRewrite(unittest.TestCase):
      opts_to_apply.append(Opt(OptOps.UPCAST, 0, 4))
      ast = si.src[0].replace(arg=KernelInfo(opts_to_apply=tuple(opts_to_apply)))
      prg = to_program(ast, Device["CPU"].renderer)
-      print(prg.src[2].arg)
+      print(prg.src[3].arg)

  def test_kernel_info(self):
    out = Tensor.arange(4).clone("NULL")
@@ -106,7 +106,7 @@ class TestRealWorld(unittest.TestCase):
  @slow
  def test_train_mnist(self):
    from examples.beautiful_mnist import Model
-    with Context(TRAINING=1):
+    with Tensor.train():
      model = Model()
      optimizer = optim.Adam(get_parameters(model))
      BS = 32
@@ -125,7 +125,7 @@ class TestRealWorld(unittest.TestCase):
  def test_forward_cifar(self):
    BS = 32
    # with training batchnorm still though
-    with Context(TRAINING=1):
+    with Tensor.train():
      model = SpeedyResNet(Tensor.ones((12,3,2,2)))
      @TinyJit
      def run(X): return model(X)
@@ -133,7 +133,7 @@ class TestRealWorld(unittest.TestCase):

  @slow
  def test_train_cifar(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      model = SpeedyResNet(Tensor.ones((12,3,2,2)))
      optimizer = optim.SGD(get_parameters(model), lr=0.01, momentum=0.8, nesterov=True, weight_decay=0.15)
      BS = 32
@@ -151,7 +151,7 @@ class TestRealWorld(unittest.TestCase):
  @unittest.skipUnless(dtypes.float16 in supported_dtypes, "need dtypes.float16")
  def test_train_cifar_hyp(self):
    dtypes.default_float = dtypes.float16
-    with Context(TRAINING=1):
+    with Tensor.train():
      model = SpeedyResNet(Tensor.ones((12,3,2,2)))
      optimizer = optim.SGD(get_parameters(model), lr=0.01, momentum=hyp['opt']['momentum'], nesterov=True, weight_decay=hyp['opt']['bias_decay'])
      initial_div_factor = hyp['opt']['initial_div_factor']
@@ -163,7 +163,7 @@ class TestRealWorld(unittest.TestCase):

  @slow
  def test_bert(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      args_tiny = {"attention_probs_dropout_prob": 0.0, "hidden_dropout_prob": 0.0, "vocab_size": 30522, "type_vocab_size": 2,
                  "max_position_embeddings": 512, "hidden_size": 128, "intermediate_size": 512, "num_attention_heads": 2, "num_hidden_layers": 2}
      model = BertForPretraining(**args_tiny)
@@ -1093,14 +1093,14 @@ class TestSchedule(unittest.TestCase):

  #@unittest.skip("may want to reconsider this")
  def test_fold_batchnorm(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(1,32,4,4)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
      out = bn(img)
      check_schedule(out, 3, nn.state.get_parameters(bn))

  def test_fold_conv_batchnorm_notrain(self):
-    with Context(TRAINING=0):
+    with Tensor.train(False):
      img = Tensor.empty(1,3,8,8)
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=True)
@@ -1108,7 +1108,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(out, 1, [c1.weight, c1.bias, *nn.state.get_parameters(bn)])

  def test_fold_conv_batchnorm_notrain_no_running_stats(self):
-    with Context(TRAINING=0):
+    with Tensor.train(False):
      img = Tensor.empty(1,3,8,8)
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
@@ -1116,7 +1116,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(out, 4, [c1.weight, c1.bias, *nn.state.get_parameters(bn)])

  def test_fold_conv_batchnorm(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(1,3,8,8)
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
@@ -1125,7 +1125,7 @@ class TestSchedule(unittest.TestCase):

  def test_fold_conv_batchnorm_optim(self, adam=False):
    optim, cnt = (nn.optim.Adam, 29) if adam else (nn.optim.SGD, 15)
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.ones(1,3,4,4)
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
@@ -1139,7 +1139,7 @@ class TestSchedule(unittest.TestCase):
  def test_fold_conv_batchnorm_optim_adam(self): self.test_fold_conv_batchnorm_optim(True)

  def test_fold_batchnorm_backward(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      x = Tensor.empty((2, 16, 8, 8)).contiguous()
      bn = nn.BatchNorm2d(16)
      fw = bn(x).contiguous_backward().relu().contiguous()
@@ -1484,7 +1484,7 @@ class TestSchedule(unittest.TestCase):
    check_schedule(out, 4)

  def test_adam_step_fusion(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      x = Tensor.empty(4, 64, 32)
      layer = nn.Linear(32, 32*4)
      _realize_weights(layer)
@@ -1494,7 +1494,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(opt.schedule_step(), 13)

  def test_adam_conv_fuse(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(2,3,4,4)
      c1 = nn.Conv2d(3,32,3)
      _realize_weights(c1)
@@ -1505,7 +1505,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(opt.schedule_step(), 13)

  def test_adam_2convs_fuse(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(2,3,4,4)
      c1 = nn.Conv2d(3,16,3,bias=False)
      c2 = nn.Conv2d(16,32,2,bias=False)
@@ -1517,7 +1517,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(opt.schedule_step(), 15)

  def test_sgd_conv_fuse(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(2,3,4,4)
      c1 = nn.Conv2d(3,32,3)
      _realize_weights(c1)
@@ -1527,7 +1527,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(opt.schedule_step(), 5) # TODO: 3?

  def test_sgd_2convs_fuse(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(2,3,4,4)
      c1 = nn.Conv2d(3,16,3,bias=False)
      c2 = nn.Conv2d(16,32,2,bias=False)
@@ -1538,7 +1538,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(opt.schedule_step(), 7)

  def test_fold_2convs_sgd_nesterov_momentum_wd(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(2,3,4,4)
      c1 = nn.Conv2d(3,16,3,bias=False)
      c2 = nn.Conv2d(16,32,2,bias=False)
@@ -1550,7 +1550,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(opt.schedule_step(), 11)

  def test_sgd_4convs_fuse(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(2,3,16,16)
      c1 = nn.Conv2d(3,4,3,bias=False)
      c2 = nn.Conv2d(4,8,3,bias=False)
@@ -1563,7 +1563,7 @@ class TestSchedule(unittest.TestCase):
      check_schedule(opt.schedule_step(), 15)

  def test_sgd_4convs_fuse_conv_bw(self):
-    with Context(TRAINING=1):
+    with Tensor.train():
      img = Tensor.empty(2,3,16,16)
      c1 = nn.Conv2d(3,4,3,bias=False)
      c2 = nn.Conv2d(4,8,3,bias=False)
@@ -1664,7 +1664,7 @@ class TestSchedule(unittest.TestCase):
    self.assertEqual(len([x for x in linear.src[0].src[0].backward_slice_with_self if x.op is Ops.REDUCE]), 0)

  def test_resnet_block(self):
-    with Context(TRAINING=0):
+    with Tensor.train(False):
      in_planes, planes = 64, 64
      conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
      bn1 = nn.BatchNorm2d(planes)
@@ -1842,7 +1842,7 @@ class TestFusionOp(unittest.TestCase):
    linear = a.schedule_linear()
    prg = to_program(linear.src[-1].src[0], renderer=Device[Device.DEFAULT].renderer)
    self.assertLess(time.perf_counter()-st, 2.0)
-    assert len(prg.src[2].arg.splitlines()) < 250
+    assert len(prg.src[3].arg.splitlines()) < 250

  def test_recursive_add_cmp(self):
    st = time.perf_counter()
@@ -1,6 +1,6 @@
 import unittest, itertools

-from tinygrad.codegen.late.devectorizer import indexing_simplify
+from tinygrad.codegen.late.devectorizer import load_store_indexing
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite
 from tinygrad.uop.symbolic import simplify_valid, sym, pm_move_where_on_load
@@ -11,22 +11,46 @@ from test.null.test_uop_symbolic import check_uop_against_string
 # symbolic-only idx + valid simplification (no late lowering of FLOORDIV/FLOORMOD)
 def simplify_valid_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move_where_on_load, name="simplify_valid_idx")
 # image-aware idx + valid simplification: adds the codegen-layer matcher that drops provably in-bounds gates
-def simplify_image_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move_where_on_load+indexing_simplify, name="simplify_image_idx")
+def simplify_image_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move_where_on_load+load_store_indexing, name="simplify_image_idx")

 def get_gated_load_uop(valid:UOp, idx:UOp):
  return UOp(Ops.LOAD, dtypes.float, (
    UOp.param(0, dtypes.float.ptr()).index(idx.valid(valid), ptr=True),
+    UOp.const(dtypes.float, 0.0)
  ))

 def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UOp]):
  return UOp(Ops.LOAD, dtypes.float.vec(4), (
    UOp.param(0, dtypes.imagef(image_shape)).index(idx[1].valid(valid), idx[0].valid(valid), ptr=True),
+    UOp(Ops.STACK, dtypes.float.vec(4), src=(UOp.const(dtypes.float, 0.0),) * 4)
  ))

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

+class TestHelpers(unittest.TestCase):
+  def test_is_increasing(self):
+    idx1 = Special("idx1", 32)
+    idx2 = Special("idx2", 64)
+    ridx0 = Variable("ridx0", 0, 5)
+    ridx1 = Variable("ridx1", 0, 2)
+    ridx2 = Variable("ridx2", 0, 2)
+    # (ridx0+(idx1*48)+(ridx2*6)+(-6)),((idx2*2)+ridx1+(-1)))
+    f0 = ((idx1*24)+(ridx2*3)+ridx0+765)%768
+    f1 = ridx0+(idx1*48)+(ridx2*6)+(-6)
+    f2 = (idx2*2)+ridx1+((idx1+((ridx2+7)//8)+31)//32)+(-2)
+    f3 = (idx2*2)+ridx1+(-1)
+
+    self.assertFalse(f0.is_increasing())
+    self.assertTrue(f1.is_increasing())
+    self.assertTrue(f2.is_increasing())
+    self.assertTrue(f3.is_increasing())
+
+    rng = UOp.range(5, 2)
+    self.assertTrue(rng.is_increasing())
+    self.assertTrue((rng+2).is_increasing())
+
 class TestValidIdxSimplification(unittest.TestCase):
  def check(self, load, sidx, svalid, extra=()):
    load = simplify_valid_idx(UOp.sink(load, *extra)).src[0]
@@ -482,20 +506,10 @@ class TestImageSimplification(unittest.TestCase):
      self.check(load, "(((lidx1<1)!=True)&(((lidx0+r0)<3)!=True)&((lidx0+r0)<11))",
                       "(lidx2+gidx0*4+lidx1*256+(lidx0*1024+r0*1024)+-3264)", "0")

-  def test_drop_non_monotonic_window(self):
-    # two-sided window valid (645 <= gidx0 < 653) on a non-monotonic index (lane split via %4 and //4):
-    # gidx0 outside the window pushes idx_x out of the (1, 48) image, so the gate is dropped
-    gidx0 = Special("gidx0", 1064)
-    r12 = Range(12, 3)
-    valid = ((gidx0 < 645).ne(True)) & (gidx0 < 653)
-    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(dtypes.weakint, 0))
-    load = get_load_image_uop((1, 48, 4), valid, idx)
-    self.check(load, None, "(r12*4+(gidx0+3)%4+(gidx0+3)//4*24+-3888)", "0")
-
 class TestDropTrueGate(unittest.TestCase):
  def test_drop_true_gate_on_index(self):
    # test that INDEX with a constant True valid gets simplified to drop the valid
-    from tinygrad.codegen.late.devectorizer import indexing_simplify
+    from tinygrad.codegen.late.devectorizer import load_store_indexing
    from tinygrad.uop.ops import graph_rewrite
    from tinygrad.uop.symbolic import sym
    buf = UOp.param(0, dtypes.int.ptr())
@@ -503,7 +517,7 @@ class TestDropTrueGate(unittest.TestCase):
    true_gate = UOp.const(dtypes.bool, True)
    index_with_gate = UOp(Ops.INDEX, dtypes.int.ptr(), (buf, idx.valid(true_gate)))
    # apply the optimization
-    result = graph_rewrite(index_with_gate, sym+indexing_simplify)
+    result = graph_rewrite(index_with_gate, sym+load_store_indexing)
    # the True valid should be dropped (INDEX should only have 2 sources)
    self.assertEqual(len(result.src), 2, "True valid should be dropped from INDEX")

@@ -12,6 +12,15 @@ x_init = np.random.randn(1,3).astype(np.float32)
 W_init = np.random.randn(3,3).astype(np.float32)
 m_init = np.random.randn(1,3).astype(np.float32)

+class TestTrainMode(unittest.TestCase):
+  def test_train_mode(self):
+    assert not Tensor.training
+    @Tensor.train()
+    def f():
+      assert Tensor.training
+    f()
+    assert not Tensor.training
+
 class TestInferenceMode(unittest.TestCase):
  def test_inference(self):
    x = Tensor(x_init)
@@ -57,7 +66,7 @@ class TestIdxUpcast(unittest.TestCase):
      if ast.op is Ops.SINK:
        renderer = Device[si.src[1].buffer.device].renderer
        prg = to_program(ast, renderer)
-        return tuple(prg.src[1].src)
+        return tuple(prg.src[2].src)

  def _assert(self, dtype: DType, a: Tensor):
    uops = self._schedule_render(a)
@@ -1,10 +1,9 @@
 import math, unittest
-from dataclasses import replace
-from tinygrad import Tensor, dtypes, Context
-from tinygrad.uop.ops import ParamArg, UOp, UPat, Ops, PatternMatcher, graph_rewrite
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, graph_rewrite

 _strip_unique_pm = PatternMatcher([
-  (UPat(Ops.BUFFER, name="b"), lambda b: b.replace(arg=replace(b.arg, slot=0)) if isinstance(b.arg, ParamArg) and b.arg.slot != 0 else None),
+  (UPat((Ops.UNIQUE, Ops.LUNIQUE), name="u"), lambda u: u.replace(arg=0) if u.arg != 0 else None),
 ])
 def _strip_unique(u: UOp) -> UOp: return graph_rewrite(u, _strip_unique_pm)

@@ -318,19 +317,6 @@ class TestTensorUOpScatterReduce(unittest.TestCase):
  def test_mean_exclude_self(self):
    self._check(_t(3, 4).float(), Tensor([[0, 1, 0, 1]]*3, dtype=dtypes.int32), Tensor.ones(3, 4).float(), reduce="mean", include_self=False)

-class TestTensorUOpMaskedSelect(unittest.TestCase):
-  # only the fixed-size path is pure
-  def _check(self, t, mask, **kw):
-    self.assertIs(t.masked_select(mask, **kw).uop, t.uop.masked_select(mask.uop, **kw))
-  def test_masked_select_1d(self): self._check(_t(6), Tensor([True, False, True, False, True, False]), size=4)
-  def test_masked_select_2d(self):
-    self._check(_t(3, 3), Tensor([[True, False, True], [False, True, False], [False, False, True]]), size=6, fill_value=-1)
-
-class TestTensorUOpNonzero(unittest.TestCase):
-  def _check(self, t, **kw): self.assertIs(t.nonzero(**kw).uop, t.uop.nonzero(**kw))
-  def test_nonzero_1d(self): self._check(_t(5), size=3)
-  def test_nonzero_2d(self): self._check(_t(2, 3), size=4)
-
 class TestTensorUOpPool(unittest.TestCase):
  def test_avg_pool2d(self):                _check(self, _t(1, 1, 5, 5).float(), lambda x: x.avg_pool2d())
  def test_avg_pool2d_padding(self):        _check(self, _t(1, 1, 5, 5).float(), lambda x: x.avg_pool2d(padding=1))
@@ -395,24 +381,6 @@ class TestTensorUOpConv2d(unittest.TestCase):
  def test_conv2d_3d(self):
    w = _t(1, 1, 2, 2, 2).float()
    _check(self, _t(1, 1, 3, 3, 3).float(), lambda x: x.conv2d(w if isinstance(x, Tensor) else w.uop))
-  def test_conv2d_winograd(self):
-    w, a = _t(2, 2, 3, 3).float(), _t(1, 2, 6, 6).float()
-    with Context(WINO=0): direct = a.conv2d(w).uop
-    with Context(WINO=1):
-      self.assertIsNot(a.conv2d(w).uop, direct)
-      _check(self, a, lambda x: x.conv2d(w if isinstance(x, Tensor) else w.uop))
-  def test_conv2d_image(self):
-    w, a = _t(4, 4, 3, 3).float(), _t(1, 4, 8, 8).float()
-    with Context(IMAGE=0): direct = a.conv2d(w).uop
-    with Context(IMAGE=1):
-      self.assertIsNot(a.conv2d(w).uop, direct)
-      _check(self, a, lambda x: x.conv2d(w if isinstance(x, Tensor) else w.uop))
-  def test_dot_image(self):
-    y, a = _t(4, 3).float(), _t(2, 4).float()
-    with Context(IMAGE=0): direct = a.dot(y).uop
-    with Context(IMAGE=1):
-      self.assertIsNot(a.dot(y).uop, direct)
-      _check(self, a, lambda x: x.dot(y if isinstance(x, Tensor) else y.uop))
  def test_conv_transpose2d_basic(self):
    w = _t(1, 1, 2, 2).float()
    _check(self, _t(1, 1, 3, 3).float(), lambda x: x.conv_transpose2d(w if isinstance(x, Tensor) else w.uop))
@@ -494,10 +462,6 @@ class TestTensorUOpCreation(unittest.TestCase):
    self.assertIs(_strip_unique(Tensor.ones(2, 3).uop), _strip_unique(UOp.ones(2, 3)))
  def test_invalids(self):
    self.assertIs(_strip_unique(Tensor.invalids(2, 3, dtype=dtypes.int8).uop), _strip_unique(UOp.invalids((2, 3), dtype=dtypes.int8)))
-  def test_empty_like(self):
-    t = Tensor.empty(2, 3, dtype=dtypes.int8)
-    self.assertIs(_strip_unique(t.empty_like().uop), _strip_unique(t.uop.empty_like()))
-    self.assertIs(_strip_unique(t.empty_like(dtype=dtypes.float, device="NULL").uop), _strip_unique(t.uop.empty_like(dtypes.float, "NULL")))
  def test_arange(self):
    self.assertIs(Tensor.arange(5).uop, UOp.arange(5))
  def test_arange_empty(self):
@@ -2,7 +2,8 @@ import unittest, math
 import numpy as np
 from tinygrad import dtypes
 from tinygrad.uop.ops import UOp
-from tinygrad.codegen.decomp.transcendental import payne_hanek_reduction, cody_waite_reduction, frexp, rintk, pow2if
+from tinygrad.uop.decompositions import TRANSCENDENTAL_DTYPES, payne_hanek_reduction, cody_waite_reduction
+from tinygrad.uop.decompositions import frexp, rintk, xpow, xexp2, xlog2, trig_poly, pow2if
 from test.helpers import eval_uop

 class TestTranscendentalFunctions(unittest.TestCase):
@@ -70,5 +71,39 @@ class TestTranscendentalFunctions(unittest.TestCase):
    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, -10), dtypes.float)), 2**-10)
    np.testing.assert_allclose(eval_uop(pow2if(UOp.const(dtypes.int, -63), dtypes.float)), 2**-63)

+class TestTranscendentalVectorizedFunctions(unittest.TestCase):
+  # given a scalar and vectorized input, check that the fxn outputs have the same
+  # scalar_dtypes, args, ops, and vcount (only for vectorized input)
+
+  def _check_uop_vcount(self, u:tuple|UOp, vcount:int):
+    # check all UOps in u are vectorized with vcount
+    if isinstance(u, UOp):
+      assert u.dtype.vcount == vcount, f'expected {vcount=} but got {u.dtype.vcount=} for UOp\n{u=}'
+    [self._check_uop_vcount(x, vcount) for x in (u if isinstance(u, tuple) else u.src)]
+
+  def _check_uops_match(self, u1:tuple|UOp, u2:tuple|UOp):
+    # check all UOps in u1, u2 have the same scalar_dtype, args, ops
+    if isinstance(u1, UOp) and isinstance(u2, UOp):
+      assert u1.dtype.scalar() == u2.dtype.scalar(), f'expected {u1.dtype.scalar()=} but got {u2.dtype.scalar()=} for UOps\n{u1=}\n{u2}'
+      assert u1.arg == u2.arg or (math.isnan(u1.arg) and math.isnan(u2.arg)), f'expected {u1.arg=} but got {u2.arg=} for UOps\n{u1=}\n{u2}'
+      assert u1.op == u2.op, f'expected {u1.op=} but got {u2.op=} for UOps\n{u1=}\n{u2}'
+    [self._check_uops_match(x1, x2) for x1, x2 in zip((u1 if isinstance(u1, tuple) else u1.src), (u2 if isinstance(u2, tuple) else u2.src))]
+
+  def _test_vectorized(self, fxn, scalar_dtypes=TRANSCENDENTAL_DTYPES, vals=[-2,1.3,194], vcounts=[1,4,19]):
+    for scalar_dtype in scalar_dtypes:
+      for val in vals:
+        for vcount in vcounts:
+          in_scalar, in_vec = UOp.const(scalar_dtype, val), UOp.const(scalar_dtype.vec(vcount), val)
+          out_scalar, out_vec = fxn(in_scalar), fxn(in_vec)
+          self._check_uops_match(out_scalar, out_vec)
+          self._check_uop_vcount(out_vec, vcount)
+
+  def test_xpow(self): return self._test_vectorized(lambda x: xpow(x, x))
+  def test_xexp2(self): return self._test_vectorized(xexp2)
+  def test_xlog2(self): return self._test_vectorized(xlog2)
+  def test_payne_hanek_reduction(self): return self._test_vectorized(payne_hanek_reduction)
+  def test_cody_waite_reduction(self): return self._test_vectorized(cody_waite_reduction)
+  def test_trig_poly(self): return self._test_vectorized(lambda x: trig_poly(x, [0.0], [1.0]))
+
 if __name__ == '__main__':
  unittest.main()
@@ -35,17 +35,19 @@ class TestGraphRewriteConst(unittest.TestCase):
    self.assertEqual(ret.arg, 4)

  def test_add_const(self):
-    v1 = UOp.const(dtypes.int, (0,1,2))
-    v2 = UOp.const(dtypes.int, (5,6,7))
+    v1 = UOp.const(dtypes.int.vec(3), (0,1,2))
+    v2 = UOp.const(dtypes.int.vec(3), (5,6,7))
    ret = graph_rewrite(v1+v2, sym)
    self.assertEqual(ret.op, Ops.STACK)
+    self.assertEqual(ret.dtype, dtypes.int.vec(3))
    self.assertEqual(const_values(ret), (5,7,9))

  def test_add_const_lose_v(self):
-    v1 = UOp.const(dtypes.int, (0,1,2))
-    v2 = UOp.const(dtypes.int, (2,1,0))
+    v1 = UOp.const(dtypes.int.vec(3), (0,1,2))
+    v2 = UOp.const(dtypes.int.vec(3), (2,1,0))
    ret = graph_rewrite(v1+v2, sym)
    self.assertEqual(ret.op, Ops.STACK)
+    self.assertEqual(ret.dtype, dtypes.int.vec(3))
    self.assertEqual(const_values(ret), (2,2,2))

 def xfail_broken_const_wraparound(fn):
@@ -769,6 +771,45 @@ class TestMovementOps(unittest.TestCase):
    self.assertEqual(result.op, Ops.INDEX)
    self.assertEqual(result.src[0].op, Ops.RESHAPE)

+class TestLoadStoreFolding(unittest.TestCase):
+  def test_gated_load_gep_preserves_alt(self):
+    """Test that LOAD(GEP, alt) preserves alt value after rewrite"""
+    from tinygrad.codegen.late.devectorizer import load_store_folding
+    buf = UOp.param(0, dtypes.float.vec(4).ptr())
+    idx = UOp.const(dtypes.int, 0)
+    gate = UOp.const(dtypes.bool, True)
+    gated_index = buf.index(idx.valid(gate))
+    gep = gated_index.gep(0)
+    alt = UOp.const(dtypes.float, 42.0)
+    gated_load = gep.load(alt)
+    self.assertEqual(len(gated_load.src), 2)  # GEP + alt
+    result = graph_rewrite(gated_load, load_store_folding, name='test')
+    # After rewrite, should still have alt value preserved
+    self.assertEqual(result.op, Ops.GEP)
+    inner_load = result.src[0]
+    self.assertEqual(inner_load.op, Ops.LOAD)
+    self.assertEqual(len(inner_load.src), 2)  # INDEX + alt
+
+  def test_gated_load_ptrcat_preserves_alt(self):
+    """Test that LOAD(PTRCAT, alt) preserves alt value after rewrite"""
+    from tinygrad.codegen.late.devectorizer import load_store_folding
+    buf1 = UOp.param(0, dtypes.float.ptr())
+    buf2 = UOp.param(1, dtypes.float.ptr())
+    idx = UOp.const(dtypes.int, 0)
+    idx1 = buf1.index(idx)
+    idx2 = buf2.index(idx)
+    ptrcat = UOp(Ops.PTRCAT, dtypes.float.ptr().vec(2), (idx1, idx2))
+    alt = UOp.const(dtypes.float.vec(2), 42.0)
+    gated_load = ptrcat.load(alt)
+    self.assertEqual(len(gated_load.src), 2)  # PTRCAT + alt
+    result = graph_rewrite(gated_load, load_store_folding, name='test')
+    # After rewrite, should be CAT of LOADs, each preserving alt
+    self.assertEqual(result.op, Ops.VCAT)
+    for inner_load in result.src:
+      self.assertEqual(inner_load.op, Ops.LOAD)
+      self.assertEqual(len(inner_load.src), 2)  # INDEX + alt
+    self.assertEqual(inner_load.src[1].arg, 42.0)  # alt value preserved
+
 class TestConstBufferize(unittest.TestCase):
  def test_const_bufferize_with_ranges(self):
    """Test that CONST.BUFFERIZE with ranges is folded correctly.
@@ -5,7 +5,7 @@ from tinygrad.tensor import Tensor
 from tinygrad.helpers import Timing, Context, cdiv
 from tinygrad.dtype import dtypes, ConstFloat  # noqa: F401
 from tinygrad.device import Device
-from tinygrad.uop.ops import Ops, ParamArg, UOp, UPat, exec_alu  # noqa: F401  # ParamArg used by eval(str(uop)) roundtrip tests
+from tinygrad.uop.ops import Ops, UOp, UPat, exec_alu
 from tinygrad.uop.spec import spec_shared
 from tinygrad.uop.symbolic import sym
 from test.helpers import eval_uop, to_uops_list
@@ -185,7 +185,7 @@ class TestStatsOptimized(unittest.TestCase):
                      renderer=Device[Device.DEFAULT].renderer)
    except KernelOptError:
      raise unittest.SkipTest("no tensor cores")
-    print(p.src[2].arg)
+    print(p.src[3].arg)
    self.check_gemm(p)

  # this is a good lesson about why UPCASTing is a good idea
@@ -7,7 +7,7 @@ from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, TrackedPatternMatch
 from tinygrad.uop.symbolic import sym
 from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.helpers import colored, ansistrip, flatten, TracingKey, ProfileRangeEvent, ProfileEvent, Context, cpu_events, profile_marker
-from tinygrad.helpers import cpu_profile, ProfilePointEvent, unwrap, VIZ, BEAM
+from tinygrad.helpers import cpu_profile, ProfilePointEvent, unwrap, VIZ
 from tinygrad.device import Buffer

 from tinygrad.uop.ops import tracked_keys, tracked_ctxs, uop_fields, active_rewrites, active_group, _name_cnt, RewriteTrace
@@ -215,23 +215,6 @@ class TestViz(unittest.TestCase):
    nop = UOp(Ops.NOOP, arg="infinite loop in fixed_point_rewrite")
    self.assertEqual(graphs[2], uop_to_json(VizData(), nop)[id(nop)])

-  def test_walk_rewrite(self):
-    from tinygrad.uop.ops import _substitute
-    with save_viz() as viz:
-      a = UOp.variable("a", 0, 10)
-      graph_rewrite(a + 4, TrackedPatternMatcher(_substitute.patterns), {a:a+1}, walk=True)
-    list(viz.get_details(0, 0))
-
-  def test_enter_calls_rewrite(self):
-    pm = PatternMatcher([(UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4))])
-    with save_viz() as viz:
-      inner = UOp.const(dtypes.int, 3)
-      func = UOp(Ops.FUNCTION, src=(UOp(Ops.SINK, src=(inner,)),))
-      call = UOp(Ops.CALL, src=(func,))
-      graph_rewrite(call, TrackedPatternMatcher(pm.patterns), enter_calls=True)
-    details = list(viz.get_details(0, 0))
-    self.assertTrue(details[-1]["change"], "viz replay should detect change inside CALL")
-
  def test_const_node_visibility(self):
    with save_viz() as viz:
      a = UOp.variable("a", 0, 10, dtype=dtypes.int)
@@ -358,28 +341,41 @@ class TestVizGC(unittest.TestCase):
 from tinygrad import Tensor, Device, TinyJit, Variable, function

 class TestVizIntegration(unittest.TestCase):
-  def test_link_sched_codegen(self):
-    c1 = Tensor.empty(4, device="NULL")
-    c2 = Tensor.empty(8, device="NULL")
-    # uniquely named A = B + 1 kernel
-    kernel_name = f"custom_add1_link_sched_codegen_{BEAM.value}"
-    def custom_add1(A:UOp, B:UOp): return A[0].store(B[0]+1).sink(arg=KernelInfo(kernel_name))
+  # codegen supports rendering of code blocks
+  def test_codegen_tracing(self):
    with save_viz() as viz:
-      c1 = Tensor.custom_kernel(c1, c2, fxn=custom_add1)[0]
-      c1.realize()
+      ast = (Tensor.empty(4)+Tensor.empty(4)).schedule_linear().src[0].src[0]
+      prg = do_to_program(ast, Device[Device.DEFAULT].renderer)
+    lst = viz.list_items()
+    self.assertEqual(len(lst), 3)
+    self.assertEqual(lst[0]["name"], "Callify 1 Buffer n1")
+    self.assertEqual(lst[1]["name"], "Schedule 1 Kernel n1")
+    self.assertEqual(lst[2]["name"], prg.arg.name)
+    input_ast = next(viz.get_details(2, 0))["graph"].values()
+    for u in input_ast:
+      if u["label"].startswith("PARAM\n"): self.assertEqual(u["addrspace"], addrspace_colors[AddrSpace.GLOBAL])
+
+  # schedule graph CALL nodes have a link to jump to codegen
+  def test_link_sched_codegen(self):
+    with save_viz() as viz:
+      c1 = Tensor.empty(4, device="NULL").add(1)
+      c2 = Tensor.empty(8, device="NULL").add(1)
+      with Context(SCACHE=0):
+        sched = c1.schedule_linear(c2)
+      from tinygrad.engine.realize import compile_linear
+      sched = compile_linear(sched)
+      with Context(NO_COLOR=0):
+        prgs = [do_to_program(si.src[0], Device[c1.device].renderer).arg.name for si in sched.src]
    lst = viz.list_items()
-    # schedule graph CALL nodes have a link to jump to codegen
    sched_idx = next(i for i,l in enumerate(lst) if l["name"].startswith("Schedule"))
    viz_kernel = next(i for i,s in enumerate(lst[sched_idx]["steps"]) if s["name"] == "View Kernel Graph")
-    graph = next(viz.get_details(sched_idx, viz_kernel))["graph"]
+    with Context(NO_COLOR=1):
+      graph = next(viz.get_details(sched_idx, viz_kernel))["graph"]
    call_nodes = [n for n in graph.values() if n["label"].startswith("CALL")]
    for i,n in enumerate(call_nodes):
      assert n["ref"] is not None
-      self.assertEqual(lst[n["ref"]]["name"], kernel_name)
-      assert kernel_name[i] in n["label"], f"CALL must contain kernel name, got {n['label']}"
-    # UOp addrspace is colored
-    for u in graph.values():
-      if u["label"].startswith("PARAM\n"): self.assertEqual(u["addrspace"], addrspace_colors[AddrSpace.GLOBAL])
+      self.assertEqual(lst[n["ref"]]["name"], prgs[i])
+      assert ansistrip(prgs[i]) in n["label"], f"CALL must contain kernel name, got {n['label']}"

  def test_link_sched_codegen_beam(self):
    with Context(BEAM=2):
@@ -497,24 +493,6 @@ class TestVizIntegration(unittest.TestCase):
    bin_render = get_render(viz.data, steps[bin_idx]["query"])["src"]
    self.assertIn(type(e.exception).__name__, bin_render)

-  def test_view_source_alt(self):
-    src = "void E_3(float* data0_3) {}"
-    binary = Device["CPU"].renderer.compiler.compile(src)
-    def custom_binary(X:UOp):
-      sink = UOp.sink(X, arg=KernelInfo("custom_binary"))
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=sink.src+(sink,)), UOp(Ops.SOURCE, arg=src),
-                                   UOp(Ops.BINARY, arg=binary)))
-    x = Tensor.custom_kernel(Tensor.empty(1, device="CPU"), fxn=custom_binary)[0]
-    with save_viz() as viz:
-      x.realize()
-    lst = viz.list_items()
-    codegen_idx = len(lst)-1
-    steps = lst[codegen_idx]["steps"]
-    src_idx = next((i for i,s in enumerate(steps) if s["name"] == "View Source"), None)
-    assert src_idx is not None, "must have source rendering in list"
-    src_render = get_render(viz.data, steps[src_idx]["query"])["src"]
-    self.assertEqual(src, src_render)
-
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEvent, ProfileGraphEntry
 from tinygrad.viz.serve import get_profile
 from tinygrad.viz.cli import decode_profile
@@ -826,7 +804,7 @@ class TestCfg(unittest.TestCase):
      lidx = UOp.special(1, "lidx0")
      gidx = UOp.special(1, "gidx0")
      sink = UOp.sink(out.base, lidx, gidx, arg=KernelInfo(name=name))
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="NULL"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
    with save_viz() as viz:
      with Context(DEV=f"NULL::{self.arch}"):
        out = Tensor.custom_kernel(Tensor.empty(1), fxn=fxn)[0]
@@ -1111,10 +1089,8 @@ class TestCLI(unittest.TestCase):
    with write_files(viz) as files, Context(NO_COLOR=1):
      flat = run_cli(*files, "-s", "NULL", "--interval", "interval_start", "interval_end")
      aggregate = run_cli(*files, "-s", "NULL", "--interval", "interval_start", "interval_end", "-t")
-      final = run_cli(*files, "-s", "NULL", "--interval", "interval_end", "-t")
    self.assertEqual([s["name"] for s in flat], ["interval_start", "target_1", "target_2", "interval_end"])
    self.assertEqual(sorted(s["name"] for s in aggregate), ["target_1", "target_2"])
-    assert all(s["name"].startswith("post_") for s in final), f"post_* kernels must be present in final, got {final}"

 if __name__ == "__main__":
  unittest.main()
@@ -13,26 +13,35 @@ class TestWinograd(unittest.TestCase):
  def test_forward_kernels(self):
    x,w = Tensor.rand(1,4,9,9).realize(), Tensor.rand(4,4,3,3).realize()
    out = Tensor.conv2d(x,w)
-    self.assertEqual(len(out.schedule_linear().src), 4)
+    self.assertEqual(len(out.schedule_linear().src), 2)

  def test_backward_kernels(self):
    x,w = Tensor.empty(1,4,9,9).realize(), Tensor.empty(4,4,3,3).realize()
    out = Tensor.conv2d(x,w, padding=1)
    out.mean().backward()
    backward_schedule = x.grad.schedule_linear(w.grad)
-    self.assertEqual(len(backward_schedule.src), 4)
+    self.assertEqual(len(backward_schedule.src), 2)

+  @unittest.skip("this requires optimizations")
  def test_counters(self):
-    IC, OC, H = 64, 64, 28
-    x,w = Tensor.empty(1,IC,H,H,device="NULL").realize(), Tensor.empty(OC,IC,3,3,device="NULL").realize()
+    IC, OC, X, Y = 4,4,9,9
+    x,w = Tensor.rand(1,IC,Y,X).realize(), Tensor.rand(OC,IC,3,3).realize()
    GlobalCounters.reset()
-    with Context(NOOPT=0, WINO=1): Tensor.conv2d(x,w).realize()
-    ops_wino = GlobalCounters.global_ops
+    with Context(WINO=1):
+      Tensor.conv2d(x,w).realize()
+    ops_wino, mem_wino = GlobalCounters.global_ops, GlobalCounters.global_mem
    GlobalCounters.reset()
-    with Context(NOOPT=0, WINO=0): Tensor.conv2d(x,w).realize()
-    ops_normal = GlobalCounters.global_ops
-    print(f"ops: normal {ops_normal} wino {ops_wino} ratio {ops_wino/ops_normal:.2f}")
-    self.assertLess(ops_wino/ops_normal, 0.6)
+    with Context(WINO=0):
+      Tensor.conv2d(x,w).realize()
+    ops_normal, mem_normal = GlobalCounters.global_ops, GlobalCounters.global_mem
+
+    ops_ratio, mem_ratio = ops_wino/ops_normal, mem_wino/mem_normal
+    print(f"ops: normal {ops_normal:9d} wino {ops_wino:9d} ratio {ops_ratio:.2f}")
+    print(f"mem: normal {mem_normal:9d} wino {mem_wino:9d} ratio {mem_ratio:.2f}")
+
+    # TODO: what's optimal on this?
+    self.assertLess(ops_ratio, 4.3)
+    self.assertLess(mem_ratio, 4)

  def test_dtype(self):
    IC, OC, X, Y = 4,4,9,9
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
George HotzandGitHub	7a214c4499	Merge branch 'master' into clean_load	2026-06-19 16:56:57 -07:00
geohot	3e16109eb6	okay w/e	2026-06-18 21:00:36 -07:00
George HotzandGitHub	f79a7fc7c6	Merge branch 'master' into clean_load	2026-06-18 20:54:45 -07:00
geohot	3526f8272b	a few fixups	2026-06-18 20:53:30 -07:00
geohot	e143904deb	cleanup loads	2026-06-18 18:24:59 -07:00