IMAGE hand_coded_optimizations update (#16720)

* switch to the new memory coalese

* move that stuff

* copy in allowed length logic

* mulitple buffers

* new coalese is better

* fine

* earlier

* fixes

* work

* work

* valid

* stack on index const

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

@@ -80,7 +80,7 @@ runs:
     - name: Cache Python packages (PR)
       if: github.event_name == 'pull_request'
       id: restore-venv-pr
-      uses: actions/cache/restore@v4
+      uses: actions/cache/restore@v5
       with:
         path: /tmp/.uv-cache
         key: uv-${{ runner.os }}-${{ runner.arch }}-python-${{ inputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
@@ -96,7 +96,7 @@ runs:
 
     - name: Cache downloads (PR)
       if: inputs.key != '' && github.event_name == 'pull_request'
-      uses: actions/cache/restore@v4
+      uses: actions/cache/restore@v5
       with:
         path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
         key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
@@ -203,7 +203,7 @@ runs:
 
     - name: Cache apt (PR)
       if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true') && github.event_name == 'pull_request'
-      uses: actions/cache/restore@v4
+      uses: actions/cache/restore@v5
       with:
         path: /var/cache/apt/archives/
         key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}

.github/workflows/benchmark.yml

@@ -99,7 +99,6 @@ jobs:
         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/weights/LLaMA weights/LLaMA
         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'
@@ -134,32 +133,10 @@ jobs:
       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 LLaMA
-      run: |
-        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run quantized LLaMA
-      run: |
-        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8
-        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4
-    - name: Run quantized LLaMA3
-      run: |
-        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8
-        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4
-    #- name: Run LLaMA 7B on 4 (virtual) GPUs
-    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run GPT2
-      run: |
-        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run OLMoE
-      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.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
 
@@ -235,9 +212,6 @@ jobs:
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
         ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
         mkdir -p extra/datasets
         ln -s /raid/datasets/imagenet extra/datasets/imagenet
@@ -279,36 +253,16 @@ jobs:
     # 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 LLaMA
-      run: |
-        BENCHMARK_LOG=llama_nojit DEV=NV JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama DEV=NV JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
-    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    - 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 quantized LLaMA3
-      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8
     # - 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
-    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
-    - name: Run GPT2
-      run: |
-        BENCHMARK_LOG=gpt2_nojit DEV=NV JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 DEV=NV JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half DEV=NV HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam DEV=NV HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - uses: actions/upload-artifact@v7
       with:
         name: Speed (NVIDIA)
@@ -402,10 +356,7 @@ jobs:
       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
         ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
         mkdir -p extra/datasets
         ln -s /raid/datasets/imagenet extra/datasets/imagenet
@@ -458,18 +409,10 @@ jobs:
       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 LLaMA 7B
-      run: |
-        BENCHMARK_LOG=llama_nojit DEV=AMD JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama DEV=AMD JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run LLaMA 7B with BEAM
-      run: BENCHMARK_LOG=llama_beam DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
-    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    - 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
@@ -478,16 +421,6 @@ jobs:
     #  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 Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral DEV=AMD python3 examples/mixtral.py --temperature 0 --count 10 --timing
-    - name: Run GPT2
-      run: |
-        BENCHMARK_LOG=gpt2_nojit DEV=AMD JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 DEV=AMD JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half DEV=AMD HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam DEV=AMD HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
@@ -594,10 +527,10 @@ jobs:
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
-  testqualcommbenchmark:
-    name: comma Benchmark
+  testcommalatest:
+    name: comma Benchmark (0.11.0)
     runs-on: [self-hosted, Linux, comma]
-    timeout-minutes: 30
+    timeout-minutes: 10
     defaults:
       run:
         shell: bash -e -o pipefail {0}
@@ -622,6 +555,27 @@ jobs:
       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
       run: BENCHMARK_LOG=openpilot_0_11_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testcommaold:
+    name: comma Benchmark (0.10.1)
+    runs-on: [self-hosted, Linux, comma]
+    timeout-minutes: 10
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - 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: DEBUG=2 openpilot compile3 0.10.1 driving_vision
       run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
     - name: openpilot compile3 0.10.1 driving_vision
@@ -630,14 +584,44 @@ jobs:
       run: BENCHMARK_LOG=openpilot_0_10_1_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/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.10.1 dmonitoring
       run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testqualcommdsp:
+    name: DSP Benchmark
+    runs-on: [self-hosted, Linux, comma4]
+    timeout-minutes: 5
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - 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: Checkout Code
+      uses: actions/checkout@v6
+    - 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: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights
         ln -s /data/home/tiny/tinygrad/extra/datasets/imagenet extra/datasets/imagenet
         ln -s /data/home/tiny/tinygrad/testsig-*.so .
-        PYTHONPATH=. CC=clang-19 DEV=CPU QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
+        PYTHONPATH=. DEV=CPU QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
         # benchmark on DSP with NOOPT=1, the devectorizer has issues
-        PYTHONPATH=. CC=clang-19 DEV=DSP NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+        PYTHONPATH=. DEV=DSP NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 

.github/workflows/test.yml

@@ -133,46 +133,26 @@ 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 METAL tensor cores
+    - name: Test emulated tensor cores
       env:
-        DEV: 'PYTHON::METAL'
+        DEBUG: 2
+        N: 64
+        CNT: 1
+        SHOULD_USE_TC: 1
       run: |
-        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'
+        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
       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 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
+        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
         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:
@@ -218,14 +198,15 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
         deps: testing_unit
         llvm: 'true'
         amd: 'true'
     - name: Run NULL backend tests
       run: DEV=NULL python -m pytest -n=auto test/null/ --durations=20
     - name: Run targeted tests on NULL backend
-      run: DEV=NULL python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+      run: |
+        DEV=NULL python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+        DEV=NULL VIZ=1 python3 -m pytest -n=auto test/null/test_viz.py
     # TODO: too slow
     # - name: Run SDXL on NULL backend
     #   run: DEV=NULL DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
@@ -249,7 +230,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
+        pydeps: "pre-commit"
         deps: testing_unit
         llvm: 'true'
     - name: Run pre-commit test hooks
@@ -266,13 +247,6 @@ 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
 
@@ -337,31 +311,6 @@ 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
@@ -378,7 +327,7 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1468 ALLOWED_GATED_READ_IMAGE=18 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          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
       - 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
@@ -421,7 +370,6 @@ 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
@@ -430,7 +378,7 @@ 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_optim.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_lr_schedule.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
@@ -468,7 +416,7 @@ jobs:
 # ****** Models Tests ******
 
   testmodels:
-    name: Models (llvm+cpu+gpu)
+    name: Models
     runs-on: *linux
     timeout-minutes: 15
     steps:
@@ -479,34 +427,12 @@ jobs:
         with:
           key: models
           deps: testing
-          opencl: 'true'
           llvm: 'true'
-      - name: Test models (llvm)
-        run: DEV=CPU:LLVM python -m pytest -n=auto test/models --durations=20
-      - name: Test models (opencl)
-        run: DEV=CL python -m pytest -n=auto test/models --durations=20
       - name: Test models (cpu)
         run: DEV=CPU python -m pytest -n=auto test/models --durations=20
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
-  testmetalmodels:
-    name: Models (metal)
-    runs-on: &macos macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: metal
-        deps: testing
-    - name: Test models (Metal)
-      run: DEV=METAL python -m pytest -n=auto test/models --durations=20
-    - name: Test LLaMA compile speed
-      run: DEV=METAL python test/external/external_test_speed_llama.py
-
 # ****** Feature Tests ******
 
   testdsp:
@@ -715,7 +641,7 @@ jobs:
 
   unittestmacos:
     name: MacOS (unit)
-    runs-on: *macos
+    runs-on: &macos macos-26
     timeout-minutes: 20
     steps:
     - name: Checkout Code

README.md

@@ -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
+from tinygrad import Tensor, nn, Context
 
 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 Tensor.train():
+with Context(TRAINING=1):
   for i in range(10):
     optim.zero_grad()
     loss = model(x).sparse_categorical_crossentropy(y).backward()

docs/quickstart.md

@@ -165,13 +165,14 @@ 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 Tensor.train()` to set the internal flag `Tensor.training` to `True` during training.
+We use `with Context(TRAINING=1)` 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 Tensor.train():
+with Context(TRAINING=1):
   for step in range(1000):
     # random sample a batch
     samp = np.random.randint(0, X_train.shape[0], size=(64))

examples/beautiful_cartpole.py

@@ -1,6 +1,6 @@
 from typing import Tuple
 import time
-from tinygrad import Tensor, TinyJit, nn
+from tinygrad import Tensor, TinyJit, nn, Context
 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 Tensor.train():
+    with Context(TRAINING=1):
       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()
 

examples/beautiful_cifar.py

@@ -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
-  @Tensor.train()
+  @Context(TRAINING=1)
   def train_step(idxs:Tensor) -> Tensor:
     X, Y = X_train[idxs], Y_train[idxs]
     if len(GPUS) > 1:

examples/beautiful_mnist.py

@@ -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
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function, Context
 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
-  @Tensor.train()
+  @Context(TRAINING=1)
   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])

examples/beautiful_mnist_multigpu.py

@@ -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
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, Device, Context
 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 Tensor.train():
+    with Context(TRAINING=1):
       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

examples/gradaccum_mnist.py

@@ -1,6 +1,6 @@
 import itertools
 from typing import Callable
-from tinygrad import nn, Tensor, dtypes, Device, TinyJit
+from tinygrad import nn, Tensor, dtypes, Device, TinyJit, Context
 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
-  @Tensor.train()
+  @Context(TRAINING=1)
   def microbatch():
     samples = Tensor.randint(BS // ACC_STEPS, high=X_train.shape[0])
     for t in params: t.grad = None

examples/hlb_cifar10.py

@@ -359,7 +359,7 @@ def train_cifar():
   i = 0
   eval_acc_pct = 0.0
   batcher = fetch_batches(X_train, Y_train, BS=BS, is_train=True)
-  with Tensor.train():
+  with Context(TRAINING=1):
     st = time.monotonic()
     while i <= STEPS:
       if i % getenv("EVAL_STEPS", STEPS) == 0 and i > 1 and not getenv("DISABLE_BACKWARD"):

examples/llm.c/train_gpt2.py

@@ -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
+from tinygrad import Tensor, nn, fetch, Device, TinyJit, GlobalCounters, Context
 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
-  @Tensor.train()
+  @Context(TRAINING=1)
   def step(x:Tensor, y:Tensor) -> Tensor:
     _, loss = model(x, y)
     optimizer.zero_grad()
@@ -204,4 +204,3 @@ if __name__ == "__main__":
     top_k = 40
     y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
     print(decode(y[0].tolist()))
-

examples/minrf.py

@@ -1,5 +1,5 @@
 # much taken from https://github.com/cloneofsimo/minRF
-from tinygrad import Tensor, nn, GlobalCounters, TinyJit
+from tinygrad import Tensor, nn, GlobalCounters, TinyJit, Context
 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
-  @Tensor.train()
+  @Context(TRAINING=1)
   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])

examples/mlperf/model_eval.py

@@ -358,7 +358,7 @@ def eval_stable_diffusion():
       batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
     return batch, unpadded_bs 
 
-  @Tensor.train(mode=False)
+  @Context(TRAINING=0)
   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

examples/mlperf/model_train.py

@@ -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
+from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes, Context
 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
@@ -614,7 +614,7 @@ def train_retinanet():
 
       if getenv("RESET_STEP", 1): _train_step.reset()
 
-      with Tensor.train(mode=False):
+      with Context(TRAINING=0):
         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
-  @Tensor.train()
+  @Context(TRAINING=1)
   def train_step(model, x, y):
     optim.zero_grad()
 
@@ -795,7 +795,7 @@ def train_unet3d():
     optim.step()
     return loss.realize()
 
-  @Tensor.train(mode=False)
+  @Context(TRAINING=0)
   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)
@@ -1282,7 +1282,7 @@ def train_bert():
         previous_step = i
 
 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE
+  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8
   from examples.llama3 import MODEL_PARAMS
   from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
   from examples.mlperf.optim import GradAccClipAdamW
@@ -1447,7 +1447,12 @@ def train_llama3():
       idx = next(j for j, p in enumerate(optim.params) if p is w)
       master = optim.master_params[idx]
       inv = w._inv_scale if w._inv_scale.device == master.device else w._inv_scale.to(master.device)
-      master.assign((master * inv.reshape(*inv.shape, *([1]*(w.ndim-inv.ndim)))).contiguous())
+      if MXFP8:
+        from extra.gemm.cdna_asm_gemm import _mx_block_scale
+        bs = _mx_block_scale(inv.reshape(-1, inv.shape[-1])).reshape(w.shape)
+        master.assign((master * bs).contiguous())
+      else:
+        master.assign((master * inv.reshape(*inv.shape, *([1]*(w.ndim-inv.ndim)))).contiguous())
 
   # realize everything here
   if optim.master_params: Tensor.realize(*optim.master_params)
@@ -1485,7 +1490,7 @@ def train_llama3():
     return lr_cpu, grad_norm_cpu
 
   @TinyJit
-  @Tensor.train(False)
+  @Context(TRAINING=0)
   def eval_step(tokens:Tensor):
     if is_dp: tokens = tokens.to(None).shard(device, 0)
     if is_mp: tokens = tokens.shard(device)
@@ -1798,7 +1803,7 @@ if __name__ == "__main__":
   elif getenv("RUNMLPERF"): bench_log_manager = WallTimeEvent(BenchEvent.MLPERF_RUN)
   else: bench_log_manager = contextlib.nullcontext()
 
-  with Tensor.train():
+  with Context(TRAINING=1):
     for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
       nm = f"train_{m}"
       if nm in globals():

examples/mlperf/models/flat_llama.py

@@ -24,6 +24,7 @@ FUSED_ADD_NORM_MUL_QUANTIZE = getenv("FUSED_ADD_NORM_MUL_QUANTIZE", 0)
 FUSED_SILU_W13 = getenv("FUSED_SILU_W13", 0)
 SPLIT_W13 = getenv("SPLIT_W13", 0)
 COLUMNWISE_WEIGHT_SCALE = getenv("COLUMNWISE_WEIGHT_SCALE", 0)
+MXFP8 = getenv("MXFP8", 0)
 
 FP8_DTYPE = dtypes.fp8e4m3
 FP8_GRAD_DTYPE = dtypes.fp8e5m2
@@ -36,35 +37,48 @@ def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
   return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal(), new_amax
 
 def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
-           x_fp8:Tensor|None=None, x_scale:Tensor|None=None, x_new_amax:Tensor|None=None,
-           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
+           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,...]:
   if not fp8:
     if ASM_GEMM:
       from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
       if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T),)
     return (x @ w.T,)
   assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
+  if 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]
+    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])
+    else:
+      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*l_shape, x_q.shape[-1])
+      out = x_phys @ (w.cast(dtypes.bfloat16) * _mx_block_scale(w_inv_scale)).T
+    return out, (amax_x.detach() if amax_x is not None else None), x_q
   if x_fp8 is None:
     if FUSED_INPUT_QUANTIZE and amax_x is not None:
       from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed
-      x_fp8, x_scale, x_new_amax, _ = quantize_fp8_delayed(x, amax_x, FP8_DTYPE)
+      x_fp8, _, x_new_amax, _ = quantize_fp8_delayed(x, amax_x, FP8_DTYPE)
     else:
-      x_fp8, x_scale, x_new_amax = quantize_fp8(x, amax_state=amax_x)
+      x_fp8, _, x_new_amax = quantize_fp8(x, amax_state=amax_x)
   if ASM_GEMM:
     from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
     if can_use_asm_gemm(x_fp8, w.T):
+      assert amax_x is not None
       if COLUMNWISE_WEIGHT_SCALE:
-        out = asm_gemm(x_fp8, w.T, x_scale=x_scale, grad_amax_state=grad_amax_state, w_post_scale=w_inv_scale)
+        out = asm_gemm(x_fp8, w.T, x_scale=amax_x, grad_amax_state=grad_amax_state, w_post_scale=w_inv_scale)
       else:
-        out = asm_gemm(x_fp8, w.T, x_scale=x_scale, w_scale=w_inv_scale, grad_amax_state=grad_amax_state)
+        out = asm_gemm(x_fp8, w.T, x_scale=amax_x, w_scale=w_inv_scale, grad_amax_state=grad_amax_state)
       return out, x_new_amax, x_fp8
-  return (x_fp8.dot(w.T, dtype=dtypes.float) * x_scale * w_inv_scale).cast(dtypes.bfloat16), x_new_amax, x_fp8
+  return (x_fp8.dot(w.T, dtype=dtypes.float) * ((amax_x.float() + 1e-8) / FP8_MAX) * w_inv_scale).cast(dtypes.bfloat16), x_new_amax, x_fp8
 
 def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor, grad_amax_state:Tensor):
   if FUSED_ADD_NORM_MUL_QUANTIZE:
     from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_rmsnorm_mul_quantize_fp8
-    x_fp8, x_inv_scale, new_amax, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE)
-    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, x_scale=x_inv_scale, x_new_amax=new_amax, grad_amax_state=grad_amax_state)
+    x_fp8, new_amax, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE)
+    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x, x_new_amax=new_amax, grad_amax_state=grad_amax_state)
     return out, x_normed, rrms, ret
   x_normed, rrms = rmsnorm(x, eps)
   out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale, grad_amax_state=grad_amax_state)
@@ -74,8 +88,8 @@ def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w
                              grad_amax_state:Tensor|None=None):
   if FUSED_ADD_NORM_MUL_QUANTIZE:
     from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
-    x_fp8, x_inv_scale, new_amax, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE)
-    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, x_scale=x_inv_scale, x_new_amax=new_amax, grad_amax_state=grad_amax_state)
+    x_fp8, new_amax, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE)
+    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x, x_new_amax=new_amax, grad_amax_state=grad_amax_state)
     return out, h, x_normed, rrms, ret
   h = x + residual
   x_normed, rrms = rmsnorm(h, eps)
@@ -87,8 +101,8 @@ def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
                              grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
   if FUSED_SILU_W13:
     from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
-    x2_fp8, x2_inv_scale, new_amax_x2 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13)
-    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, x_scale=x2_inv_scale, x_new_amax=new_amax_x2, grad_amax_state=grad_amax_xout)
+    x2_fp8, new_amax_x2 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13)
+    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, amax_x=amax_x2, x_new_amax=new_amax_x2, grad_amax_state=grad_amax_xout)
     return out, ret
   hidden = x_w13.shape[-1] // 2
   x_w1, x_w3 = x_w13[..., :hidden], x_w13[..., hidden:]
@@ -140,12 +154,17 @@ class FlatTransformer:
     self._fp8_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
     w_scales = [("wqkv", s_qkv), ("wo", s_o), ("w2", s_2)]
     w_scales += [("w1", s_1), ("w3", s_3)] if SPLIT_W13 else [("w13", s_13)]
-    self._fp8_inv_scale = {name: s.float().contiguous().is_param_(False) for name, s in w_scales}
-    self._fp8_next_inv_scale = {name: s.float().contiguous().is_param_(False) for name, s in w_scales}
+    self._fp8_inv_scale = {name: (s if MXFP8 else s.float()).contiguous().is_param_(False) for name, s in w_scales}
+    self._fp8_next_inv_scale = {name: (s if MXFP8 else s.float()).contiguous().is_param_(False) for name, s in w_scales}
 
-  def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02):
-    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)
+  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)
+    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))
+      return w_q.reshape(self.n_layers, out_features, in_features), w_e8.reshape(self.n_layers, out_features, in_features // 32)
     amax = (w.abs().max(axis=2) if COLUMNWISE_WEIGHT_SCALE else w.abs().flatten(1).max(1)).detach()
     scale = FP8_MAX / (amax + 1e-8)
     inv_scale = (amax + 1e-8) / FP8_MAX
@@ -171,7 +190,7 @@ class FlatTransformer:
     xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)
     if getenv("HK_FLASH_ATTENTION"):
       from extra.thunder.amd.fa import flash_attention
-      attn, *save = flash_attention(xq, xk, xv, is_causal=True)
+      attn, *save = flash_attention(xq, xk, xv, is_causal=True, write_flat=True)
       saves.extend(save)
     else:
       xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
@@ -197,8 +216,15 @@ 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])
-      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"])
+      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"])
       amaxs.append(new_amax)
       saves.extend([*s, out])
     else:
@@ -228,20 +254,30 @@ 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):
-        getattr(self, name).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(getattr(self, name), self._fp8_inv_scale[name], self._fp8_next_inv_scale[name])
+      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)
       _shard_fp8("wqkv", 1)          # (n_layers, out, dim) shard out
-      _shard_fp8("wo", 2)            # (n_layers, dim, in) shard in
+      _shard_fp8("wo", 2, sstd)      # (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)            # (n_layers, dim, hidden) shard in
+      _shard_fp8("w2", 2, sstd)      # (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()

examples/mlperf/optim.py

@@ -8,6 +8,7 @@ STOCHASTIC_ROUND = getenv("STOCHASTIC_ROUND", 0)
 MASTER_WEIGHTS = getenv("MASTER_WEIGHTS", 0)
 FP8_AMAX_MARGIN = getenv("FP8_AMAX_MARGIN", 1.1)
 IMMEDIATE_SCALE = getenv("IMMEDIATE_SCALE", 0)
+MXFP8 = getenv("MXFP8", 0)
 
 def stochastic_round_bf16(x:Tensor) -> Tensor:
   bits = x.bitcast(dtypes.uint32)
@@ -90,6 +91,13 @@ class GradAccClipAdamW(Optimizer):
       out = stochastic_round_bf16(new_w)
       return out.shard_like(t) if offloaded else out
     if t.dtype in dtypes.fp8s:
+      if MXFP8:
+        from extra.gemm.cdna_asm_gemm import quantize_mxfp8
+        w_q, w_e8, _ = quantize_mxfp8(new_w.reshape(-1, new_w.shape[-1]))
+        new_e8 = w_e8.reshape(t._inv_scale.shape)
+        t._inv_scale.assign(new_e8.shard_like(t._inv_scale) if offloaded else new_e8)
+        ret = w_q.reshape(new_w.shape)
+        return ret.shard_like(t) if offloaded else ret
       from examples.mlperf.models.flat_llama import FP8_MAX
       if IMMEDIATE_SCALE:
         amax_axis = tuple(range(t._inv_scale.ndim, new_w.ndim))

examples/mnist_gan.py

@@ -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
+from tinygrad.helpers import trange, Context
 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 Tensor.train():
+  with Context(TRAINING=1):
     for epoch in (t := trange(epochs)):
       loss_g, loss_d = 0.0, 0.0
       for _ in range(n_steps):

examples/stunning_mnist.py

@@ -5,7 +5,7 @@
 #  - symbolic removal
 
 from examples.beautiful_mnist import Model
-from tinygrad import Tensor, nn, getenv, GlobalCounters, Variable
+from tinygrad import Tensor, nn, getenv, GlobalCounters, Variable, Context
 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 Tensor.train():
+  with Context(TRAINING=1):
     """
     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)

extra/export_model.py

@@ -1,5 +1,5 @@
 from typing import Tuple, Dict, List, Optional
-from tinygrad.dtype import DType, dtypes
+from tinygrad.dtype import DType, dtypes, AddrSpace
 from tinygrad.tensor import Tensor
 from tinygrad.device import Device, Buffer
 from tinygrad.engine.jit import TinyJit
@@ -39,7 +39,7 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li
     prg = to_program(call.src[0], Device[arg_uops[0].device].renderer)
     info = prg.arg
     functions[info.function_name] = prg.src[3].arg
-    cargs = [name_of(bu, i == 0) for i, bu in enumerate(arg_uops)] + [v for v in info.vars if v.op is Ops.DEFINE_VAR]
+    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))
 
   return functions, statements, {name:(size, dtype, key) for name, size, dtype, key in bufs.values()}, bufs_to_save
@@ -253,17 +253,18 @@ def export_model(model, target:str, *inputs, model_name: Optional[str] = "model"
   symbolic_vars = OrderedDict()
   for i, (_, args, global_size, _) in enumerate(statements):
     for j, var in enumerate(args):
-      if getattr(var, "op", None) is Ops.DEFINE_VAR and isinstance(getattr(var, "arg", None), tuple) and isinstance(var.arg[0], str):
+      if getattr(var, "op", None) is Ops.PARAM and var.addrspace is AddrSpace.ALU and var.arg.name is not None:
         if var not in symbolic_vars:
-          symbolic_vars[var] = var.arg[0]
+          symbolic_vars[var] = var.expr
           bufs[symbolic_vars[var]] = (var.dtype.itemsize, var.dtype, symbolic_vars[var])
         statements[i][1][j] = symbolic_vars[var]
 
     if global_size:
       for j, dim in enumerate(global_size):
-        if getattr(dim, "op", None) is Ops.ADD and len(dim.src) == 2 and {dim.src[0].op, dim.src[1].op} == {Ops.DEFINE_VAR, Ops.CONST}:
+        if getattr(dim, "op", None) is Ops.ADD and len(dim.src) == 2 and \
+           any(s.op is Ops.PARAM and s.addrspace is AddrSpace.ALU for s in dim.src) and any(s.op is Ops.CONST for s in dim.src):
           name, val = dim.src if dim.src[1].op is Ops.CONST else reversed(dim.src)
-          global_size[j] = f"_{name.arg[0]}[0] + {val.arg}"
+          global_size[j] = f"_{name.expr}[0] + {val.arg}"
 
   prg = ""
   if target == "clang":

extra/gemm/amd_asm_matmul.py

@@ -458,7 +458,8 @@ def test_matmul():
   def asm_kernel(A:UOp, B:UOp, C:UOp) -> UOp:
     gidxs = [UOp.special(n, f"gidx{i}") for i,n in enumerate(grid)]
     lidxs = [UOp.special(n, f"lidx{i}") for i,n in enumerate(local)]
-    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=max(LDS_SIZE, 65536//getenv("LIMIT_OCC", 65536)), addrspace=AddrSpace.LOCAL), (), 'lds')
+    lds_size = max(LDS_SIZE, 65536//getenv("LIMIT_OCC", 65536))
+    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.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

extra/gemm/amd_copy_matmul.py

@@ -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()))
+  acc = acc.after(acc.store(acc.zeros_like(buffer=False)))
 
   if use_wmma:
     k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)

extra/gemm/amd_flash_attention.py

@@ -126,7 +126,7 @@ def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
   P_lds = QP_lds[:, :BLOCK_N]
   P_write = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TN, LANES_PER_WAVE_N)
   P_write = P_write.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TN)
-  # TODO: P_write[tid].store(S_reg.cast(dtypes.half)) — shaped store fails due to RESHAPE(DEFINE_LOCAL) surviving linearization
+  # TODO: P_write[tid].store(S_reg.cast(dtypes.half)) -- shaped store fails due to RESHAPE(local BUFFER) surviving linearization
   rw1 = UOp.range(TM, 296, AxisType.LOOP)
   rw2 = UOp.range(TN, 297, AxisType.LOOP)
   P_store = P_write[tid, rw1, rw2].store(S_reg[rw1, rw2].cast(dtypes.half)).end(rw1, rw2)

extra/gemm/cdna_asm_gemm.py

@@ -2619,7 +2619,7 @@ def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
   lidx = UOp.special(WORKGROUP_SIZE, "lidx0")
   gidx = UOp.special(NUM_WG, "gidx0")
   insts = build_kernel(batch, M, N, K, A.dtype.base)
-  lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=133_120, addrspace=AddrSpace.LOCAL), (), 'lds')
+  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.DEVICE, arg=dname),
@@ -2630,7 +2630,7 @@ def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
 @functools.cache
 def custom_hk_fp8_gemm(C:UOp, A:UOp, B:UOp, *args:UOp, dname:str, scale_mode:int=3) -> UOp:
   # scale_mode: 0=no scale, 1=x only, 2=w only, 3=both
-  n_scales = (1 if scale_mode & 1 else 0) + (1 if scale_mode & 2 else 0)
+  n_scales = (1 if scale_mode & 1 else 0) + (1 if scale_mode & 2 else 0) + (1 if scale_mode & 4 else 0)
   scales, extra = args[:n_scales], args[n_scales:]
   M, K = A.shape[0]*A.shape[1], A.shape[2]
   N, K2 = B.shape[(1 if B.ndim == 3 else 0):]
@@ -2660,7 +2660,9 @@ def custom_hk_mxfp8_gemm(C:UOp, A:UOp, B:UOp, scale_A:UOp, scale_B:UOp, *extra:U
   block_size = 256
   threads = UOp.special(64 * 8, "lidx0")
   workgroups = UOp.special((M // block_size) * (N // block_size), "gidx0")
-  sink_inputs = (C.base, A.base, B.base, scale_A.base, scale_B.base, threads, workgroups)
+  e_a = extra[0].base if len(extra) >= 1 else scale_A.base
+  e_b = extra[1].base if len(extra) >= 2 else scale_B.base
+  sink_inputs = (C.base, A.base, B.base, scale_A.base, scale_B.base, e_a, e_b, threads, workgroups)
   sink = UOp.sink(*sink_inputs,
                   arg=KernelInfo(f"hk_mxfp8_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K, mem=(M*K+N*K)*A.dtype.itemsize+M*N*C.dtype.itemsize)))
   kittens_path = pathlib.Path(__file__).parent.parent/"thunder"/"amd"
@@ -2672,14 +2674,14 @@ def custom_hk_mxfp8_gemm(C:UOp, A:UOp, B:UOp, scale_A:UOp, scale_B:UOp, *extra:U
 
 def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
   # 1x32 block scaling along the last axis
-  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)
+  *batch, K = x.shape
+  scale_K = K // 32
+  amax = x.detach().float().reshape(*batch, 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(rows, scale_K, 1).expand(rows, scale_K, 32).reshape(rows, K)
+  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(*batch, scale_K, 1).expand(*batch, scale_K, 32).reshape(*batch, 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)
+  return x_clamped.cast(FP8_DTYPE), e8, (mx_pack(e8) if len(batch) == 1 else None)
 
 def mx_pack(e8:Tensor) -> Tensor:
   rows, scale_K = e8.shape
@@ -2750,7 +2752,8 @@ def custom_hk_bf16_gemm(C:UOp, A:UOp, B:UOp, *args:UOp, dname:str) -> UOp:
   assert M % block_m == 0 and N % block_n == 0 and K % block_k == 0, f"invalid bf16 tile {(block_m, block_n, block_k)} for {(M, N, K)}"
   threads = UOp.special(64 * num_warps, "lidx0")
   workgroups = UOp.special((M // block_m) * (N // block_n), "gidx0")
-  sink = UOp.sink(C.base, A.base, B.base, threads, workgroups,
+  b_extra = args[0].base if len(args) >= 1 else B.base
+  sink = UOp.sink(C.base, A.base, B.base, b_extra, threads, workgroups,
                   arg=KernelInfo(f"hk_bf16_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K, mem=(M*K+N*K+M*N)*A.dtype.itemsize)))
   kittens_path = pathlib.Path(__file__).parent.parent/"thunder"/"amd"
   src = (kittens_path/"gemm_bf16.cpp").read_text()
@@ -2785,15 +2788,21 @@ def hk_bf16_atb_gemm(a:Tensor, b:Tensor) -> Tensor:
   assert M % TILE_M == 0 and N % TILE_N == 0 and (batch * rows) % TILE_K == 0, \
     f"atb shape {a.shape} {b.shape} must produce (M,N,K) multiples of ({TILE_M},{TILE_N},{TILE_K})"
   is_multi = isinstance(a.device, tuple)
+  reduce_out = False
   if is_multi:
-    out = Tensor(Tensor.invalids(1, M, N, dtype=a.dtype, device=a.device).uop.multi(0), device=a.device)
+    ndev = len(a.device)
+    if a.uop.axis in (0, 1) or b.uop.axis in (0, 1): inv, out_axis, reduce_out = Tensor.invalids(1, M, N, dtype=a.dtype, device=a.device), 0, True
+    elif b.uop.axis == 2: inv, out_axis = Tensor.invalids(1, M, N // ndev, dtype=a.dtype, device=a.device), 2
+    elif a.uop.axis == 2: inv, out_axis = Tensor.invalids(1, M // ndev, N, dtype=a.dtype, device=a.device), 1
+    else: inv, out_axis, reduce_out = Tensor.invalids(1, M, N, dtype=a.dtype, device=a.device), 0, True
+    out = Tensor(inv.uop.multi(out_axis), device=a.device)
     dname = a.device[0]
   else:
     out = Tensor.invalids(1, M, N, dtype=a.dtype, device=a.device)
     dname = a.device
   dname = dname.split(":")[0]
   out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_hk_bf16_atb_gemm, dname=dname))[0]
-  if is_multi: out = out.sum(0)
+  if reduce_out: out = out.sum(0)
   return out.squeeze(0) if out.ndim == 3 else out
 
 
@@ -2805,13 +2814,15 @@ def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=
     out, a, b = inputs[:3]
     i = 3
     s_x = inputs[i]; i += 1
-    has_w = n_scales == 2
+    has_w = n_scales >= 2
     s_w = inputs[i] if has_w else None; i += has_w
+    s_g = inputs[i] if n_scales == 3 else None; i += (n_scales == 3)
     grad_amax_state = inputs[i] if has_grad_amax else None; i += has_grad_amax
     w_post = inputs[i] if has_w_post else None
     a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
     s_x_t = Tensor(s_x, device=a.device)
     s_w_t = Tensor(s_w, device=a.device) if has_w else None
+    s_g_t = Tensor(s_g, device=a.device) if s_g is not None else None
     w_post_t = Tensor(w_post, device=a.device) if has_w_post else None
     g_t = g_t[:a.shape[0]]
     from extra.llama_kernels.cast_amax import _grad_fp8_mailbox
@@ -2836,7 +2847,8 @@ def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=
         store_effect = grad_amax_state.store(new_grad_amax.uop)
         g_fp8 = Tensor(g_fp8.contiguous().uop.after(store_effect), device=a.device)
     # dgrad: uses g_scale * x_scale * w_scale (only when scalar)
-    grad_a = asm_gemm(g_fp8, b_t, x_scale=g_scale * s_x_t, w_scale=s_w_t) if has_w else asm_gemm(g_fp8, b_t, x_scale=g_scale * s_x_t)
+    if s_g_t is not None: g_scale = g_scale * s_g_t
+    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])
     if getenv("FAST_FP8_TRANSPOSE", 0) and g_fp8_2d.shape[0] % 64 == 0 and g_fp8_2d.shape[1] % 64 == 0:
@@ -2844,7 +2856,7 @@ def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=
       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=g_scale * s_x_t)
+    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:
       grad_b = grad_b / w_post_t.reshape(*w_post_t.shape, *([1]*(grad_b.ndim - w_post_t.ndim)))
@@ -2876,7 +2888,7 @@ def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=
 
 # ** mxfp8 gemm backward
 
-def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool):
+def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=False):
   inputs = kernel.src[1:]  # (out, a_q, b_q, a_si, b_si, a_e8, b_e8, [w_post])
   aq, bq = Tensor(inputs[1], device=inputs[1].device), Tensor(inputs[2], device=inputs[2].device)
   ae8, be8 = Tensor(inputs[5], device=inputs[5].device), Tensor(inputs[6], device=inputs[6].device)
@@ -2890,14 +2902,14 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool):
   grad_b = asm_gemm(g.T, a_phys, mx=True)
 
   grad_a = (grad_a * _mx_block_scale(ae8)).reshape(aq.shape)
-  grad_b = grad_b * _mx_block_scale(be8)
+  if not w_stored: grad_b = grad_b * _mx_block_scale(be8)
   if wp is not None: grad_b = grad_b / wp.reshape(-1, 1)
   return (None, grad_a.uop, grad_b.uop) + tuple(None for _ in inputs[3:])
 
 # ** 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) -> 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
@@ -2939,13 +2951,13 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
         b_q, b_e8, b_si = quantize_mxfp8(b.T)
       has_w_post = w_post_scale is not None
       fxn = functools.partial(custom_hk_mxfp8_gemm, dname=dname)
-      grad_fxn = functools.partial(custom_mx_gemm_bw, has_w_post=has_w_post)
+      grad_fxn = functools.partial(custom_mx_gemm_bw, has_w_post=has_w_post, w_stored=mx_w_stored)
       extra = [w_post_scale] if w_post_scale is not None else []
       out = Tensor.custom_kernel(out, a_q.reshape(a.shape), b_q, a_si, b_si, a_e8, b_e8, *extra, fxn=fxn, grad_fxn=grad_fxn)[0]
     # fp8 gemm computes a@b.T, kernel multiplies output by x_scale * w_scale before bf16 store
     elif a.dtype == FP8_DTYPE:
-      scales = tuple(s for s in (x_scale, w_scale) if s is not None)
-      scale_mode = (1 if x_scale is not None else 0) | (2 if w_scale is not None else 0)
+      scales = tuple(s for s in (x_scale, w_scale, g_scale) if s is not None)
+      scale_mode = (1 if x_scale is not None else 0) | (2 if w_scale is not None else 0) | (4 if g_scale is not None else 0)
       extra = ([grad_amax_state] if grad_amax_state is not None else []) + ([w_post_scale] if w_post_scale is not None else [])
       fxn = functools.partial(custom_hk_fp8_gemm, dname=dname, scale_mode=scale_mode)
       bw = functools.partial(custom_gemm_bw, n_scales=len(scales), has_grad_amax=grad_amax_state is not None, has_w_post=w_post_scale is not None)

extra/gemm/rdna4_asm_matmul.py

@@ -219,7 +219,8 @@ def test_matmul():
   def asm_kernel(A, B, C):
     gidxs = [UOp.special(n, f"gidx{i}") for i,n in enumerate(grid)]
     lidxs = [UOp.special(THREADS, "lidx0")]
-    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=max(LDS_SIZE, 65536//getenv("LIMIT_OCC",2)), addrspace=AddrSpace.LOCAL), (), 'lds')
+    lds_size = max(LDS_SIZE, 65536//getenv("LIMIT_OCC",2))
+    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.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

extra/hcq2/graph/hcq.py

@@ -1,145 +0,0 @@
-from __future__ import annotations
-import time
-from typing import cast
-from tinygrad.device import Buffer, BufferSpec, Compiled, Device, MultiBuffer
-from tinygrad.dtype import dtypes
-from tinygrad.engine.jit import GraphRunner
-from tinygrad.engine.realize import get_call_outs_ins, get_runtime
-from tinygrad.helpers import round_up, ceildiv
-from tinygrad.runtime.support.memory import BumpAllocator
-from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, graph_rewrite
-from extra.hcq2.hcq2 import HCQ2Compiled, HCQ2DeviceCtx, HCQ2LowerCtx, pm_prep_runtime, pm_lower_ops
-from extra.hcq2.hcq2 import pm_split_into_queues, pm_add_barriers, pm_add_signals
-from extra.hcq2.hcq2 import pm_bufferize, pm_lift_patches_to_cmdbuf, pm_resolve_patches, pm_parametrize_host_buffers
-from extra.hcq2.hcq2 import pm_add_timeline_inc, pm_callify, pm_calc_kernargs_sizes
-
-# **************** insert deps ****************
-
-def insert_deps(ctx:HCQ2Graph, linear:UOp) -> UOp:
-  src = []
-  for j, call in enumerate(linear.src):
-    call = call.replace(tag=j)
-    _, _, bufs, _ = ctx.calls[j]
-    outs, ins = get_call_outs_ins(call)
-    deps = ctx._access_resources([bufs[i] for i in outs + ins], list(range(len(outs))), call)
-    src.append(UOp(Ops.AFTER, call.dtype, (call, *deps), tag=call.tag))
-  return linear.replace(src=tuple(src))
-pm_insert_deps = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), insert_deps)])
-
-pm_replace_params = PatternMatcher([
-  (UPat(Ops.PARAM, name="p"), lambda ctx, p: ctx.input_addrs_uop.index(UOp.const(dtypes.int, p.arg))),
-  (UPat(Ops.SLICE, src=(UPat(Ops.INDEX, name="addr"), UPat(Ops.CONST, dtype=dtypes.weakint, name="off")), name="bv"),
-    lambda ctx, bv, addr, off: addr.cast(dtypes.uint64) + UOp.const(dtypes.uint64, off.arg * ctx.input_uops[addr.src[1].arg].dtype.itemsize)),
-])
-
-# **************** graph-only passes ****************
-
-def alloc_queue_sig(ctx:HCQ2Graph, q:UOp) -> None:
-  if q.arg in ctx.queue_sigs: return None
-  dev = q.arg[0][0]  # TODO: multi device
-  buf = Buffer(dev, 0x100, dtypes.uint8, options=BufferSpec(host=True, uncached=True, cpu_access=True), preallocate=True)
-  ctx.queue_sig_bufs.append(buf)
-  ctx.queue_sigs[q.arg] = UOp.from_buffer(buf, dev)
-  return None
-pm_alloc_queue_sigs = PatternMatcher([(UPat(Ops.LINEAR, src=UPat({Ops.PROGRAM, Ops.COPY}), name="q"), alloc_queue_sig)])
-
-def lower_queue_deps(ctx:HCQ2Graph, after:UOp) -> UOp:
-  wrapper, deps, call_idx = after.src[0], after.src[1:], after.tag
-  def store(q_arg, v): return ctx.queue_sigs[q_arg].store(UOp.const(dtypes.uint32, v))
-  waits = tuple(UOp(Ops.WAIT, dtypes.void, (ctx.queue_sigs[dep.src[0].arg], UOp.const(dtypes.uint32, dep.tag),
-                                            store(dep.src[0].arg, dep.tag))) for dep in deps)
-  return wrapper.replace(src=tuple(q.replace(src=(*waits, *q.src, store(q.arg, call_idx))) for q in wrapper.src))
-pm_lower_queue_deps = PatternMatcher([(UPat(Ops.AFTER, src=UPat(Ops.LINEAR), name="after"), lower_queue_deps)])
-
-def optimize_queue_deps(ctx:HCQ2Graph, queue:UOp) -> UOp|None:
-  src, seen, pending, queue_sig = [], {}, {}, ctx.queue_sigs[queue.arg]
-  for x in queue.src:
-    if x.op is Ops.WAIT:
-      sig, val = x.src[0], x.src[1]
-      if sig is queue_sig or seen.get(sig, -1) >= val.arg: continue
-      if (old:=pending.get(sig)) is None or old.src[1].arg < val.arg: pending[sig] = x
-      continue
-    for wait in pending.values():
-      src.append(wait)
-      seen[wait.src[0]] = wait.src[1].arg
-    pending.clear()
-    src.append(x)
-  src += pending.values()
-  return queue.replace(src=tuple(src)) if tuple(src) != queue.src else None
-pm_optimize_queue_deps = PatternMatcher([
-  (UPat(Ops.LINEAR, src=UPat({Ops.BARRIER, Ops.WAIT, Ops.STORE, Ops.PROGRAM, Ops.COPY}), name="queue"), optimize_queue_deps),
-])
-
-def drop_dead_stores(ctx:HCQ2Graph, outer:UOp) -> UOp:
-  live = {u.src[2] for u in outer.toposort() if u.op is Ops.WAIT}
-  return outer.replace(src=tuple(q.replace(src=tuple(x for x in q.src if x.op is not Ops.STORE or x in live)) for q in outer.src))
-pm_drop_dead_stores = PatternMatcher([(UPat(Ops.LINEAR, src=UPat(Ops.LINEAR), name="outer"), drop_dead_stores)])
-
-def add_queue_sig_resets(ctx:HCQ2Graph, x:UOp, cmdbuf:UOp) -> UOp|None:
-  if not ctx.queue_sig_bufs or cmdbuf.tag not in ("compute", "copy"): return None
-  resets = tuple((b:=UOp.from_buffer(sig)).index(UOp.const(dtypes.int, 0), dtype=b.dtype.ptr())
-                 .cast(dtypes.uint64.ptr()).store(UOp.const(dtypes.uint64, 0)) for sig in ctx.queue_sig_bufs)
-  return x.replace(src=x.src + resets)
-pm_add_queue_sig_resets = PatternMatcher([(UPat(Ops.AFTER, src=(UPat(Ops.BUFFER, name="cmdbuf"),), allow_any_len=True, name="x"),
-                                           add_queue_sig_resets)])
-
-# **************** Graph ****************
-
-class HCQ2Graph(GraphRunner):
-  def __init__(self, linear:UOp, input_uops:tuple[UOp, ...]=()):
-    super().__init__(linear, input_uops)
-    self.dev = cast(HCQ2Compiled, Device[self.device])
-    self.hcq_ctx = HCQ2LowerCtx(name="hcq_graph")
-
-    self.input_addrs = Buffer("CPU", max(len(input_uops), 1), dtypes.uint64, preallocate=True)
-    self.input_addrs_uop = UOp.from_buffer(self.input_addrs, "CPU")
-
-    self.linear = graph_rewrite(self.linear, pm_insert_deps, ctx=self, name="hcq: insert deps", walk=True)
-    self.linear = graph_rewrite(self.linear, pm_replace_params, ctx=self, name="hcq: replace params", walk=True)
-    self.linear = graph_rewrite(self.linear, pm_prep_runtime, ctx=self.hcq_ctx, name="hcq: prepare runtime")
-    self.linear = graph_rewrite(self.linear, pm_lower_ops, ctx=self.hcq_ctx, name="hcq: lower ops")
-
-    # per-queue signal state — populated as a side-effect by pm_alloc_queue_sigs walking the lowered linear.
-    self.queue_sig_bufs:list[Buffer] = []
-    self.queue_sigs:dict[tuple[str, str], UOp] = {}
-    graph_rewrite(self.linear, pm_alloc_queue_sigs, ctx=self, name="hcq: alloc queue sigs", walk=True)
-
-    self.linear = graph_rewrite(self.linear, pm_lower_queue_deps, ctx=self, name="hcq: lower queue deps")
-    self.linear = graph_rewrite(self.linear, pm_split_into_queues, ctx=self.hcq_ctx, name="hcq: split into queues")
-    self.linear = graph_rewrite(self.linear, pm_add_barriers, ctx=self.hcq_ctx, name="hcq: add barriers", walk=True)
-    self.linear = graph_rewrite(self.linear, pm_optimize_queue_deps, ctx=self, name="hcq: optimize queue deps", walk=True)
-    self.linear = graph_rewrite(self.linear, pm_drop_dead_stores, ctx=self, name="hcq: drop dead stores")
-    self.linear = graph_rewrite(self.linear, pm_add_signals, ctx=self.hcq_ctx, name="hcq: add signals", walk=True)
-    self.linear = graph_rewrite(self.linear, pm_add_timeline_inc, ctx=self.hcq_ctx, name="hcq: add submit", walk=True)
-    self.linear = graph_rewrite(self.linear, self.dev.pm_lower, ctx=self.hcq_ctx, name=f"hcq: encode cmdbuf {self.dev.device}", walk=True)
-
-    graph_rewrite(self.linear, pm_calc_kernargs_sizes, ctx=(sizes:={}), name=None)
-    for dev_name, sz in sizes.items():
-      buf = Buffer(dev_name, sz, dtypes.uint8, options=BufferSpec(cpu_access=True), preallocate=True)
-      self.hcq_ctx.dev_ctx[dev_name] = HCQ2DeviceCtx(dev_name, UOp.from_buffer(buf, dev_name), UOp.const(dtypes.uint64, buf._buf.va_addr))
-
-    self.linear = graph_rewrite(self.linear, pm_bufferize, ctx=self.hcq_ctx, bottom_up=True, name="realize binaries")
-    self.linear = graph_rewrite(self.linear, pm_lift_patches_to_cmdbuf, ctx=self.hcq_ctx, bottom_up=False, name="lift patches to cmdbuf")
-    self.linear = graph_rewrite(self.linear, pm_resolve_patches, ctx=self.hcq_ctx, bottom_up=False, name="simplify patches")
-    self.linear = graph_rewrite(self.linear, pm_add_queue_sig_resets, ctx=self, name="hcq: add queue sig resets", walk=True)
-    self.linear = graph_rewrite(self.linear, pm_parametrize_host_buffers, ctx=self.hcq_ctx, bottom_up=True, name="parametrize host buffers")
-    self.host_call = graph_rewrite(self.linear, pm_callify, ctx=self.hcq_ctx, name="hcq: callify")
-
-    self.host_rt, self.host_globals = get_runtime("CPU", self.host_call.src[0]), self.host_call.src[0].arg.globals
-
-  def __call__(self, input_uops:tuple[UOp, ...], var_vals:dict[str, int], wait=False) -> float|None:
-    addrs = self.input_addrs.as_memoryview(force_zero_copy=True).cast('Q')
-    for i, u in enumerate(input_uops):
-      buf = next(b for b in u.buffer.bufs if b.device == self.dev.device) if isinstance(u.buffer, MultiBuffer) else u.buffer
-      addrs[i] = buf._buf.va_addr
-    self.host_rt(*[self.hcq_ctx.inputs[i].get_buf("CPU") for i in self.host_globals], vals=self.host_call.src[0].arg.vals(var_vals), wait=True)
-    if wait:
-      st = time.perf_counter()
-      self.dev.synchronize()
-      return time.perf_counter() - st
-    return None
-
-  @staticmethod
-  def supports_uop(batch_devs:list[Compiled], new_call:UOp) -> bool:
-    all_devs = GraphRunner._all_devs(batch_devs, new_call)
-    return new_call.src[0].op in (Ops.PROGRAM, Ops.COPY) and len(all_devs) == 1 and isinstance(all_devs[0], HCQ2Compiled)

extra/hcq2/hcq2.py

@@ -1,14 +1,13 @@
 from __future__ import annotations
 from typing import cast, Callable, TypeVar, Generic, Any
 import struct, functools, time, collections, importlib, itertools, weakref
-from dataclasses import replace
-from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, mv_address, DEBUG, dedup, pluralize, to_tuple
+from dataclasses import replace, dataclass, field
+from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, DEBUG, dedup, flatten, pluralize
+from tinygrad.helpers import to_tuple, round_up
 from tinygrad.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer
 from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, track_rewrites, GroupOp
 from tinygrad.uop.symbolic import symbolic_simple, symbolic
-from tinygrad.dtype import dtypes, DType
-from dataclasses import dataclass, field
-from tinygrad.runtime.support.memory import BumpAllocator
+from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.runtime.support.hcq import MMIOInterface
 from tinygrad.renderer import Renderer, Estimates
 from tinygrad.engine.realize import to_program, get_call_arg_uops, get_call_name, get_call_outs_ins, estimate_uop, pm_flatten_linear
@@ -28,7 +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.arg, b.dtype, options=BufferSpec(host=True, 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)
@@ -137,21 +136,24 @@ def unwrap_after(uop):
   while uop.op is Ops.AFTER: uop = uop.src[0]
   return uop
 
-def make_getaddr(u, dev=None):
+def make_getaddr(u, device=None):
   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=dev or to_tuple(u.device)[0])))
+  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 if len(devs) > 1 else devs[0], len(blob), dtypes.uint8).rtag(tag)
-  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)
+  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])
 
 def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))
 
@@ -172,6 +174,10 @@ class HCQInfo:
   name:str = ""
   estimates:Estimates = Estimates()
   outs:tuple[int, ...] = ()
+  devs:tuple[str, ...] = ()
+
+  params:tuple[int, ...] = ()
+  inputs:int|None = None
 
   @staticmethod
   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])
@@ -208,15 +214,11 @@ 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 = 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:])
+  (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, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], dev_uop))) 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:])
 
 pm_prep_runtime = PatternMatcher([
   # bind generic PROGRAM device to the call's actual dev(s), then run device-specific lowering
@@ -269,6 +271,14 @@ class DepsCtx:
   deps:DepsTracker = field(default_factory=DepsTracker)
   opid:itertools.count = field(default_factory=lambda: itertools.count(0))
   last_per_queue:weakref.WeakValueDictionary[tuple[Any, str], UOp] = field(default_factory=weakref.WeakValueDictionary)
+  params:dict[tuple[int, int], Buffer] = field(default_factory=dict)
+
+def get_dep_buf(ctx:DepsCtx, u:UOp, lane:int) -> Buffer:
+  # TODO: should this be a part of DepsTracker?
+  if u.op is Ops.PARAM: return ctx.params.setdefault((u.arg.slot, lane), Buffer("NULL", u.max_numel(), u.dtype.base))
+  if u.op is Ops.MSTACK: return get_dep_buf(ctx, u.src[lane], 0)
+  if u.op in (Ops.SLICE, Ops.MSELECT): return get_dep_buf(ctx, u.src[0], u.arg if u.op is Ops.MSELECT else lane)
+  return b.bufs[lane] if isinstance(b:=u.buffer, MultiBuffer) else b
 
 def schedule_inner_sync(ctx:DepsCtx, linear:UOp) -> UOp:
   new_src = []
@@ -277,16 +287,26 @@ def schedule_inner_sync(ctx:DepsCtx, linear:UOp) -> UOp:
       new_src.append(call)
       continue
 
-    q = get_submit(call.src[0]).src[0]
-    new_q = ctx.last_per_queue[q.arg] = q.rtag(next(ctx.opid))
+    new_q = ctx.last_per_queue[q.arg] = (q:=get_submit(call.src[0]).src[0]).rtag(next(ctx.opid))
+    qdevs, refs = to_tuple(new_q.arg[0]), get_call_arg_uops(call)
 
-    deps = []
-    refs = [b.buffer for b in get_call_arg_uops(call)]
-    for lane in range(len(refs[0].bufs) if isinstance(refs[0], MultiBuffer) else 1):
-      deps += ctx.deps.access_resources([b.bufs[lane] if isinstance(b, MultiBuffer) else b for b in refs], call.arg.aux.outs, new_q)
+    # per-lane deps, tracked per (device, queue). skip self
+    dep_lanes:list[tuple[UOp, int]] = []
+    for lane, d in enumerate(qdevs):
+      for dep in ctx.deps.access_resources([get_dep_buf(ctx, b, lane) for b in refs], call.arg.aux.outs, new_q.replace(arg=(d, new_q.arg[1]))):
+        if dep.tag != new_q.tag: dep_lanes.append((dep, lane))
 
-    new_q = new_q.after(*dps).rtag("deps") if (dps:=dedup(deps)) else new_q
-    new_src.append(call.replace(src=(call.src[0].substitute({q:new_q}), *call.src[1:])))
+    # drop self-queue waits, queue self-orders
+    if qdevs[0].split(":")[0] in {"AMD", "QCOM"} or new_q.arg[1].startswith("COPY"):
+      dep_lanes = [(dep, lane) for dep, lane in dep_lanes if dep.arg != (qdevs[lane], new_q.arg[1])]
+
+    # keep latest dep per lane, group lanes
+    latest = {(dep.arg, lane): dep for dep, lane in sorted(dep_lanes, key=lambda x: x[0].tag)}
+    deps:dict[UOp, tuple[int, ...]] = collections.defaultdict(tuple)
+    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}),)))
   return linear.replace(src=tuple(new_src))
 pm_schedule_inner_sync = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), schedule_inner_sync)])
 
@@ -298,11 +318,16 @@ def make_finalizer(queues:list[UOp], nbump:int) -> UOp:
   zero = UOp.const(dtypes.int, 0)
   tl = make_signal_value(devs)
 
-  submit = make_submit(make_signal(devs).store(tl.index(zero) + 1), devs=devs, queue="COMPUTE:0")
+  # queue is inc with deps
+  submit = make_submit(make_signal(devs).store(tl.index(zero)), devs=devs, queue="COMPUTE:0")
+
+  # split each (multi-device) queue into per-device deps so each finalizer lane waits on the matching device's signal
+  lane_queues = [(q.replace(arg=(d, q.arg[1])), (devs.index(d),)) for q in queues for d in to_tuple(q.arg[0])]
+  submit = submit.replace(src=(submit.src[0].after(*(q for q, _ in lane_queues), arg=tuple(l for _, l in lane_queues)).rtag("deps"),))
 
   upd = [(tl, 1)] + [(make_signal_value(devs, queue=qn), nbump) for qn in dedup([q.arg[1] for q in queues])]
-  return UOp.barrier(*[s.after(submit).index(zero, dtype=s.dtype.ptr()).store(s.index(zero) + inc) for s, inc in upd]) \
-    .sink().call(aux=HCQInfo("hcq finalizer")).rtag("hcq")
+  patches = [s.after(submit).index(zero, dtype=s.dtype.ptr()).store(s.index(zero) + inc) for s, inc in upd]
+  return UOp.barrier(*patches).sink().call(aux=HCQInfo("hcq finalizer")).rtag("hcq")
 
 def add_finalizer(ctx:DepsCtx, linear:UOp) -> UOp:
   parts:dict[str, list[UOp]] = collections.defaultdict(list)
@@ -319,21 +344,22 @@ def add_loads(ctx:set[int], deps:UOp) -> UOp:
   cur_devs = to_tuple((cur:=deps.src[0]).arg[0])
 
   waits = []
-  for dep in deps.src[1:]:
-    devs, queue = dep.arg
+  for lanes, dep in zip(deps.arg, deps.src[1:]):
+    dep_dev, queue = dep.arg # dep_dev is a single device (deps are recorded per-device)
     ctx.add(dep.tag) # mark op to update signal.
 
-    sig = make_mstack([make_signal(d, queue=queue, sentinel=d not in devs) for d in cur_devs])
-    val = make_signal_value(cur_devs, queue=queue).index(UOp.const(dtypes.int, 0))
+    # for lanes that need this dep, wait on the dep device's signal/value; other lanes get a passing sentinel
+    lanes = set(lanes)
+    sig = make_mstack([make_signal(dep_dev if j in lanes else d, queue=queue, sentinel=j not in lanes) for j, d in enumerate(cur_devs)])
+    val = make_mstack([make_signal_value(dep_dev if j in lanes else d, queue=queue) for j, d in enumerate(cur_devs)]).index(UOp.const(dtypes.int, 0))
     waits.append(sig.wait(val + dep.tag))
   return cur.replace(src=tuple(waits) + cur.src)
 pm_add_inner_loads = PatternMatcher([(UPat(Ops.AFTER, tag="deps", name="deps"), add_loads)])
 
-def add_stores(ctx:set[int], submit:UOp, q:UOp) -> UOp:
-  src = q.src
-  if q.tag in ctx:
-    devs, queue = q.arg
-    src += (make_signal(devs, queue=queue).store(make_signal_value(devs, queue=queue).index(UOp.const(dtypes.int, 0)) + q.tag),)
+def add_stores(ctx:set[int], submit:UOp, q:UOp) -> UOp|None:
+  if q.tag not in ctx: return None
+  devs, queue = q.arg
+  src = q.src + (make_signal(devs, queue=queue).store(make_signal_value(devs, queue=queue).index(UOp.const(dtypes.int, 0)) + q.tag),)
   return submit.replace(src=(q.replace(src=src, tag=None),))
 pm_add_inner_stores = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_stores)])
 
@@ -342,34 +368,36 @@ pm_add_inner_stores = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit", s
 
 def get_submit(ast:UOp) -> UOp: return next(u for u in ast.toposort() if u.op is Ops.CUSTOM_FUNCTION and u.arg == "submit")
 
-def merge_sinks(old_sink:UOp, new_sink:UOp) -> UOp:
-  old_submit, new_submit = get_submit(old_sink), get_submit(new_sink)
-  old_queue, new_queue = old_submit.src[0], new_submit.src[0]
-  merged_submit = new_submit.replace(src=(new_queue.replace(src=old_queue.src + new_queue.src),))
-  old_root = old_sink.src[0].substitute({old_submit: merged_submit})
-  new_anchor = merged_submit if old_sink.src[0] is old_submit else old_root
-  return new_sink.substitute({new_submit: new_anchor})
+def merge_sink(sinks:list[UOp]) -> UOp:
+  if len(sinks) == 1: return sinks[0]
+  submits = [get_submit(sink) for sink in sinks]
+  queues = [submit.src[0] for submit in submits]
+  anchor = submits[-1].replace(src=(queues[-1].replace(src=tuple(x for q in queues for x in q.src)),))
+  for sink, submit in zip(sinks[:-1], submits[:-1]):
+    if sink.src[0] is not submit: anchor = sink.src[0].substitute({submit: anchor}, walk=True)
+  return sinks[-1].substitute({submits[-1]: anchor}, walk=True)
 
 def merge_queues(linear:UOp) -> UOp:
   new_src:list[UOp] = []
-  opened_qs:dict[tuple[tuple[str, ...], str], tuple[UOp, HCQInfo]] = {} # (devs, queue) -> (sink, aux), kept in submit order
+  opened_qs:dict[tuple[tuple[str, ...], str], tuple[list[UOp], HCQInfo]] = {} # (devs, queue) -> (sinks, aux), kept in submit order
 
   for call in linear.src:
-    if call.tag != "hcq":
-      new_src += [(sa:=opened_qs.pop(k))[0].call(aux=sa[1]).rtag('hcq') for k in list(opened_qs)] + [call]
+    # finalizer cannot be merged, since it bumps inner signal (this introduces race when multidevs).
+    if call.tag != "hcq" or (call.tag == "hcq" and call.arg.aux.name == "hcq finalizer"):
+      new_src += [merge_sink((sa:=opened_qs.pop(k))[0]).call(aux=sa[1]).rtag("hcq") for k in list(opened_qs)] + [call]
       continue
 
     devs, queue = get_submit(new_sink:=call.src[0]).src[0].arg
-    aux = call.arg.aux
+    new_rec = ([new_sink], call.arg.aux)
     if (old:=opened_qs.pop((devs, queue), None)) is not None:
-      new_sink = merge_sinks(old[0], new_sink) # exact same queue: merge, and re-insert at the end
-      aux = replace(aux, name=f"{queue.lower()} submit", estimates=old[1].estimates + aux.estimates)
+      new_rec = (old[0] + [new_sink], replace(new_rec[1], name=f"{queue.lower()} submit", estimates=old[1].estimates + new_rec[1].estimates))
     else:
       # no such queue opened: close every open submit on this queue that shares a device, so submit order is kept
-      new_src += [(sa:=opened_qs.pop(k))[0].call(aux=sa[1]).rtag('hcq') for k in [k for k in opened_qs if k[1] == queue and set(k[0]) & set(devs)]]
-    opened_qs[(devs, queue)] = (new_sink, aux)
+      closing = [k for k in opened_qs if k[1] == queue and set(k[0]) & set(devs)]
+      new_src += [merge_sink((sa:=opened_qs.pop(k))[0]).call(aux=sa[1]).rtag("hcq") for k in closing]
+    opened_qs[(devs, queue)] = new_rec
 
-  return linear.replace(src=tuple(new_src + [sink.call(aux=aux).rtag('hcq') for sink, aux in opened_qs.values()]))
+  return linear.replace(src=tuple(new_src + [merge_sink(sinks).call(aux=aux).rtag("hcq") for sinks, aux in opened_qs.values()]))
 pm_merge_queues = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), merge_queues)])
 
 # *****************
@@ -384,6 +412,29 @@ def add_global_sync(ctx:set[tuple[str, ...]], submit:UOp, q:UOp) -> UOp|None:
   return submit.replace(src=(q.replace(src=(UOp(Ops.BARRIER, dtypes.void), wait, *q.src)),))
 pm_add_global_sync = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_global_sync)])
 
+# *****************
+# 4.3. annotate exec devs
+
+pm_annotate_devs = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"),
+  lambda call: call.replace(arg=replace(call.arg, aux=replace(call.arg.aux, devs=get_submit(call.src[0]).src[0].arg[0]))))])
+
+# *****************
+# 4.4. replace params with per-submit input address loads
+
+def replace_params(call:UOp) -> UOp|None:
+  if not (params:={u:u.arg.slot for u in call.src[0].toposort() if u.op is Ops.PARAM and u.addrspace is AddrSpace.GLOBAL}): return None
+
+  # fill new info
+  hcqinfo = replace(call.arg.aux, params=tuple(sorted(set(params.values()))), inputs=len(get_call_arg_uops(call)))
+
+  inputs = UOp.new_buffer(get_submit(call.src[0]).src[0].arg[0], len(hcqinfo.params), dtypes.uint64).rtag("inputs")
+
+  slot2idx = {s:i for i,s in enumerate(hcqinfo.params)}
+  body = call.src[0].substitute({u:inputs.index(UOp.const(dtypes.int, slot2idx[s])).load() for u,s in params.items()})
+
+  return call.replace(src=(body, *call.src[1:], inputs), arg=replace(call.arg, aux=hcqinfo))
+pm_replace_params = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), replace_params)])
+
 # *****************
 # 5.1. encode cmdbufs
 
@@ -411,16 +462,28 @@ pm_lift_patches_to_cmdbuf = PatternMatcher([
 ])
 
 # *****************
-# 6. bufferize placeholders: replace placeholders with real buffers.
+# 5.3. pack placeholders buffers
 
-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), dev) for dev in to_tuple(buf.src[1].arg))
-  return make_mstack(uops)
-pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, name="buf"), bufferize_buf)])
+def pack_hcq_placeholders(call:UOp) -> UOp|None:
+  bufs = [b for b in call.src[0].toposort() if b.op is Ops.BUFFER and b.tag in (maxtags:={"scratch"}) | (sumtags:={"program", "kernargs"})]
+
+  off_per_buf:dict[UOp, int] = {}
+  size_per_tag:dict[str, int] = {}
+  for b in bufs:
+    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] + 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.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)])
 
 # *****************
-# 7.1. capture buffers reachable from each hcq call as BIND, so we don't drop their refs
+# 5.4. capture buffers reachable from each hcq call as BIND, so we don't drop their refs
 
 def hold_call_buffers(call:UOp) -> UOp|None:
   if not (bufs:=tuple(dedup(u for u in call.src[0].toposort() if u.op is Ops.BUFFER and u not in call.src))): return None
@@ -428,57 +491,83 @@ def hold_call_buffers(call:UOp) -> UOp|None:
 pm_hold_call_buffers = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), hold_call_buffers)])
 
 # *****************
-# 7.2. resolve patches
+# 6. bufferize placeholders: replace placeholders with real buffers.
+
+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.src[1].arg))
+  return make_mstack(uops)
+pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, name="buf"), bufferize_buf)])
+
+# *****************
+# 7. resolve patches
 
 def push_stack(op, s): return UOp(Ops.STACK, op.dtype.scalar().vec(len(s.src)),
   tuple(op.replace(dtype=op.dtype.scalar(), src=tuple(x if y is s else y for y in op.src)) for x in s.src))
 
 def fold_blob_store(buf:UOp, blob:UOp) -> UOp:
-  for b in (buf.src if buf.op is Ops.MSTACK else (buf,)): b.buffer.ensure_allocated()._buf.cpu_view().mv.cast('B')[:len(blob.arg)] = blob.arg
+  for b in (mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)): b.ensure_allocated()._buf.cpu_view().mv.cast('B')[:len(blob.arg)] = blob.arg
   return UOp(Ops.NOOP)
 
 def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
-  for b, v in zip((buf.src if buf.op is Ops.MSTACK else (buf,)), (val.src if val.op is Ops.STACK else (val,))):
-    struct.pack_into(f'<{v.dtype.fmt}', b.buffer.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * b.dtype.base.itemsize, v.arg)
+  for b, v in zip((bs:=mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)), val.src if val.op is Ops.STACK else (val,)*len(bs)):
+    struct.pack_into(f'<{v.dtype.fmt}', b.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * buf.dtype.base.itemsize, v.arg)
   return UOp(Ops.NOOP)
 
 def resolve_getaddr(buf:UOp, g:UOp) -> UOp:
-  if isinstance(b:=buf.buffer, Buffer): return UOp.const(dtypes.uint64, b.get_buf(g.src[1].arg).va_addr)
-  return UOp(Ops.STACK, dtypes.uint64.vec(len(b.bufs)), tuple(UOp.const(dtypes.uint64, x.ensure_allocated()._buf.va_addr) for x in b.bufs))
+  if buf.op not in (Ops.BUFFER, Ops.MSTACK, Ops.MSELECT): return buf
+  devs, b = to_tuple(g.src[1].arg), buf.buffer
+  bufs = tuple(cast(Buffer, x.buffer) for x in buf.src) if buf.op is Ops.MSTACK else tuple(b.bufs if isinstance(b, MultiBuffer) else (b,)*len(devs))
+  assert len(bufs) == len(devs), f"can't resolve {len(bufs)} buffers on {len(devs)} devices"
+  addrs = tuple(UOp.const(dtypes.uint64, x.get_buf(d).va_addr) for x, d in zip(bufs, devs))
+  return addrs[0] if len(addrs) == 1 else UOp(Ops.STACK, dtypes.uint64.vec(len(addrs)), addrs)
+
+def resolve_getaddr_slice(bv:UOp, dev:UOp) -> UOp:
+  itemsize = bv.src[0].dtype.itemsize if unwrap_after(bv.src[0]).op in (Ops.BUFFER, Ops.SLICE, Ops.MSTACK, Ops.MSELECT) else bv.dtype.itemsize
+  return UOp(Ops.GETADDR, dtypes.uint64, src=(bv.src[0], dev)) + UOp.const(dtypes.uint64, bv.src[1].arg * itemsize)
 
 pm_resolve_patches = PatternMatcher([
   # multi
   (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]))),
+
   # getaddr
-  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"), UPat(Ops.DEVICE, name="dev"))), # getaddr(slice(x)) -> offset+getaddr(x)
-    lambda bv, dev: UOp(Ops.GETADDR, dtypes.uint64, src=(bv.src[0], dev)) + UOp.const(dtypes.uint64, bv.src[1].arg * bv.src[0].dtype.itemsize)),
-  (UPat(Ops.GETADDR, src=(UPat({Ops.BUFFER, Ops.MSTACK, Ops.MSELECT}, name="buf"), UPat(Ops.DEVICE)), name="g"), resolve_getaddr),
+  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"), UPat(Ops.DEVICE, name="dev"))), resolve_getaddr_slice), # getaddr(slice(x)) -> offset+getaddr(x)
+  (UPat(Ops.GETADDR, src=(UPat(name="buf"), UPat(Ops.DEVICE)), name="g"), resolve_getaddr),
 
   # folders
-  (UPat({Ops.BUFFER, Ops.MSTACK}, name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
-  (UPat({Ops.BUFFER, Ops.MSTACK}, name="buf").index(UPat.cvar("off")).or_casted().store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))),
-    fold_const_store),
+  (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),
 ]) + symbolic_simple
 
 # *****************
 # 8. callify hcq programs
 
 def to_param(bufs:list[UOp], ref:UOp) -> UOp:
-  bufs.append(ref)
-  return UOp.placeholder((ref.buffer.size,), ref.dtype, len(bufs)-1)
+  if ref not in bufs: bufs.append(ref)
+  return UOp.placeholder((ref.buffer.size,), ref.dtype, bufs.index(ref))
 pm_to_param = PatternMatcher([(UPat({Ops.MSELECT, Ops.MSTACK, Ops.BUFFER}, name="r"), lambda ctx, r: to_param(ctx, r))])
 
 def parametrize_host_buffers(call:UOp) -> UOp:
-  body = graph_rewrite(call.src[0], pm_to_param, ctx=(bufs:=[]), bottom_up=True, name="parametrize host buffers")
-  return call.replace(src=(body, *bufs) + call.src[1:], tag="hcq_param")
+  # preserve original order of args
+  body = graph_rewrite(call.src[0], pm_to_param, ctx=(bufs:=list(get_call_arg_uops(call))), bottom_up=True, name="parametrize host buffers")
+
+  # move vars to new slots
+  var_slots = {nm:len(bufs)+i for i,nm in enumerate(sorted({v.expr for v in body.variables() if v.op is Ops.PARAM}))}
+  body = body.substitute({v:v.replace(arg=replace(v.arg, slot=var_slots[v.expr])) for v in body.variables() if v.op is Ops.PARAM})
+
+  return call.replace(src=(body, *bufs) + tuple(x for x in call.src[1:] if x.op is Ops.BIND))
 pm_parametrize_host_buffers = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), parametrize_host_buffers)])
 
 def callify_hcq(call:UOp) -> UOp:
   prg = to_program(call.src[0].sink(arg=KernelInfo("hcq_submit"), tag=1), Device["CPU"].renderer)
   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_param", name="call"), callify_hcq)])
+pm_callify_hcq = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), callify_hcq)])
 
 @track_rewrites(lambda _,ret: f"HCQ Schedule {pluralize('Kernel', len(ret.src))}")
 def hcq_schedule(linear:UOp) -> UOp:
@@ -492,14 +581,17 @@ def hcq_schedule(linear:UOp) -> UOp:
   linear = graph_rewrite(linear, pm_add_inner_stores, ctx=waited, walk=True, name="add stores", enter_calls=True)
   linear = graph_rewrite(linear, pm_merge_queues, name="merge queues")
   linear = graph_rewrite(linear, pm_add_global_sync, ctx=set(), walk=True, name="add global sync", enter_calls=True)
+  linear = graph_rewrite(linear, pm_annotate_devs, name="annotate devs")
+  linear = graph_rewrite(linear, pm_replace_params, name="replace params")
   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")
+  linear = graph_rewrite(linear, pm_hold_call_buffers, walk=True, name="hold call buffers")
 
   # realize starts from here
-  linear = graph_rewrite(linear, pm_bufferize, bottom_up=True, name="bufferize placeholders", enter_calls=True)
-  linear = graph_rewrite(linear, pm_hold_call_buffers, walk=True, name="hold call buffers")
+  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, name="parametrize host buffers")
+  linear = graph_rewrite(linear, pm_parametrize_host_buffers, walk=True, name="parametrize host buffers")
   linear = graph_rewrite(linear, pm_callify_hcq, name="callify hcq")
 
   return linear

extra/hcq2/ops_amd2.py

@@ -1,5 +1,5 @@
 from __future__ import annotations
-from typing import cast
+from typing import cast, Any, Callable
 import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit
 assert sys.platform != 'win32'
 from dataclasses import dataclass
@@ -89,25 +89,25 @@ def memory_barrier(ctx):
                  reg_done=getattr(ctx.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_DONE').addr[0], value=0xffffffff),
     acquire_mem(ctx)))
 
-def pm4_wait(ctx, dst, val): return wait_reg_mem(ctx, val, mem=make_getaddr(dst, ctx.device))
+def pm4_wait(ctx, dst, val): return wait_reg_mem(ctx, val, mem=make_getaddr(dst, ctx.devs))
 
 def pm4_barrier(ctx): return memory_barrier(ctx)
 
 def pm4_store(ctx, dst, val):
   if val.op is Ops.BINARY: return None
-  return release_mem(ctx, make_getaddr(dst, ctx.device), val, ctx.pm4.data_sel__mec_release_mem__send_32_bit_low,
+  return release_mem(ctx, make_getaddr(dst, ctx.devs), val, ctx.pm4.data_sel__mec_release_mem__send_32_bit_low,
                      ctx.pm4.int_sel__mec_release_mem__send_interrupt_after_write_confirm, cache_flush=True)
 
 def pm4_timestamp(ctx, dst):
-  return release_mem(ctx, make_getaddr(dst, ctx.device), 0, ctx.pm4.data_sel__mec_release_mem__send_gpu_clock_counter,
+  return release_mem(ctx, make_getaddr(dst, ctx.devs), 0, ctx.pm4.data_sel__mec_release_mem__send_gpu_clock_counter,
                      ctx.pm4.int_sel__mec_release_mem__none)
 
 def pm4_program(ctx, prg):
   data, info = prg.arg
   lib_gpu, args = prg.src
-  prog_addr = make_getaddr(lib_gpu, ctx.device) + data.entry_point_offset
-  scratch_addr = make_getaddr(UOp.new_buffer(lib_gpu.device, data.private_segment_size, dtypes.uint8).rtag("scratch"), ctx.device)
-  args_addr = make_getaddr(args, ctx.device)
+  prog_addr = make_getaddr(lib_gpu, ctx.devs) + data.entry_point_offset
+  scratch_addr = make_getaddr(UOp.new_buffer(lib_gpu.device, data.private_segment_size, dtypes.uint8).rtag("scratch"), ctx.devs)
+  args_addr = make_getaddr(args, ctx.devs)
 
   user_regs = []
   if data.enable_private_segment_sgpr:
@@ -174,7 +174,7 @@ pm_pm4_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"),
 class SDMAOps(FastEnum): COPY = auto(); POLL_REGMEM = auto(); FENCE = auto(); TRAP = auto(); TIMESTAMP = auto()  # noqa: E702
 
 def sdma_copy(ctx, dst, src, copy):
-  src_addr, dst_addr = make_getaddr(src, ctx.device), make_getaddr(dst, ctx.device)
+  src_addr, dst_addr = make_getaddr(src, ctx.devs), make_getaddr(dst, ctx.devs)
   return UOp(Ops.LINEAR, dtypes.void, tuple([make_ins(SDMAOps.COPY,
      ctx.sdma.SDMA_OP_COPY | ctx.sdma.SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_COPY_LINEAR),
      ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(copy.arg - off, ctx.max_copy_size) - 1), 0,
@@ -183,17 +183,17 @@ def sdma_copy(ctx, dst, src, copy):
 def sdma_wait(ctx, dst, val):
   op = ctx.sdma.SDMA_OP_POLL_REGMEM | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_FUNC(WAIT_REG_MEM_FUNCTION_GEQ) \
      | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)
-  return make_ins(SDMAOps.POLL_REGMEM, op, *data64_le(make_getaddr(dst, ctx.device)), val, 0xffffffff,
+  return make_ins(SDMAOps.POLL_REGMEM, op, *data64_le(make_getaddr(dst, ctx.devs)), val, 0xffffffff,
     ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(0x04) | ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff))
 
 def sdma_store(ctx, dst, val):
   op = ctx.sdma.SDMA_OP_FENCE | (ctx.sdma.SDMA_PKT_FENCE_HEADER_MTYPE(3) if ctx.target[0] != 9 else 0)
   return UOp(Ops.LINEAR, dtypes.void, (
-    make_ins(SDMAOps.FENCE, op, *data64_le(make_getaddr(dst, ctx.device)), val), make_ins(SDMAOps.TRAP, ctx.sdma.SDMA_OP_TRAP, 0)))
+    make_ins(SDMAOps.FENCE, op, *data64_le(make_getaddr(dst, ctx.devs)), val), make_ins(SDMAOps.TRAP, ctx.sdma.SDMA_OP_TRAP, 0)))
 
 def sdma_timestamp(ctx, dst):
   op = ctx.sdma.SDMA_OP_TIMESTAMP | ctx.sdma.SDMA_PKT_TIMESTAMP_GET_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_TIMESTAMP_GET_GLOBAL)
-  return make_ins(SDMAOps.TIMESTAMP, op, *data64_le(make_getaddr(dst, ctx.device)))
+  return make_ins(SDMAOps.TIMESTAMP, op, *data64_le(make_getaddr(dst, ctx.devs)))
 
 pm_sdma_opsel = PatternMatcher([
   (UPat(Ops.BARRIER), lambda: UOp(Ops.NOOP, dtypes.void, ())),
@@ -516,11 +516,17 @@ class PCIIface(PCIIfaceBase):
 
 def _mock(iface, name=None): return type(name or f"MOCK{iface.__name__}", (iface,), {})
 
+@dataclass(frozen=True)
+class AMDEncodeCtx:  # encode-time constants for one queue: devs (every cmdbuf address resolves into these) + gfx version + packet/ip modules
+  devs: tuple[str, ...]; target: tuple[int, ...]; pm4: Any; sdma: Any; soc: Any  # noqa: E702
+  gc: AMDIP; nbio: AMDIP; xccs: int; max_copy_size: int; tmpring_size: Callable  # noqa: E702
+
 def encode_queue(q:UOp) -> UOp|None:
   if not (isinstance(q.arg, tuple) and len(q.arg) == 2 and isinstance(q.arg[1], str) and q.arg[1].startswith(("COMPUTE", "COPY"))): return None
-  devs = to_tuple(q.arg[0])
+  d = Device[(devs:=to_tuple(q.arg[0]))[0]]
+  ctx = AMDEncodeCtx(devs, d.target, d.pm4, d.sdma, d.soc, d.gc, d.nbio, d.xccs, d.max_copy_size, d.tmpring_size)
   opsel, submit = (pm_pm4_opsel, pm_pm4_submit) if q.arg[1].startswith("COMPUTE") else (pm_sdma_opsel, pm_sdma_submit)
-  return submit.rewrite(graph_rewrite(q, opsel + pm_flatten_linear, walk=True, ctx=Device[devs[0]], name=f"{q.arg[1]} opsel"))
+  return submit.rewrite(graph_rewrite(q, opsel + pm_flatten_linear, walk=True, ctx=ctx, name=f"{q.arg[1]} opsel"))
 
 pm_lower = PatternMatcher([
   (UPat(Ops.CUSTOM_FUNCTION, arg="submit", src=(UPat(Ops.LINEAR, name="q"),)), encode_queue),

extra/llama_kernels/cast_amax/__init__.py

@@ -59,8 +59,8 @@ def _fused_quantize_bwd_w13(gradient:UOp, kernel:UOp):
   _grad_fp8_mailbox[grad_xw13_uop] = (grad_xw13_fp8_uop, inv_scale.uop)
   return (None, None, grad_xw13_uop, None, None)
 
-def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_state:Tensor) -> tuple[Tensor, Tensor, Tensor]:
-  # NOTE: silu(xw1)*xw3 -> fp8 + amax over fused xw13 layout. Returns (fp8, inv_scale, new_amax)
+def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_state:Tensor) -> tuple[Tensor, Tensor]:
+  # NOTE: silu(xw1)*xw3 -> fp8 + amax over fused xw13 layout. Returns (fp8, new_amax)
   # grad_amax_state: delayed amax for grad_xw13 fp8 quantization in the backward.
   assert xw13.dtype == dtypes.bfloat16, f"expected bf16, got {xw13.dtype}"
   MBS, SEQ, H2 = xw13.shape
@@ -72,5 +72,4 @@ def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_
   fxn = functools.partial(_custom_fused_cast_amax_w13, dname=dname_of(xw13.device))
   fp8_out, amax_buf, *_ = Tensor.custom_kernel(fp8_out, amax_buf, xw13, amax_state, grad_amax_state,
                                                 fxn=fxn, grad_fxn=_fused_quantize_bwd_w13)
-  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
-  return fp8_out, inv_scale, scalar_amax(amax_buf)
+  return fp8_out, scalar_amax(amax_buf)

extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/__init__.py

@@ -112,8 +112,8 @@ def _fused_add_bwd(*args, **kwargs):
   grad_h, grad_w = _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel)
   return (None, None, None, None, None, grad_h, grad_h, grad_w, None)
 
-def fused_rmsnorm_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor, Tensor]:
-  # NOTE: rmsnorm(x) * weight -> fp8 + amax. Returns (fp8, inv_scale, new_amax, x_normed, rrms).
+def fused_rmsnorm_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor]:
+  # NOTE: rmsnorm(x) * weight -> fp8 + amax. Returns (fp8, new_amax, x_normed, rrms).
   # x_normed + rrms are saved for the rmsnorm backward (also recomputed here from x regs).
   assert x.dtype == dtypes.bfloat16 and weight.dtype == dtypes.bfloat16
   assert x.shape[-1] == weight.shape[-1], f"HIDDEN mismatch: x={x.shape}, weight={weight.shape}"
@@ -127,13 +127,12 @@ def fused_rmsnorm_mul_quantize_fp8(x:Tensor, weight:Tensor, amax_state:Tensor, e
   fxn = functools.partial(_custom_fwd, dname=dname_of(x.device), eps_val=eps)
   fp8_out, x_normed_out, rrms_out, amax_buf, *_ = Tensor.custom_kernel(
     fp8_out, x_normed_out, rrms_out, amax_buf, x, weight, amax_state, fxn=fxn, grad_fxn=_fused_bwd)
-  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
-  return fp8_out, inv_scale, scalar_amax(amax_buf), x_normed_out, rrms_out
+  return fp8_out, scalar_amax(amax_buf), x_normed_out, rrms_out
 
 def fused_add_rmsnorm_mul_quantize_fp8(x:Tensor, residual:Tensor, weight:Tensor, amax_state:Tensor,
-                                       eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]:
+                                       eps:float, fp8_dtype) -> tuple[Tensor, Tensor, Tensor, Tensor, Tensor]:
   # NOTE: h = x + residual; y_normed = rmsnorm(h); fp8 = quantize(y_normed * weight).
-  # Returns (fp8, inv_scale, new_amax, h, x_normed, rrms). h is also written so downstream can
+  # Returns (fp8, new_amax, h, x_normed, rrms). h is also written so downstream can
   # reuse it without recomputing x+residual — eliminates the separate residual-add kernel.
   assert x.dtype == dtypes.bfloat16 and residual.dtype == dtypes.bfloat16 and weight.dtype == dtypes.bfloat16
   assert x.shape == residual.shape
@@ -149,5 +148,4 @@ def fused_add_rmsnorm_mul_quantize_fp8(x:Tensor, residual:Tensor, weight:Tensor,
   fp8_out, h_out, x_normed_out, rrms_out, amax_buf, *_ = Tensor.custom_kernel(
     fp8_out, h_out, x_normed_out, rrms_out, amax_buf, x, residual, weight, amax_state,
     fxn=fxn, grad_fxn=_fused_add_bwd)
-  inv_scale = (amax_state.float() + 1e-8) / FP8_MAX
-  return fp8_out, inv_scale, scalar_amax(amax_buf), h_out, x_normed_out, rrms_out
+  return fp8_out, scalar_amax(amax_buf), h_out, x_normed_out, rrms_out

extra/llama_kernels/fused_silu_mul_quantize_mxfp8/__init__.py

@@ -0,0 +1,104 @@
+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

extra/llama_kernels/quantize_fp8_delayed/__init__.py

@@ -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].load(UOp.const(dtypes.float, 0.0), active)
-    lds = lds.after(lds[tid].store(lds[tid].maximum(other), gate=active).barrier())
+    other = lds[(tid + step).valid(active)].load()
+    lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)).barrier())
     step //= 2
 
   amax_store = amax_partial[tid.eq(0).where(wg, UOp.invalid())].store(lds[0])

extra/llama_kernels/quantize_mxfp8_fused/__init__.py

@@ -0,0 +1,71 @@
+import functools
+from tinygrad import Tensor, dtypes
+from tinygrad.helpers import prod
+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
+
+@functools.cache
+def _custom_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x:UOp) -> UOp:
+  rows, K = x.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 = x.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
+
+  x_f = x[idx].cast(dtypes.float)
+  abs_x = (x_f < 0.0).where(-x_f, x_f)
+  blk_max = abs_x.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 = (x_f * 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)
+
+  # pack the 4 e8 of this super-block into one uint32 (little-endian: byte sb), write transposed (sk4, row)
+  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"quantize_mxfp8_{n_elems}", opts_to_apply=()))
+
+def _quantize_mxfp8_fused_bwd(gradient:UOp, kernel:UOp):
+  _, e8_out, _, x = kernel.src[1:]
+  device = x.device
+  rows, K = x.shape
+  scale_K = K // BLK
+  e8 = Tensor(e8_out, device=device).reshape(rows, scale_K)
+  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(rows, scale_K, 1).expand(rows, scale_K, BLK).reshape(rows, K)
+  grad_x = (Tensor(gradient, device=device).float() * qscale).cast(dtypes.bfloat16)
+  return (None, None, None, grad_x.uop)
+
+def quantize_mxfp8_fused(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  assert x.dtype == dtypes.bfloat16, f"expected bf16, got {x.dtype}"
+  assert x.ndim == 2, f"expected 2d (rows, K), got {x.shape}"
+  from extra.gemm.cdna_asm_gemm import FP8_DTYPE
+  rows, K = x.shape
+  scale_K = K // BLK
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  fp8_out = alloc_like((rows, K), FP8_DTYPE, x.device, axis)
+  e8_out = alloc_like((rows, scale_K), dtypes.uint8, x.device, axis)
+  si_out = alloc_like((scale_K // PACK, rows), dtypes.uint32, x.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, fxn=_custom_quantize_mxfp8, grad_fxn=_quantize_mxfp8_fused_bwd)
+  return fp8_out, e8_out, si_out

extra/sqtt/examples/generate_examples.py

@@ -9,8 +9,8 @@ EXAMPLES = {
   "empty":"test/backend/test_custom_kernel.py TestCustomKernel.test_empty",
   "plus":"test/test_tiny.py TestTiny.test_plus",
   "gemm":"-c \"from tinygrad import Tensor; (Tensor.empty(N:=32, N)@Tensor.empty(N, N)).realize()\"",
-  "sync":"test/amd/test_custom_kernel.py TestCustomKernel.test_lds_sync",
-  "handwritten":"test/amd/test_custom_kernel.py TestCustomKernel.test_handwritten",
+  "sync":"test/amd/test_asm_kernel.py TestAsmKernel.test_lds_sync",
+  "handwritten":"test/amd/test_asm_kernel.py TestAsmKernel.test_handwritten",
 }
 
 if __name__ == "__main__":

extra/testsig/generate_testsig.py

@@ -255,7 +255,7 @@ def _sign(serial_num, out_dir):
   final = _build_elf(ehdr, allph, segs)
 
   os.makedirs(out_dir, exist_ok=True)
-  out = os.path.join(out_dir, "testsig-0x%08X.so" % serial_num)
+  out = os.path.join(out_dir, "testsig-0x%08x.so" % serial_num)
   with open(out, 'wb') as f:
     f.write(final)
   print("Signing complete! Output saved at %s" % out)

extra/thunder/amd/fa.py

@@ -51,7 +51,8 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
     return None, None, dq.uop, dk.uop, dv.uop
   return grad
 
-def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False):
+# TODO: remove write_flat once scheduler can remove reshapes between custom_kernel. TestCustomKernel.test_simple_reshape
+def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False):
   assert attn_mask is None, "attn_mask not supported"
   assert is_causal, "only causal attention supported"
 
@@ -73,6 +74,7 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
   arch = Device[single_device].renderer.target.arch
 
   attn = _sharded_empty_like(xq, axis=shard_axis)
+  attn = _sharded_empty((B, N, H * D), xq, axis=shard_axis) if write_flat else _sharded_empty_like(xq, axis=shard_axis)
   l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
 
   grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch)

extra/thunder/amd/gemm_bf16.cpp

@@ -26,7 +26,7 @@ constexpr int NUM_THREADS = WARP_THREADS * NUM_WARPS;
 
 using G = kittens::group<NUM_WARPS>;
 
-__global__ __launch_bounds__(NUM_THREADS, 2) void hk_bf16_gemm(bf16 *C_ptr, bf16 *A_ptr, bf16 *B_ptr) {
+__global__ __launch_bounds__(NUM_THREADS, 2) void hk_bf16_gemm(bf16 *C_ptr, bf16 *A_ptr, bf16 *B_ptr, bf16 *b_unused) {
     constexpr int M = GEMM_M, N = GEMM_N, K = GEMM_K;
     static_assert(M % BLOCK_SIZE == 0 && N % BLOCK_SIZE == 0 && K % K_STEP == 0);
 

extra/thunder/amd/gemm_fp8.cpp

@@ -99,12 +99,14 @@ using G = kittens::group<NUM_WARPS>;
 #endif
 
 __global__ __launch_bounds__(512, 2) void hk_fp8_gemm(bf16 *C_ptr, fp8e4m3 *A_ptr, fp8e4m3 *B_ptr
-#if SCALE_MODE == 1
+#if SCALE_MODE & 1
     , float *x_scale_ptr
-#elif SCALE_MODE == 2
+#endif
+#if SCALE_MODE & 2
     , float *w_scale_ptr
-#elif SCALE_MODE == 3
-    , float *x_scale_ptr, float *w_scale_ptr
+#endif
+#if SCALE_MODE & 4
+    , float *g_scale_ptr
 #endif
 ) {
     constexpr int M = GEMM_M, N = GEMM_N, K = GEMM_K;
@@ -346,20 +348,19 @@ __global__ __launch_bounds__(512, 2) void hk_fp8_gemm(bf16 *C_ptr, fp8e4m3 *A_pt
     }
 
     // apply x_scale * w_scale before bf16 store to prevent overflow
-#if SCALE_MODE == 1
-    float scale = *x_scale_ptr;
-    mul(cA, cA, scale);
-    mul(cB, cB, scale);
-    mul(cC, cC, scale);
-    mul(cD, cD, scale);
-#elif SCALE_MODE == 2
-    float scale = *w_scale_ptr;
-    mul(cA, cA, scale);
-    mul(cB, cB, scale);
-    mul(cC, cC, scale);
-    mul(cD, cD, scale);
-#elif SCALE_MODE == 3
-    float scale = *x_scale_ptr * *w_scale_ptr;
+#if SCALE_MODE != 0
+    float scale = 1.0f;
+#if SCALE_MODE & 1
+    float x_scale = (*x_scale_ptr + 1e-08f) * (1.0f / 448.0f);
+    scale *= x_scale;
+#endif
+#if SCALE_MODE & 2
+    scale *= *w_scale_ptr;
+#endif
+#if SCALE_MODE & 4
+    scale *= *g_scale_ptr;
+#endif
+
     mul(cA, cA, scale);
     mul(cB, cB, scale);
     mul(cC, cC, scale);

extra/thunder/amd/gemm_mxfp8.cpp

@@ -26,9 +26,9 @@ constexpr int REG_N      = BLOCK_COL / WARPS_COL / 2;
 
 using G = kittens::group<NUM_WARPS>;
 
-__global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3 *A_ptr, fp8e4m3 *B_ptr,
-    const uint32_t *__restrict__ scale_A_iter,
-    const uint32_t *__restrict__ scale_B_iter) {
+__global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3 *A_ptr, fp8e4m3 *B_ptr, fp8e8m0 *scale_A_ptr, fp8e8m0 *scale_B_ptr,
+    const uint8_t *__restrict__ a_e8_unused,
+    const uint8_t *__restrict__ b_e8_unused) {
     constexpr int M = GEMM_M, N = GEMM_N, K = GEMM_K;
 
     kittens::gl<fp8e4m3, 1, 1, M, K> A{A_ptr, nullptr, nullptr, nullptr, nullptr};
@@ -38,15 +38,19 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
     constexpr int k_iters      = K / BLOCK_K;
     constexpr int NUM_THREADS  = NUM_WARPS * WARP_THREADS;
 
-    using ST_A = st_fp8e4m3<HALF_ROW, BLOCK_K, st_16x128_s>;
-    using ST_B = st_fp8e4m3<HALF_COL, BLOCK_K, st_16x128_s>;
-    using RT_A = rt_fp8e4m3<REG_M, BLOCK_K>;
-    using RT_B = rt_fp8e4m3<REG_N, BLOCK_K>;
-    using RT_C = rt_fl<REG_M, REG_N, col_l, rt_16x16_s>;
+    kittens::gl<fp8e8m0, k_iters * (M / BLOCK_ROW), 1, 16, 64> scale_A_gl{scale_A_ptr, nullptr, nullptr, nullptr, nullptr};
+    kittens::gl<fp8e8m0, k_iters * (N / BLOCK_COL), 1, 16, 64> scale_B_gl{scale_B_ptr, nullptr, nullptr, nullptr, nullptr};
+
+    using ST_A     = st_fp8e4m3<HALF_ROW, BLOCK_K, st_16x128_s>;
+    using ST_B     = st_fp8e4m3<HALF_COL, BLOCK_K, st_16x128_s>;
+    using ST_Scale = st<fp8e8m0, 16, 64, st_16x64_s>;
+    using RT_A     = rt_fp8e4m3<REG_M, BLOCK_K>;
+    using RT_B     = rt_fp8e4m3<REG_N, BLOCK_K>;
+    using RT_C     = rt_fl<REG_M, REG_N, col_l, rt_16x16_s>;
 
     __shared__ ST_A As[2][2];
     __shared__ ST_B Bs[2][2];
-    __shared__ uint8_t smem_scales[2048];
+    __shared__ ST_Scale scale_A_smem[2], scale_B_smem[2];
 
     RT_A a;
     RT_B b0, b1;
@@ -79,54 +83,87 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
     G::prefill_swizzled_offsets(As[0][0], A, sw_A);
     G::prefill_swizzled_offsets(Bs[0][0], B, sw_B);
 
+    const T *a_base = (const T *)&A[{0, 0, 0, 0}];
+    const T *b_base = (const T *)&B[{0, 0, 0, 0}];
+    const int a_row_stride = A.template stride<2>() * sizeof(T);
+    const int b_row_stride = B.template stride<2>() * sizeof(T);
+    i32x4 a_srd = make_srsrc(a_base, M * a_row_stride, a_row_stride);
+    i32x4 b_srd = make_srsrc(b_base, N * b_row_stride, b_row_stride);
+
+    const int wid = warpid() % NUM_WARPS;
+    constexpr int elem_per_warp = (16 / sizeof(T)) * kittens::WARP_THREADS;
+    uint32_t a_lds_00 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[0][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t a_lds_01 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[0][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t a_lds_10 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[1][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t a_lds_11 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[1][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_00 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[0][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_01 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[0][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_10 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[1][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_11 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[1][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+
     int a_row_h0 = warp_m * REG_M;
     int a_row_h1 = HALF_ROW + warp_m * REG_M;
     int b_row_h0 = warp_n * REG_N;
     int b_row_h1 = HALF_COL + warp_n * REG_N;
 
     int tic = 0, toc = 1;
+    int tic_scales = 0, toc_scales = 1;
 
-    G::load(Bs[tic][0], B, {0, 0, block_col * 2, 0}, sw_B);
-    G::load(As[tic][0], A, {0, 0, block_row * 2, 0}, sw_A);
-    G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, 0}, sw_B);
-    G::load(As[tic][1], A, {0, 0, block_row * 2 + 1, 0}, sw_A);
+    uint32_t a_lds[2][2] = {{a_lds_00, a_lds_01}, {a_lds_10, a_lds_11}};
+    uint32_t b_lds[2][2] = {{b_lds_00, b_lds_01}, {b_lds_10, b_lds_11}};
+
+    if constexpr (k_iters >= 6 && (k_iters % 2 == 0)) {
+    G::load(Bs[tic][0], B, {0, 0, block_col * 2, 0}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][0]));
+    G::load(As[tic][0], A, {0, 0, block_row * 2, 0}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[tic][0]));
+    G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, 0}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][1]));
+    G::load(As[tic][1], A, {0, 0, block_row * 2 + 1, 0}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[tic][1]));
 
     if (warp_m == 1) __builtin_amdgcn_s_barrier();
     asm volatile("s_waitcnt vmcnt(4)");
     __builtin_amdgcn_s_barrier();
 
-    G::load(As[toc][0], A, {0, 0, block_row * 2, 1}, sw_A);
-    G::load(Bs[toc][0], B, {0, 0, block_col * 2, 1}, sw_B);
-    G::load(Bs[toc][1], B, {0, 0, block_col * 2 + 1, 1}, sw_B);
-
+    G::load(As[toc][0], A, {0, 0, block_row * 2, 1}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[toc][0]));
+    G::load(Bs[toc][0], B, {0, 0, block_col * 2, 1}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[toc][0]));
+    G::load(Bs[toc][1], B, {0, 0, block_col * 2 + 1, 1}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[toc][1]));
     asm volatile("s_waitcnt vmcnt(6)");
     __builtin_amdgcn_s_barrier();
 
-    #pragma unroll 2
-    for (int k = 0; k < k_iters - 2; k++, tic ^= 1, toc ^= 1) {
-        load_scales_to_lds(smem_scales, scale_A_iter, scale_B_iter, block_m, block_n, k, M, N);
+    G::load(scale_A_smem[0], scale_A_gl, {0 * tiles_M + block_row, 0, 0, 0});
+    G::load(scale_B_smem[0], scale_B_gl, {0 * tiles_N + block_col, 0, 0, 0});
+    asm volatile("s_waitcnt vmcnt(0)");
+    asm volatile("s_waitcnt lgkmcnt(0)");
+    __builtin_amdgcn_s_barrier();
 
+    int sa_idx = block_row, sb_idx = block_col;
+
+    #pragma unroll 2
+    for (int k = 0; k < k_iters - 2; k++, tic ^= 1, toc ^= 1, tic_scales ^= 1, toc_scales ^= 1) {
+        if (k + 1 < k_iters) {
+            sa_idx += tiles_M; sb_idx += tiles_N;
+            G::load(scale_A_smem[toc_scales], scale_A_gl, {sa_idx, 0, 0, 0});
+            G::load(scale_B_smem[toc_scales], scale_B_gl, {sb_idx, 0, 0, 0});
+        }
         auto bs0 = subtile_inplace<REG_N, BLOCK_K>(Bs[tic][0], {warp_n, 0});
         load(b0, bs0);
         auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[tic][0], {warp_m, 0});
         load(a, as0);
-        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A);
+        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[toc][1]));
         asm volatile("s_waitcnt lgkmcnt(8)");
         __builtin_amdgcn_s_barrier();
 
-        fp8e8m0_4 sa_h0 = pack_scales(smem_scales, 0, a_row_h0);
-        fp8e8m0_4 sb_h0 = pack_scales(smem_scales, 1024, b_row_h0);
-
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[tic_scales].data, a_row_h0);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[tic_scales].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[tic_scales].data, b_row_h1);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[tic_scales].data, a_row_h1);
         __builtin_amdgcn_s_setprio(2);
         mma_ABt_scaled(cA, a, b0, cA, &sa_h0, &sb_h0);
         __builtin_amdgcn_s_setprio(0);
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
 
-        fp8e8m0_4 sb_h1 = pack_scales(smem_scales, 1024, b_row_h1);
         auto bs1 = subtile_inplace<REG_N, BLOCK_K>(Bs[tic][1], {warp_n, 0});
         load(b1, bs1);
-        G::load(As[tic][0], A, {0, 0, block_row * 2, k + 2}, sw_A);
+        G::load(As[tic][0], A, {0, 0, block_row * 2, k + 2}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[tic][0]));
         asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
 
@@ -135,10 +172,9 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_s_setprio(0);
         __builtin_amdgcn_s_barrier();
 
-        fp8e8m0_4 sa_h1 = pack_scales(smem_scales, 0, a_row_h1);
         auto as1 = subtile_inplace<REG_M, BLOCK_K>(As[tic][1], {warp_m, 0});
         load(a, as1);
-        G::load(Bs[tic][0], B, {0, 0, block_col * 2, k + 2}, sw_B);
+        G::load(Bs[tic][0], B, {0, 0, block_col * 2, k + 2}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][0]));
         asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
 
@@ -148,7 +184,7 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
 
-        G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, k + 2}, sw_B);
+        G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, k + 2}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][1]));
         asm volatile("s_waitcnt vmcnt(6)");
         __builtin_amdgcn_s_barrier();
 
@@ -160,19 +196,24 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
 
     { // Epilogue k = k_iters - 2
         int k = k_iters - 2;
-        load_scales_to_lds(smem_scales, scale_A_iter, scale_B_iter, block_m, block_n, k, M, N);
+        if (k + 1 < k_iters) {
+            sa_idx += tiles_M; sb_idx += tiles_N;
+            G::load(scale_A_smem[toc_scales], scale_A_gl, {sa_idx, 0, 0, 0});
+            G::load(scale_B_smem[toc_scales], scale_B_gl, {sb_idx, 0, 0, 0});
+        }
+        asm volatile("s_waitcnt vmcnt(0)");
+        asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
-
-        fp8e8m0_4 sa_h0 = pack_scales(smem_scales, 0, a_row_h0);
-        fp8e8m0_4 sa_h1 = pack_scales(smem_scales, 0, a_row_h1);
-        fp8e8m0_4 sb_h0 = pack_scales(smem_scales, 1024, b_row_h0);
-        fp8e8m0_4 sb_h1 = pack_scales(smem_scales, 1024, b_row_h1);
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[tic_scales].data, a_row_h0);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[tic_scales].data, a_row_h1);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[tic_scales].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[tic_scales].data, b_row_h1);
 
         auto bs0 = subtile_inplace<REG_N, BLOCK_K>(Bs[tic][0], {warp_n, 0});
         load(b0, bs0);
         auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[tic][0], {warp_m, 0});
         load(a, as0);
-        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A);
+        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[toc][1]));
         __builtin_amdgcn_s_barrier();
 
         asm volatile("s_waitcnt lgkmcnt(0)");
@@ -215,17 +256,17 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_sched_barrier(0);
 
         tic ^= 1; toc ^= 1;
+        tic_scales ^= 1; toc_scales ^= 1;
     }
 
     { // Epilogue k = k_iters - 1 (final)
-        int k = k_iters - 1;
-        load_scales_to_lds(smem_scales, scale_A_iter, scale_B_iter, block_m, block_n, k, M, N);
+        asm volatile("s_waitcnt vmcnt(0)");
+        asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
-
-        fp8e8m0_4 sa_h0 = pack_scales(smem_scales, 0, a_row_h0);
-        fp8e8m0_4 sa_h1 = pack_scales(smem_scales, 0, a_row_h1);
-        fp8e8m0_4 sb_h0 = pack_scales(smem_scales, 1024, b_row_h0);
-        fp8e8m0_4 sb_h1 = pack_scales(smem_scales, 1024, b_row_h1);
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[tic_scales].data, a_row_h0);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[tic_scales].data, a_row_h1);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[tic_scales].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[tic_scales].data, b_row_h1);
 
         auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[tic][0], {warp_m, 0});
         load(a, as0);
@@ -261,6 +302,38 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_s_barrier();
     }
 
+    } else {
+    #pragma unroll 1
+    for (int kk = 0; kk < k_iters; kk++) {
+        G::load(As[0][0], A, {0, 0, block_row * 2,     kk}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[0][0]));
+        G::load(As[0][1], A, {0, 0, block_row * 2 + 1, kk}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[0][1]));
+        G::load(Bs[0][0], B, {0, 0, block_col * 2,     kk}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[0][0]));
+        G::load(Bs[0][1], B, {0, 0, block_col * 2 + 1, kk}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[0][1]));
+        G::load(scale_A_smem[0], scale_A_gl, {kk * tiles_M + block_row, 0, 0, 0});
+        G::load(scale_B_smem[0], scale_B_gl, {kk * tiles_N + block_col, 0, 0, 0});
+        asm volatile("s_waitcnt vmcnt(0)");
+        asm volatile("s_waitcnt lgkmcnt(0)");
+        __builtin_amdgcn_s_barrier();
+
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[0].data, a_row_h0);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[0].data, a_row_h1);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[0].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[0].data, b_row_h1);
+
+        auto bs0 = subtile_inplace<REG_N, BLOCK_K>(Bs[0][0], {warp_n, 0}); load(b0, bs0);
+        auto bs1 = subtile_inplace<REG_N, BLOCK_K>(Bs[0][1], {warp_n, 0}); load(b1, bs1);
+        auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[0][0], {warp_m, 0}); load(a, as0);
+        asm volatile("s_waitcnt lgkmcnt(0)");
+        mma_ABt_scaled(cA, a, b0, cA, &sa_h0, &sb_h0);
+        mma_ABt_scaled(cB, a, b1, cB, &sa_h0, &sb_h1);
+        auto as1 = subtile_inplace<REG_M, BLOCK_K>(As[0][1], {warp_m, 0}); load(a, as1);
+        asm volatile("s_waitcnt lgkmcnt(0)");
+        mma_ABt_scaled(cC, a, b0, cC, &sa_h1, &sb_h0);
+        mma_ABt_scaled(cD, a, b1, cD, &sa_h1, &sb_h1);
+        __builtin_amdgcn_s_barrier();
+    }
+    }
+
     store(C, cA, {0, 0, block_row * WARPS_ROW * 2 + warp_m, block_col * WARPS_COL * 2 + warp_n});
     store(C, cB, {0, 0, block_row * WARPS_ROW * 2 + warp_m, block_col * WARPS_COL * 2 + WARPS_COL + warp_n});
     store(C, cC, {0, 0, block_row * WARPS_ROW * 2 + WARPS_ROW + warp_m, block_col * WARPS_COL * 2 + warp_n});

extra/thunder/amd/include/cdna4/common/macros.cuh

@@ -438,6 +438,35 @@ __device__ __forceinline__ T buffer_load_dwordx2(
   return result;
 }
 
+template<int GPR_START>
+__device__ __forceinline__ void buffer_load_dwordx3(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
+  if constexpr (GPR_START >= 256) {
+    asm volatile("buffer_load_dwordx3 a[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START - 256), "n"(GPR_START + 2 - 256), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  } else {
+    asm volatile("buffer_load_dwordx3 v[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START), "n"(GPR_START + 2), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  }
+}
+
+// Returns a u32x3 (ext_vector_type(3)). Note sizeof(u32x3) == 16 due to vector
+// alignment padding, but the compiler still allocates exactly 3 consecutive VGPRs
+// for the "=v" operand and emits buffer_load_dwordx3 v[N:N+2].
+template<typename T = u32x3>
+__device__ __forceinline__ T buffer_load_dwordx3(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  T result;
+  asm volatile("buffer_load_dwordx3 %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
 template<int GPR_START>
 __device__ __forceinline__ void buffer_load_dwordx4(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   if constexpr (GPR_START >= 256) {
@@ -465,6 +494,116 @@ __device__ __forceinline__ T buffer_load_dwordx4(
   return result;
 }
 
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_ubyte(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_ubyte %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_sbyte(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_sbyte %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_ushort(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_ushort %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_sshort(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_sshort %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+// d16 variants merge the loaded sub-dword into the low (or high, for _hi) 16 bits
+// of the destination register, leaving the other half untouched: read-modify-write.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_ubyte_d16(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_ubyte_d16 %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_ubyte_d16_hi(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_ubyte_d16_hi %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_sbyte_d16(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_sbyte_d16 %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_sbyte_d16_hi(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_sbyte_d16_hi %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_short_d16(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_short_d16 %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_short_d16_hi(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_short_d16_hi %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
 template<int GPR>
 __device__ __forceinline__ void buffer_store_dword(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   // AGPRS
@@ -519,6 +658,32 @@ __device__ __forceinline__ void buffer_store_dwordx2(
     : "memory");
 }
 
+template<int GPR_START>
+__device__ __forceinline__ void buffer_store_dwordx3(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
+  // AGPRS
+  if constexpr (GPR_START >= 256) {
+    asm volatile("buffer_store_dwordx3 a[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START - 256), "n"(GPR_START + 2 - 256), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  // VGPRS
+  } else {
+    asm volatile("buffer_store_dwordx3 v[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START), "n"(GPR_START + 2), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  }
+}
+
+template<typename T = u32x3>
+__device__ __forceinline__ void buffer_store_dwordx3(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  asm volatile("buffer_store_dwordx3 %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
 template<int GPR_START>
 __device__ __forceinline__ void buffer_store_dwordx4(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   // AGPRS
@@ -546,6 +711,50 @@ __device__ __forceinline__ void buffer_store_dwordx4(
     : "memory");
 }
 
+// Stores the low 8 bits of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_byte(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_byte %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+// Stores the high 16 bits' low 8 bits (bits [23:16]) of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_byte_d16_hi(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_byte_d16_hi %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+// Stores the low 16 bits of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_short(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_short %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+// Stores the high 16 bits of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_short_d16_hi(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_short_d16_hi %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
 template<int GPR>
 __device__ __forceinline__ void buffer_atomic_pk_add_bf16(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   if constexpr (GPR >= 256) {

extra/thunder/amd/include/cdna4/common/util.cuh

@@ -52,6 +52,8 @@ __device__ __forceinline__ int laneid() { return threadIdx.x & 0x3f; }
 
 using i32x2 = int32_t __attribute__((ext_vector_type(2)));
 using u32x2 = uint32_t __attribute__((ext_vector_type(2)));
+using i32x3 = int32_t __attribute__((ext_vector_type(3)));
+using u32x3 = uint32_t __attribute__((ext_vector_type(3)));
 using i32x4 = int32_t __attribute__((ext_vector_type(4)));
 using u32x4 = uint32_t __attribute__((ext_vector_type(4)));
 

extra/thunder/amd/include/cdna4/includes.cuh

@@ -0,0 +1,10 @@
+/**
+ * @file
+ * @brief Aggregate header for CDNA4 ThunderKittens core include tree.
+ */
+
+#pragma once
+
+#include "common/common.cuh"
+#include "types/types.cuh"
+#include "ops/ops.cuh"

extra/thunder/amd/include/cdna4/ops/warp/memory/tile/shared_to_register.cuh

@@ -248,7 +248,7 @@ __device__ inline static void load(RT &dst, const ST &src) {
                                         "ds_read_b64_tr_b16 %0, %2 offset:%3\n"
                                         "ds_read_b64_tr_b16 %1, %2 offset:%4\n"
                                         // "s_waitcnt lgkmcnt(0)\n"
-                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])), 
+                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])),
                                         "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx + 2]))
                                         : "v"(addr), "i"(offset), "i"(offset + 4 * ST::underlying_subtile_row_bytes)
                                         : "memory"
@@ -259,7 +259,7 @@ __device__ inline static void load(RT &dst, const ST &src) {
                                         "ds_read_b64_tr_b16 %0, %2 offset:%4\n"
                                         "ds_read_b64_tr_b16 %1, %3 offset:%4\n"
                                         // "s_waitcnt lgkmcnt(0)\n"
-                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])), 
+                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])),
                                         "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx + 2]))
                                         : "v"(addr), "v"(next_addr), "i"(offset)
                                         : "memory"
@@ -688,4 +688,88 @@ __device__ inline static void store(ST &dst, const RT &src) {
     }
 }
 
+template<ducks::rt::col_layout RT, ducks::st::all ST>
+__device__ inline static void load(RT &dst, const ST &src, int col_offset) {
+    static_assert(RT::cols == ST::rows,
+                  "col_layout load with col_offset: ST.rows must equal RT::cols");
+    static_assert(RT::width == 1, "col_layout load with col_offset: only width==1 supported");
+
+    using T2 = RT::dtype;
+    using U  = ST::dtype;
+    using U2 = base_types::packing<U>::packed_type;
+
+    const int laneid = kittens::laneid();
+    const uint32_t src_ptr = (uint32_t)(uintptr_t)(&src.data[0]);
+    constexpr int subtile_bytes = ST::underlying_subtile_bytes;
+
+    if constexpr (std::is_same_v<T2, fp8e4m3_4>) {
+        const int block_id = laneid / 16;
+        const int l_within = laneid % 16;
+        const int tr_k_grp = l_within / 2;
+        const int m_half   = l_within & 1;
+
+        #pragma unroll
+        for (int i = 0; i < RT::height; i++) {
+            const int m_col = col_offset + i * RT::base_tile_rows + m_half * 8;
+
+            const int subtile_base = block_id;
+            uint32_t addr = src_ptr
+                          + (uint32_t)(subtile_base * subtile_bytes)
+                          + src.swizzle({tr_k_grp, m_col});
+
+            asm volatile(
+                "ds_read_b64_tr_b8 %0, %2 offset:0\n"
+                "ds_read_b64_tr_b8 %1, %2 offset:%3\n"
+                : "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[0])),
+                  "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[4]))
+                : "v"(addr), "i"(4 * subtile_bytes)
+                : "memory"
+            );
+
+            addr ^= 1088u;
+
+            asm volatile(
+                "ds_read_b64_tr_b8 %0, %2 offset:0\n"
+                "ds_read_b64_tr_b8 %1, %2 offset:%3\n"
+                : "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[2])),
+                  "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[6]))
+                : "v"(addr), "i"(4 * subtile_bytes)
+                : "memory"
+            );
+        }
+    } else if constexpr (std::is_same_v<U2, bf16_2> || std::is_same_v<U2, half_2>) {
+        const int row_offset = ((laneid % 16) / 4) + ((laneid / 16) * 4);
+        const int col_offset_lane = (laneid % 4) * 4;
+
+        #pragma unroll
+        for (int i = 0; i < RT::height; i++) {
+            const int m_col = col_offset + i * RT::base_tile_rows + col_offset_lane;
+            const uint32_t addr = src_ptr + src.swizzle({row_offset, m_col});
+            const uint32_t next_addr = src_ptr + src.swizzle({row_offset + 4, m_col});
+
+            if constexpr (RT::base_tile_stride == 8) {
+                asm volatile(
+                    "ds_read_b64_tr_b16 %0, %2 offset:0\n"
+                    "ds_read_b64_tr_b16 %1, %3 offset:0\n"
+                    : "=v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[0])),
+                      "=v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[2]))
+                    : "v"(addr), "v"(next_addr)
+                    : "memory"
+                );
+            } else if constexpr (RT::base_tile_stride == 4) {
+                asm volatile(
+                    "ds_read_b64_tr_b16 %0, %1 offset:0\n"
+                    : "=v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[0]))
+                    : "v"(addr)
+                    : "memory"
+                );
+            } else {
+                static_assert(false, "Unsupported stride for col_layout load with col_offset");
+            }
+        }
+    } else {
+        static_assert(false, "Unsupported type for col_layout load with col_offset");
+    }
+}
+
 } // namespace kittens

extra/thunder/amd/include/cdna4/ops/warp/register/tile/mma.cuh

@@ -201,10 +201,10 @@ __device__ static inline void mma_AB_base(rt_base<float, ducks::rt_layout::col,
  * @param[in] b The second input rt_base<bf16_2, row_layout> matrix in row-major mode.
  * @param[in] c The input rt_base<float2, row_layout> accumulator matrix.
  */
-template<ducks::rt_shape::all D_shape, ducks::rt_shape::all A_shape, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T=bf16>
+template<ducks::rt_shape::all D_shape, ducks::rt_layout::all A_layout, ducks::rt_shape::all A_shape, ducks::rt_layout::all B_layout, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T=bf16>
 __device__ static inline void mma_ABt_base(rt_base<float, ducks::rt_layout::col, D_shape> &d,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, A_shape> &a,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, B_shape> &b, // in row-major mode
+    const rt_base<MM_Operand_T, A_layout, A_shape> &a,
+    const rt_base<MM_Operand_T, B_layout, B_shape> &b,
     const rt_base<float, ducks::rt_layout::col, C_shape> &c) {
 
     static_assert(std::is_same_v<D_shape, C_shape>, "D and C must have the same shape");
@@ -243,6 +243,32 @@ __device__ static inline void mma_ABt_base(rt_base<float, ducks::rt_layout::col,
     }
 }
 
+/**
+ * @brief Pack 4 E8M0 scale bytes from LDS into one fp8e8m0_4 register.
+ *
+ * @param smem_scales LDS pointer to scale tile data.
+ * @param row_offset Starting row within the scale region (warp's tile offset).
+ * @return fp8e8m0_4 with 4 scale bytes packed for MFMA opsel.
+ */
+__device__ __forceinline__ fp8e8m0_4 pack_scales(
+    const fp8e8m0 *smem_scales, int row_offset) {
+    int lid   = laneid();
+    int r16   = lid % 16;
+    int k_sub = lid / 16;
+
+    const fp8e8m0_4 *s4 = (const fp8e8m0_4 *)smem_scales;
+    fp8e8m0_4 w0 = s4[row_offset + 0 * 16 + r16];
+    fp8e8m0_4 w1 = s4[row_offset + 1 * 16 + r16];
+    fp8e8m0_4 w2 = s4[row_offset + 2 * 16 + r16];
+    fp8e8m0_4 w3 = s4[row_offset + 3 * 16 + r16];
+
+    fp8e8m0_4 sel = 0x0C0C0000u | (k_sub << 8) | (4u + k_sub);
+    fp8e8m0_4 lo  = __builtin_amdgcn_perm(w0, w1, sel);
+    fp8e8m0_4 hi  = __builtin_amdgcn_perm(w2, w3, sel);
+
+    return lo | (hi << 16);
+}
+
 /**
  * @brief Base dot product operation for row layout.
  *
@@ -254,10 +280,10 @@ __device__ static inline void mma_ABt_base(rt_base<float, ducks::rt_layout::col,
  * @param[in] b The second input rt_base<Operand_T, row_layout> matrix.
  * @param[in] c The input rt_base<float, col_layout> accumulator matrix.
  */
-template<int opsel_a, int opsel_b, int cbsz = 0, int blgp = 0, ducks::rt_shape::all D_shape, ducks::rt_shape::all A_shape, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T>
+template<int opsel_a, int opsel_b, int cbsz = 0, int blgp = 0, ducks::rt_shape::all D_shape, ducks::rt_layout::all A_layout, ducks::rt_shape::all A_shape, ducks::rt_layout::all B_layout, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T>
 __device__ static inline void mma_ABt_base_scaled(rt_base<float, ducks::rt_layout::col, D_shape> &d,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, A_shape> &a,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, B_shape> &b,
+    const rt_base<MM_Operand_T, A_layout, A_shape> &a,
+    const rt_base<MM_Operand_T, B_layout, B_shape> &b,
     const rt_base<float, ducks::rt_layout::col, C_shape> &c,
     const fp8e8m0_4 *scale_a,
     const fp8e8m0_4 *scale_b) {
@@ -441,7 +467,7 @@ __device__ static inline void mma_AB(D &d,
  * @param[in] b The second input rt_bf<M, K, row_layout> matrix in row-major mode.
  * @param[in] c The input rt_fl<N, M, row_layout> accumulator matrix.
  */
-template<ducks::rt::col_layout D, ducks::rt::row_layout A, ducks::rt::row_layout B, ducks::rt::col_layout C>
+template<ducks::rt::col_layout D, ducks::rt::all A, ducks::rt::all B, ducks::rt::col_layout C>
 __device__ static inline void mma_ABt(D &d,
                                 const A &a,
                                 const B &b, // notice row and (M, K) instead of col and (K, M)
@@ -499,7 +525,7 @@ __device__ static inline void mma_ABt(D &d,
  * @param[in] scale_a Pointer to the packed E8M0 scale for the A matrix.
  * @param[in] scale_b Pointer to the packed E8M0 scale for the B matrix.
  */
-template<int cbsz = 0, int blgp = 0, ducks::rt::col_layout D, ducks::rt::row_layout A, ducks::rt::row_layout B, ducks::rt::col_layout C>
+template<int cbsz = 0, int blgp = 0, ducks::rt::col_layout D, ducks::rt::all A, ducks::rt::all B, ducks::rt::col_layout C>
 __device__ static inline void mma_ABt_scaled(D &d,
                                 const A &a,
                                 const B &b,

extra/thunder/amd/include/cdna4/ops/warp/register/tile/tile.cuh

@@ -9,6 +9,5 @@
 #include "maps.cuh"
 #include "reductions.cuh"
 #include "mma.cuh"
-#include "scales.cuh"
 
 #include "assembly/tile.cuh"