assign

remove bad contiguous usage in torch backend
remove assign contiguous hack
2025-10-14 16:15:54 +08:00 · 2025-10-14 16:11:26 +08:00 · 2025-10-14 14:54:39 +08:00 · 2025-10-14 14:22:01 +08:00 · 2025-10-14 13:05:26 +08:00 · 2025-10-14 12:35:20 +08:00
433 changed files with 43929 additions and 4018 deletions
@@ -28,7 +28,7 @@ jobs:
      # since sudo is required for usbgpu on macos, move the cache to a new location, as some of the files are owned by root
      PYTHONPYCACHEPREFIX: /tmp/tiny_python_pycache
    runs-on: [self-hosted, macOS]
-    timeout-minutes: 20
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -52,14 +52,16 @@ jobs:
    - name: reset process replay
      run: python3.11 test/external/process_replay/reset.py
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=500 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=800 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=700 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=900 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
    - name: Run Stable Diffusion v2
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=1600 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
+      # TODO: very slow step time
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=10000 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
    # process replay can't capture this, the graph is too large
-    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+    # TODO: too slow
+    # - name: Run SDXL
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run model inference benchmark
      run: METAL=1 python3.11 test/external/external_model_benchmark.py
    - name: Test speed vs torch
@@ -99,7 +101,7 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=8 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
@@ -108,14 +110,19 @@ jobs:
      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
    - name: Train MNIST
      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
-    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=320 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
-    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=385 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
+
+    # NOTE: this is failing in CI. it is not failing on my machine and I don't really have a way to debug it
+    # the error is "RuntimeError: Internal Error (0000000e:Internal Error)"
+    #- name: Run 10 CIFAR training steps
+    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+    #- name: Run 10 CIFAR training steps w HALF
+    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
+
    #- name: Run 10 CIFAR training steps w BF16
    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
-    - name: Run 10 CIFAR training steps w winograd
-      run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    # TODO: too slow
+    # - name: Run 10 CIFAR training steps w winograd
+    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: UsbGPU boot time
      run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU tiny tests
@@ -160,7 +167,7 @@ jobs:
  testnvidiabenchmark:
    name: tinybox green Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 30
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -213,8 +220,9 @@ jobs:
      run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
    - name: Run Stable Diffusion
      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
-    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+    # TODO: too slow
+    # - name: Run SDXL
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA
      run: |
        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
@@ -238,9 +246,9 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
@@ -274,7 +282,7 @@ jobs:
  testmorenvidiabenchmark:
    name: tinybox green Training Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 20
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -299,24 +307,27 @@ jobs:
        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: Fuzz Padded Tensor Core GEMM (NV)
-      run: NV=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
-    - name: Fuzz Padded Tensor Core GEMM (PTX)
-      run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    # TODO: too slow
+    # - name: Fuzz Padded Tensor Core GEMM (NV)
+    #   run: NV=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    # TODO: too slow
+    # - name: Fuzz Padded Tensor Core GEMM (PTX)
+    #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    - name: Train MNIST
      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=85 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=270 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=68 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=310 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=75 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
-    - name: Run 10 CIFAR training steps w winograd
-      run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=35 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=310 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    # TODO: too slow
+    # - name: Run 10 CIFAR training steps w winograd
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    - name: Run MLPerf resnet eval on training data
      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
@@ -346,7 +357,7 @@ jobs:
  testamdbenchmark:
    name: tinybox red Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 20
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -415,9 +426,10 @@ jobs:
    - name: Test AM warm start time
      run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=450 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
-    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=1400 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+    # TODO: too slow
+    # - name: Run SDXL
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA 7B
      run: |
        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
@@ -476,7 +488,7 @@ jobs:
  testmoreamdbenchmark:
    name: tinybox red Training Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 30
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -508,19 +520,20 @@ jobs:
    - name: Train MNIST
      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=85 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=188 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
-    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
-    - name: Run 10 CIFAR training steps w winograd
-      run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+    # - name: Run 10 CIFAR training steps w BF16
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    # TODO: too slow
+    # - name: Run 10 CIFAR training steps w winograd
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    #- name: Run full CIFAR training steps w 6 GPUS
-    #  run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+    #  run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    #- name: Run full CIFAR training steps w 6 GPUS (REMOTE)
-    #  run: time BENCHMARK_LOG=cifar_6gpu_remote REMOTE=1 REMOTEDEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu_remote.txt
+    #  run: time BENCHMARK_LOG=cifar_6gpu_remote REMOTE=1 REMOTEDEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu_remote.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD Training)
@@ -539,7 +552,7 @@ jobs:
  testmlperfamdbenchmark:
    name: tinybox red MLPerf Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 30
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -606,21 +619,21 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: benchmark openpilot 0.9.9 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_9_9_vision ASSERT_MIN_STEP_TIME=30 PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+      run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
    - name: benchmark openpilot 0.9.9 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_9_9_policy ASSERT_MIN_STEP_TIME=45 PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
    - name: benchmark openpilot 0.9.9 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring ASSERT_MIN_STEP_TIME=70 PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: openpilot compile3 0.9.9 driving_vision
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=18 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
    - name: openpilot compile3 0.9.9 driving_policy
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=7 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
    - name: openpilot compile3 0.9.9 dmonitoring
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=12 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: openpilot compile3 Space Lab policy + vision
      run: |
-        PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
-        PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
+        PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
+        PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
    - name: benchmark MobileNetV2 on DSP
      run: |
        # generate quantized weights
@@ -645,7 +658,7 @@ jobs:
  testreddriverbenchmark:
    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 15
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -695,7 +708,7 @@ jobs:
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
    # TODO: enable
    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
@@ -716,7 +729,7 @@ jobs:
  testgreendriverbenchmark:
    name: NV Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 15
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -758,7 +771,7 @@ jobs:
    - name: Test LLAMA-3
      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
    #  run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
@@ -30,8 +30,6 @@ jobs:
        key: llvm-speed
        deps: testing_minimal
        llvm: 'true'
-    - name: External Benchmark Schedule
-      run: python3 test/external/external_benchmark_schedule.py
    - name: Speed Test
      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
@@ -48,7 +46,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        deps: docs
-        pydeps: "capstone"
+        pydeps: "capstone torch"
    - name: Build wheel and show size
      run: |
        pip install build
@@ -79,6 +77,8 @@ jobs:
      run: |
        python docs/abstractions2.py
        python docs/abstractions3.py
+    - name: Test README
+      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
    - name: Test Quickstart
      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' docs/quickstart.md > quickstart.py &&  python quickstart.py
    - name: Test DEBUG
@@ -144,7 +144,7 @@ jobs:
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: SPLIT_REDUCEOP=0 FUSE_ARANGE=1 CPU=1 CPU_LLVM=1 TARGET_EVAL_ACC_PCT=96.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+      run: CPU=1 CPU_LLVM=1 TARGET_EVAL_ACC_PCT=96.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
    - name: Test some torch tests (expect failure)
      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true

@@ -259,21 +259,28 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-12
-        pydeps: "pillow"
+        pydeps: "pillow numpy ftfy regex"
        deps: testing_unit
-    - name: Test README
-      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
+    - name: Check Device.DEFAULT
+      run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
    - name: Run unit tests
-      run: python -m pytest -n=auto test/unit/ --durations=20
+      run: CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+    - name: Check SPEC=1
+      run: SPEC=1 python3 test/test_tiny.py
    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 test/test_multitensor.py TestMultiTensor.test_data_parallel_resnet_train_step
-    - name: Run SDXL on NULL backend
-      run: MAX_BUFFER_SIZE=0 NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+      run: NULL=1 python3 -m unittest test.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
+    # TODO: too slow
+    # - name: Run SDXL on NULL backend
+    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+    - name: Run Clip tests for SD MLPerf on NULL backend
+      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
    # TODO: support fake weights
    #- name: Run LLaMA 7B on 4 fake devices
    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
    - name: Run GC tests
      run: python test/external/external_uop_gc.py
+    - name: External Benchmark Schedule
+      run: python3 test/external/external_benchmark_schedule.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
    - name: Regen dataset on test_tiny
@@ -303,16 +310,16 @@ jobs:
    - name: Fuzz Test fast idiv
      run: python test/external/fuzz_fast_idiv.py
    - name: Fuzz Test shapetracker
-      run: |
-        python test/external/fuzz_shapetracker.py
-        python test/external/fuzz_shapetracker_math.py
+      run: CNT=50 python test/external/fuzz_shapetracker.py
+    - name: Fuzz Test shapetracker math
+      run: CNT=200 python test/external/fuzz_shapetracker_math.py
    - name: Fuzz Test shape ops
      run: python test/external/fuzz_shape_ops.py

  testopenclimage:
-    name: 'CL IMAGE Tests'
+    name: CL IMAGE Tests
    runs-on: ubuntu-22.04
-    timeout-minutes: 10
+    timeout-minutes: 15
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -330,7 +337,7 @@ jobs:
        uses: ./.github/actions/process-replay

  testgpumisc:
-    name: 'CL Misc tests'
+    name: CL Misc tests
    runs-on: ubuntu-22.04
    timeout-minutes: 10
    steps:
@@ -355,7 +362,7 @@ jobs:
          path: /tmp/sops.gz

  testopenpilot:
-    name: 'openpilot Compile Tests'
+    name: openpilot Compile Tests
    runs-on: ubuntu-22.04
    timeout-minutes: 15
    steps:
@@ -370,7 +377,7 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=208 ALLOWED_READ_IMAGE=2175 ALLOWED_GATED_READ_IMAGE=16 FLOAT16=0 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx
+          ALLOWED_KERNEL_COUNT=190 ALLOWED_READ_IMAGE=2081 ALLOWED_GATED_READ_IMAGE=28 FLOAT16=0 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx
      - name: Test openpilot alt model correctness (float32)
        run: FLOAT16=0 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/3799fe46b3a629e491d4b8498b8ae83e4c88c304/selfdrive/modeld/models/supercombo.onnx
      - name: Test openpilot fastvits model correctness (float32)
@@ -387,7 +394,7 @@ jobs:
 # ****** ONNX Tests ******

  testonnxcpu:
-    name: 'ONNX (CPU) Tests'
+    name: ONNX (CPU) Tests
    runs-on: ubuntu-22.04
    timeout-minutes: 20

@@ -415,7 +422,7 @@ jobs:
        uses: ./.github/actions/process-replay

  testopencl:
-    name: 'ONNX (GPU)+Optimization Tests'
+    name: ONNX (CL)+Optimization Tests
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    steps:
@@ -439,8 +446,12 @@ jobs:
        run: CL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
+      - name: NULL=1 beautiful_mnist_multigpu
+        run: NULL=1 python examples/beautiful_mnist_multigpu.py
+      - name: Test Bert training
+        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
-        run: MAX_BUFFER_SIZE=0 DEV=NULL SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=8 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=8 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -503,39 +514,6 @@ jobs:

 # ****** Feature Tests ******

-  testrangeify:
-    name: Linux (rangeify)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: rangeify-minimal-llvm
-        deps: testing_minimal
-        opencl: 'true'
-        llvm: "true"
-    - name: Test CPU=1 RANGEIFY=1
-      # TODO: add more passing tests here
-      # test_symbolic_arange_sym_step is passing now
-      # test_threefry_doesnt_use_long is because there's a contig after the long now
-      run: |
-        CPU=1 CPU_LLVM=0 RANGEIFY=1 python3 -m pytest -n auto --durations 20 \
-          -k "not test_symbolic_arange_sym_step and not test_threefry_doesnt_use_long" \
-          test/test_tiny.py test/test_rangeify.py test/test_ops.py test/test_tensor_variable.py \
-          test/test_outerworld_range.py test/test_sample.py test/test_randomness.py
-    - name: Test multitensor
-      run: RANGEIFY=1 PYTHONPATH="." python3 test/test_multitensor.py TestMultiTensor.test_matmul_shard_1_1 TestMultiTensor.test_simple_add_W
-    - name: Test GPU=1 RANGEIFY=1
-      run: GPU=1 RANGEIFY=1 pytest -n auto test/test_ops.py
-    - name: Test CPU=1 RANGEIFY=2
-      run: CPU=1 CPU_LLVM=0 RANGEIFY=2 python3 -m pytest -n auto test/test_tiny.py test/test_rangeify.py test/test_ops.py --durations 20
-    # slow (and still wrong on beautiful_mnist)
-    #- name: Test LLVM=1 RANGEIFY=1 (slow tests)
-    #  run: CPU=1 CPU_LLVM=1 RANGEIFY=1 python3 -m pytest -n auto test/models/test_mnist.py --durations 20
-
  testdevectorize:
    name: Linux (devectorize)
    runs-on: ubuntu-24.04
@@ -555,7 +533,7 @@ jobs:
    - name: Test LLVM=1 DEVECTORIZE=0 for model
      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
    - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 FUSE_ARANGE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
+      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"

  testdsp:
    name: Linux (DSP)
@@ -656,7 +634,7 @@ jobs:
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
      - name: Run TestOps.test_add with SQTT
        run: |
-          PROFILE=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
@@ -30,10 +30,6 @@ persistent=yes
 # Specify a configuration file.
 #rcfile=

-# When enabled, pylint would attempt to guess common misconfiguration and emit
-# user-friendly hints instead of false-positive error messages
-suggestion-mode=yes
-
 # Allow loading of arbitrary C extensions. Extensions are imported into the
 # active Python interpreter and may run arbitrary code.
 unsafe-load-any-extension=no
@@ -414,10 +414,29 @@ generate_sqtt() {
  clang2py -k cdefstum \
    extra/sqtt/sqtt.h \
    -o $BASE/sqtt.py
-
  fixup $BASE/sqtt.py
  sed -i "s\import ctypes\import ctypes, os\g" $BASE/sqtt.py
  python3 -c "import tinygrad.runtime.autogen.sqtt"
+
+  ROCPROF_COMMIT_HASH=dd0485100971522cc4cd8ae136bdda431061a04d
+  ROCPROF_SRC=/tmp/rocprof-trace-decoder-$ROCPROF_COMMIT_HASH
+  if [ ! -d "$ROCPROF_SRC" ]; then
+    git clone https://github.com/ROCm/rocprof-trace-decoder $ROCPROF_SRC
+    pushd .
+    cd $ROCPROF_SRC
+    git reset --hard $ROCPROF_COMMIT_HASH
+    popd
+  fi
+
+  clang2py -k cdefstum \
+    $ROCPROF_SRC/include/rocprof_trace_decoder.h \
+    $ROCPROF_SRC/include/trace_decoder_instrument.h \
+    $ROCPROF_SRC/include/trace_decoder_types.h \
+    -o extra/sqtt/rocprof/rocprof.py
+  fixup extra/sqtt/rocprof/rocprof.py
+  sed -i '1s/^/# pylint: skip-file\n/' extra/sqtt/rocprof/rocprof.py
+  sed -i "s/import ctypes/import ctypes, ctypes.util/g" extra/sqtt/rocprof/rocprof.py
+  sed -i "s|FunctionFactoryStub()|ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder'))|g" extra/sqtt/rocprof/rocprof.py
 }

 generate_webgpu() {
@@ -42,7 +42,6 @@ import struct
 from tinygrad.dtype import dtypes
 from tinygrad.device import Buffer, Device
 from tinygrad.uop.ops import UOp, Ops
-from tinygrad.shape.shapetracker import ShapeTracker

 # allocate some buffers + load in values
 out = Buffer(DEVICE, 1, dtypes.int32).allocate()
@@ -51,13 +50,14 @@ b = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struc
 # NOTE: a._buf is the same as the return from cpu.allocator.alloc

 # describe the computation
+idx = UOp.const(dtypes.index, 0)
 buf_1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 1)
 buf_2 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 2)
-ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1.view(ShapeTracker.from_shape((1,))),))
-ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2.view(ShapeTracker.from_shape((1,))),))
+ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1.index(idx),))
+ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2.index(idx),))
 alu = ld_1 + ld_2
 output_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
-st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.view(ShapeTracker.from_shape((1,))), alu))
+st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.index(idx), alu))
 s = UOp(Ops.SINK, dtypes.void, (st_0,))

 # convert the computation to a "linearized" format (print the format)
@@ -80,7 +80,7 @@ print("******** third, the UOp ***********")

 from tinygrad.engine.realize import run_schedule
 from tinygrad.engine.schedule import create_schedule_with_vars
-from tinygrad.schedule.kernelize import get_kernelize_map
+from tinygrad.schedule.rangeify import get_rangeify_map

 # allocate some values + load in values
 a = UOp.new_buffer(DEVICE, 1, dtypes.int32)
@@ -93,10 +93,10 @@ out = a + b
 s = UOp(Ops.SINK, dtypes.void, (out,))

 # group the computation into kernels
-becomes_map = get_kernelize_map(s)
+becomes_map = get_rangeify_map(s)

 # the compute maps to an assign
-assign = becomes_map[a+b]
+assign = becomes_map[a+b].base

 # the first source is the output buffer (data)
 assert assign.src[0].op is Ops.BUFFER
@@ -10,7 +10,7 @@ Directories are listed in order of how they are processed.

 Group UOps into kernels.

-::: tinygrad.schedule.kernelize.get_kernelize_map
+::: tinygrad.schedule.rangeify.get_rangeify_map
    options:
        members: false
        show_labels: false
@@ -41,8 +41,6 @@ BEAM                | [#]        | number of beams in kernel beam search
 DEFAULT_FLOAT       | [HALF, ...]| specify the default float dtype (FLOAT32, HALF, BFLOAT16, FLOAT64, ...), default to FLOAT32
 IMAGE               | [1-2]      | enable 2d specific optimizations
 FLOAT16             | [1]        | use float16 for images instead of float32
-PTX                 | [1]        | enable the specialized [PTX](https://docs.nvidia.com/cuda/parallel-thread-execution/) assembler for Nvidia GPUs. If not set, defaults to generic CUDA codegen backend.
-PROFILE             | [1]        | enable profiling. This feature is supported in NV, AMD, QCOM and METAL backends.
 VISIBLE_DEVICES     | [list[int]]| restricts the NV/AMD devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
 JIT                 | [0-2]      | 0=disabled, 1=[jit enabled](quickstart.md#jit) (default), 2=jit enabled, but graphs are disabled
 VIZ                 | [1]        | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
@@ -2,17 +2,17 @@

 tinygrad supports various runtimes, enabling your code to scale across a wide range of devices. The default runtime can be automatically selected based on the available hardware, or you can force a specific runtime to be default using environment variables (e.g., `CPU=1`).

-| Runtime | Description | Requirements |
-|---------|-------------|--------------|
-| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | Ampere/Ada series GPUs |
-| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | RDNA2/RDNA3/RDNA4 series GPUs. You can select one of the interfaces for communication by setting `AMD_IFACE=(KFD|PCI)`. See [AMD interfaces](#amd-interfaces) for more details. |
-| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | 6xx series GPUs |
-| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | M1+ Macs; Metal 3.0+ for `bfloat` support |
-| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | NVIDIA GPU with CUDA support |
-| [OpenCL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | OpenCL 2.0 compatible device |
-| [CPU (C Code)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang compiler | `clang` compiler in system `PATH` |
-| [LLVM (LLVM IR)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_llvm.py) | Runs on CPU using the LLVM compiler infrastructure | llvm libraries installed and findable |
-| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | Dawn library installed and findable. Download binaries [here](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0). |
+| Runtime | Description | Compiler Options | Requirements |
+|---------|-------------|------------------|--------------|
+| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`NV_PTX=1`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via `NV_IFACE=(NVK\|PCI)`. See [NV interfaces](#nv-interfaces) for details. |
+| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`AMD_LLVM=1`)<br>HIP/COMGR (`AMD_HIP=1`) | RDNA2 or newer GPUs.<br>You can select an interface via `AMD_IFACE=(KFD\|PCI\|USB)`. See [AMD interfaces](#amd-interfaces) for details. |
+| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
+| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for `bfloat` support |
+| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`CUDA_PTX=1`) | NVIDIA GPU with CUDA support |
+| [CL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | - | OpenCL 2.0 compatible device |
+| [CPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang or llvm compiler | Clang JIT (default)<br>LLVM IR (`CPU_LLVM=1`) | `clang` compiler in system `PATH` |
+| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | - | Dawn library installed and discoverable. Binaries: [pydawn v0.3.0](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0) |
+

 ## Interoperability

@@ -70,5 +70,12 @@ AMD backend supports several interfaces for communicating with devices:

 * `KFD`: uses the amdgpu driver
 * `PCI`: uses the [AM driver](developer/am.md)
+* `USB`: USB3 interafce for asm24xx chips.

 You can force an interface by setting `AMD_IFACE` to one of these values. In the case of `AMD_IFACE=PCI`, this may unbind your GPU from the amdgpu driver.
+
+## NV Interfaces
+NV backend supports several interfaces for communicating with devices:
+
+* `NVK`: uses the nvidia driver
+* `PCI`: uses the [NV driver](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/support/nv/nvdev.py)
@@ -10,7 +10,7 @@ GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]

 # override tinygrad defaults
 dtypes.default_float = dtypes.half
-Context(FUSE_ARANGE=1, FUSE_OPTIM=1).__enter__()
+Context(FUSE_OPTIM=1).__enter__()

 # from https://github.com/tysam-code/hlb-CIFAR10/blob/main/main.py
 batchsize = getenv("BS", 1024)
@@ -1,6 +1,6 @@
 import sys, time
 from tinygrad import TinyJit, GlobalCounters, fetch, getenv
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.nn.onnx import OnnxRunner
 from extra.onnx_helpers import get_example_inputs, validate

 def load_onnx_model(onnx_file):
@@ -8,7 +8,7 @@ import numpy as np
 import subprocess
 import tensorflow as tf
 import tf2onnx
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.nn.onnx import OnnxRunner
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import to_mv
 from extra.export_model import export_model_clang, compile_net, jit_model
@@ -26,8 +26,8 @@ class Attention:
      start_pos = start_pos.val

    if HALF: x = x.half()
-    xqkv = self.c_attn(x)
-    xq, xk, xv = [xqkv.shrink((None, None, (i*self.dim, (i+1)*self.dim))).reshape(None, None, self.n_heads, self.head_dim) for i in range(3)]
+    xqkv = self.c_attn(x).reshape(None, None, 3, self.n_heads, self.head_dim)
+    xq, xk, xv = [xqkv[:, :, i, :, :] for i in range(3)]
    bsz, seqlen, _, _ = xq.shape

    # create kv cache
@@ -35,11 +35,11 @@ class Attention:
      self.cache_kv = Tensor.zeros(2, bsz, MAX_CONTEXT, self.n_heads, self.head_dim, dtype=x.dtype).contiguous().realize()

    # update the cache
-    self.cache_kv.shrink((None, None,(start_pos,start_pos+seqlen),None,None)).assign(Tensor.stack(xk, xv)).realize()
+    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()

    if start_pos > 0:
-      keys = self.cache_kv[0].shrink((None, (0, start_pos+seqlen), None, None))
-      values = self.cache_kv[1].shrink((None, (0, start_pos+seqlen), None, None))
+      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
+      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
    else:
      keys = xk
      values = xv
@@ -64,7 +64,7 @@ class TransformerBlock:

  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]):
    h = x + self.attn(self.ln_1(x), start_pos, mask).float()
-    return (h + self.mlp(self.ln_2(h)))
+    return (h + self.mlp(self.ln_2(h))).contiguous()

 class Transformer:
  def __init__(self, dim, n_heads, n_layers, norm_eps, vocab_size, max_seq_len=1024):
@@ -145,7 +145,6 @@ hyp = {
  },
 }

-@Context(FUSE_ARANGE=getenv("FUSE_ARANGE", 1))
 def train_cifar():

  def set_seed(seed):
@@ -229,7 +228,8 @@ def train_cifar():
    if getenv("RANDOM_CROP", 1):
      X = random_crop(X, crop_size=32)
    if getenv("RANDOM_FLIP", 1):
-      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X) # flip LR
+      # NOTE: RANGEIFY=1 needs this contiguous or the X[perms] is very slow
+      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X).contiguous() # flip LR
    X, Y = X[perms], Y[perms]
    return X, Y, *cutmix(X, Y, perms, mask_size=hyp['net']['cutmix_size'])

@@ -511,6 +511,33 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
      # happens with BENCHMARK set
      pass

+# stable diffusion callbacks to match mlperf ref; declared here because they're pickled
+def filter_dataset(sample:dict): return {k:v for k,v in sample.items() if k in {'npy', 'txt'}}
+def collate(batch:list[dict]):
+  ret = {"npy": [], "txt": [], "__key__": []}
+  for sample in batch:
+    for k,v in sample.items():
+      ret[k].append(v)
+  return ret
+def collate_fn(batch): return batch
+
+# Reference (code): https://github.com/mlcommons/training/blob/2f4a93fb4888180755a8ef55f4b977ef8f60a89e/stable_diffusion/ldm/data/webdatasets.py, Line 55
+# Reference (params): https://github.com/mlcommons/training/blob/ab4ae1ca718d7fe62c369710a316dff18768d04b/stable_diffusion/configs/train_01x08x08.yaml, Line 107
+def batch_load_train_stable_diffusion(urls:str, BS:int):
+  import webdataset
+  dataset = webdataset.WebDataset(urls=urls, resampled=True, cache_size=-1, cache_dir=None)
+  dataset = dataset.shuffle(size=1000)
+  dataset = dataset.decode()
+  dataset = dataset.map(filter_dataset)
+  dataset = dataset.batched(BS, partial=False, collation_fn=collate)
+  dataset = webdataset.WebLoader(dataset, batch_size=None, shuffle=False, num_workers=1, persistent_workers=True, collate_fn=collate_fn)
+
+  for x in dataset:
+    assert isinstance(x, dict) and all(isinstance(k, str) for k in x.keys()) and all(isinstance(v, list) for v in x.values())
+    assert all(isinstance(moment_mean_logvar, np.ndarray) and moment_mean_logvar.shape==(1,8,64,64) for moment_mean_logvar in x["npy"])
+    assert all(isinstance(caption, str) for caption in x["txt"])
+    yield x
+
 # llama3

 class BinIdxDataset:
@@ -2,7 +2,9 @@ import math
 from typing import Union

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

 # rejection sampling truncated randn
 def rand_truncn(*shape, dtype=None, truncstds=2, **kwargs) -> Tensor:
@@ -17,6 +19,10 @@ def he_normal(*shape, a: float = 0.00, **kwargs) -> Tensor:
  std = math.sqrt(2.0 / (1 + a ** 2)) / math.sqrt(prod(argfix(*shape)[1:])) / 0.87962566103423978
  return std * rand_truncn(*shape, **kwargs)

+# Stable Diffusion v2 training uses default torch gelu, which doesn't use tanh approximation
+def gelu_erf(x:Tensor) -> Tensor:
+  return 0.5 * x * (1.0 + (x / 1.4142135623730951).erf())
+
 class Conv2dHeNormal(nn.Conv2d):
  def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
    super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
@@ -127,3 +133,59 @@ class Conv2dRetinaNet(nn.Conv2d):
  def __call__(self, x:Tensor) -> Tensor:
    return x.conv2d(self.weight.cast(dtypes.default_float), self.bias.cast(dtypes.default_float) if self.bias is not None else None,
                    groups=self.groups, stride=self.stride, dilation=self.dilation, padding=self.padding)
+
+# copy torch AMP: isolate mixed precision to just the below autocast ops, instead of using dtypes.default_float which affects all new Tensors
+class AutocastLinear(nn.Linear):
+  cast_dtype=dtypes.bfloat16 # enable monkeypatching of the mixed precision dtype
+  def __call__(self, x:Tensor) -> Tensor:
+    dtype = type(self).cast_dtype
+    return x.cast(dtype).linear(self.weight.cast(dtype).transpose(), self.bias.cast(dtype) if self.bias is not None else None)
+
+class AutocastConv2d(nn.Conv2d):
+  cast_dtype=dtypes.bfloat16
+  def __call__(self, x:Tensor) -> Tensor:
+    dtype = type(self).cast_dtype
+    return x.cast(dtype).conv2d(self.weight.cast(dtype), self.bias.cast(dtype), self.groups, self.stride, self.dilation, self.padding)
+
+# copy torch AMP: upcast to float32 before GroupNorm and LayerNorm
+class AutocastGroupNorm(nn.GroupNorm):
+  def __call__(self, x:Tensor) -> Tensor:
+    return super().__call__(x.cast(dtypes.float32))
+
+class AutocastLayerNorm(nn.LayerNorm):
+  def __call__(self, x:Tensor) -> Tensor:
+    return super().__call__(x.cast(dtypes.float32))
+
+def zero_module(module):
+  for p in get_parameters(module): p.assign(Tensor.zeros_like(p).contiguous())
+
+# Stable Diffusion mlperf reference doesn't call scaled_dot_product_attention
+# copy torch AMP: upcast to float32 before softmax on CUDA
+def attn_f32_softmax(q:Tensor, k:Tensor, v:Tensor) -> Tensor:
+  return (q.matmul(k.transpose(-2,-1), dtype=dtypes.float32) / math.sqrt(q.shape[-1])).softmax(-1).cast(q.dtype) @ v
+
+def init_stable_diffusion(version:str, pretrained:str, devices:list[str]):
+  from examples.stable_diffusion import StableDiffusion
+  from tinygrad.nn.state import safe_load, safe_save, load_state_dict, get_state_dict
+  from tempfile import TemporaryDirectory
+  model = StableDiffusion(version=version, pretrained=pretrained)
+  unet:UNetModel = model.model.diffusion_model
+
+  # this prevents extra consumption of memory, enabling much larger BS
+  Tensor.realize(*get_parameters(unet))
+  with TemporaryDirectory(prefix="unet_init") as tmp:
+    safe_save(get_state_dict(unet), init_fn:=f"{tmp}/init_model.safetensors")
+    load_state_dict(unet, safe_load(init_fn))
+
+  sqrt_alphas_cumprod = model.alphas_cumprod.sqrt().realize()
+  sqrt_one_minus_alphas_cumprod = (1 - model.alphas_cumprod).sqrt().realize()
+
+  if len(devices) > 1:
+    to_move = [sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod]
+    if version == "v2-mlperf-train": to_move += get_parameters(unet) + get_parameters(model.cond_stage_model)
+    for p in to_move:
+      p.to_(devices)
+    with Context(BEAM=0):
+      Tensor.realize(*to_move)
+
+  return model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod
@@ -1,8 +1,9 @@
 import math
-from tinygrad import dtypes
+from tinygrad import dtypes, Tensor
 from tinygrad.nn.optim import Optimizer

 from extra.lr_scheduler import LR_Scheduler
+from typing import Callable

 # https://github.com/mlcommons/training/blob/e237206991d10449d9675d95606459a3cb6c21ad/image_classification/tensorflow2/lars_util.py
 class PolynomialDecayWithWarmup(LR_Scheduler):
@@ -36,4 +37,24 @@ class CosineAnnealingLRWithWarmup(LR_Scheduler):
  def get_lr(self):
    warmup_lr = ((self.epoch_counter+1) / self.warmup_steps) * self.base_lr
    decay_lr = self.end_lr + 0.5 * (self.base_lr-self.end_lr) * (1 + (((self.epoch_counter+1-self.warmup_steps)/self.decay_steps) * math.pi).cos())
-    return (self.epoch_counter < self.warmup_steps).where(warmup_lr, decay_lr).cast(self.optimizer.lr.dtype)
+    return (self.epoch_counter < self.warmup_steps).where(warmup_lr, decay_lr).cast(self.optimizer.lr.dtype)
+
+# Reference: https://github.com/mlcommons/training/blob/64b14a9abc74e08779a175abca7d291f8c957632/stable_diffusion/ldm/lr_scheduler.py, Lines 36-97
+class LambdaLinearScheduler:
+  def __init__(self, warm_up_steps:int, f_min:float, f_max:float, f_start:float, cycle_lengths:int):
+    self.lr_warm_up_steps, self.f_min, self.f_max, self.f_start, self.cycle_lengths = warm_up_steps, f_min, f_max, f_start, cycle_lengths
+
+  def schedule(self, n:Tensor) -> Tensor:
+    warm_up = (n < self.lr_warm_up_steps)
+    f_warm_up = (self.f_max - self.f_start) / self.lr_warm_up_steps * n + self.f_start
+    return warm_up.where(f_warm_up, self.f_min + (self.f_max - self.f_min) * (self.cycle_lengths - n) / (self.cycle_lengths))
+
+# based on torch.optim.lr_scheduler.LambdaLR
+class LambdaLR(LR_Scheduler):
+  def __init__(self, optimizer:Optimizer, base_lr:Tensor, lr_lambda:Callable):
+    super().__init__(optimizer)
+    self.base_lr, self.lr_lambda = base_lr, lr_lambda
+    self.step()
+
+  def get_lr(self):
+    return self.base_lr * self.lr_lambda(self.epoch_counter - 1)
@@ -1,10 +1,10 @@
-import time, math
+import time, math, os
 start = time.perf_counter()
 from pathlib import Path
 import numpy as np
 from tinygrad import Tensor, Device, dtypes, GlobalCounters, TinyJit
 from tinygrad.nn.state import get_parameters, load_state_dict, safe_load
-from tinygrad.helpers import getenv
+from tinygrad.helpers import getenv, Context, prod
 from extra.bench_log import BenchEvent, WallTimeEvent
 def tlog(x): print(f"{x:25s}  @ {time.perf_counter()-start:5.2f}s")

@@ -287,6 +287,256 @@ def eval_llama3():
  log_perplexity = np.mean(losses)
  print(f"Log Perplexity: {log_perplexity}")

+# NOTE: BEAM hangs on 8xmi300x with DECODE_BS=384 in final realize below; function is declared here for external testing
+@TinyJit
+def vae_decode(x:Tensor, vae, disable_beam=False) -> Tensor:
+  from examples.stable_diffusion import AutoencoderKL
+  assert isinstance(vae, AutoencoderKL)
+  x = vae.post_quant_conv(1./0.18215 * x)
+
+  x = vae.decoder.conv_in(x)
+  x = vae.decoder.mid(x)
+  for i, l in enumerate(vae.decoder.up[::-1]):
+    print("decode", x.shape)
+    for b in l['block']: x = b(x)
+    if 'upsample' in l:
+      bs,c,py,px = x.shape
+      x = x.reshape(bs, c, py, 1, px, 1).expand(bs, c, py, 2, px, 2).reshape(bs, c, py*2, px*2)
+      x = l['upsample']['conv'](x)
+    if i == len(vae.decoder.up) - 1 and disable_beam:
+      with Context(BEAM=0): x.realize()
+    else: x.realize()
+  x = vae.decoder.conv_out(vae.decoder.norm_out(x).swish())
+
+  x = ((x + 1.0) / 2.0).clip(0.0, 1.0)
+  return x
+
+def eval_stable_diffusion():
+  import csv, PIL, sys
+  from tqdm import tqdm
+  from examples.mlperf.initializers import init_stable_diffusion, gelu_erf
+  from examples.stable_diffusion import AutoencoderKL
+  from extra.models.unet import UNetModel
+  from tinygrad.nn.state import load_state_dict, torch_load
+  from tinygrad.helpers import BEAM
+  from extra.models import clip
+  from extra.models.clip import FrozenOpenClipEmbedder
+  from extra.models.clip import OpenClipEncoder
+  from extra.models.inception import FidInceptionV3
+
+  config = {}
+  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
+  for x in GPUS: Device[x]
+  print(f"running eval on {GPUS}")
+  seed               = config["seed"]                   = getenv("SEED", 12345)
+  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
+  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
+  CONTEXT_BS         = config["CONTEXT_BS"]             = getenv("CONTEXT_BS", 1 * len(GPUS))
+  DENOISE_BS         = config["DENOISE_BS"]             = getenv("DENOISE_BS", 1 * len(GPUS))
+  DECODE_BS          = config["DECODE_BS"]              = getenv("DECODE_BS", 1 * len(GPUS))
+  INCEPTION_BS       = config["INCEPTION_BS"]           = getenv("INCEPTION_BS", 1 * len(GPUS))
+  CLIP_BS            = config["CLIP_BS"]                = getenv("CLIP_BS", 1 * len(GPUS))
+  EVAL_CKPT_DIR      = config["EVAL_CKPT_DIR"]          = getenv("EVAL_CKPT_DIR", "")
+  STOP_IF_CONVERGED  = config["STOP_IF_CONVERGED"]      = getenv("STOP_IF_CONVERGED", 0)
+
+  if (WANDB := getenv("WANDB", "")):
+    import wandb
+    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
+
+  assert EVAL_CKPT_DIR != "", "provide a directory with checkpoints to be evaluated"
+  print(f"running eval on checkpoints in {EVAL_CKPT_DIR}\nSEED={seed}")
+  eval_queue:list[tuple[int, Path]] = []
+  for p in Path(EVAL_CKPT_DIR).iterdir():
+    if p.name.endswith(".safetensors"):
+      ckpt_iteration = p.name.split(".safetensors")[0]
+      assert ckpt_iteration.isdigit(), f"invalid checkpoint name: {p.name}, expected <digits>.safetensors"
+      eval_queue.append((int(ckpt_iteration), p))
+  assert len(eval_queue), f'no files ending with ".safetensors" were found in {EVAL_CKPT_DIR}'
+  print(sorted(eval_queue, reverse=True))
+
+  Tensor.manual_seed(seed)  # seed for weight initialization
+  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-eval", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
+
+  # load prompts for generating images for validation; 2 MB of data total
+  with open(DATADIR / "coco2014" / "val2014_30k.tsv") as f:
+    reader = csv.DictReader(f, delimiter="\t")
+    eval_inputs:list[dict] = [{"image_id": int(row["image_id"]), "id": int(row["id"]), "caption": row["caption"]} for row in reader]
+  assert len(eval_inputs) == 30_000
+  # NOTE: the clip weights are the same between model.cond_stage_model and clip_encoder
+  eval_timesteps = list(reversed(range(1, 1000, 20)))
+
+  original_device, Device.DEFAULT = Device.DEFAULT, "CPU"
+  # The choice of alphas_prev[0] = alphas_cumprod[0] seems arbitrary, but it's how the mlperf ref does it:
+  #   alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+  eval_alphas_prev = model.alphas_cumprod[0:1].cat(model.alphas_cumprod[list(range(1, 1000, 20))[:-1]]).to(GPUS).realize()
+  inception = FidInceptionV3().load_from_pretrained(CKPTDIR / "inception" / "pt_inception-2015-12-05-6726825d.pth")
+  vision_cfg = {'width': 1280, 'layers': 32, 'd_head': 80, 'image_size': 224, 'patch_size': 14}
+  text_cfg = {'width': 1024, 'n_heads': 16, 'layers': 24, 'vocab_size': 49408, 'ctx_length': 77}
+  clip.gelu = gelu_erf
+  clip_encoder = OpenClipEncoder(1024, text_cfg, vision_cfg)
+  loaded = torch_load(CKPTDIR / "clip" / "open_clip_pytorch_model.bin")
+  loaded.update({"attn_mask": clip_encoder.attn_mask, "mean": clip_encoder.mean, "std": clip_encoder.std})
+  load_state_dict(clip_encoder, loaded)
+  Device.DEFAULT=original_device
+
+  @TinyJit
+  def denoise_step(x:Tensor, x_x:Tensor, t_t:Tensor, uc_c:Tensor, sqrt_alphas_cumprod_t:Tensor, sqrt_one_minus_alphas_cumprod_t:Tensor,
+                    alpha_prev:Tensor, unet:UNetModel, GPUS) -> Tensor:
+    out_uncond, out = unet(x_x, t_t, uc_c).to("CPU").reshape(-1, 2, 4, 64, 64).chunk(2, dim=1)
+    out_uncond = out_uncond.squeeze(1).shard(GPUS,axis=0)
+    out = out.squeeze(1).shard(GPUS,axis=0)
+    v_t = out_uncond + 8.0 * (out - out_uncond)
+    e_t = sqrt_alphas_cumprod_t * v_t + sqrt_one_minus_alphas_cumprod_t * x
+    pred_x0 = sqrt_alphas_cumprod_t * x - sqrt_one_minus_alphas_cumprod_t * v_t
+    dir_xt = (1. - alpha_prev).sqrt() * e_t
+    x_prev = alpha_prev.sqrt() * pred_x0 + dir_xt
+    return x_prev.realize()
+
+  def shard_tensor(t:Tensor) -> Tensor: return t.shard(GPUS, axis=0) if len(GPUS) > 1 else t.to(GPUS[0])
+  def get_batch(whole:Tensor, i:int, bs:int) -> tuple[Tensor, int]:
+    batch = whole[i: i + bs].to("CPU")
+    if (unpadded_bs:=batch.shape[0]) < bs:
+      batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
+    return batch, unpadded_bs 
+
+  @Tensor.train(mode=False)
+  def eval_unet(eval_inputs:list[dict], unet:UNetModel, cond_stage:FrozenOpenClipEmbedder, first_stage:AutoencoderKL,
+                inception:FidInceptionV3, clip:OpenClipEncoder) -> tuple[float, float]:
+    # Eval is divided into 5 jits, one per model
+    # It doesn't make sense to merge these jits, e.g. unet repeats 50 times in isolation; images fork to separate inception/clip
+    # We're generating and scoring 30,000 images per eval, and all the data can flow through one jit at a time
+    # To maximize throughput for each jit, we have only one model/jit on the GPU at a time, and pool outputs from each jit off-GPU
+    for model in (unet, first_stage, inception, clip):
+      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
+
+    uc_written = False
+    models = (cond_stage, unet, first_stage, inception, clip)
+    jits = (jit_context:=TinyJit(cond_stage.embed_tokens), denoise_step, vae_decode, jit_inception:=TinyJit(inception),
+            jit_clip:=TinyJit(clip.get_clip_score))
+    all_bs = (CONTEXT_BS, DENOISE_BS, DECODE_BS, INCEPTION_BS, CLIP_BS)
+    if (EVAL_SAMPLES:=getenv("EVAL_SAMPLES", 0)) and EVAL_SAMPLES > 0:
+      eval_inputs = eval_inputs[0:EVAL_SAMPLES]
+    output_shapes = [(ns:=len(eval_inputs),77), (ns,77,1024), (ns,4,64,64), (ns,3,512,512), (ns,2048), (ns,)]
+    # Writing progress to disk lets us resume eval if we crash
+    stages = ["tokens", "embeds", "latents", "imgs", "inception", "clip"]
+    disk_tensor_names, disk_tensor_shapes = stages + ["end", "uc"], output_shapes + [(6,), (1,77,1024)]
+    if not all(os.path.exists(f"{EVAL_CKPT_DIR}/{name}.bytes") for name in disk_tensor_names):
+      for name, shape in zip(disk_tensor_names, disk_tensor_shapes):
+        file = Path(f"{EVAL_CKPT_DIR}/{name}.bytes")
+        file.unlink(missing_ok=True)
+        with file.open("wb") as f: f.truncate(prod(shape) * 4)
+    progress = {name: Tensor.empty(*shape, device=f"disk:{EVAL_CKPT_DIR}/{name}.bytes", dtype=dtypes.int if name in {"tokens", "end"} else dtypes.float)
+                for name, shape in zip(disk_tensor_names, disk_tensor_shapes)}
+
+    def embed_tokens(tokens:Tensor) -> Tensor:
+      nonlocal uc_written
+      if not uc_written:
+        with Context(BEAM=0): progress["uc"].assign(cond_stage.embed_tokens(cond_stage.tokenize("").to(GPUS)).to("CPU").realize()).realize()
+        uc_written = True
+      return jit_context(shard_tensor(tokens))
+
+    def generate_latents(embeds:Tensor) -> Tensor:
+      uc_c = Tensor.stack(progress["uc"].to("CPU").expand(bs, 77, 1024), embeds, dim=1).reshape(-1, 77, 1024)
+      uc_c = shard_tensor(uc_c)
+      x = shard_tensor(Tensor.randn(bs,4,64,64))
+      for step_idx, timestep in enumerate(tqdm(eval_timesteps)):
+        reversed_idx = Tensor([50 - step_idx - 1], device=GPUS)
+        alpha_prev = eval_alphas_prev[reversed_idx]
+        ts = Tensor.full(bs, fill_value=timestep, dtype=dtypes.int, device="CPU")
+        ts_ts = shard_tensor(ts.cat(ts))
+        ts = shard_tensor(ts)
+        sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
+        sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
+        x_x = shard_tensor(Tensor.stack(x.to("CPU"), x.to("CPU"), dim=1).reshape(-1, 4, 64, 64))
+        x.assign(denoise_step(x, x_x, ts_ts, uc_c, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t, alpha_prev, unet, GPUS)).realize()
+      return x
+
+    def decode_latents(latents:Tensor) -> Tensor: return vae_decode(shard_tensor(latents), first_stage, disable_beam=True)
+    def generate_inception(imgs:Tensor) -> Tensor: return jit_inception(shard_tensor(imgs))[:,:,0,0]
+
+    def calc_clip_scores(batch:Tensor, batch_tokens:Tensor) -> Tensor:
+      # Tensor.interpolate does not yet support bicubic, so we use PIL
+      batch = (batch.to(GPUS[0]).permute(0,2,3,1) * 255).clip(0, 255).cast(dtypes.uint8).numpy()
+      batch = [np.array(PIL.Image.fromarray(batch[i]).resize((224,224), PIL.Image.BICUBIC)) for i in range(bs)]
+      batch = shard_tensor(Tensor(np.stack(batch, axis=0).transpose(0,3,1,2), device="CPU").realize())
+      batch = batch.cast(dtypes.float) / 255
+      batch = (batch - model.mean) / model.std
+      batch = jit_clip(shard_tensor(batch_tokens), batch)
+      return batch
+
+    callbacks = (embed_tokens, generate_latents, decode_latents, generate_inception, calc_clip_scores)
+
+    # save every forward pass output to disk; NOTE: this needs ~100 GB disk space because 30k images are large
+    def stage_progress(stage_idx:int) -> int: return progress["end"].to("CPU")[stage_idx].item()
+    if stage_progress(0) < len(eval_inputs):
+      tokens = []
+      for i in tqdm(range(0, len(eval_inputs), CONTEXT_BS)):
+        subset = [cond_stage.tokenize(row["caption"], device="CPU") for row in eval_inputs[i: i+CONTEXT_BS]]
+        tokens.append(Tensor.cat(*subset, dim=0).realize())
+      progress["tokens"].assign(Tensor.cat(*tokens, dim=0).realize()).realize()
+      progress["end"][0:1].assign(Tensor([len(eval_inputs)], dtype=dtypes.int)).realize()
+    prev_stage = "tokens"
+    tokens = progress["tokens"]
+
+    # wrapper code for every model
+    for stage_idx, model, jit, bs, callback in zip(range(1,6), models, jits, all_bs, callbacks):
+      stage = stages[stage_idx]
+      if stage_progress(stage_idx) >= len(eval_inputs):
+        prev_stage = stage
+        continue # use cache
+      t0 = time.perf_counter()
+      print(f"starting eval with model: {model}")
+      if stage_idx == 1: inputs = tokens
+      elif stage_idx == 5: inputs = progress["imgs"]
+      else: inputs = progress[prev_stage]
+
+      Tensor.realize(*[p.to_(GPUS) for p in get_parameters(model)])
+      for batch_idx in tqdm(range(stage_progress(stage_idx), inputs.shape[0], bs)):
+        t1 = time.perf_counter()
+        batch, unpadded_bs = get_batch(inputs, batch_idx, bs)
+        if isinstance(model, OpenClipEncoder): batch = callback(batch, get_batch(tokens, batch_idx, bs)[0].realize())
+        else: batch = callback(batch)
+        # to(GPUS[0]) is necessary for this to work, without that the result is still on GPUS, probably due to a bug
+        batch = batch.to(GPUS[0]).to("CPU")[0:unpadded_bs].realize()
+        progress[stage][batch_idx: batch_idx + bs].assign(batch).realize()
+        # keep track of what our last output was, so we can resume from there if we crash in this loop
+        progress["end"][stage_idx: stage_idx + 1].assign(Tensor([batch_idx + bs], dtype=dtypes.int)).realize()
+        print(f"model: {model}, batch_idx: {batch_idx}, elapsed: {(time.perf_counter() - t1):.2f}")
+      del batch
+        
+      jit.reset()
+      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
+      print(f"done with model: {model}, elapsed: {(time.perf_counter() - t0):.2f}")
+      prev_stage = stage
+
+    inception_stats_fn = str(DATADIR / "coco2014" / "val2014_30k_stats.npz")
+    fid_score = inception.compute_score(progress["inception"].to("CPU"), inception_stats_fn)
+    clip_score = progress["clip"].to(GPUS[0]).mean().item()
+    for name in disk_tensor_names:
+      Path(f"{EVAL_CKPT_DIR}/{name}.bytes").unlink(missing_ok=True)
+    
+    if EVAL_SAMPLES and BEAM:
+      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
+      sys.exit() # Don't eval additional models; we don't care about clip/fid scores when running BEAM on eval sample subset
+
+    return clip_score, fid_score
+
+  # evaluate checkpoints in reverse chronological order
+  for ckpt_iteration, p in sorted(eval_queue, reverse=True):
+    unet_ckpt = safe_load(p)
+    load_state_dict(unet, unet_ckpt)
+    clip_score, fid_score = eval_unet(eval_inputs, unet, model.cond_stage_model, model.first_stage_model, inception, clip_encoder)
+    converged = True if clip_score >= 0.15 and fid_score <= 90 else False
+    print(f"eval results for {EVAL_CKPT_DIR}/{p.name}: clip={clip_score}, fid={fid_score}, converged={converged}")
+    if WANDB:
+      wandb.log({"eval/ckpt_iteration": ckpt_iteration, "eval/clip_score": clip_score, "eval/fid_score": fid_score})
+    if converged and STOP_IF_CONVERGED:
+      print(f"Convergence detected, exiting early before evaluating other checkpoints due to STOP_IF_CONVERGED={STOP_IF_CONVERGED}")
+      sys.exit()
+
+  # for testing
+  return clip_score, fid_score, ckpt_iteration
+
 if __name__ == "__main__":
  # inference only
  Tensor.training = False
@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing

 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
-from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, FUSE_CONV_BW, Profiling
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
 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

@@ -707,7 +707,7 @@ def train_unet3d():
  ```BASEDIR=<folder_path> ./examples/mlperf/scripts/setup_kits19_dataset.sh```

  2) To start training the model, run the following:
-  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 FUSE_CONV_BW=1 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
+  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
  """
  from examples.mlperf.losses import dice_ce_loss
  from examples.mlperf.metrics import dice_score
@@ -749,7 +749,6 @@ def train_unet3d():
    "train_beam": TRAIN_BEAM,
    "eval_beam": EVAL_BEAM,
    "wino": WINO.value,
-    "fuse_conv_bw": FUSE_CONV_BW.value,
    "gpus": GPUS,
    "default_float": dtypes.default_float.name
  }
@@ -1189,7 +1188,9 @@ def train_bert():
      if MLLOGGER and RUNMLPERF:
        MLLOGGER.start(key=mllog_constants.EVAL_START, value=None, metadata={"epoch_num": i*GBS, "step_num": i})
      if getenv("RESET_STEP"): train_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE", 1) and train_step_bert.captured is not None: train_step_bert.captured.free_intermediates()
+      elif getenv("FREE_INTERMEDIATE", 0) and train_step_bert.captured is not None:
+        # TODO: FREE_INTERMEDIATE nan'ed after jit step 2
+        train_step_bert.captured.free_intermediates()
      eval_lm_losses = []
      eval_clsf_losses = []
      eval_lm_accs = []
@@ -1223,7 +1224,7 @@ def train_bert():
          return

      if getenv("RESET_STEP"): eval_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE", 1) and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()
+      elif getenv("FREE_INTERMEDIATE", 0) and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()

      del eval_data
      avg_lm_loss = sum(eval_lm_losses) / len(eval_lm_losses)
@@ -1309,7 +1310,7 @@ def train_llama3():
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)

-  # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 FUSE_ARANGE=1 JITBEAM=2 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=512 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py
+  # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 JITBEAM=2 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=512 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py
  # trains to 7

  opt_adamw_beta_1 = 0.9
@@ -1493,6 +1494,144 @@ def train_llama3():
          safe_save(get_state_dict(model), fn)
        break

+def train_stable_diffusion():
+  from extra.models.unet import UNetModel
+  from examples.mlperf.dataloader import batch_load_train_stable_diffusion
+  from examples.mlperf.lr_schedulers import LambdaLR, LambdaLinearScheduler
+  from examples.mlperf.initializers import init_stable_diffusion
+  from examples.mlperf.helpers import get_training_state
+  import numpy as np
+
+  config = {}
+  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
+  seed               = config["seed"]                   = getenv("SEED", 12345)
+  # ** hyperparameters **
+  BS                 = config["BS"]                     = getenv("BS", 1 * len(GPUS))
+  BASE_LR            = config["LEARNING_RATE"]          = getenv("LEARNING_RATE", 2.5e-7)
+  # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
+  # "Checkpoint must be collected every 512,000 images. CEIL(512000 / global_batch_size) if 512000 is not divisible by GBS."
+  # NOTE: It's inferred that "steps" is the unit for the output of the CEIL formula, based on all other cases of CEIL in the rules
+  CKPT_STEP_INTERVAL = config["CKPT_STEP_INTERVAL"]     = getenv("CKPT_STEP_INTERVAL", math.ceil(512_000 / BS))
+  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
+  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
+  UNET_CKPTDIR       = config["UNET_CKPTDIR"]           = Path(getenv("UNET_CKPTDIR", "./checkpoints"))
+  TOTAL_CKPTS        = config["TOTAL_CKPTS"]            = getenv("TOTAL_CKPTS", 0)
+
+  print(f"training on {GPUS}")
+  lr = BS * BASE_LR
+  print(f"BS={BS}, BASE_LR={BASE_LR}, lr={lr}")
+  print(f"CKPT_STEP_INTERVAL = {CKPT_STEP_INTERVAL}")
+  for x in GPUS: Device[x]
+  if (WANDB := getenv("WANDB", "")):
+    import wandb
+    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
+
+  Tensor.manual_seed(seed)  # seed for weight initialization
+  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-train", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
+
+  optimizer = AdamW(get_parameters(unet))
+  lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
+  lr_scheduler = LambdaLR(optimizer, Tensor(lr, dtype=dtypes.float, device=optimizer.device), lambda_lr_callback)
+
+  @TinyJit
+  def train_step(mean:Tensor, logvar:Tensor, tokens:Tensor, unet:UNetModel, optimizer:LAMB, lr_scheduler:LambdaLR) -> Tensor:
+    optimizer.zero_grad()
+
+    timestep = Tensor.randint(BS, low=0, high=model.alphas_cumprod.shape[0], dtype=dtypes.int, device=GPUS[0])
+    latent_randn = Tensor.randn(*mean.shape, device=GPUS[0])
+    noise = Tensor.randn(*mean.shape, device=GPUS[0])
+    for t in (mean, logvar, tokens, timestep, latent_randn, noise):
+      t.shard_(GPUS, axis=0)
+
+    std = Tensor.exp(0.5 * logvar.clamp(-30.0, 20.0))
+    latent = (mean + std * latent_randn) * 0.18215
+
+    sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
+    sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
+    latent_with_noise = sqrt_alphas_cumprod_t * latent + sqrt_one_minus_alphas_cumprod_t * noise
+    v_true = sqrt_alphas_cumprod_t * noise - sqrt_one_minus_alphas_cumprod_t * latent
+
+    context = model.cond_stage_model.embed_tokens(tokens)
+
+    out = unet(latent_with_noise, timestep, context)
+    loss = ((out - v_true) ** 2).mean()
+    del mean, logvar, std, latent, noise, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t
+    del out, v_true, context, latent_randn, tokens, timestep
+    loss.backward()
+
+    optimizer.step()
+    lr_scheduler.step()
+    loss, out_lr = loss.detach().to("CPU"), optimizer.lr.to("CPU")
+    Tensor.realize(loss, out_lr)
+    return loss, out_lr
+    
+  # checkpointing takes ~9 minutes without this, and ~1 minute with this
+  @TinyJit
+  def ckpt_to_cpu():
+    ckpt = get_training_state(unet, optimizer, lr_scheduler)
+    # move to CPU first so more GPU bufs aren't created (can trigger OOM)
+    for k,v in ckpt.items(): ckpt[k] = v.detach().to("CPU")
+    Tensor.realize(*[v for v in ckpt.values()])
+    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype.base).contiguous()
+    Tensor.realize(*[v for v in ckpt.values()])
+    return ckpt
+
+  # training loop
+  dl = batch_load_train_stable_diffusion(f'{DATADIR}/laion-400m/webdataset-moments-filtered/{{00000..00831}}.tar', BS)
+  # for tests
+  saved_checkpoints = []
+
+  train_start_time = time.perf_counter()
+  t0 = t6 = time.perf_counter()
+  for i, batch in enumerate(dl, start=1):
+    loop_time = time.perf_counter() - t0
+    t0 = time.perf_counter()
+    dl_time = t0 - t6
+    GlobalCounters.reset()
+
+    mean, logvar = np.split(np.concatenate(batch["npy"], axis=0), 2, axis=1)
+    mean, logvar = Tensor(mean, dtype=dtypes.float32, device="CPU"), Tensor(logvar, dtype=dtypes.float32, device="CPU")
+    tokens = []
+    for text in batch['txt']: tokens += model.cond_stage_model.tokenizer.encode(text, pad_with_zeros=True)
+    tokens = Tensor(tokens, dtype=dtypes.int32, device="CPU").reshape(-1, 77)
+
+    t1 = time.perf_counter()
+    loss, lr = train_step(mean, logvar, tokens, unet, optimizer, lr_scheduler)
+    loss_item, lr_item = loss.item(), lr.item()
+    t2 = time.perf_counter()
+
+    if i == 3:
+      for _ in range(3): ckpt_to_cpu() # do this at the beginning of run to prevent OOM surprises when checkpointing
+      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
+      
+    total_train_time = time.perf_counter() - train_start_time
+    if WANDB:
+      wandb.log({"train/loss": loss_item, "train/lr": lr_item, "train/loop_time_prev": loop_time, "train/dl_time": dl_time, "train/step": i,
+                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (t2-t1), "train/input_prep_time": t1-t0,
+                    "train/train_step_time": t2-t1, "train/total_time": total_train_time})
+
+      if i == 1 and wandb.run is not None:
+        with open(f"{UNET_CKPTDIR}/wandb_run_id_{wandb.run.id}", "w") as f:
+          f.write(f"wandb.run.id = {wandb.run.id}")
+
+    if i % CKPT_STEP_INTERVAL == 0:
+      # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
+      # "evaluation is done offline, the time is not counted towards the submission time."
+      fn = f"{UNET_CKPTDIR}/{i}.safetensors"
+      print(f"saving unet checkpoint at {fn}")
+      saved_checkpoints.append(fn)
+      safe_save({k.replace("model.", ""):v for k,v in ckpt_to_cpu().items() if k.startswith("model.")}, fn)
+      if TOTAL_CKPTS and i == TOTAL_CKPTS * CKPT_STEP_INTERVAL:
+        print(f"ending run after {i} steps ({TOTAL_CKPTS} checkpoints collected)")
+        return saved_checkpoints
+
+    t3 = time.perf_counter()
+    print(f"""step {i}: {GlobalCounters.global_ops * 1e-9 / (t2-t1):9.2f} GFLOPS, mem_used: {GlobalCounters.mem_used / 1e9:.2f} GB,
+  loop_time_prev: {loop_time:.2f}, dl_time: {dl_time:.2f}, input_prep_time: {t1-t0:.2f}, train_step_time: {t2-t1:.2f},
+  t3-t2: {t3-t2:.4f}, loss:{loss_item:.5f}, lr:{lr_item:.3e}, total_train_time:{total_train_time:.2f}
+  """)
+    t6 = time.perf_counter()
+
 if __name__ == "__main__":
  multiprocessing.set_start_method('spawn')

@@ -1501,7 +1640,7 @@ if __name__ == "__main__":
  else: bench_log_manager = contextlib.nullcontext()

  with Tensor.train():
-    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn").split(","):
+    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
      nm = f"train_{m}"
      if nm in globals():
        print(f"training {m}")
@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+# adapted from https://github.com/mlcommons/training/blob/4bdf5c8ed218ad76565a2ba1ac27c919ccc6d689/stable_diffusion/README.md
+
+# setup dirs
+
+DATA=/raid/datasets/stable_diffusion
+
+LAION=$DATA/laion-400m/webdataset-moments-filtered 
+COCO=$DATA/coco2014
+mkdir -p $LAION $COCO
+
+CKPT=/raid/weights/stable_diffusion
+mkdir -p $CKPT/clip $CKPT/sd $CKPT/inception
+
+# download data
+
+# if rclone isn't installed system-wide / in your PATH, put the executable path in quotes below
+#RCLONE=""
+RCLONE="rclone"
+
+## VAE-encoded image latents, from 6.1M image subset of laion-400m
+## about 1 TB for whole download
+$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
+$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/ ${LAION} --include="*.tar" -P
+$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/sha512sums.txt ${LAION} -P
+cd $LAION && grep -E '\.tar$' sha512sums.txt | sha512sum -c --quiet - && \
+  echo "All .tar files verified" || { echo "Checksum failure when validating downloaded Laion moments"; exit 1; }
+
+## prompts and FID statistics from 30k image subset of coco2014
+## 33 MB
+$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
+$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k.tsv ${COCO} -P
+
+$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
+$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k_stats.npz ${COCO} -P
+
+# download checkpoints
+
+## clip (needed for text and vision encoders for validation)
+CLIP_WEIGHTS_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin"
+CLIP_WEIGHTS_SHA256="9a78ef8e8c73fd0df621682e7a8e8eb36c6916cb3c16b291a082ecd52ab79cc4"
+CLIP_CONFIG_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/raw/main/open_clip_config.json"
+wget -N -P ${CKPT}/clip ${CLIP_WEIGHTS_URL}
+wget -N -P ${CKPT}/clip ${CLIP_CONFIG_URL}
+echo "${CLIP_WEIGHTS_SHA256}  ${CKPT}/clip/open_clip_pytorch_model.bin"                    | sha256sum -c
+
+## sd (needed for latent->image decoder for validation, also has clip text encoder for training)
+SD_WEIGHTS_URL='https://huggingface.co/stabilityai/stable-diffusion-2-base/resolve/main/512-base-ema.ckpt'
+SD_WEIGHTS_SHA256="d635794c1fedfdfa261e065370bea59c651fc9bfa65dc6d67ad29e11869a1824"
+wget -N -P ${CKPT}/sd ${SD_WEIGHTS_URL}
+echo "${SD_WEIGHTS_SHA256}  ${CKPT}/sd/512-base-ema.ckpt"                    | sha256sum -c
+
+## inception (needed for validation)
+FID_WEIGHTS_URL='https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth'
+FID_WEIGHTS_SHA1="bd836944fd6db519dfd8d924aa457f5b3c8357ff"
+wget -N -P ${CKPT}/inception ${FID_WEIGHTS_URL}
+echo "${FID_WEIGHTS_SHA1}  ${CKPT}/inception/pt_inception-2015-12-05-6726825d.pth"                    | sha1sum -c
@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+
+DATETIME=${2:-$(date "+%m%d%H%M")}
+LOGFILE="${HOME}/logs/sd_mi300x_${DATETIME}.log"
+# UNET_CKPTDIR must be set: training saves checkpoints to this path, then a separate eval process scans this path to know which checkpoints to eval
+export UNET_CKPTDIR="${HOME}/stable_diffusion/training_checkpoints/${DATETIME}"
+mkdir -p "${HOME}/logs" "$UNET_CKPTDIR"
+
+# run this script in isolation when using the --bg flag
+if [[ "${1:-}" == "--bg" ]]; then
+  echo "logging output to $LOGFILE"
+  echo "saving UNet checkpoints to $UNET_CKPTDIR"
+  script_path="$(readlink -f "${BASH_SOURCE[0]}")"
+  nohup bash "$script_path" run "$DATETIME" >"$LOGFILE" 2>&1 & disown $!
+  exit 0
+fi
+
+# venv management
+if [[ -d .venv-sd-mlperf ]]; then
+  . .venv-sd-mlperf/bin/activate
+else
+  python3 -m venv .venv-sd-mlperf && . .venv-sd-mlperf/bin/activate
+  pip install --index-url https://download.pytorch.org/whl/cpu torch && pip install tqdm numpy ftfy regex pillow scipy wandb webdataset
+fi
+pip list
+apt list --installed | grep amdgpu
+rocm-smi --version
+modinfo amdgpu | grep version
+
+export BEAM=2 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 IGNORE_JIT_FIRST_BEAM=1 HCQDEV_WAIT_TIMEOUT_MS=300000
+export AMD_LLVM=0 # bf16 seems to require this
+export DATADIR="/raid/datasets/stable_diffusion"
+export CKPTDIR="/raid/weights/stable_diffusion"
+export EVAL_CKPT_DIR=$UNET_CKPTDIR
+export MODEL="stable_diffusion" PYTHONPATH="."
+export GPUS=8 BS=304
+export CONTEXT_BS=816 DENOISE_BS=600 DECODE_BS=384 INCEPTION_BS=560 CLIP_BS=240
+export WANDB=1
+export PARALLEL=4
+export PYTHONUNBUFFERED=1
+sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setperfdeterminism 1500 || exit 1
+
+# Retry BEAM search if script fails before BEAM COMPLETE is printed, but don't retry after that
+run_retry(){ local try=0 max=5 code tmp py pgid kids
+  while :; do
+    tmp=$(mktemp)
+    setsid bash -c 'exec env "$@"' _ "$@" > >(tee -a "$LOGFILE" | tee "$tmp") 2>&1 &
+    py=$!; pgid=$(ps -o pgid= -p "$py" | tr -d ' ')
+    wait "$py"; code=$?
+    [[ -n "$pgid" ]] && { kill -TERM -"$pgid" 2>/dev/null; sleep 1; kill -KILL -"$pgid" 2>/dev/null; }
+    kids=$(pgrep -P "$py" || true)
+    while [[ -n "$kids" ]]; do
+      kill -TERM $kids 2>/dev/null; sleep 0.5
+      kids=$(for k in $kids; do pgrep -P "$k" || true; done)
+    done
+    grep -q 'BEAM COMPLETE' "$tmp" && { rm -f "$tmp"; return 1; }
+    rm -f "$tmp"
+    ((code==0)) && return 0
+    ((try>=max)) && return 2
+    ((try++)); sleep 90; echo "try = ${try}"
+  done
+}
+
+# Power limiting to 400W is only needed if GPUs fall out of sync (causing 2.2x increased train time) at higher power, which has been observed at 450W
+sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setpoweroverdrive 750 && \
+run_retry TOTAL_CKPTS=7 python3 examples/mlperf/model_train.py; (( $? == 2 )) && { echo "training failed before BEAM completion"; exit 2; }
+sleep 90
+
+run_retry EVAL_SAMPLES=600 python3 examples/mlperf/model_eval.py; (( $? == 2 )) && { echo "eval failed before BEAM completion"; exit 2; }
+# Checkpoints will be evaluated in reverse chronological order, even if above training crashed early
+# STOP_IF_CONVERGED=1: Stop the eval after the first time convergence is detected; no more checkpoints will be evaluated after that.
+STOP_IF_CONVERGED=1 python3 examples/mlperf/model_eval.py
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
+
+export IGNORE_OOB=1
+
+export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+# export BEAM_LOG_SURPASS_MAX=1
+# export BASEDIR="/raid/datasets/wiki"
+
+export RESET_STEP=1
+export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,69 @@
+# 1. Problem
+
+This problem uses BERT for NLP.
+
+## Requirements
+
+Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+Also install gdown (for dataset), numpy, tqdm and tensorflow.
+```
+pip install gdown numpy tqdm tensorflow
+```
+
+### tinybox_green
+Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
+This is the default on production tinybox green.
+
+# 2. Directions
+
+## Steps to download and verify data
+
+### 1. Download raw data
+
+```
+BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
+```
+
+### 2. Preprocess train and validation data
+
+Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
+
+#### Training:
+```
+BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
+```
+
+Generating a specific topic (Between 0 and 499)
+```
+BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
+```
+
+#### Validation:
+```
+BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
+```
+## Running
+
+### tinybox_green
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
+```
+
+### tinybox_red
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
+```
+### tinybox_8xMI300X
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
+```
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+
+export BENCHMARK=10 BERT_LAYERS=2
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+
+# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+export TRAIN_STEPS=3900
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -0,0 +1,31 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export SUBMISSION_PLATFORM="tinybox_8xMI300X"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+
+# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+export TRAIN_STEPS=3900
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=1
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="bert_8xMI300x_${DATETIME}_${SEED}.log"
+
+BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,69 @@
+# 1. Problem
+
+This problem uses BERT for NLP.
+
+## Requirements
+
+Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+Also install gdown (for dataset), numpy, tqdm and tensorflow.
+```
+pip install gdown numpy tqdm tensorflow
+```
+
+### tinybox_green
+Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
+This is the default on production tinybox green.
+
+# 2. Directions
+
+## Steps to download and verify data
+
+### 1. Download raw data
+
+```
+BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
+```
+
+### 2. Preprocess train and validation data
+
+Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
+
+#### Training:
+```
+BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
+```
+
+Generating a specific topic (Between 0 and 499)
+```
+BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
+```
+
+#### Validation:
+```
+BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
+```
+## Running
+
+### tinybox_green
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
+```
+
+### tinybox_red
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
+```
+### tinybox_8xMI300X
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
+```
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BEAM_LOG_SURPASS_MAX=1
+export BASEDIR="/raid/datasets/wiki"
+
+export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,16 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/wiki"
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -0,0 +1,28 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." NV=1
+export MODEL="bert"
+export SUBMISSION_PLATFORM="tinybox_green"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/wiki"
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=1
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="bert_green_${DATETIME}_${SEED}.log"
+
+# init
+BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,69 @@
+# 1. Problem
+
+This problem uses BERT for NLP.
+
+## Requirements
+
+Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+Also install gdown (for dataset), numpy, tqdm and tensorflow.
+```
+pip install gdown numpy tqdm tensorflow
+```
+
+### tinybox_green
+Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
+This is the default on production tinybox green.
+
+# 2. Directions
+
+## Steps to download and verify data
+
+### 1. Download raw data
+
+```
+BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
+```
+
+### 2. Preprocess train and validation data
+
+Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
+
+#### Training:
+```
+BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
+```
+
+Generating a specific topic (Between 0 and 499)
+```
+BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
+```
+
+#### Validation:
+```
+BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
+```
+## Running
+
+### tinybox_green
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
+```
+
+### tinybox_red
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
+```
+### tinybox_8xMI300X
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
+```
@@ -0,0 +1,18 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BEAM_LOG_SURPASS_MAX=1
+export BASEDIR="/raid/datasets/wiki"
+
+export RESET_STEP=1
+export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,16 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/wiki"
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -0,0 +1,31 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export SUBMISSION_PLATFORM="tinybox_red"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+
+export IGNORE_OOB=1
+export REWRITE_STACK_LIMIT=500000
+
+export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/wiki"
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=1
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="bert_red_${DATETIME}_${SEED}.log"
+
+export HCQDEV_WAIT_TIMEOUT_MS=100000  # prevents hang?
+
+# init
+sleep 5 && sudo rmmod amdgpu || true
+BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,50 @@
+# 1. Problem
+
+This problem uses the ResNet-50 CNN to do image classification.
+
+## Requirements
+
+Install tinygrad and mlperf-logging from master.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+
+### tinybox_green
+Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
+This is the default on production tinybox green.
+
+### tinybox_red
+Disable cwsr
+This is the default on production tinybox red.
+```
+sudo vi /etc/modprobe.d/amdgpu.conf
+cat <<EOF > /etc/modprobe.d/amdgpu.conf
+options amdgpu cwsr_enable=0
+EOF
+sudo update-initramfs -u
+sudo reboot
+
+# validate
+sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
+```
+
+# 2. Directions
+
+## Steps to download and verify data
+
+```
+IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
+```
+
+## Steps for one time setup
+
+### tinybox_red
+```
+examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
+```
+
+## Steps to run benchmark
+```
+examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
+```
@@ -0,0 +1,13 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="resnet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
+
+export RESET_STEP=0
+
+export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
+
+export BENCHMARK=10 DEBUG=2
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="resnet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
+
+export RESET_STEP=0
+
+export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
+
+export EVAL_START_EPOCH=3 EVAL_FREQ=4
+
+export WANDB=1 PARALLEL=0
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,25 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." NV=1
+export MODEL="resnet"
+export SUBMISSION_PLATFORM="tinybox_green"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
+
+export RESET_STEP=0
+
+export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=${LOGMLPERF:-1}
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="resnet_green_${DATETIME}_${SEED}.log"
+
+# init
+BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,50 @@
+# 1. Problem
+
+This problem uses the ResNet-50 CNN to do image classification.
+
+## Requirements
+
+Install tinygrad and mlperf-logging from master.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+
+### tinybox_green
+Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
+This is the default on production tinybox green.
+
+### tinybox_red
+Disable cwsr
+This is the default on production tinybox red.
+```
+sudo vi /etc/modprobe.d/amdgpu.conf
+cat <<EOF > /etc/modprobe.d/amdgpu.conf
+options amdgpu cwsr_enable=0
+EOF
+sudo update-initramfs -u
+sudo reboot
+
+# validate
+sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
+```
+
+# 2. Directions
+
+## Steps to download and verify data
+
+```
+IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
+```
+
+## Steps for one time setup
+
+### tinybox_red
+```
+examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
+```
+
+## Steps to run benchmark
+```
+examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
+```
@@ -0,0 +1,13 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="resnet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
+
+export RESET_STEP=0
+
+export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export BENCHMARK=10 DEBUG=${DEBUG:-2}
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="resnet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
+
+export RESET_STEP=0
+
+export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export EVAL_START_EPOCH=3 EVAL_FREQ=4
+
+export WANDB=1 PARALLEL=0
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,26 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." AMD=1
+export MODEL="resnet"
+export SUBMISSION_PLATFORM="tinybox_red"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
+
+export RESET_STEP=0
+
+export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=${LOGMLPERF:-1}
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="resnet_red_${DATETIME}_${SEED}.log"
+
+# init
+sleep 5 && sudo rmmod amdgpu || true
+BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,8 @@
+#!/bin/bash
+
+rocm-smi --setprofile compute
+rocm-smi --setmclk 3
+rocm-smi --setperflevel high
+
+# power cap to 350W
+echo "350000000" | sudo tee /sys/class/drm/card{1..6}/device/hwmon/hwmon*/power1_cap
@@ -0,0 +1,38 @@
+# 1. Problem
+
+This problem uses RetinaNet for SSD.
+
+## Requirements
+
+Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+
+Also install the following dependencies:
+```
+pip install tqdm numpy pycocotools boto3 pandas torch torchvision
+```
+
+### tinybox_green
+Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
+This is the default on production tinybox green.
+
+# 2. Directions
+
+## Steps to download data
+
+Run the following:
+```
+BASEDIR=/raid/datasets/openimages python3 extra/datasets/openimages.py
+```
+
+## Running
+
+### tinybox_green
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh
+```
@@ -0,0 +1,14 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export BENCHMARK=5 DEBUG=2
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export WANDB=1 PARALLEL=0
+export RUNMLPERF=1
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,25 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." NV=1
+export MODEL="retinanet"
+export SUBMISSION_PLATFORM="tinybox_green"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export TRAIN_BEAM=2 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/openimages"
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=1
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="retinanet_green_${DATETIME}_${SEED}.log"
+
+# init
+BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,14 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export BENCHMARK=5 DEBUG=2
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export WANDB=1 PARALLEL=0
+export RUNMLPERF=1
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,38 @@
+{
+    "submitter": "tinycorp",
+    "division": "closed",
+    "status": "Available on-premise",
+    "system_name": "tinybox 8xMI300X",
+    "number_of_nodes": "1",
+    "host_processors_per_node": "2",
+    "host_processor_model_name": "AMD EPYC 9354",
+    "host_processor_core_count": "32",
+    "host_processor_vcpu_count": "64",
+    "host_processor_frequency": "",
+    "host_processor_caches": "",
+    "host_processor_interconnect": "",
+    "host_memory_capacity": "2304GB",
+    "host_storage_type": "NVMe SSD",
+    "host_storage_capacity": "3x 4TB raid array",
+    "host_networking": "",
+    "host_networking_topology": "",
+    "host_memory_configuration": "24x 96GB DDR5",
+    "accelerators_per_node": "8",
+    "accelerator_model_name": "AMD Instinct MI300X 192GB HBM3",
+    "accelerator_host_interconnect": "PCIe 5.0 x16",
+    "accelerator_frequency": "",
+    "accelerator_on-chip_memories": "",
+    "accelerator_memory_configuration": "HBM3",
+    "accelerator_memory_capacity": "192GB",
+    "accelerator_interconnect": "",
+    "accelerator_interconnect_topology": "",
+    "cooling": "air",
+    "hw_notes": "",
+    "framework": "tinygrad, branch mlperf_training_v5.0",
+    "other_software_stack": {
+        "python": "3.10.16",
+        "ROCm": "3.0.0+94441cb"
+    },
+    "operating_system": "Ubuntu 24.04.1 LTS",
+    "sw_notes": ""
+  }
@@ -0,0 +1,38 @@
+{
+  "submitter": "tinycorp",
+  "division": "closed",
+  "status": "Available on-premise",
+  "system_name": "tinybox green",
+  "number_of_nodes": "1",
+  "host_processors_per_node": "1",
+  "host_processor_model_name": "AMD EPYC 7532",
+  "host_processor_core_count": "32",
+  "host_processor_vcpu_count": "64",
+  "host_processor_frequency": "",
+  "host_processor_caches": "",
+  "host_processor_interconnect": "",
+  "host_memory_capacity": "128GB",
+  "host_storage_type": "NVMe SSD",
+  "host_storage_capacity": "4 TB raid array + 1 TB boot",
+  "host_networking": "",
+  "host_networking_topology": "",
+  "host_memory_configuration": "8x 16GB DDR4",
+  "accelerators_per_node": "6",
+  "accelerator_model_name": "NVIDIA GeForce RTX 4090",
+  "accelerator_host_interconnect": "PCIe 4.0 x16",
+  "accelerator_frequency": "",
+  "accelerator_on-chip_memories": "",
+  "accelerator_memory_configuration": "GDDR6X",
+  "accelerator_memory_capacity": "24GB",
+  "accelerator_interconnect": "",
+  "accelerator_interconnect_topology": "",
+  "cooling": "air",
+  "hw_notes": "",
+  "framework": "tinygrad, branch mlperf_training_v5.0",
+  "other_software_stack": {
+    "python": "3.10.12",
+    "CUDA": "12.4"
+  },
+  "operating_system": "Ubuntu 22.04.4",
+  "sw_notes": ""
+}
@@ -0,0 +1,37 @@
+{
+  "submitter": "tinycorp",
+  "division": "closed",
+  "status": "Available on-premise",
+  "system_name": "tinybox red",
+  "number_of_nodes": "1",
+  "host_processors_per_node": "1",
+  "host_processor_model_name": "AMD EPYC 7532",
+  "host_processor_core_count": "32",
+  "host_processor_vcpu_count": "64",
+  "host_processor_frequency": "",
+  "host_processor_caches": "",
+  "host_processor_interconnect": "",
+  "host_memory_capacity": "128GB",
+  "host_storage_type": "NVMe SSD",
+  "host_storage_capacity": "4 TB raid array + 1 TB boot",
+  "host_networking": "",
+  "host_networking_topology": "",
+  "host_memory_configuration": "8x 16GB DDR4",
+  "accelerators_per_node": "6",
+  "accelerator_model_name": "AMD Radeon RX 7900 XTX",
+  "accelerator_host_interconnect": "PCIe 4.0 x16",
+  "accelerator_frequency": "",
+  "accelerator_on-chip_memories": "",
+  "accelerator_memory_configuration": "GDDR6",
+  "accelerator_memory_capacity": "24GB",
+  "accelerator_interconnect": "",
+  "accelerator_interconnect_topology": "",
+  "cooling": "air",
+  "hw_notes": "",
+  "framework": "tinygrad, branch mlperf_training_v5.0",
+  "other_software_stack": {
+    "python": "3.10.12"
+  },
+  "operating_system": "Ubuntu 22.04.4",
+  "sw_notes": ""
+}
@@ -1,4 +1,4 @@
-import os, sys, pickle, time
+import os, sys, pickle, time, re
 import numpy as np
 if "FLOAT16" not in os.environ: os.environ["FLOAT16"] = "1"
 if "IMAGE" not in os.environ: os.environ["IMAGE"] = "2"
@@ -10,7 +10,7 @@ from tinygrad.helpers import DEBUG, getenv
 from tinygrad.engine.realize import CompiledRunner

 import onnx
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.nn.onnx import OnnxRunner

 OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
 OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
@@ -52,6 +52,8 @@ def compile(onnx_file):
      kernel_count += 1
      read_image_count += ei.prg.p.src.count("read_image")
      gated_read_image_count += ei.prg.p.src.count("?read_image")
+      for v in [m.group(1) for m in re.finditer(r'(val\d+)\s*=\s*read_imagef\(', ei.prg.p.src)]:
+        if len(re.findall(fr'[\?\:]{v}\.[xyzw]', ei.prg.p.src)) > 0: gated_read_image_count += 1
  print(f"{kernel_count=},  {read_image_count=}, {gated_read_image_count=}")
  if (allowed_kernel_count:=getenv("ALLOWED_KERNEL_COUNT", -1)) != -1:
    assert kernel_count == allowed_kernel_count, f"different kernels! {kernel_count=}, {allowed_kernel_count=}"
@@ -77,13 +79,20 @@ def test_vs_compile(run, new_inputs, test_val=None):
            **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}

  # run 20 times
+  step_times = []
  for _ in range(20):
    st = time.perf_counter()
    out = run(**inputs)
    mt = time.perf_counter()
    val = out.numpy()
    et = time.perf_counter()
-    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {(et-st)*1e3:6.2f} ms")
+    step_times.append((et-st)*1e3)
+    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {step_times[-1]:6.2f} ms")
+
+  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+    min_time = min(step_times)
+    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
+
  print(out, val.shape, val.dtype)
  if test_val is not None: np.testing.assert_equal(test_val, val)
  print("**** test done ****")
@@ -1,8 +1,8 @@
 import sys
 from tinygrad import Tensor, fetch, GlobalCounters, dtypes
 from tinygrad.uop.ops import UOp
-from tinygrad.frontend.onnx import OnnxRunner
-from tinygrad.schedule.kernelize import get_kernelize_map
+from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.schedule.rangeify import get_rangeify_map
 from tinygrad.engine.schedule import create_schedule_with_vars
 from tinygrad.engine.realize import run_schedule

@@ -33,7 +33,7 @@ if __name__ == "__main__":
        if not in_target_path[s]:
          independent_set[s] = None
  independent = UOp.sink(*independent_set.keys())
-  kernelized = get_kernelize_map(independent)
+  kernelized = get_rangeify_map(independent)
  independent = independent.substitute(kernelized)
  schedule, var_vals = create_schedule_with_vars(independent)
  run_schedule(schedule)
@@ -27,7 +27,7 @@ class Model(nn.Module):

 if __name__ == "__main__":
  if getenv("TINY_BACKEND"):
-    import tinygrad.frontend.torch  # noqa: F401
+    import tinygrad.nn.torch  # noqa: F401
    device = torch.device("tiny")
  else:
    device = torch.device({"METAL":"mps","NV":"cuda"}.get(Device.DEFAULT, "cpu"))
@@ -9,11 +9,13 @@ from typing import Dict, Any
 from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
-from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm
+from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten
 from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import torch_load, load_state_dict, get_state_dict
-from extra.models.clip import Closed, Tokenizer
+from extra.models.clip import Closed, Tokenizer, FrozenOpenClipEmbedder
+from extra.models import unet, clip
 from extra.models.unet import UNetModel
+from examples.mlperf.initializers import AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm, zero_module, attn_f32_softmax, gelu_erf
 from extra.bench_log import BenchEvent, WallTimeEvent

 class AttnBlock:
@@ -154,12 +156,46 @@ unet_params: Dict[str,Any] = {
  "use_linear": False,
 }

+mlperf_params: Dict[str,Any] = {"adm_in_ch": None, "in_ch": 4, "out_ch": 4, "model_ch": 320, "attention_resolutions": [4, 2, 1], "num_res_blocks": 2,
+                                "channel_mult": [1, 2, 4, 4], "d_head": 64, "transformer_depth": [1, 1, 1, 1], "ctx_dim": 1024, "use_linear": True,
+                                "num_groups":16, "st_norm_eps":1e-6}
+
 class StableDiffusion:
-  def __init__(self):
+  def __init__(self, version:str|None=None, pretrained:str|None=None):
    self.alphas_cumprod = get_alphas_cumprod()
-    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_params))
-    self.first_stage_model = AutoencoderKL()
-    self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))
+    if version != "v2-mlperf-train":
+      self.first_stage_model = AutoencoderKL() # only needed for decoding generated latents to images; not needed in mlperf training from preprocessed moments
+
+    if not version:
+      self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))
+      unet_init_params = unet_params
+    elif version in {"v2-mlperf-train", "v2-mlperf-eval"}:
+      unet_init_params = mlperf_params
+      clip.gelu = gelu_erf
+      self.cond_stage_model = FrozenOpenClipEmbedder(**{"dims": 1024, "n_heads": 16, "layers": 24, "return_pooled": False, "ln_penultimate": True,
+                                                        "clip_tokenizer_version": "sd_mlperf_v5_0"})
+      unet.Linear, unet.Conv2d, unet.GroupNorm, unet.LayerNorm = AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm
+      unet.attention, unet.gelu, unet.mixed_precision_dtype = attn_f32_softmax, gelu_erf, dtypes.bfloat16
+      if pretrained:
+        print("loading text encoder")
+        weights: dict[str,Tensor] = {k.replace("cond_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("cond_stage_model.")}
+        weights["model.attn_mask"] = Tensor.full((77, 77), fill_value=float("-inf")).triu(1)
+        load_state_dict(self.cond_stage_model, weights)
+        # only the eval model needs the decoder
+        if version == "v2-mlperf-eval":
+          print("loading image latent encoder")
+          weights = {k.replace("first_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("first_stage_model.")}
+          load_state_dict(self.first_stage_model, weights)
+
+    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_init_params))
+    if version == "v2-mlperf-train":
+      # the mlperf reference inits certain weights as zeroes
+      for bb in flatten(self.model.diffusion_model.input_blocks) + self.model.diffusion_model.middle_block + flatten(self.model.diffusion_model.output_blocks):
+        if isinstance(bb, unet.ResBlock):
+          zero_module(bb.out_layers[3])
+        elif isinstance(bb, unet.SpatialTransformer):
+          zero_module(bb.proj_out)
+      zero_module(self.model.diffusion_model.out[2])

  def get_x_prev_and_pred_x0(self, x, e_t, a_t, a_prev):
    temperature = 1
@@ -233,12 +269,14 @@ if __name__ == "__main__":

  # load in weights
  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], strict=False)
+    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], verbose=False, strict=False, realize=False)

    if args.fp16:
      for k,v in get_state_dict(model).items():
        if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16).realize())
+          v.replace(v.cast(dtypes.float16))
+
+    Tensor.realize(*get_state_dict(model).values())

  # run through CLIP to get context
  tokenizer = Tokenizer.ClipTokenizer()
@@ -32,7 +32,7 @@ if __name__ == "__main__":

  lr = 5e-3
  transform = ComposeTransforms([
-    lambda x: [Image.fromarray(xx, mode='L').resize((64, 64)) for xx in x],
+    lambda x: [Image.fromarray(xx).resize((64, 64)) for xx in x],
    lambda x: np.stack([np.asarray(xx) for xx in x], 0),
    lambda x: x / 255.0,
    lambda x: np.tile(np.expand_dims(x, 1), (1, 3, 1, 1)).astype(np.float32),
@@ -109,7 +109,7 @@ class TextDecoder:

  def forward(self, x:Tensor, pos:Union[Variable, Literal[0]], encoded_audio:Tensor):
    seqlen = x.shape[-1]
-    x = self.token_embedding(x) + self.positional_embedding.shrink(((pos, pos+seqlen), None, None))
+    x = self.token_embedding(x) + self.positional_embedding.shrink(((pos, pos+seqlen), None))
    for block in self.blocks: x = block(x, xa=encoded_audio, mask=self.mask, len=pos)
    return self.output_tok(x)

@@ -2,7 +2,7 @@
 import os
 from ultralytics import YOLO
 from pathlib import Path
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.nn.onnx import OnnxRunner
 from extra.onnx_helpers import get_example_inputs

 os.chdir("/tmp")
@@ -49,8 +49,7 @@ def rangeify_kernel3():
  b = Tensor.empty(N,N)
  c = a@b
  #c = c.reshape((32,2,16,4,32,2,16,4)).contiguous()
-  with Context(RANGEIFY=1):
-    sink = c.schedule()[-1].ast
+  sink = c.schedule()[-1].ast
  #print(sink)

  opts  = [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UPCAST, 0, 2)]
@@ -329,7 +328,7 @@ if __name__ == "__main__":
  elif HL == 1: hprg = hl_spec_kernel3()
  else: hprg = hand_spec_kernel3()
  if HL == 3:
-    with Context(RANGEIFY=1, BLOCK_REORDER=0):
+    with Context(BLOCK_REORDER=0):
      prg = get_program(hprg, Device.default.renderer)
  else:
    prg = get_program(hprg, Device.default.renderer)
@@ -7,7 +7,6 @@ bert_train_params = {
  "GPUS": 6,
  "BS": 96,
  "EVAL_BS": 96,
-  "FUSE_ARANGE": 1,
  "BASEDIR": "/raid/datasets/wiki",
 }

@@ -50,7 +50,7 @@ def ioctls_from_header():
  hdr = (pathlib.Path(__file__).parent / "kfd_ioctl.h").read_text().replace("\\\n", "")
  pattern = r'#define\s+(AMDKFD_IOC_[A-Z0-9_]+)\s+AMDKFD_IOW?R?\((0x[0-9a-fA-F]+),\s+struct\s([A-Za-z0-9_]+)\)'
  matches = re.findall(pattern, hdr, re.MULTILINE)
-  return {int(nr, 0x10):(name, getattr(kfd_ioctl, "struct_"+sname)) for name, nr, sname in matches}
+  return {int(nr, 0x10):(name, getattr(kfd_ioctl, "struct_"+sname, None)) for name, nr, sname in matches}
 nrs = ioctls_from_header()

@ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_int, ctypes.c_ulong, ctypes.c_void_p)
@@ -1,7 +1,7 @@
 import onnx, yaml, tempfile, time, argparse, json
 from pathlib import Path
 from typing import Any
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.nn.onnx import OnnxRunner
 from extra.onnx_helpers import validate, get_example_inputs
 from extra.huggingface_onnx.huggingface_manager import DOWNLOADS_DIR, snapshot_download_with_retry

@@ -9,6 +9,9 @@ from PIL import Image
 import numpy as np
 import re, gzip

+# Allow for monkeypatching for mlperf.
+gelu = Tensor.gelu
+
@lru_cache()
 def default_bpe():
  # Clip tokenizer, taken from https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py (MIT license)
@@ -53,8 +56,8 @@ class Tokenizer:
    cs = [chr(n) for n in cs]
    return dict(zip(bs, cs))
  class ClipTokenizer:
-    def __init__(self):
-      self.byte_encoder = Tokenizer.bytes_to_unicode()
+    def __init__(self, version=None):
+      self.byte_encoder, self.version = Tokenizer.bytes_to_unicode(), version
      merges = gzip.open(default_bpe()).read().decode("utf-8").split('\n')
      merges = merges[1:49152-256-2+1]
      merges = [tuple(merge.split()) for merge in merges]
@@ -62,11 +65,17 @@ class Tokenizer:
      vocab = vocab + [v+'</w>' for v in vocab]
      for merge in merges:
        vocab.append(''.join(merge))
-      vocab.extend(['<|startoftext|>', '<|endoftext|>'])
+      if self.version == "sd_mlperf_v5_0":
+        import regex
+        vocab.extend(['<start_of_text>', '<end_of_text>'])
+        self.cache = {'<start_of_text>': '<start_of_text>', '<end_of_text>': '<end_of_text>'}
+        self.pat = regex.compile(r"""<start_of_text>|<end_of_text>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", regex.IGNORECASE)
+      else:
+        vocab.extend(['<|startoftext|>', '<|endoftext|>'])
+        self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
+        self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[^\s]+""", re.IGNORECASE)
      self.encoder = dict(zip(vocab, range(len(vocab))))
      self.bpe_ranks = dict(zip(merges, range(len(merges))))
-      self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
-      self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[^\s]+""", re.IGNORECASE)

    def bpe(self, token):
      if token in self.cache:
@@ -110,8 +119,17 @@ class Tokenizer:

    def encode(self, text:str, pad_with_zeros:bool=False) -> List[int]:
      bpe_tokens: List[int] = []
-      text = Tokenizer.whitespace_clean(text.strip()).lower()
-      for token in re.findall(self.pat, text):
+      if self.version == "sd_mlperf_v5_0":
+        import regex, ftfy, html
+        text = ftfy.fix_text(text)
+        text = html.unescape(html.unescape(text)).strip()
+        text = Tokenizer.whitespace_clean(text).lower()
+        re_module = regex
+      else:
+        text = Tokenizer.whitespace_clean(text.strip()).lower()
+        re_module = re
+
+      for token in re_module.findall(self.pat, text):
        token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
        bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
      # Truncation, keeping two slots for start and end tokens.
@@ -252,10 +270,8 @@ class Open:
      q,k,v = [y.reshape(T, B*self.n_heads, self.d_head).transpose(0, 1).reshape(B, self.n_heads, T, self.d_head) for y in proj.chunk(3)]

      attn_output = Tensor.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
-      attn_output = attn_output.permute(2, 0, 1, 3).reshape(T*B, C)
-
+      attn_output = attn_output.permute(2, 0, 1, 3).reshape(T, B, C)
      attn_output = self.out_proj(attn_output)
-      attn_output = attn_output.reshape(T, B, C)

      return attn_output

@@ -263,9 +279,10 @@ class Open:
    def __init__(self, dims, hidden_dims):
      self.c_fc   = Linear(dims, hidden_dims)
      self.c_proj = Linear(hidden_dims, dims)
+      self.gelu = gelu

    def __call__(self, x:Tensor) -> Tensor:
-      return x.sequential([self.c_fc, Tensor.gelu, self.c_proj])
+      return x.sequential([self.c_fc, self.gelu, self.c_proj])

  # https://github.com/mlfoundations/open_clip/blob/58e4e39aaabc6040839b0d2a7e8bf20979e4558a/src/open_clip/transformer.py#L210
  class ResidualAttentionBlock:
@@ -350,15 +367,15 @@ class Open:
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/encoders/modules.py#L396
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/encoders/modules.py#L498
 class FrozenOpenClipEmbedder(Embedder):
-  def __init__(self, dims:int, n_heads:int, layers:int, return_pooled:bool, ln_penultimate:bool=False):
-    self.tokenizer = Tokenizer.ClipTokenizer()
+  def __init__(self, dims:int, n_heads:int, layers:int, return_pooled:bool, ln_penultimate:bool=False, clip_tokenizer_version=None):
+    self.tokenizer = Tokenizer.ClipTokenizer(version=clip_tokenizer_version)
    self.model = Open.ClipTextTransformer(dims, n_heads, layers)
    self.return_pooled = return_pooled
    self.input_key = "txt"
    self.ln_penultimate = ln_penultimate

  def tokenize(self, text:str, device:Optional[str]=None) -> Tensor:
-    return Tensor(self.tokenizer.encode(text, pad_with_zeros=True), dtype=dtypes.int64, device=device).reshape(1,-1)
+    return Tensor(self.tokenizer.encode(text, pad_with_zeros=True), dtype=dtypes.int32, device=device).reshape(1,-1)

  def text_transformer_forward(self, x:Tensor, attn_mask:Optional[Tensor]=None):
    for r in self.model.transformer.resblocks:
@@ -449,7 +466,7 @@ class OpenClipEncoder:
    x = x + self.positional_embedding
    x = self.transformer(x, attn_mask=self.attn_mask)
    x = self.ln_final(x)
-    x = x[:, tokens.argmax(axis=-1)]
+    x = x[Tensor.arange(x.shape[0], device=x.device), tokens.argmax(axis=-1)]
    x = x @ self.text_projection
    return x

@@ -270,8 +270,10 @@ class FidInceptionV3:
    self.Mixed_7b = inception.Mixed_7b
    self.Mixed_7c = inception.Mixed_7c

-  def load_from_pretrained(self):
-    state_dict = torch_load(str(fetch("https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth", "pt_inception-2015-12-05-6726825d.pth")))
+  def load_from_pretrained(self, path=None):
+    if path is None:
+      path = fetch("https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth", "pt_inception-2015-12-05-6726825d.pth")
+    state_dict = torch_load(str(path))
    for k,v in state_dict.items():
      if k.endswith(".num_batches_tracked"):
        state_dict[k] = v.reshape(1)
@@ -1,21 +1,24 @@
-from tinygrad import Tensor, dtypes
-from tinygrad.nn import Linear, Conv2d, GroupNorm, LayerNorm
+from tinygrad import Tensor, dtypes, nn
 from tinygrad.device import is_dtype_supported
-from typing import Optional, Union, List, Any, Tuple
+from typing import Optional, Union, List, Any, Tuple, Callable
 import math

+# allow for monkeypatching
+Linear, Conv2d, GroupNorm, LayerNorm = nn.Linear, nn.Conv2d, nn.GroupNorm, nn.LayerNorm
+attention, gelu, mixed_precision_dtype = Tensor.scaled_dot_product_attention, Tensor.gelu, dtypes.float16
+
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/diffusionmodules/util.py#L207
 def timestep_embedding(timesteps:Tensor, dim:int, max_period=10000):
  half = dim // 2
  freqs = (-math.log(max_period) * Tensor.arange(half, device=timesteps.device) / half).exp()
  args = timesteps.unsqueeze(1) * freqs.unsqueeze(0)
  out = Tensor.cat(args.cos(), args.sin(), dim=-1)
-  return out.cast(dtypes.float16) if is_dtype_supported(dtypes.float16) else out
+  return out.cast(mixed_precision_dtype) if is_dtype_supported(mixed_precision_dtype) else out

 class ResBlock:
-  def __init__(self, channels:int, emb_channels:int, out_channels:int):
+  def __init__(self, channels:int, emb_channels:int, out_channels:int, num_groups:int=32):
    self.in_layers = [
-      GroupNorm(32, channels),
+      GroupNorm(num_groups, channels),
      Tensor.silu,
      Conv2d(channels, out_channels, 3, padding=1),
    ]
@@ -24,7 +27,7 @@ class ResBlock:
      Linear(emb_channels, out_channels),
    ]
    self.out_layers = [
-      GroupNorm(32, out_channels),
+      GroupNorm(num_groups, out_channels),
      Tensor.silu,
      lambda x: x,  # needed for weights loading code to work
      Conv2d(out_channels, out_channels, 3, padding=1),
@@ -45,35 +48,37 @@ class CrossAttention:
    self.to_v = Linear(ctx_dim,   n_heads*d_head, bias=False)
    self.num_heads = n_heads
    self.head_size = d_head
+    self.attn = attention
    self.to_out = [Linear(n_heads*d_head, query_dim)]

  def __call__(self, x:Tensor, ctx:Optional[Tensor]=None) -> Tensor:
    ctx = x if ctx is None else ctx
    q,k,v = self.to_q(x), self.to_k(ctx), self.to_v(ctx)
    q,k,v = [y.reshape(x.shape[0], -1, self.num_heads, self.head_size).transpose(1,2) for y in (q,k,v)]
-    attention = Tensor.scaled_dot_product_attention(q, k, v).transpose(1,2)
+    attention = self.attn(q, k, v).transpose(1,2)
    h_ = attention.reshape(x.shape[0], -1, self.num_heads * self.head_size)
    return h_.sequential(self.to_out)

 class GEGLU:
  def __init__(self, dim_in:int, dim_out:int):
    self.proj = Linear(dim_in, dim_out * 2)
+    self.gelu = gelu
    self.dim_out = dim_out

  def __call__(self, x:Tensor) -> Tensor:
    x, gate = self.proj(x).chunk(2, dim=-1)
-    return x * gate.gelu()
+    return x * self.gelu(gate)

 class FeedForward:
  def __init__(self, dim:int, mult:int=4):
-    self.net = [
+    self.net: tuple[GEGLU, Callable, nn.Linear] = (
      GEGLU(dim, dim*mult),
      lambda x: x,  # needed for weights loading code to work
      Linear(dim*mult, dim)
-    ]
+    )

  def __call__(self, x:Tensor) -> Tensor:
-    return x.sequential(self.net)
+    return x.sequential(list(self.net))

 class BasicTransformerBlock:
  def __init__(self, dim:int, ctx_dim:int, n_heads:int, d_head:int):
@@ -92,12 +97,13 @@ class BasicTransformerBlock:

 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/attention.py#L619
 class SpatialTransformer:
-  def __init__(self, channels:int, n_heads:int, d_head:int, ctx_dim:Union[int,List[int]], use_linear:bool, depth:int=1):
+  def __init__(self, channels:int, n_heads:int, d_head:int, ctx_dim:Union[int,List[int]], use_linear:bool, depth:int=1,
+               norm_eps:float=1e-5):
    if isinstance(ctx_dim, int):
      ctx_dim = [ctx_dim]*depth
    else:
      assert isinstance(ctx_dim, list) and depth == len(ctx_dim)
-    self.norm = GroupNorm(32, channels)
+    self.norm = GroupNorm(32, channels, eps=norm_eps)
    assert channels == n_heads * d_head
    self.proj_in  = Linear(channels, channels) if use_linear else Conv2d(channels, channels, 1)
    self.transformer_blocks = [BasicTransformerBlock(channels, ctx_dim[d], n_heads, d_head) for d in range(depth)]
@@ -134,7 +140,9 @@ class Upsample:

 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/diffusionmodules/openaimodel.py#L472
 class UNetModel:
-  def __init__(self, adm_in_ch:Optional[int], in_ch:int, out_ch:int, model_ch:int, attention_resolutions:List[int], num_res_blocks:int, channel_mult:List[int], transformer_depth:List[int], ctx_dim:Union[int,List[int]], use_linear:bool=False, d_head:Optional[int]=None, n_heads:Optional[int]=None):
+  def __init__(self, adm_in_ch:Optional[int], in_ch:int, out_ch:int, model_ch:int, attention_resolutions:List[int], num_res_blocks:int,
+               channel_mult:List[int], transformer_depth:List[int], ctx_dim:Union[int,List[int]], use_linear:bool=False, d_head:Optional[int]=None,
+               n_heads:Optional[int]=None, num_groups:int=32, st_norm_eps:float=1e-5):
    self.model_ch = model_ch
    self.num_res_blocks = [num_res_blocks] * len(channel_mult)

@@ -174,12 +182,12 @@ class UNetModel:
    for idx, mult in enumerate(channel_mult):
      for _ in range(self.num_res_blocks[idx]):
        layers: List[Any] = [
-          ResBlock(ch, time_embed_dim, model_ch*mult),
+          ResBlock(ch, time_embed_dim, model_ch*mult, num_groups),
        ]
        ch = mult * model_ch
        if ds in attention_resolutions:
          d_head, n_heads = get_d_and_n_heads(ch)
-          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx]))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))

        self.input_blocks.append(layers)
        input_block_channels.append(ch)
@@ -193,9 +201,9 @@ class UNetModel:

    d_head, n_heads = get_d_and_n_heads(ch)
    self.middle_block: List = [
-      ResBlock(ch, time_embed_dim, ch),
-      SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[-1]),
-      ResBlock(ch, time_embed_dim, ch),
+      ResBlock(ch, time_embed_dim, ch, num_groups),
+      SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[-1], norm_eps=st_norm_eps),
+      ResBlock(ch, time_embed_dim, ch, num_groups),
    ]

    self.output_blocks = []
@@ -203,13 +211,13 @@ class UNetModel:
      for i in range(self.num_res_blocks[idx] + 1):
        ich = input_block_channels.pop()
        layers = [
-          ResBlock(ch + ich, time_embed_dim, model_ch*mult),
+          ResBlock(ch + ich, time_embed_dim, model_ch*mult, num_groups),
        ]
        ch = model_ch * mult

        if ds in attention_resolutions:
          d_head, n_heads = get_d_and_n_heads(ch)
-          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx]))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))

        if idx > 0 and i == self.num_res_blocks[idx]:
          layers.append(Upsample(ch))
@@ -217,7 +225,7 @@ class UNetModel:
        self.output_blocks.append(layers)

    self.out = [
-      GroupNorm(32, ch),
+      GroupNorm(num_groups, ch),
      Tensor.silu,
      Conv2d(model_ch, out_ch, 3, padding=1),
    ]
@@ -230,10 +238,10 @@ class UNetModel:
      assert y.shape[0] == x.shape[0]
      emb = emb + y.sequential(self.label_emb[0])

-    if is_dtype_supported(dtypes.float16):
-      emb = emb.cast(dtypes.float16)
-      ctx = ctx.cast(dtypes.float16)
-      x   = x  .cast(dtypes.float16)
+    if is_dtype_supported(mixed_precision_dtype):
+      emb = emb.cast(mixed_precision_dtype)
+      ctx = ctx.cast(mixed_precision_dtype)
+      x   = x  .cast(mixed_precision_dtype)

    def run(x:Tensor, bb) -> Tensor:
      if isinstance(bb, ResBlock): x = bb(x, emb)
@@ -1,6 +1,6 @@
 from tinygrad import Tensor
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.frontend.onnx import OnnxRunner, OnnxValue
+from tinygrad.nn.onnx import OnnxRunner, OnnxValue
 import numpy as np
 import onnxruntime as ort

@@ -81,7 +81,7 @@ def lin_to_feats(lin:Kernel, use_sts=True):
  ret = [float(x) for x in ret]

  if use_sts:
-    my_sts = dedup([(x.shape == lin.full_shape, x.real_strides(), any(v.mask is not None for v in x.views), len(x.views)) for x in lin.sts])
+    my_sts = dedup([(x.shape == lin.full_shape, x.is_expanded(), any(v.mask is not None for v in x.views), len(x.views)) for x in lin.sts])
    assert len(my_sts) < MAX_BUFS
    sts_len = 3 + 5*MAX_DIMS
    for s in my_sts:
@@ -50,7 +50,7 @@ class TestBeamSearch(unittest.TestCase):
  def test_variable_shrink_prime_number(self):
    v = Variable("v", 1, 400).bind(367)
    a = rand(400, 367)
-    b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
+    b = (a.shrink(((0,v), None))+1)[:367,:367].realize()
    np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)

  def test_no_mutate_rawbuffers(self):
@@ -930,7 +930,7 @@ impl<'a> Thread<'a> {

            let op = ((instr >> 16) & 0x3ff) as u32;
            match op {
-                764 | 765 | 288 | 289 | 290 | 766 | 768 | 769 => {
+                764 | 765 | 288 | 289 | 290 | 766 | 767 | 768 | 769 => {
                    let vdst = (instr & 0xff) as usize;
                    let sdst = ((instr >> 8) & 0x7f) as usize;
                    let f = |i: u32| -> usize { ((instr >> i) & 0x1ff) as usize };
@@ -944,6 +944,16 @@ impl<'a> Thread<'a> {
                    assert_eq!(clmp, 0);

                    let vcc = match op {
+                        767 => {
+                            let (s0, s1, s2): (u32, u32, u64) = (self.val(s0), self.val(s1), self.val(s2));
+                            let (mul_result, overflow_mul) = (s0 as i64).overflowing_mul(s1 as i64);
+                            let (ret, overflow_add) = mul_result.overflowing_add(s2 as i64);
+                            let overflowed = overflow_mul || overflow_add;
+                            if self.exec.read() {
+                                self.vec_reg.write64(vdst, ret as u64);
+                            }
+                            overflowed
+                        },
                        766 => {
                            let (s0, s1, s2): (u32, u32, u64) = (self.val(s0), self.val(s1), self.val(s2));
                            let (mul_result, overflow_mul) = (s0 as u64).overflowing_mul(s1 as u64);
@@ -4,7 +4,7 @@

 Only supported on 7900XTX, requires either AM (`rmmod amdgpu`) or disabling power gating on AMD (`ppfeaturemask=0xffff3fff`, don't forget to rebuild initramfs)

-SQTT is implemented on top of normal tinygrad PROFILE=1, `PROFILE=1 SQTT=1` to get profile pickle with sqtt data embedded in it.
+SQTT is implemented on top of normal tinygrad profiling, `VIZ=1 SQTT=1` to get profile pickle with sqtt data embedded in it.

 `SQTT_BUFFER_SIZE=X` to change size of SQTT buffer (per shader engine, 6 SEs on 7900xtx) in megabytes, default 256.

@@ -0,0 +1,68 @@
+import ctypes
+from dataclasses import dataclass
+import tinygrad.runtime.autogen.comgr as comgr
+from tinygrad.runtime.support.compiler_amd import check
+
+@dataclass
+class InstrCtx:
+  pc:int=0
+  inst:str=""
+
+@comgr.amd_comgr_create_disassembly_info.argtypes[2]
+def instr_cb(text, user_data):
+  c = ctypes.cast(user_data, ctypes.POINTER(ctypes.py_object)).contents.value
+  c.inst = ctypes.string_at(text).decode("utf-8","replace").strip()
+  return comgr.AMD_COMGR_STATUS_SUCCESS
+
+# nop callback
+@comgr.amd_comgr_create_disassembly_info.argtypes[3]
+def addr_cb(*args): return comgr.AMD_COMGR_STATUS_SUCCESS
+
+def comgr_get_address_table(lib:bytes) -> dict[int, tuple[str, int]]:
+  check(comgr.amd_comgr_create_data(comgr.AMD_COMGR_DATA_KIND_EXECUTABLE, ctypes.byref(data_src:=comgr.amd_comgr_data_t())))
+  lib_buf = ctypes.create_string_buffer(lib, len(lib))
+  check(comgr.amd_comgr_set_data(data_src, len(lib), lib_buf))
+  check(comgr.amd_comgr_get_data_isa_name(data_src, isa_sz:=ctypes.c_size_t(128), isa:=(ctypes.c_char*isa_sz.value)()))
+
+  @comgr.amd_comgr_create_disassembly_info.argtypes[1]
+  def memory_cb(from_addr, to, size, _):
+    base, buf_len = ctypes.addressof(lib_buf), len(lib_buf)
+    start = int(from_addr) - base
+    if start < 0 or start >= buf_len: return 0
+    ctypes.memmove(to, base + start, n:=min(int(size), buf_len - start))
+    return n
+
+  info_src = comgr.amd_comgr_disassembly_info_t()
+  check(comgr.amd_comgr_create_disassembly_info(ctypes.cast(isa, ctypes.POINTER(ctypes.c_char)), memory_cb, instr_cb, addr_cb, info_src))
+
+  @comgr.amd_comgr_iterate_symbols.argtypes[1]
+  def sym_callback(sym, udata):
+    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_TYPE, ctypes.byref(sym_type:=ctypes.c_int())))
+    if sym_type.value != comgr.AMD_COMGR_SYMBOL_TYPE_FUNC: return comgr.AMD_COMGR_STATUS_SUCCESS
+    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_VALUE, ctypes.byref(vaddr:=ctypes.c_uint64())))
+    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_SIZE, ctypes.byref(size:=ctypes.c_uint64())))
+    check(comgr.amd_comgr_map_elf_virtual_address_to_code_object_offset(data_src, vaddr.value, ctypes.byref(offset:=ctypes.c_uint64()),
+                                                                        ctypes.byref(ctypes.c_uint64()), ctypes.byref(nobits:=ctypes.c_bool())))
+    check(nobits.value)
+    base = ctypes.addressof(lib_buf)
+    pc = base + offset.value
+    end = pc + size.value
+    addr_table = ctypes.cast(udata, ctypes.POINTER(ctypes.py_object)).contents.value
+    instr_ref = ctypes.py_object(ctx:=InstrCtx())
+    instr_ptr = ctypes.cast(ctypes.pointer(instr_ref), ctypes.c_void_p)
+    while pc < end:
+      size_read = ctypes.c_uint64(0)
+      ctx.pc = pc
+      st = comgr.amd_comgr_disassemble_instruction(info_src, ctypes.c_uint64(pc), instr_ptr, ctypes.byref(size_read))
+      if st == comgr.AMD_COMGR_STATUS_SUCCESS and size_read.value:
+        rel = (pc - base) - offset.value
+        addr_table[vaddr.value + rel] = (ctx.inst, int(size_read.value))
+        pc += size_read.value
+      else: # don't inf loop if comgr fails
+        b = ctypes.c_ubyte.from_buffer(lib_buf, pc - base).value
+        addr_table[vaddr.value + (pc - base - offset.value)] = (f"DISASSEMBLER ISSUE 0x{b:02x}", 1)
+        pc += 1
+    return comgr.AMD_COMGR_STATUS_SUCCESS
+  addr_table:dict[int, tuple[str, int]] = {}
+  check(comgr.amd_comgr_iterate_symbols(data_src, sym_callback, ctypes.cast(ctypes.pointer(ctypes.py_object(addr_table)), ctypes.c_void_p)))
+  return addr_table
@@ -155,6 +155,10 @@ class RGP:
      device_event = device_events[device]
    sqtt_events = [x for x in profile if isinstance(x, ProfileSQTTEvent) and x.device == device_event.device]
    if len(sqtt_events) == 0: raise RuntimeError(f"Device {device_event.device} doesn't contain SQTT data")
+    device_props = sqtt_events[0].props
+    gfx_ver = device_props['gfx_target_version'] // 10000
+    gfx_iplvl = getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}_{(device_props['gfx_target_version']//100)%100}",
+                        getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}", None))
    sqtt_itrace_enabled = any([event.itrace for event in sqtt_events])
    sqtt_itrace_masked = not all_same([event.itrace for event in sqtt_events])
    sqtt_itrace_se_mask = functools.reduce(lambda a,b: a|b, [int(event.itrace) << event.se for event in sqtt_events], 0) if sqtt_itrace_masked else 0
@@ -192,21 +196,21 @@ class RGP:
        flags=0,
        trace_shader_core_clock=0x93f05080,
        trace_memory_clock=0x4a723a40,
-        device_id=0x744c,
+        device_id={110000: 0x744c, 110003: 0x7480, 120001: 0x7550}[device_props['gfx_target_version']],
        device_revision_id=0xc8,
        vgprs_per_simd=1536,
        sgprs_per_simd=128*16,
-        shader_engines=6,
-        compute_unit_per_shader_engine=16,
-        simd_per_compute_unit=2,
-        wavefronts_per_simd=16,
+        shader_engines=device_props['array_count'] // device_props['simd_arrays_per_engine'],
+        compute_unit_per_shader_engine=device_props['simd_count'] // device_props['simd_per_cu'] // (device_props['array_count'] // device_props['simd_arrays_per_engine']),
+        simd_per_compute_unit=device_props['simd_per_cu'],
+        wavefronts_per_simd=device_props['max_waves_per_simd'],
        minimum_vgpr_alloc=4,
        vgpr_alloc_granularity=8,
        minimum_sgpr_alloc=128,
        sgpr_alloc_granularity=128,
        hardware_contexts=8,
        gpu_type=sqtt.SQTT_GPU_TYPE_DISCRETE,
-        gfxip_level=sqtt.SQTT_GFXIP_LEVEL_GFXIP_11_0,
+        gfxip_level=gfx_iplvl,
        gpu_index=0,
        gds_size=0,
        gds_per_shader_engine=0,
@@ -218,7 +222,7 @@ class RGP:
        vram_bus_width=384, # 384-bit
        l2_cache_size=6 * 1024 * 1024, # 6 MB
        l1_cache_size=32 * 1024, # 32 KB per SIMD (?)
-        lds_size=65536, # 64 KB per CU
+        lds_size=device_props['lds_size_in_kb'] * 1024,
        gpu_name=b'NAVI31',
        alu_per_clock=0,
        texture_per_clock=0,
@@ -257,7 +261,7 @@ class RGP:
            major_version=0, minor_version=2,
          ),
          shader_engine_index=sqtt_event.se,
-          sqtt_version=sqtt.SQTT_VERSION_3_2,
+          sqtt_version={11: sqtt.SQTT_VERSION_3_2, 12: sqtt.SQTT_VERSION_3_3}.get(gfx_ver),
          _0=sqtt.union_sqtt_file_chunk_sqtt_desc_0(
            v1=sqtt.struct_sqtt_file_chunk_sqtt_desc_0_v1(
              instrumentation_spec_version=1,
@@ -0,0 +1,96 @@
+import ctypes, pathlib, argparse, pickle, re, functools, dataclasses
+from extra.sqtt.rocprof import rocprof
+from extra.sqtt.disasm import comgr_get_address_table
+from tinygrad.helpers import temp, DEBUG
+from tinygrad.device import ProfileEvent, ProfileProgramEvent
+from tinygrad.runtime.ops_amd import ProfileSQTTEvent
+
+@dataclasses.dataclass
+class InstInfo:
+  typ:str=""
+  inst:str=""
+  hit:int=0
+  lat:int=0
+  stall:int=0
+  def __str__(self): return f"{self.inst:>20} hits:{self.typ:>6} hits:{self.hit:>6} latency:{self.lat:>6} stall:{self.stall:>6}"
+
+  def on_ev(self, ev):
+    self.hit, self.lat, self.stall = self.hit + 1, self.lat + ev.duration, self.stall + ev.stall
+
+class _ROCParseCtx:
+  def __init__(self, sqtt_evs:list[ProfileSQTTEvent], prog_evs:list[ProfileProgramEvent]):
+    self.sqtt_evs, self.prog_evs = iter(sqtt_evs), prog_evs
+    self.wave_events, self.disasms, self.addr2prg = {}, {}, {}
+
+    for prog in prog_evs:
+      for addr, info in comgr_get_address_table(prog.lib).items():
+        self.disasms[prog.base + addr] = info
+        self.addr2prg[prog.base + addr] = prog
+
+  def next_sqtt(self): return next(self.sqtt_evs, None)
+  def find_program(self, addr): return self.addr2prg[addr]
+
+  def on_occupancy_ev(self, ev):
+    if DEBUG >= 4: print("OCC", ev.time, ev.cu, ev.simd, ev.wave_id, ev.start)
+
+  def on_wave_ev(self, ev):
+    if DEBUG >= 4: print("WAVE", ev.wave_id, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
+
+    asm = {}
+    for j in range(ev.instructions_size):
+      inst_ev = ev.instructions_array[j]
+      inst_typ = rocprof.rocprofiler_thread_trace_decoder_inst_category_t__enumvalues[inst_ev.category]
+      asm.setdefault(inst_ev.pc.address, InstInfo(typ=inst_typ, inst=self.disasms[inst_ev.pc.address][0]))
+      asm[inst_ev.pc.address].on_ev(inst_ev)
+
+    self.wave_events[(self.find_program(ev.instructions_array[0].pc.address).name, ev.wave_id, ev.cu, ev.simd)] = asm
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser()
+  parser.add_argument('--profile', type=pathlib.Path, help='Path to profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
+  args = parser.parse_args()
+
+  with args.profile.open("rb") as f: profile = pickle.load(f)
+  sqtt_events:list[ProfileSQTTEvent] = []
+  prog_events:list[ProfileProgramEvent] = []
+  for e in profile:
+    if isinstance(e, ProfileSQTTEvent): sqtt_events.append(e)
+    if isinstance(e, ProfileProgramEvent) and e.device.startswith("AMD"): prog_events.append(e)
+
+  ROCParseCtx = _ROCParseCtx(sqtt_events, prog_events)
+
+  @rocprof.rocprof_trace_decoder_se_data_callback_t
+  def copy_cb(buf, buf_size, data_ptr):
+    if (prof:=ROCParseCtx.next_sqtt()) is None: return 0
+    buf[0] = ctypes.cast((ctypes.c_ubyte * len(prof.blob)).from_buffer_copy(prof.blob), ctypes.POINTER(ctypes.c_ubyte))
+    buf_size[0] = len(prof.blob)
+    return len(prof.blob)
+
+  @rocprof.rocprof_trace_decoder_trace_callback_t
+  def trace_cb(record_type, events_ptr, n, data_ptr):
+    match record_type:
+      case rocprof.ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY:
+        for ev in (rocprof.rocprofiler_thread_trace_decoder_occupancy_t * n).from_address(events_ptr): ROCParseCtx.on_occupancy_ev(ev)
+      case rocprof.ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE:
+        for ev in (rocprof.rocprofiler_thread_trace_decoder_wave_t * n).from_address(events_ptr): ROCParseCtx.on_wave_ev(ev)
+      case _:
+        if DEBUG >= 2: print(rocprof.rocprofiler_thread_trace_decoder_record_type_t__enumvalues[record_type], events_ptr, n)
+    return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS
+
+  @rocprof.rocprof_trace_decoder_isa_callback_t
+  def isa_cb(instr_ptr, mem_size_ptr, size_ptr, pc, data_ptr):
+    instr, mem_size_ptr[0] = ROCParseCtx.disasms[pc.address]
+
+    # this is the number of bytes to next instruction, set to 0 for end_pgm
+    if instr == "s_endpgm": mem_size_ptr[0] = 0
+    if (max_sz:=size_ptr[0]) == 0: return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES
+
+    # truncate the instr if it doesn't fit
+    if (str_sz:=len(instr_bytes:=instr.encode()))+1 > max_sz: str_sz = max_sz
+    ctypes.memmove(instr_ptr, instr_bytes, str_sz)
+    size_ptr[0] = str_sz
+
+    return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS
+
+  rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
+  print(ROCParseCtx.wave_events.keys())
@@ -0,0 +1,18 @@
+#!/usr/bin/env python3
+import os, shutil
+from pathlib import Path
+from tinygrad.helpers import fetch, OSX
+
+DEST = Path("/usr/local/lib")
+DEST.mkdir(exist_ok=True)
+
+if __name__ == "__main__":
+  if OSX:
+    fp = fetch("https://github.com/ROCm/rocprof-trace-decoder/releases/download/0.1.4/rocprof-trace-decoder-macos-arm64-0.1.4-Darwin.sh")
+    lib = fp.parent/"rocprof-trace-decoder-macos-arm64-0.1.4-Darwin"/"lib"/"librocprof-trace-decoder.dylib"
+    os.chmod(fp, 0o755)
+    os.system(f"sudo {fp} --prefix={fp.parent} --include-subdir")
+  else:
+    lib = fetch("https://github.com/ROCm/rocprof-trace-decoder/raw/5420409ad0963b2d76450add067b9058493ccbd0/releases/linux_glibc_2_28_x86_64/librocprof-trace-decoder.so", name="librocprof-trace-decoder.so")
+  shutil.copy2(lib, DEST)
+  print(f"Installed {lib.name} to", DEST)
@@ -0,0 +1,656 @@
+# pylint: skip-file
+# mypy: ignore-errors
+# -*- coding: utf-8 -*-
+#
+# TARGET arch is: []
+# WORD_SIZE is: 8
+# POINTER_SIZE is: 8
+# LONGDOUBLE_SIZE is: 16
+#
+import ctypes, ctypes.util
+
+
+class AsDictMixin:
+    @classmethod
+    def as_dict(cls, self):
+        result = {}
+        if not isinstance(self, AsDictMixin):
+            # not a structure, assume it's already a python object
+            return self
+        if not hasattr(cls, "_fields_"):
+            return result
+        # sys.version_info >= (3, 5)
+        # for (field, *_) in cls._fields_:  # noqa
+        for field_tuple in cls._fields_:  # noqa
+            field = field_tuple[0]
+            if field.startswith('PADDING_'):
+                continue
+            value = getattr(self, field)
+            type_ = type(value)
+            if hasattr(value, "_length_") and hasattr(value, "_type_"):
+                # array
+                if not hasattr(type_, "as_dict"):
+                    value = [v for v in value]
+                else:
+                    type_ = type_._type_
+                    value = [type_.as_dict(v) for v in value]
+            elif hasattr(value, "contents") and hasattr(value, "_type_"):
+                # pointer
+                try:
+                    if not hasattr(type_, "as_dict"):
+                        value = value.contents
+                    else:
+                        type_ = type_._type_
+                        value = type_.as_dict(value.contents)
+                except ValueError:
+                    # nullptr
+                    value = None
+            elif isinstance(value, AsDictMixin):
+                # other structure
+                value = type_.as_dict(value)
+            result[field] = value
+        return result
+
+
+class Structure(ctypes.Structure, AsDictMixin):
+
+    def __init__(self, *args, **kwds):
+        # We don't want to use positional arguments fill PADDING_* fields
+
+        args = dict(zip(self.__class__._field_names_(), args))
+        args.update(kwds)
+        super(Structure, self).__init__(**args)
+
+    @classmethod
+    def _field_names_(cls):
+        if hasattr(cls, '_fields_'):
+            return (f[0] for f in cls._fields_ if not f[0].startswith('PADDING'))
+        else:
+            return ()
+
+    @classmethod
+    def get_type(cls, field):
+        for f in cls._fields_:
+            if f[0] == field:
+                return f[1]
+        return None
+
+    @classmethod
+    def bind(cls, bound_fields):
+        fields = {}
+        for name, type_ in cls._fields_:
+            if hasattr(type_, "restype"):
+                if name in bound_fields:
+                    if bound_fields[name] is None:
+                        fields[name] = type_()
+                    else:
+                        # use a closure to capture the callback from the loop scope
+                        fields[name] = (
+                            type_((lambda callback: lambda *args: callback(*args))(
+                                bound_fields[name]))
+                        )
+                    del bound_fields[name]
+                else:
+                    # default callback implementation (does nothing)
+                    try:
+                        default_ = type_(0).restype().value
+                    except TypeError:
+                        default_ = None
+                    fields[name] = type_((
+                        lambda default_: lambda *args: default_)(default_))
+            else:
+                # not a callback function, use default initialization
+                if name in bound_fields:
+                    fields[name] = bound_fields[name]
+                    del bound_fields[name]
+                else:
+                    fields[name] = type_()
+        if len(bound_fields) != 0:
+            raise ValueError(
+                "Cannot bind the following unknown callback(s) {}.{}".format(
+                    cls.__name__, bound_fields.keys()
+            ))
+        return cls(**fields)
+
+
+class Union(ctypes.Union, AsDictMixin):
+    pass
+
+
+
+c_int128 = ctypes.c_ubyte*16
+c_uint128 = c_int128
+void = None
+if ctypes.sizeof(ctypes.c_longdouble) == 16:
+    c_long_double_t = ctypes.c_longdouble
+else:
+    c_long_double_t = ctypes.c_ubyte*16
+
+def string_cast(char_pointer, encoding='utf-8', errors='strict'):
+    value = ctypes.cast(char_pointer, ctypes.c_char_p).value
+    if value is not None and encoding is not None:
+        value = value.decode(encoding, errors=errors)
+    return value
+
+
+def char_pointer_cast(string, encoding='utf-8'):
+    if encoding is not None:
+        try:
+            string = string.encode(encoding)
+        except AttributeError:
+            # In Python3, bytes has no encode attribute
+            pass
+    string = ctypes.c_char_p(string)
+    return ctypes.cast(string, ctypes.POINTER(ctypes.c_char))
+
+
+
+class FunctionFactoryStub:
+    def __getattr__(self, _):
+      return ctypes.CFUNCTYPE(lambda y:y)
+
+# libraries['FIXME_STUB'] explanation
+# As you did not list (-l libraryname.so) a library that exports this function
+# This is a non-working stub instead.
+# You can either re-run clan2py with -l /path/to/library.so
+# Or manually fix this by comment the ctypes.CDLL loading
+_libraries = {}
+_libraries['FIXME_STUB'] = ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder')) #  ctypes.CDLL('FIXME_STUB')
+
+
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_info_t'
+rocprofiler_thread_trace_decoder_info_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE = 0
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST = 1
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE = 2
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE = 3
+ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST = 4
+rocprofiler_thread_trace_decoder_info_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_pc_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_pc_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_pc_t._fields_ = [
+    ('address', ctypes.c_uint64),
+    ('code_object_id', ctypes.c_uint64),
+]
+
+rocprofiler_thread_trace_decoder_pc_t = struct_rocprofiler_thread_trace_decoder_pc_t
+class struct_rocprofiler_thread_trace_decoder_perfevent_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_perfevent_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_perfevent_t._fields_ = [
+    ('time', ctypes.c_int64),
+    ('events0', ctypes.c_uint16),
+    ('events1', ctypes.c_uint16),
+    ('events2', ctypes.c_uint16),
+    ('events3', ctypes.c_uint16),
+    ('CU', ctypes.c_ubyte),
+    ('bank', ctypes.c_ubyte),
+    ('PADDING_0', ctypes.c_ubyte * 6),
+]
+
+rocprofiler_thread_trace_decoder_perfevent_t = struct_rocprofiler_thread_trace_decoder_perfevent_t
+class struct_rocprofiler_thread_trace_decoder_occupancy_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_occupancy_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_occupancy_t._fields_ = [
+    ('pc', rocprofiler_thread_trace_decoder_pc_t),
+    ('time', ctypes.c_uint64),
+    ('reserved', ctypes.c_ubyte),
+    ('cu', ctypes.c_ubyte),
+    ('simd', ctypes.c_ubyte),
+    ('wave_id', ctypes.c_ubyte),
+    ('start', ctypes.c_uint32, 1),
+    ('_rsvd', ctypes.c_uint32, 31),
+]
+
+rocprofiler_thread_trace_decoder_occupancy_t = struct_rocprofiler_thread_trace_decoder_occupancy_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_wstate_type_t'
+rocprofiler_thread_trace_decoder_wstate_type_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY = 0
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE = 1
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC = 2
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT = 3
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL = 4
+ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST = 5
+rocprofiler_thread_trace_decoder_wstate_type_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_wave_state_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_wave_state_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_wave_state_t._fields_ = [
+    ('type', ctypes.c_int32),
+    ('duration', ctypes.c_int32),
+]
+
+rocprofiler_thread_trace_decoder_wave_state_t = struct_rocprofiler_thread_trace_decoder_wave_state_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_inst_category_t'
+rocprofiler_thread_trace_decoder_inst_category_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS',
+    6: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU',
+    7: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP',
+    8: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT',
+    9: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED',
+    10: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT',
+    11: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE',
+    12: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH',
+    13: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE = 0
+ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM = 1
+ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU = 2
+ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM = 3
+ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT = 4
+ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS = 5
+ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU = 6
+ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP = 7
+ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT = 8
+ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED = 9
+ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT = 10
+ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE = 11
+ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH = 12
+ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST = 13
+rocprofiler_thread_trace_decoder_inst_category_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_inst_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_inst_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_inst_t._fields_ = [
+    ('category', ctypes.c_uint32, 8),
+    ('stall', ctypes.c_uint32, 24),
+    ('duration', ctypes.c_int32),
+    ('time', ctypes.c_int64),
+    ('pc', rocprofiler_thread_trace_decoder_pc_t),
+]
+
+rocprofiler_thread_trace_decoder_inst_t = struct_rocprofiler_thread_trace_decoder_inst_t
+class struct_rocprofiler_thread_trace_decoder_wave_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_wave_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_wave_t._fields_ = [
+    ('cu', ctypes.c_ubyte),
+    ('simd', ctypes.c_ubyte),
+    ('wave_id', ctypes.c_ubyte),
+    ('contexts', ctypes.c_ubyte),
+    ('_rsvd1', ctypes.c_uint32),
+    ('_rsvd2', ctypes.c_uint32),
+    ('_rsvd3', ctypes.c_uint32),
+    ('begin_time', ctypes.c_int64),
+    ('end_time', ctypes.c_int64),
+    ('timeline_size', ctypes.c_uint64),
+    ('instructions_size', ctypes.c_uint64),
+    ('timeline_array', ctypes.POINTER(struct_rocprofiler_thread_trace_decoder_wave_state_t)),
+    ('instructions_array', ctypes.POINTER(struct_rocprofiler_thread_trace_decoder_inst_t)),
+]
+
+rocprofiler_thread_trace_decoder_wave_t = struct_rocprofiler_thread_trace_decoder_wave_t
+class struct_rocprofiler_thread_trace_decoder_realtime_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_realtime_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_realtime_t._fields_ = [
+    ('shader_clock', ctypes.c_int64),
+    ('realtime_clock', ctypes.c_uint64),
+    ('reserved', ctypes.c_uint64),
+]
+
+rocprofiler_thread_trace_decoder_realtime_t = struct_rocprofiler_thread_trace_decoder_realtime_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_shaderdata_flags_t'
+rocprofiler_thread_trace_decoder_shaderdata_flags_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM = 0
+ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV = 1
+rocprofiler_thread_trace_decoder_shaderdata_flags_t = ctypes.c_uint32 # enum
+class struct_rocprofiler_thread_trace_decoder_shaderdata_t(Structure):
+    pass
+
+struct_rocprofiler_thread_trace_decoder_shaderdata_t._pack_ = 1 # source:False
+struct_rocprofiler_thread_trace_decoder_shaderdata_t._fields_ = [
+    ('time', ctypes.c_int64),
+    ('value', ctypes.c_uint64),
+    ('cu', ctypes.c_ubyte),
+    ('simd', ctypes.c_ubyte),
+    ('wave_id', ctypes.c_ubyte),
+    ('flags', ctypes.c_ubyte),
+    ('reserved', ctypes.c_uint32),
+]
+
+rocprofiler_thread_trace_decoder_shaderdata_t = struct_rocprofiler_thread_trace_decoder_shaderdata_t
+
+# values for enumeration 'rocprofiler_thread_trace_decoder_record_type_t'
+rocprofiler_thread_trace_decoder_record_type_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG',
+    6: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA',
+    7: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME',
+    8: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY',
+    9: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP = 0
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY = 1
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT = 2
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE = 3
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO = 4
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG = 5
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA = 6
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME = 7
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY = 8
+ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST = 9
+rocprofiler_thread_trace_decoder_record_type_t = ctypes.c_uint32 # enum
+
+# values for enumeration 'c__EA_rocprofiler_thread_trace_decoder_status_t'
+c__EA_rocprofiler_thread_trace_decoder_status_t__enumvalues = {
+    0: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS',
+    1: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR',
+    2: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES',
+    3: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT',
+    4: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA',
+    5: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST',
+}
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS = 0
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR = 1
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES = 2
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT = 3
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA = 4
+ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST = 5
+c__EA_rocprofiler_thread_trace_decoder_status_t = ctypes.c_uint32 # enum
+rocprofiler_thread_trace_decoder_status_t = c__EA_rocprofiler_thread_trace_decoder_status_t
+rocprofiler_thread_trace_decoder_status_t__enumvalues = c__EA_rocprofiler_thread_trace_decoder_status_t__enumvalues
+rocprof_trace_decoder_trace_callback_t = ctypes.CFUNCTYPE(c__EA_rocprofiler_thread_trace_decoder_status_t, rocprofiler_thread_trace_decoder_record_type_t, ctypes.POINTER(None), ctypes.c_uint64, ctypes.POINTER(None))
+rocprof_trace_decoder_isa_callback_t = ctypes.CFUNCTYPE(c__EA_rocprofiler_thread_trace_decoder_status_t, ctypes.POINTER(ctypes.c_char), ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(ctypes.c_uint64), struct_rocprofiler_thread_trace_decoder_pc_t, ctypes.POINTER(None))
+rocprof_trace_decoder_se_data_callback_t = ctypes.CFUNCTYPE(ctypes.c_uint64, ctypes.POINTER(ctypes.POINTER(ctypes.c_ubyte)), ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(None))
+try:
+    rocprof_trace_decoder_parse_data = _libraries['FIXME_STUB'].rocprof_trace_decoder_parse_data
+    rocprof_trace_decoder_parse_data.restype = rocprofiler_thread_trace_decoder_status_t
+    rocprof_trace_decoder_parse_data.argtypes = [rocprof_trace_decoder_se_data_callback_t, rocprof_trace_decoder_trace_callback_t, rocprof_trace_decoder_isa_callback_t, ctypes.POINTER(None)]
+except AttributeError:
+    pass
+try:
+    rocprof_trace_decoder_get_info_string = _libraries['FIXME_STUB'].rocprof_trace_decoder_get_info_string
+    rocprof_trace_decoder_get_info_string.restype = ctypes.POINTER(ctypes.c_char)
+    rocprof_trace_decoder_get_info_string.argtypes = [rocprofiler_thread_trace_decoder_info_t]
+except AttributeError:
+    pass
+try:
+    rocprof_trace_decoder_get_status_string = _libraries['FIXME_STUB'].rocprof_trace_decoder_get_status_string
+    rocprof_trace_decoder_get_status_string.restype = ctypes.POINTER(ctypes.c_char)
+    rocprof_trace_decoder_get_status_string.argtypes = [rocprofiler_thread_trace_decoder_status_t]
+except AttributeError:
+    pass
+rocprofiler_thread_trace_decoder_debug_callback_t = ctypes.CFUNCTYPE(None, ctypes.c_int64, ctypes.POINTER(ctypes.c_char), ctypes.POINTER(ctypes.c_char), ctypes.POINTER(None))
+uint64_t = ctypes.c_uint64
+try:
+    rocprof_trace_decoder_dump_data = _libraries['FIXME_STUB'].rocprof_trace_decoder_dump_data
+    rocprof_trace_decoder_dump_data.restype = rocprofiler_thread_trace_decoder_status_t
+    rocprof_trace_decoder_dump_data.argtypes = [ctypes.POINTER(ctypes.c_char), uint64_t, rocprofiler_thread_trace_decoder_debug_callback_t, ctypes.POINTER(None)]
+except AttributeError:
+    pass
+class union_rocprof_trace_decoder_gfx9_header_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_gfx9_header_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_gfx9_header_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_gfx9_header_t_0._fields_ = [
+    ('legacy_version', ctypes.c_uint64, 13),
+    ('gfx9_version2', ctypes.c_uint64, 3),
+    ('DSIMDM', ctypes.c_uint64, 4),
+    ('DCU', ctypes.c_uint64, 5),
+    ('reserved1', ctypes.c_uint64, 1),
+    ('SEID', ctypes.c_uint64, 6),
+    ('reserved2', ctypes.c_uint64, 32),
+]
+
+union_rocprof_trace_decoder_gfx9_header_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_gfx9_header_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_gfx9_header_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_gfx9_header_t_0),
+    ('raw', ctypes.c_uint64),
+]
+
+rocprof_trace_decoder_gfx9_header_t = union_rocprof_trace_decoder_gfx9_header_t
+class union_rocprof_trace_decoder_instrument_enable_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_instrument_enable_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_instrument_enable_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_instrument_enable_t_0._fields_ = [
+    ('char1', ctypes.c_uint32, 8),
+    ('char2', ctypes.c_uint32, 8),
+    ('char3', ctypes.c_uint32, 8),
+    ('char4', ctypes.c_uint32, 8),
+]
+
+union_rocprof_trace_decoder_instrument_enable_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_instrument_enable_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_instrument_enable_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_instrument_enable_t_0),
+    ('u32All', ctypes.c_uint32),
+]
+
+rocprof_trace_decoder_instrument_enable_t = union_rocprof_trace_decoder_instrument_enable_t
+class union_rocprof_trace_decoder_packet_header_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_packet_header_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_packet_header_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_packet_header_t_0._fields_ = [
+    ('opcode', ctypes.c_uint32, 8),
+    ('type', ctypes.c_uint32, 4),
+    ('data20', ctypes.c_uint32, 20),
+]
+
+union_rocprof_trace_decoder_packet_header_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_packet_header_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_packet_header_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_packet_header_t_0),
+    ('u32All', ctypes.c_uint32),
+]
+
+rocprof_trace_decoder_packet_header_t = union_rocprof_trace_decoder_packet_header_t
+
+# values for enumeration 'rocprof_trace_decoder_packet_opcode_t'
+rocprof_trace_decoder_packet_opcode_t__enumvalues = {
+    4: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ',
+    5: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP',
+    6: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO',
+}
+ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ = 4
+ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP = 5
+ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO = 6
+rocprof_trace_decoder_packet_opcode_t = ctypes.c_uint32 # enum
+
+# values for enumeration 'rocprof_trace_decoder_agent_info_type_t'
+rocprof_trace_decoder_agent_info_type_t__enumvalues = {
+    0: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ',
+    1: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL',
+    2: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST',
+}
+ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ = 0
+ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL = 1
+ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST = 2
+rocprof_trace_decoder_agent_info_type_t = ctypes.c_uint32 # enum
+class union_rocprof_trace_decoder_codeobj_marker_tail_t(Union):
+    pass
+
+class struct_rocprof_trace_decoder_codeobj_marker_tail_t_0(Structure):
+    pass
+
+struct_rocprof_trace_decoder_codeobj_marker_tail_t_0._pack_ = 1 # source:False
+struct_rocprof_trace_decoder_codeobj_marker_tail_t_0._fields_ = [
+    ('isUnload', ctypes.c_uint32, 1),
+    ('bFromStart', ctypes.c_uint32, 1),
+    ('legacy_id', ctypes.c_uint32, 30),
+]
+
+union_rocprof_trace_decoder_codeobj_marker_tail_t._pack_ = 1 # source:False
+union_rocprof_trace_decoder_codeobj_marker_tail_t._anonymous_ = ('_0',)
+union_rocprof_trace_decoder_codeobj_marker_tail_t._fields_ = [
+    ('_0', struct_rocprof_trace_decoder_codeobj_marker_tail_t_0),
+    ('raw', ctypes.c_uint32),
+]
+
+rocprof_trace_decoder_codeobj_marker_tail_t = union_rocprof_trace_decoder_codeobj_marker_tail_t
+
+# values for enumeration 'rocprof_trace_decoder_codeobj_marker_type_t'
+rocprof_trace_decoder_codeobj_marker_type_t__enumvalues = {
+    0: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL',
+    1: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO',
+    2: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO',
+    3: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI',
+    4: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI',
+    5: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO',
+    6: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI',
+    7: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST',
+}
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL = 0
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO = 1
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO = 2
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI = 3
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI = 4
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO = 5
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI = 6
+ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST = 7
+rocprof_trace_decoder_codeobj_marker_type_t = ctypes.c_uint32 # enum
+__all__ = \
+    ['ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU',
+    'ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA',
+    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM',
+    'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL',
+    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT',
+    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL',
+    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST',
+    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO',
+    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL',
+    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO',
+    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ',
+    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP',
+    'c__EA_rocprofiler_thread_trace_decoder_status_t',
+    'rocprof_trace_decoder_agent_info_type_t',
+    'rocprof_trace_decoder_codeobj_marker_tail_t',
+    'rocprof_trace_decoder_codeobj_marker_type_t',
+    'rocprof_trace_decoder_dump_data',
+    'rocprof_trace_decoder_get_info_string',
+    'rocprof_trace_decoder_get_status_string',
+    'rocprof_trace_decoder_gfx9_header_t',
+    'rocprof_trace_decoder_instrument_enable_t',
+    'rocprof_trace_decoder_isa_callback_t',
+    'rocprof_trace_decoder_packet_header_t',
+    'rocprof_trace_decoder_packet_opcode_t',
+    'rocprof_trace_decoder_parse_data',
+    'rocprof_trace_decoder_se_data_callback_t',
+    'rocprof_trace_decoder_trace_callback_t',
+    'rocprofiler_thread_trace_decoder_debug_callback_t',
+    'rocprofiler_thread_trace_decoder_info_t',
+    'rocprofiler_thread_trace_decoder_inst_category_t',
+    'rocprofiler_thread_trace_decoder_inst_t',
+    'rocprofiler_thread_trace_decoder_occupancy_t',
+    'rocprofiler_thread_trace_decoder_pc_t',
+    'rocprofiler_thread_trace_decoder_perfevent_t',
+    'rocprofiler_thread_trace_decoder_realtime_t',
+    'rocprofiler_thread_trace_decoder_record_type_t',
+    'rocprofiler_thread_trace_decoder_shaderdata_flags_t',
+    'rocprofiler_thread_trace_decoder_shaderdata_t',
+    'rocprofiler_thread_trace_decoder_status_t',
+    'rocprofiler_thread_trace_decoder_status_t__enumvalues',
+    'rocprofiler_thread_trace_decoder_wave_state_t',
+    'rocprofiler_thread_trace_decoder_wave_t',
+    'rocprofiler_thread_trace_decoder_wstate_type_t',
+    'struct_rocprof_trace_decoder_codeobj_marker_tail_t_0',
+    'struct_rocprof_trace_decoder_gfx9_header_t_0',
+    'struct_rocprof_trace_decoder_instrument_enable_t_0',
+    'struct_rocprof_trace_decoder_packet_header_t_0',
+    'struct_rocprofiler_thread_trace_decoder_inst_t',
+    'struct_rocprofiler_thread_trace_decoder_occupancy_t',
+    'struct_rocprofiler_thread_trace_decoder_pc_t',
+    'struct_rocprofiler_thread_trace_decoder_perfevent_t',
+    'struct_rocprofiler_thread_trace_decoder_realtime_t',
+    'struct_rocprofiler_thread_trace_decoder_shaderdata_t',
+    'struct_rocprofiler_thread_trace_decoder_wave_state_t',
+    'struct_rocprofiler_thread_trace_decoder_wave_t', 'uint64_t',
+    'union_rocprof_trace_decoder_codeobj_marker_tail_t',
+    'union_rocprof_trace_decoder_gfx9_header_t',
+    'union_rocprof_trace_decoder_instrument_enable_t',
+    'union_rocprof_trace_decoder_packet_header_t']
@@ -43,6 +43,7 @@ enum sqtt_version
   SQTT_VERSION_2_3 = 0x6, /* GFX9 */
   SQTT_VERSION_2_4 = 0x7, /* GFX10+ */
   SQTT_VERSION_3_2 = 0xb, /* GFX11+ */
+   SQTT_VERSION_3_3 = 0xc, /* GFX12+ */
 };

 enum sqtt_file_chunk_type
@@ -144,6 +145,8 @@ enum sqtt_gfxip_level
   SQTT_GFXIP_LEVEL_GFXIP_10_1 = 0x7,
   SQTT_GFXIP_LEVEL_GFXIP_10_3 = 0x9,
   SQTT_GFXIP_LEVEL_GFXIP_11_0 = 0xc,
+   SQTT_GFXIP_LEVEL_GFXIP_11_5 = 0xd,
+   SQTT_GFXIP_LEVEL_GFXIP_12 = 0x10,
 };

 enum sqtt_memory_type
@@ -427,6 +430,8 @@ enum elf_gfxip_level
   EF_AMDGPU_MACH_AMDGCN_GFX1010 = 0x033,
   EF_AMDGPU_MACH_AMDGCN_GFX1030 = 0x036,
   EF_AMDGPU_MACH_AMDGCN_GFX1100 = 0x041,
+   EF_AMDGPU_MACH_AMDGCN_GFX1150 = 0x043,
+   EF_AMDGPU_MACH_AMDGCN_GFX1200 = 0x04e,
 };

 struct sqtt_file_chunk_spm_db {
@@ -1,40 +0,0 @@
-import time
-from extra.optimization.helpers import load_worlds, ast_str_to_ast
-from tinygrad import Device
-from tinygrad.codegen.lowerer import pm_lowerer, get_index
-from tinygrad.uop.ops import graph_rewrite
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.postrange import Scheduler
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.helpers import getenv
-
-if __name__ == "__main__":
-  renderer = Device.default.renderer
-  ast_strs = load_worlds()
-  if (n:=getenv("N", -1)) != -1: ast_strs = ast_strs[n:n+1]
-  good = 0
-  for i, ast_str in enumerate(ast_strs):
-    ast = ast_str_to_ast(ast_str)
-
-    st = time.perf_counter()
-    lin = Kernel(ast, renderer)
-    opt1 = hand_coded_optimizations(lin)
-    et_lin = time.perf_counter() - st
-
-    lowered = graph_rewrite(ast, pm_lowerer, ctx=get_index(ast), bottom_up=True)
-    st = time.perf_counter()
-    sch = Scheduler(lowered, renderer)
-    sch.convert_loop_to_global()
-    sch.simplify_merge_adjacent()
-    opt2 = hand_coded_optimizations(sch)
-    et_sch = time.perf_counter() - st
-
-    if opt1 != opt2:
-      print(f"******* {i:6d}")
-      print("Kernel:    ", lin.colored_shape(), "->", lin.apply_opts(opt1).colored_shape())
-      print("Scheduler: ", sch.colored_shape(), "->", sch.apply_opts(opt2).colored_shape())
-      print(opt1)
-      print(opt2)
-    else:
-      good += 1
-      print(f"******* {i:6d} MATCH {good/(i+1)*100:.2f}% -- {et_lin/et_sch:4.2f}x speedup")
@@ -0,0 +1,400 @@
+/**
+ * @file
+ * @brief Basic operations on generic types.
+ */
+
+#pragma once
+
+#include <cuda_bf16.h>
+#include <limits>
+#include "base_types.cuh"
+
+namespace kittens {
+
+/**
+ * @namespace base_ops
+ *
+ * @brief A namespace for operations on basic data types.
+ */
+namespace base_ops {
+
+/* ----------  CONST OPS  ---------- */
+
+/**
+ * @brief Represents the zero constant operation.
+ *
+ * This operation returns the zero value of the specified type.
+ *
+ * @tparam T The data type for which to return the zero value.
+ * @return The zero value of type T.
+ */
+struct zero {
+    template<typename T, typename... args> __device__ static inline constexpr T op(args... _) { return base_types::constants<T>::zero();      }
+};
+/**
+ * @brief Represents the one constant operation.
+ *
+ * This operation returns the one value of the specified type.
+ *
+ * @tparam T The data type for which to return the one value.
+ * @return The one value of type T.
+ */
+struct one {
+    template<typename T, typename... args> __device__ static inline constexpr T op(args... _) { return base_types::constants<T>::one();       }
+};
+/**
+ * @brief Represents the positive infinity constant operation.
+ *
+ * This operation returns the positive infinity value of the specified type.
+ *
+ * @tparam T The data type for which to return the positive infinity value.
+ * @return The positive infinity value of type T.
+ */
+struct pos_infty {
+    template<typename T, typename... args> __device__ static inline constexpr T op(args... _) { return base_types::constants<T>::pos_infty(); }
+};
+/**
+ * @brief Represents the negative infinity constant operation.
+ *
+ * This operation returns the negative infinity value of the specified type.
+ *
+ * @tparam T The data type for which to return the negative infinity value.
+ * @return The negative infinity value of type T.
+ */
+struct neg_infty {
+    template<typename T, typename... args> __device__ static inline constexpr T op(args... _) { return base_types::constants<T>::neg_infty(); }
+};
+
+
+/* ----------  UNARY OPS  ---------- */
+
+/**
+ * @brief Exponential function operation.
+ *
+ * This operation calculates the exponential of the input value.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param x[in] The input value.
+ * @return The exponential of the input value.
+ */
+struct exp {
+    template<typename T> static __device__ inline T op(const T &x) { return exp(x); }
+};
+template<> __device__ inline float  exp::op<float> (const float &x ) { return __expf(x);                        }
+template<> __device__ inline float2 exp::op<float2>(const float2 &x) { return float2{__expf(x.x), __expf(x.y)}; }
+template<> __device__ inline bf16   exp::op<bf16>  (const bf16 &x  ) { return hexp(x);                          }
+template<> __device__ inline bf16_2 exp::op<bf16_2>(const bf16_2 &x) { return h2exp(x);                         }
+template<> __device__ inline half   exp::op<half>  (const half &x  ) { return hexp(x);                          }
+template<> __device__ inline half_2 exp::op<half_2>(const half_2 &x) { return h2exp(x);                         }
+
+/**
+ * @brief Exponential function operation, in base 2
+ *
+ * This operation calculates the exponential of the input value, in base 2.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param x[in] The input value.
+ * @return The exponential of the input value.
+ */
+struct exp2 {
+    template<typename T> static __device__ inline T op(const T &x) { return exp2f(x); }
+};
+template<> __device__ inline float  exp2::op<float> (const float &x ) { return exp2f(x);                        }
+template<> __device__ inline float2 exp2::op<float2>(const float2 &x) { return float2{exp2f(x.x), exp2f(x.y)}; }
+template<> __device__ inline bf16   exp2::op<bf16>  (const bf16 &x  ) { return hexp2(x);                          }
+template<> __device__ inline bf16_2 exp2::op<bf16_2>(const bf16_2 &x) { return h2exp2(x);                         }
+template<> __device__ inline half   exp2::op<half>  (const half &x  ) { return hexp2(x);                          }
+template<> __device__ inline half_2 exp2::op<half_2>(const half_2 &x) { return h2exp2(x);                         }
+/**
+ * @brief Natural log function operation.
+ *
+ * This operation calculates the natural logarithm of the input value.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param x[in] The input value.
+ * @return The natural logarithm of the input value.
+ */
+struct log {
+    template<typename T> static __device__ inline T op(const T &x) { return log(x); }
+};
+template<> __device__ inline float  log::op<float> (const float &x ) { return __logf(x);                        }
+template<> __device__ inline float2 log::op<float2>(const float2 &x) { return float2{__logf(x.x), __logf(x.y)}; }
+template<> __device__ inline bf16   log::op<bf16>  (const bf16 &x  ) { return hlog(x);                          }
+template<> __device__ inline bf16_2 log::op<bf16_2>(const bf16_2 &x) { return h2log(x);                         }
+template<> __device__ inline half   log::op<half>  (const half &x  ) { return hlog(x);                          }
+template<> __device__ inline half_2 log::op<half_2>(const half_2 &x) { return h2log(x);                         }
+/**
+ * @brief Logarithm base 2 operation.
+ *
+ * This operation calculates the logarithm base 2 of the input value.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param x[in] The input value.
+ * @return The logarithm base 2 of the input value.
+ */
+struct log2 {
+    template<typename T> static __device__ inline T op(const T &x) { return log2(x); }
+};
+template<> __device__ inline float  log2::op<float> (const float &x ) { return __log2f(x);                        }
+template<> __device__ inline float2 log2::op<float2>(const float2 &x) { return float2{__log2f(x.x), __log2f(x.y)}; }
+template<> __device__ inline bf16   log2::op<bf16>  (const bf16 &x  ) { return hlog2(x);                          }
+template<> __device__ inline bf16_2 log2::op<bf16_2>(const bf16_2 &x) { return h2log2(x);                         }
+template<> __device__ inline half   log2::op<half>  (const half &x  ) { return hlog2(x);                          }
+template<> __device__ inline half_2 log2::op<half_2>(const half_2 &x) { return h2log2(x);                         }
+/**
+ * @brief Absolute value operation.
+ *
+ * This operation calculates the absolute value of the input.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param x[in] The input value.
+ * @return The absolute value of the input.
+ */
+struct abs {
+    template<typename T> static __device__ inline T op(const T &x) { return abs(x); }
+};
+template<> __device__ inline float  abs::op<float> (const float &x ) { return fabsf(x);                       }
+template<> __device__ inline float2 abs::op<float2>(const float2 &x) { return float2{fabsf(x.x), fabsf(x.y)}; }
+template<> __device__ inline bf16   abs::op<bf16>  (const bf16 &x  ) { return __habs(x);                      }
+template<> __device__ inline bf16_2 abs::op<bf16_2>(const bf16_2 &x) { return __habs2(x);                     }
+template<> __device__ inline half   abs::op<half>  (const half &x  ) { return __habs(x);                      }
+template<> __device__ inline half_2 abs::op<half_2>(const half_2 &x) { return __habs2(x);                     }
+/**
+ * @brief Rectified Linear Unit (ReLU) operation.
+ *
+ * This operation applies the ReLU function to the input, which is the
+ * maximum of zero and the input value.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param x[in] The input value.
+ * @return The result of ReLU function applied to the input.
+ */
+struct relu {
+    template<typename T> static __device__ inline T op(const T &x) { return max(x, base_types::constants<T>::zero()); }
+};
+template<> __device__ inline float  relu::op<float> (const float &x ) { return max(x, 0.f);                                  }
+template<> __device__ inline float2 relu::op<float2>(const float2 &x) { return float2{max(x.x, 0.f), max(x.y, 0.f)};         }
+template<> __device__ inline bf16   relu::op<bf16>  (const bf16 &x  ) { return __hmax(x, base_types::constants<bf16>::zero());    }
+template<> __device__ inline bf16_2 relu::op<bf16_2>(const bf16_2 &x) { return __hmax2(x, base_types::constants<bf16_2>::zero()); }
+template<> __device__ inline half   relu::op<half>  (const half &x  ) { return __hmax(x, base_types::constants<half>::zero());    }
+template<> __device__ inline half_2 relu::op<half_2>(const half_2 &x) { return __hmax2(x, base_types::constants<half_2>::zero()); }
+/**
+ * @brief Copy operation.
+ *
+ * This operation returns the input value unchanged.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The input value.
+ * @return The same value as the input.
+ */
+struct copy { // for non-compile-time setters.
+    template<typename T> static __device__ inline T op(const T &a) { return a; }
+};
+
+
+/* ----------  BINARY OPS  ---------- */
+
+/**
+ * @brief Copy2 operation.
+ *
+ * This operation returns the second input value unchanged.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value (ignored).
+ * @param b[in] The second input value.
+ * @return The same value as the second input.
+ */
+struct copy2 { // this turns out to be a slightly hacky op that makes some code cleaner :/
+    template<typename T> static __device__ inline T op(const T &a, const T &b) { return b; }
+};
+/**
+ * @brief Sum operation.
+ *
+ * This operation calculates the sum of two input values.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The second input value.
+ * @return The sum of the input values.
+ */
+struct sum {
+    template<typename T> static __device__ inline T op(const T &a, const T &b) { return a+b; }
+};
+template<> __device__ inline float2 sum::op<float2>(const float2 &a, const float2 &b) {
+#ifdef KITTENS_BLACKWELL
+    float2 c;
+    asm volatile("add.f32x2 %0, %1, %2;" : "=l"(*(uint64_t*)&c) : "l"(*(uint64_t*)&a), "l"(*(uint64_t*)&b));
+    return c;
+#else
+    return float2{a.x+b.x, a.y+b.y};
+#endif
+}
+template<> __device__ inline bf16   sum::op<bf16>  (const bf16   &a, const bf16   &b) { return __hadd(a, b);             }
+template<> __device__ inline bf16_2 sum::op<bf16_2>(const bf16_2 &a, const bf16_2 &b) { return __hadd2(a, b);            }
+template<> __device__ inline half   sum::op<half>  (const half   &a, const half   &b) { return __hadd(a, b);             }
+template<> __device__ inline half_2 sum::op<half_2>(const half_2 &a, const half_2 &b) { return __hadd2(a, b);            }
+/**
+ * @brief Subtraction operation.
+ *
+ * This operation calculates the difference between two input values.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The second input value.
+ * @return The difference between the input values.
+ */
+struct sub {
+    template<typename T> static __device__ inline T op(const T &a, const T &b) { return a-b; }
+};
+template<> __device__ inline float2 sub::op<float2>(const float2 &a, const float2 &b) { 
+#ifdef KITTENS_BLACKWELL
+    float2 c;
+    asm volatile("sub.f32x2 %0, %1, %2;" : "=l"(*(uint64_t*)&c) : "l"(*(uint64_t*)&a), "l"(*(uint64_t*)&b));
+    return c;
+#else
+    return float2{a.x-b.x, a.y-b.y}; 
+#endif
+}
+template<> __device__ inline bf16   sub::op<bf16>  (const bf16   &a, const bf16   &b) { return __hsub(a, b);             }
+template<> __device__ inline bf16_2 sub::op<bf16_2>(const bf16_2 &a, const bf16_2 &b) { return __hsub2(a, b);            }
+template<> __device__ inline half   sub::op<half>  (const half   &a, const half   &b) { return __hsub(a, b);             }
+template<> __device__ inline half_2 sub::op<half_2>(const half_2 &a, const half_2 &b) { return __hsub2(a, b);            }
+/**
+ * @brief Multiplication operation.
+ *
+ * This operation calculates the product of two input values.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The second input value.
+ * @return The product of the input values.
+ */
+struct mul {
+    template<typename T> static __device__ inline T op(const T &a, const T &b) { return a*b; }
+};
+template<> __device__ inline float2 mul::op<float2>(const float2 &a, const float2 &b) { 
+#ifdef KITTENS_BLACKWELL
+    float2 c;
+    asm volatile("mul.f32x2 %0, %1, %2;" : "=l"(*(uint64_t*)&c) : "l"(*(uint64_t*)&a), "l"(*(uint64_t*)&b));
+    return c;
+#else
+    return float2{a.x*b.x, a.y*b.y}; 
+#endif
+}
+template<> __device__ inline bf16   mul::op<bf16>  (const bf16   &a, const bf16   &b) { return __hmul(a, b);             }
+template<> __device__ inline bf16_2 mul::op<bf16_2>(const bf16_2 &a, const bf16_2 &b) { return __hmul2(a, b);            }
+template<> __device__ inline half   mul::op<half>  (const half   &a, const half   &b) { return __hmul(a, b);             }
+template<> __device__ inline half_2 mul::op<half_2>(const half_2 &a, const half_2 &b) { return __hmul2(a, b);            }
+/**
+ * @brief Division operation.
+ *
+ * This operation calculates the quotient of two input values.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The second input value.
+ * @return The quotient of the input values.
+ */
+struct div {
+    template<typename T> static __device__ inline T op(const T &a, const T &b) { return a/b; }
+};
+template<> __device__ inline float2 div::op<float2>(const float2 &a, const float2 &b) { return float2{a.x/b.x, a.y/b.y}; }
+template<> __device__ inline bf16   div::op<bf16>  (const bf16   &a, const bf16   &b) { return __hdiv(a, b);             }
+template<> __device__ inline bf16_2 div::op<bf16_2>(const bf16_2 &a, const bf16_2 &b) { return __h2div(a, b);            } // this op is a special snowflake
+template<> __device__ inline half   div::op<half>  (const half   &a, const half   &b) { return __hdiv(a, b);             }
+template<> __device__ inline half_2 div::op<half_2>(const half_2 &a, const half_2 &b) { return __h2div(a, b);            }
+/**
+ * @brief Maximum operation.
+ *
+ * This operation calculates the maximum of two input values.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The second input value.
+ * @return The maximum of the input values.
+ */
+ struct max {
+    template<typename T> static __device__ inline T op(const T &a, const T &b) { return ::max(a, b); }
+};
+template<>  __device__ inline float2 max::op<float2>(const float2 &a, const float2 &b) { return float2{::max(a.x, b.x), ::max(a.y, b.y)}; }
+template<>  __device__ inline bf16   max::op<bf16>  (const bf16   &a, const bf16   &b) { return __hmax(a, b);                             }
+template<>  __device__ inline bf16_2 max::op<bf16_2>(const bf16_2 &a, const bf16_2 &b) { return __hmax2(a, b);                            }
+template<>  __device__ inline half   max::op<half>  (const half   &a, const half   &b) { return __hmax(a, b);                             }
+template<>  __device__ inline half_2 max::op<half_2>(const half_2 &a, const half_2 &b) { return __hmax2(a, b);                            }
+/**
+ * @brief Minimum operation.
+ *
+ * This operation calculates the minimum of two input values.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The second input value.
+ * @return The minimum of the input values.
+ */
+struct min {
+    template<typename T> static __device__ inline T op(const T &a, const T &b) { return ::min(a, b); }
+};
+template<>  __device__ inline float2 min::op<float2>(const float2 &a, const float2 &b) { return float2{::min(a.x, b.x), ::min(a.y, b.y)}; }
+template<>  __device__ inline bf16   min::op<bf16>  (const bf16   &a, const bf16   &b) { return __hmin(a, b);                         }
+template<>  __device__ inline bf16_2 min::op<bf16_2>(const bf16_2 &a, const bf16_2 &b) { return __hmin2(a, b);                        }
+template<>  __device__ inline half   min::op<half>  (const half   &a, const half   &b) { return __hmin(a, b);                         }
+template<>  __device__ inline half_2 min::op<half_2>(const half_2 &a, const half_2 &b) { return __hmin2(a, b);                        }
+
+
+/* ----------  TERNARY OPS  ---------- */
+
+/**
+ * @brief Fused multiply-add operation A * B + C.
+ *
+ * This operation performs a fused multiply-add, computing (A * B) + C with only one rounding.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The second input value.
+ * @param c[in] The third input value to be added.
+ * @return The result of the fused multiply-add operation.
+ */
+struct fma_AxBtC {
+    template<typename T> static __device__ inline T op(const T &a, const T &b, const T &c) {
+        return sum::op<T>(mul::op<T>(a, b), c);
+    }
+};
+template<> __device__ inline float2 fma_AxBtC::op<float2>(const float2 &a, const float2 &b, const float2 &c) {
+#ifdef KITTENS_BLACKWELL
+    float2 d;
+    asm volatile("fma.rn.f32x2 %0, %1, %2, %3;" : "=l"(*(uint64_t*)&d) : "l"(*(uint64_t*)&a), "l"(*(uint64_t*)&b), "l"(*(uint64_t*)&c));
+    return d;
+#else
+    return float2{a.x*b.x+c.x, a.y*b.y+c.y};
+#endif
+}
+/**
+ * @brief Fused multiply-add operation A * C + B.
+ *
+ * This operation performs a fused multiply-add, computing (A * C) + B with only one rounding.
+ * This is particularly useful for attention mechanisms in neural networks.
+ *
+ * @tparam T The data type of the input and output values.
+ * @param a[in] The first input value.
+ * @param b[in] The third input value to be added.
+ * @param c[in] The second input value.
+ * @return The result of the fused multiply-add operation.
+ */
+struct fma_AxCtB { // this is the one needed for attention
+    template<typename T> static __device__ inline T op(const T &a, const T &b, const T &c) {
+        return sum::op<T>(mul::op<T>(a, c), b);
+    }
+};
+template<> __device__ inline float2 fma_AxCtB::op<float2>(const float2 &a, const float2 &b, const float2 &c) {
+#ifdef KITTENS_BLACKWELL
+    float2 d;
+    asm volatile("fma.rn.f32x2 %0, %1, %2, %3;" : "=l"(*(uint64_t*)&d) : "l"(*(uint64_t*)&a), "l"(*(uint64_t*)&c), "l"(*(uint64_t*)&b));
+    return d;
+#else
+    return float2{a.x*c.x+b.x, a.y*c.y+b.y};
+#endif
+}
+
+} // namespace base_ops
+
+} // namespace kittens
@@ -0,0 +1,519 @@
+/**
+ * @file
+ * @brief Declarations, manipulations, and wrappers for basic types.
+ * 
+ * This file is a bunch of utilities for going back and forth between different types.
+ * 
+ * Many of them are for the compiler, so as to clean up the code. It unfortunately
+ * seems necessary when we have types we really care about that are less than word width.
+ */
+
+#pragma once
+
+#ifdef KITTENS_HOPPER
+#include <cuda_fp8.h>
+#endif
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <string>
+#include <bit>
+
+
+namespace kittens {
+
+/**
+ * @brief Bfloat16 floating-point type.
+ */
+using bf16 = __nv_bfloat16;
+/**
+ * @brief Half-precision floating-point type.
+ */
+using half = __half;
+/**
+ * @brief Packed word of two bfloat16 floating-point values.
+ */
+using bf16_2 = __nv_bfloat162;
+/**
+ * @brief Packed word of two half-precision floating-point values.
+ */
+using half_2 = __half2;
+#ifdef KITTENS_HOPPER
+/**
+ * @brief float8 floating-point type.
+ */
+using fp8e4m3 = __nv_fp8_e4m3;
+using fp8e5m2 = __nv_fp8_e5m2;
+#ifdef KITTENS_BLACKWELL
+using fp8e8m0 = __nv_fp8_e8m0;
+#endif
+/**
+ * @brief 2-packed float8 floating-point type.
+ */
+using fp8e4m3_2 = __nv_fp8x2_e4m3;
+using fp8e5m2_2 = __nv_fp8x2_e5m2;
+#ifdef KITTENS_BLACKWELL
+using fp8e8m0_2 = __nv_fp8x2_e8m0;
+#endif
+/**
+ * @brief 4-packed float8 floating-point type.
+ */
+using fp8e4m3_4 = __nv_fp8x4_e4m3;
+using fp8e5m2_4 = __nv_fp8x4_e5m2;
+#ifdef KITTENS_BLACKWELL
+using fp8e8m0_4 = __nv_fp8x4_e8m0;
+#endif
+#endif
+
+namespace ducks {
+/**
+ * @namespace base_types
+ *
+ * @brief A namespace for concepts for basic data types.
+ */
+namespace base_types {
+
+#ifdef KITTENS_HOPPER
+#ifdef KITTENS_BLACKWELL
+template<typename T>
+concept T2 = std::is_same_v<T, float2> || std::is_same_v<T, bf16_2> || std::is_same_v<T, half_2> || std::is_same_v<T, fp8e4m3_4> || std::is_same_v<T, fp8e5m2_4> || std::is_same_v<T, fp8e8m0_4>; // could add half_2 later if implemented.
+template<typename T>
+concept T1 = std::is_same_v<T, float>  || std::is_same_v<T, bf16  > || std::is_same_v<T, half> || std::is_same_v<T, fp8e4m3> || std::is_same_v<T, fp8e5m2> || std::is_same_v<T, fp8e8m0>; // could add half_2 later if implemented.
+#else
+template<typename T>
+concept T2 = std::is_same_v<T, float2> || std::is_same_v<T, bf16_2> || std::is_same_v<T, half_2> || std::is_same_v<T, fp8e4m3_4> || std::is_same_v<T, fp8e5m2_4>;
+template<typename T>
+concept T1 = std::is_same_v<T, float>  || std::is_same_v<T, bf16  > || std::is_same_v<T, half> || std::is_same_v<T, fp8e4m3> || std::is_same_v<T, fp8e5m2>;
+#endif
+#else
+template<typename T>
+concept T2 = std::is_same_v<T, float2> || std::is_same_v<T, bf16_2> || std::is_same_v<T, half_2>;
+template<typename T>
+concept T1 = std::is_same_v<T, float>  || std::is_same_v<T, bf16  > || std::is_same_v<T, half>;
+#endif
+
+} // namespace base_types
+} // namespace ducks
+
+/**
+ * @namespace base_types
+ *
+ * @brief A namespace for ThunderKittens basic data types.
+ */
+namespace base_types {
+
+/**
+ * @brief Provides compile-time constants for different types.
+ *
+ * @tparam T The type for which to provide constants.
+ */
+template<typename T> struct constants {
+    /**
+     * @brief Zero
+     * @return Constexpr zero with type T
+     */
+    static __device__ inline constexpr T zero()      { return T{0}; }
+    /**
+     * @brief One
+     * @return Constexpr one with type T
+     */
+    static __device__ inline constexpr T one()       { return T{1}; }
+    /**
+     * @brief Positive infinity. Particularly useful for initializing before a min op.
+     * @return Constexpr positive infinity with type T
+     */
+    static __device__ inline constexpr T pos_infty() { return T{INFINITY}; } // I'll find a better way at some point but this appears to work.
+    /**
+     * @brief Negative infinity. Particularly useful for initializing before a max op.
+     * @return Constexpr negative infinity with type T
+     */
+    static __device__ inline constexpr T neg_infty() { return T{-INFINITY}; }
+};
+template<> struct constants<float2> {
+    static __device__ inline constexpr float2 zero()      { return float2{0.f, 0.f}; }
+    static __device__ inline constexpr float2 one()       { return float2{1.f, 1.f}; }
+    static __device__ inline constexpr float2 pos_infty() { return float2{constants<float>::pos_infty(), constants<float>::pos_infty()}; }
+    static __device__ inline constexpr float2 neg_infty() { return float2{constants<float>::neg_infty(), constants<float>::neg_infty()}; }
+};
+template<> struct constants<bf16> {
+    static __device__ inline constexpr bf16 zero()      { return std::bit_cast<__nv_bfloat16>(uint16_t(0x0000)); } // unfortunately __float2bf16_rn is not constexpr
+    static __device__ inline constexpr bf16 one()       { return std::bit_cast<__nv_bfloat16>(uint16_t(0x3F80)); }
+    static __device__ inline constexpr bf16 pos_infty() { return std::bit_cast<__nv_bfloat16>(uint16_t(0x7F80)); }
+    static __device__ inline constexpr bf16 neg_infty() { return std::bit_cast<__nv_bfloat16>(uint16_t(0xFF80)); }
+};
+template<> struct constants<bf16_2> {
+    static __device__ inline constexpr bf16_2 zero()      { return bf16_2{constants<bf16>::zero(),      constants<bf16>::zero()};      }
+    static __device__ inline constexpr bf16_2 one()       { return bf16_2{constants<bf16>::one(),       constants<bf16>::one()};       }
+    static __device__ inline constexpr bf16_2 pos_infty() { return bf16_2{constants<bf16>::pos_infty(), constants<bf16>::pos_infty()}; }
+    static __device__ inline constexpr bf16_2 neg_infty() { return bf16_2{constants<bf16>::neg_infty(), constants<bf16>::neg_infty()}; }
+};
+template<> struct constants<half> {
+    static __device__ inline constexpr half zero()      { return std::bit_cast<__half>(uint16_t(0x0000)); }
+    static __device__ inline constexpr half one()       { return std::bit_cast<__half>(uint16_t(0x3C00)); }
+    static __device__ inline constexpr half pos_infty() { return std::bit_cast<__half>(uint16_t(0x7C00)); }
+    static __device__ inline constexpr half neg_infty() { return std::bit_cast<__half>(uint16_t(0xFC00)); }
+};
+template<> struct constants<half_2> {
+    static __device__ inline constexpr half_2 zero()      { return half_2{constants<half>::zero(),      constants<half>::zero()};      }
+    static __device__ inline constexpr half_2 one()       { return half_2{constants<half>::one(),       constants<half>::one()};       }
+    static __device__ inline constexpr half_2 pos_infty() { return half_2{constants<half>::pos_infty(), constants<half>::pos_infty()}; }
+    static __device__ inline constexpr half_2 neg_infty() { return half_2{constants<half>::neg_infty(), constants<half>::neg_infty()}; }
+};
+#ifdef KITTENS_HOPPER
+template<> struct constants<fp8e4m3> {
+    static __device__ inline constexpr fp8e4m3 zero() { return std::bit_cast<__nv_fp8_e4m3>(uint8_t(0x00)); }
+    static __device__ inline constexpr fp8e4m3 one() { return std::bit_cast<__nv_fp8_e4m3>(uint8_t(0x38)); }
+};
+template<> struct constants<fp8e4m3_2> {
+    static __device__ inline constexpr fp8e4m3_2 zero() { return std::bit_cast<fp8e4m3_2>(uint16_t(0x0000)); }
+    static __device__ inline constexpr fp8e4m3_2 one() { return std::bit_cast<fp8e4m3_2>(uint16_t(0x3838)); }
+};
+template<> struct constants<fp8e4m3_4> {
+    static __device__ inline constexpr fp8e4m3_4 zero() { return std::bit_cast<fp8e4m3_4>(uint32_t(0x00000000)); }
+    static __device__ inline constexpr fp8e4m3_4 one() { return std::bit_cast<fp8e4m3_4>(uint32_t(0x38383838)); }
+};
+template<> struct constants<fp8e5m2> {
+    static __device__ inline constexpr fp8e5m2 zero() { return std::bit_cast<__nv_fp8_e5m2>(uint8_t(0x00)); }
+    static __device__ inline constexpr fp8e5m2 one() { return std::bit_cast<__nv_fp8_e5m2>(uint8_t(0x3C)); }
+};
+template<> struct constants<fp8e5m2_2> {
+    static __device__ inline constexpr fp8e5m2_2 zero() { return std::bit_cast<fp8e5m2_2>(uint16_t(0x0000)); }
+    static __device__ inline constexpr fp8e5m2_2 one() { return std::bit_cast<fp8e5m2_2>(uint16_t(0x3C3C)); }
+};
+template<> struct constants<fp8e5m2_4> {
+    static __device__ inline constexpr fp8e5m2_4 zero() { return std::bit_cast<fp8e5m2_4>(uint32_t(0x00000000)); }
+    static __device__ inline constexpr fp8e5m2_4 one() { return std::bit_cast<fp8e5m2_4>(uint32_t(0x3C3C3C3C)); }
+};
+#endif
+
+template<> struct constants<int> {
+    static __device__ inline constexpr int zero()      { return 0; }
+    static __device__ inline constexpr int one()       { return 1; }
+};
+template<> struct constants<int2> {
+    static __device__ inline constexpr int2 zero()      { return int2{0, 0}; }
+    static __device__ inline constexpr int2 one()       { return int2{1, 1}; }
+};
+
+/**
+ * @brief Provides information about packing of elements for a given type.
+ *
+ * @tparam T The type for which to provide packing information.
+ */
+template<typename T> struct packing {
+    /**
+     * @brief The number of elements packed together.
+     *
+     * @return constexpr int representing number of elements within the type.
+     */
+    static __device__ inline constexpr int num() { return 1; }
+    /**
+     * @brief Packs a single T element twice (replicated) into its packed type.
+     *
+     * @param i[in] The element to pack.
+     * @return The packed type.
+     */
+    static __device__ inline constexpr T pack(const bf16 &i);
+};
+template<> struct packing<bf16> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = bf16;
+    using packed_type = bf16_2;
+    static __device__ inline constexpr bf16_2 pack(const bf16 &i) { return bf16_2{i, i}; }
+};
+template<> struct packing<bf16_2> {
+    static __device__ inline constexpr int num() { return 2; }
+    using unpacked_type = bf16;
+    using packed_type = bf16_2;
+    static __device__ inline constexpr bf16_2 pack(const bf16 &i) { return bf16_2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<half> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = half;
+    using packed_type = half_2;
+    static __device__ inline constexpr half_2 pack(const half &i) { return half_2{i, i}; }
+};
+template<> struct packing<half_2> {
+    static __device__ inline constexpr int num() { return 2; }
+    using unpacked_type = half;
+    using packed_type = half_2;
+    static __device__ inline constexpr half_2 pack(const half &i) { return half_2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<float> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = float;
+    using packed_type = float2;
+    static __device__ inline constexpr float2 pack(const float &i) { return float2{i, i}; }
+};
+template<> struct packing<float2> {
+    static __device__ inline constexpr int num() { return 2; }
+    using unpacked_type = float;
+    using packed_type = float2;
+    static __device__ inline constexpr float2 pack(const float &i) { return float2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<char> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = char;
+    using packed_type = char2;
+    static __device__ inline constexpr char2 pack(const char &i) { return char2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<char2> {
+    static __device__ inline constexpr int num() { return 2; }
+    using unpacked_type = char;
+    using packed_type = char2;
+    static __device__ inline constexpr char2 pack(const char &i) { return char2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<int> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = int;
+    using packed_type = int2;
+    static __device__ inline constexpr int2 pack(const int &i) { return int2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<int2> {
+    static __device__ inline constexpr int num() { return 2; }
+    using unpacked_type = int;
+    using packed_type = int2;
+    static __device__ inline constexpr int2 pack(const int &i) { return int2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<uint> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = uint;
+    using packed_type = uint2;
+    static __device__ inline constexpr uint2 pack(const uint &i) { return uint2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<uint2> {
+    static __device__ inline constexpr int num() { return 2; }
+    using unpacked_type = uint;
+    using packed_type = uint2;
+    static __device__ inline constexpr uint2 pack(const uint &i) { return uint2{i, i}; } // this replication makes code cleaner later.
+};
+struct uint64_2 { uint64_t x, y; };
+template<> struct packing<uint64_t> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = uint64_t;
+    using packed_type = uint64_2;
+    static __device__ inline constexpr uint64_2 pack(const uint64_t &i) { return uint64_2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<uint64_2> {
+    static __device__ inline constexpr int num() { return 2; }
+    using unpacked_type = uint64_t;
+    using packed_type = uint64_2;
+    static __device__ inline constexpr uint64_2 pack(const uint64_t &i) { return uint64_2{i, i}; } // this replication makes code cleaner later.
+};
+template<> struct packing<float4> {
+    static __device__ inline constexpr int num() { return 4; }
+};
+template<> struct packing<int4> {
+    static __device__ inline constexpr int num() { return 4; }
+};
+#ifdef KITTENS_HOPPER
+template<> struct packing<fp8e4m3> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = fp8e4m3;
+    using packed_type = fp8e4m3_4;
+};
+template<> struct packing<fp8e4m3_4> {
+    static __device__ inline constexpr int num() { return 4; }
+    using unpacked_type = fp8e4m3;
+    using packed_type = fp8e4m3_4;
+};
+template<> struct packing<fp8e5m2> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = fp8e5m2;
+    using packed_type = fp8e5m2_4;
+};
+template<> struct packing<fp8e5m2_4> {
+    static __device__ inline constexpr int num() { return 4; }
+    using unpacked_type = fp8e5m2;
+    using packed_type = fp8e5m2_4;
+};
+#ifdef KITTENS_BLACKWELL
+template<> struct packing<fp8e8m0> {
+    static __device__ inline constexpr int num() { return 1; }
+    using unpacked_type = fp8e8m0;
+    using packed_type = fp8e8m0_4;
+};
+template<> struct packing<fp8e8m0_4> {
+    static __device__ inline constexpr int num() { return 4; }
+    using unpacked_type = fp8e8m0;
+    using packed_type = fp8e8m0_4;
+};
+#endif
+#endif
+
+
+/**
+ * @brief Provides templated functionality to convert between different types.
+ *
+ * @tparam T The target type for conversion.
+ * @tparam U The source type for conversion.
+ */
+template<typename T, typename U> struct convertor {
+    /**
+     * @brief Converts a value of type U to type T.
+     *
+     * @param u[in] The value of type U to convert.
+     * @return T The converted value of type T.
+     */
+    static __host__ __device__ inline T convert(const U & u) {
+        return (T)u;
+    }
+};
+template<> struct convertor<float, bf16> {
+    static __host__ __device__ inline float convert(const bf16 & u) {
+        return 	__bfloat162float(u);
+    }
+};
+template<> struct convertor<bf16, float> {
+    static __host__ __device__ inline bf16 convert(const float & u) {
+        return 	__float2bfloat16_rn(u);
+    }
+};
+template<> struct convertor<float2, bf16_2> {
+    static __host__ __device__ inline float2 convert(const bf16_2 & u) {
+        return 	__bfloat1622float2(u);
+    }
+};
+template<> struct convertor<bf16_2, float2> {
+    static __host__ __device__ inline bf16_2 convert(const float2 & u) {
+        return 	__float22bfloat162_rn(u);
+    }
+};
+template<> struct convertor<float, half> {
+    static __host__ __device__ inline float convert(const half & u) {
+        return __half2float(u);
+    }
+};
+template<> struct convertor<half, float> {
+    static __host__ __device__ inline half convert(const float & u) {
+        return __float2half(u);
+    }
+};
+template<> struct convertor<float2, half_2> {
+    static __host__ __device__ inline float2 convert(const half_2 & u) {
+        return __half22float2(u);
+    }
+};
+template<> struct convertor<half_2, float2> {
+    static __host__ __device__ inline half_2 convert(const float2 & u) {
+        return __float22half2_rn(u);
+    }
+};
+template<> struct convertor<bf16, half> {
+    static __host__ __device__ inline bf16 convert(const half & u) {
+        return __float2bfloat16_rn(__half2float(u));
+    }
+};
+template<> struct convertor<half, bf16> {
+    static __host__ __device__ inline half convert(const bf16 & u) {
+        return __float2half(__bfloat162float(u));
+    }
+};
+template<> struct convertor<bf16_2, half_2> {
+    static __host__ __device__ inline bf16_2 convert(const half_2 & u) {
+        return __float22bfloat162_rn(__half22float2(u));
+    }
+};
+template<> struct convertor<half_2, bf16_2> {
+    static __host__ __device__ inline half_2 convert(const bf16_2 & u) {
+        return __float22half2_rn(__bfloat1622float2(u));
+    }
+};
+#ifdef KITTENS_HOPPER
+// fp8e4m3
+template<> struct convertor<fp8e4m3_4, float4> {
+    static __host__ __device__ inline fp8e4m3_4 convert(const float4& u) {
+        return __nv_fp8x4_e4m3(u); 
+    }
+};
+template<> struct convertor<float4, fp8e4m3_4> {
+    static __host__ __device__ inline float4 convert(const fp8e4m3_4& u) {
+        __nv_fp8_e4m3 *vals = reinterpret_cast<__nv_fp8_e4m3*>(const_cast<__nv_fp8x4_e4m3*>(&u));
+        return make_float4(float(vals[0]), float(vals[1]), float(vals[2]), float(vals[3]));
+    }
+};
+template<> struct convertor<fp8e4m3_2, float2> {
+    static __host__ __device__ inline fp8e4m3_2 convert(const float2& u) {
+        return __nv_fp8x2_e4m3(u); 
+    }
+};
+template<> struct convertor<float2, fp8e4m3_2> {
+    static __host__ __device__ inline float2 convert(const fp8e4m3_2& u) {
+        __nv_fp8_e4m3 *vals = reinterpret_cast<__nv_fp8_e4m3*>(const_cast<__nv_fp8x2_e4m3*>(&u));
+        return make_float2(float(vals[0]), float(vals[1]));
+    }
+};
+template<> struct convertor<fp8e4m3, float> {
+    static __host__ __device__ inline fp8e4m3 convert(const float & u) {
+        return __nv_fp8_e4m3(u);
+    }
+};
+template<> struct convertor<float, fp8e4m3> {
+    static __host__ __device__ inline float convert(const fp8e4m3 & u) {
+        return float(u);
+    }
+};
+template<> struct convertor<bf16_2, fp8e4m3_4> {
+    static __host__ __device__ inline bf16_2 convert(const fp8e4m3_4 & u) {
+        float4 f4 = convertor<float4, fp8e4m3_4>::convert(u);
+        float2 f2 = make_float2(f4.x, f4.y);
+        return __float22bfloat162_rn(f2);
+    }
+};
+template<> struct convertor<fp8e4m3_4, bf16_2> {
+    static __host__ __device__ inline fp8e4m3_4 convert(const bf16_2 & u) {
+        float2 f2 = __bfloat1622float2(u);
+        float4 f4 = make_float4(f2.x, f2.y, 0.0f, 0.0f);
+        return __nv_fp8x4_e4m3(f4);
+    }
+};
+// fp8e5m2
+template<> struct convertor<fp8e5m2_4, float4> {
+    static __host__ __device__ inline fp8e5m2_4 convert(const float4& u) {
+        return __nv_fp8x4_e5m2(u); 
+    }
+};
+template<> struct convertor<float4, fp8e5m2_4> {
+    static __host__ __device__ inline float4 convert(const fp8e5m2_4& u) {
+        __nv_fp8_e5m2 *vals = reinterpret_cast<__nv_fp8_e5m2*>(const_cast<__nv_fp8x4_e5m2*>(&u));
+        return make_float4(float(vals[0]), float(vals[1]), float(vals[2]), float(vals[3]));
+    }
+};
+template<> struct convertor<fp8e5m2_2, float2> {
+    static __host__ __device__ inline fp8e5m2_2 convert(const float2& u) {
+        return __nv_fp8x2_e5m2(u); 
+    }
+};
+template<> struct convertor<float2, fp8e5m2_2> {
+    static __host__ __device__ inline float2 convert(const fp8e5m2_2& u) {
+        __nv_fp8_e5m2 *vals = reinterpret_cast<__nv_fp8_e5m2*>(const_cast<__nv_fp8x2_e5m2*>(&u));
+        return make_float2(float(vals[0]), float(vals[1]));
+    }
+};
+template<> struct convertor<fp8e5m2, float> {
+    static __host__ __device__ inline fp8e5m2 convert(const float & u) {
+        return __nv_fp8_e5m2(u);
+    }
+};
+template<> struct convertor<float, fp8e5m2> {
+    static __host__ __device__ inline float convert(const fp8e5m2 & u) {
+        return float(u);
+    }
+};
+template<> struct convertor<bf16_2, fp8e5m2_4> {
+    static __host__ __device__ inline bf16_2 convert(const fp8e5m2_4 & u) {
+        float4 f4 = convertor<float4, fp8e5m2_4>::convert(u);
+        float2 f2 = make_float2(f4.x, f4.y);
+        return __float22bfloat162_rn(f2);
+    }
+};
+template<> struct convertor<fp8e5m2_4, bf16_2> {
+    static __host__ __device__ inline fp8e5m2_4 convert(const bf16_2 & u) {
+        float2 f2 = __bfloat1622float2(u);
+        float4 f4 = make_float4(f2.x, f2.y, 0.0f, 0.0f);
+        return __nv_fp8x4_e5m2(f4);
+    }
+};
+#endif
+}
+}
@@ -0,0 +1,11 @@
+/**
+ * @file
+ * @brief A collection of common resources on which ThunderKittens depends.
+ */
+ 
+
+#pragma once
+
+#include "util.cuh"
+#include "base_types.cuh"
+#include "base_ops.cuh"
@@ -0,0 +1,56 @@
+#pragma once
+
+// Reset
+#define TK_RESET "\033[0m"
+
+// Foreground colors
+#define TK_FG_BLACK "\033[30m"
+#define TK_FG_RED "\033[31m"
+#define TK_FG_GREEN "\033[32m"
+#define TK_FG_YELLOW "\033[33m"
+#define TK_FG_BLUE "\033[34m"
+#define TK_FG_MAGENTA "\033[35m"
+#define TK_FG_CYAN "\033[36m"
+#define TK_FG_WHITE "\033[37m"
+
+// Background colors
+#define TK_BG_BLACK "\033[40m"
+#define TK_BG_RED "\033[41m"
+#define TK_BG_GREEN "\033[42m"
+#define TK_BG_YELLOW "\033[43m"
+#define TK_BG_BLUE "\033[44m"
+#define TK_BG_MAGENTA "\033[45m"
+#define TK_BG_CYAN "\033[46m"
+#define TK_BG_WHITE "\033[47m"
+
+// Bright foreground colors
+#define TK_FG_BRIGHT_BLACK "\033[90m"
+#define TK_FG_BRIGHT_RED "\033[91m"
+#define TK_FG_BRIGHT_GREEN "\033[92m"
+#define TK_FG_BRIGHT_YELLOW "\033[93m"
+#define TK_FG_BRIGHT_BLUE "\033[94m"
+#define TK_FG_BRIGHT_MAGENTA "\033[95m"
+#define TK_FG_BRIGHT_CYAN "\033[96m"
+#define TK_FG_BRIGHT_WHITE "\033[97m"
+
+// Bright background colors
+#define TK_BG_BRIGHT_BLACK "\033[100m"
+#define TK_BG_BRIGHT_RED "\033[101m"
+#define TK_BG_BRIGHT_GREEN "\033[102m"
+#define TK_BG_BRIGHT_YELLOW "\033[103m"
+#define TK_BG_BRIGHT_BLUE "\033[104m"
+#define TK_BG_BRIGHT_MAGENTA "\033[105m"
+#define TK_BG_BRIGHT_CYAN "\033[106m"
+#define TK_BG_BRIGHT_WHITE "\033[107m"
+
+// Text styles
+#define TK_BOLD "\033[1m"
+#define TK_DIM "\033[2m"
+#define TK_ITALIC "\033[3m"
+#define TK_UNDERLINE "\033[4m"
+#define TK_BLINK "\033[5m"
+#define TK_REVERSE "\033[7m"
+#define TK_HIDDEN "\033[8m"
+
+// Macro to combine styles
+#define TK_STYLE(...)  "\033[" #__VA_ARGS__ "m"
@@ -0,0 +1,314 @@
+/**
+ * @file
+ * @brief General utilities for ThunderKittens.
+ */
+
+#pragma once
+
+#include <stdint.h>
+#include <type_traits>
+#include <concepts>
+#include <memory>
+
+// CUDA driver API
+#define CUCHECK(cmd) do {                                     \
+    CUresult err = cmd;                                       \
+    if (err != CUDA_SUCCESS) {                                \
+        const char *errStr;                                   \
+        cuGetErrorString(err, &errStr);                       \
+        fprintf(stderr, "Failed: CUDA error %s:%d '%s'\n",    \
+            __FILE__, __LINE__, errStr);                      \
+        exit(EXIT_FAILURE);                                   \
+    }                                                         \
+} while(0)
+
+// CUDA runtime API
+#define CUDACHECK(cmd) do {                                   \
+    cudaError_t err = cmd;                                    \
+    if (err != cudaSuccess) {                                 \
+        fprintf(stderr, "Failed: CUDA error %s:%d '%s'\n",    \
+            __FILE__, __LINE__, cudaGetErrorString(err));     \
+        exit(EXIT_FAILURE);                                   \
+    }                                                         \
+} while(0)
+
+/**
+ * @namespace kittens
+ *
+ * @brief The main namespace of ThunderKittens.
+ */
+namespace kittens {
+
+/* ----------  GENERAL CONSTANTS FOR KITTENS  ---------- */
+
+/**
+ * @brief Tile dimension constant.
+ */
+template<typename T> constexpr int TILE_COL_DIM = sizeof(T) == 1 ? 32 : 16;
+template<typename T> constexpr int TILE_ROW_DIM = 16;
+/**
+ * @brief Tile num elements constant calculated as TILE_DIM squared.
+ */
+template<typename T> constexpr int TILE_ELEMENTS{TILE_COL_DIM<T>*TILE_ROW_DIM<T>};
+/**
+ * @brief Constant representing number of threads in a warp.
+ */
+constexpr int WARP_THREADS{32};
+/**
+ * @brief Constant representing number of threads in a warpgroup of four warps.
+ */
+constexpr int WARPGROUP_THREADS{128};
+/**
+
+ * @brief Constant representing number of warps in a warpgroup of four warps.
+ */
+constexpr int WARPGROUP_WARPS{4};
+/**
+
+ * @brief Get the warp ID of the current thread.
+ * @return The warp ID.
+ */
+__device__ static __forceinline__ int warpid() {
+    // uint32_t wid;
+    // asm volatile("mov.u32 %0, %warpid;" : "=r"(wid));
+    // return wid;
+    return threadIdx.x >> 5;
+}
+/**
+ * @brief Get the warpgroup ID of the current thread.
+ * @return The warpgroup ID.
+ */
+__device__ static __forceinline__ int warpgroupid() { return warpid() >> 2; }
+/**
+ * @brief Get the lane ID of the current thread within its warp.
+ * @return The lane ID.
+ */
+__device__ static __forceinline__ int laneid() {
+    // uint32_t lid;
+    // asm volatile("mov.u32 %0, %laneid;" : "=r"(lid));
+    // return lid;
+    return threadIdx.x & 31;
+}
+
+#if defined(KITTENS_HOPPER)
+constexpr int MAX_SHARED_MEMORY = 227000;
+#elif defined(KITTENS_A100)
+constexpr int MAX_SHARED_MEMORY = 164000;
+#elif defined(KITTENS_4090)
+constexpr int MAX_SHARED_MEMORY = 100000;
+#endif
+
+struct transpose {
+    static constexpr int N = 0; // not transposed
+    static constexpr int T = 1; // transposed
+};
+struct axis {
+    static constexpr int ROW = 0; // row axis of a tile
+    static constexpr int COL = 1; // column axis of a tile
+};
+
+/* ----------  TYPE HELPERS  ---------- */
+
+/**
+ * @namespace ducks
+ *
+ * @brief ThunderKittens' namespace for template metaprogramming..
+ * 
+ * This includes primarily dummy types and concept wrappers, along
+ * with a few additional utilities.
+ */
+namespace ducks {
+
+/**
+ * @brief A type representing an empty default for a template.
+ */
+struct default_type {};
+
+// This macro can't be done as a template, so it doesn't really have a location in kittens.
+#define typeof(A) typename std::remove_const<typename std::remove_reference<decltype(A)>::type>::type
+
+}
+
+/* ----------  SHUFFLE UTILS  ---------- */
+
+/**
+ * @brief Mask constant for all active threads in a warp.
+ */
+static constexpr uint32_t MASK_ALL = 0xFFFFFFFF;
+
+/**
+ * @brief Perform a shuffle down operation on a packed type synchronously across a warp.
+ * @tparam T The type of the value to be shuffled.
+ * @param mask[in] The mask of active threads.
+ * @param f[in] The value to be shuffled.
+ * @param delta[in] The number of positions to shuffle down.
+ * @return The result of the shuffle operation.
+ */
+template<typename T>
+__device__ static inline T packed_shfl_down_sync(uint32_t mask, const T &f, int delta) {
+    return __shfl_down_sync(mask, f, delta);
+}
+template<>
+__device__ inline float2 packed_shfl_down_sync<float2>(uint32_t mask, const float2 &f, int delta) {
+    float2 r;
+    r.x = __shfl_down_sync(mask, f.x, delta);
+    r.y = __shfl_down_sync(mask, f.y, delta);
+    return r;
+}
+/**
+ * @brief Perform a packed shuffle operation synchronously across a warp.
+ * @tparam T The type of the value to be shuffled.
+ * @param mask[in] The mask of active threads.
+ * @param f[in] The value to be shuffled.
+ * @param src[in] The source lane from which to shuffle.
+ * @return The result of the shuffle operation.
+ */
+template<typename T>
+__device__ static inline T packed_shfl_sync(uint32_t mask, const T &f, int src) {
+    return __shfl_sync(mask, f, src);
+}
+template<>
+__device__ inline float2 packed_shfl_sync<float2>(uint32_t mask, const float2 &f, int src) {
+    float2 r;
+    r.x = __shfl_sync(mask, f.x, src);
+    r.y = __shfl_sync(mask, f.y, src);
+    return r;
+}
+
+/* ----------  SHARED MEMORY UTILS  ---------- */
+
+// namespace ducks {
+// namespace sb {
+// struct identifier {};
+// }
+// }
+
+// template<typename Args...>
+// struct sb {
+//     using identifier = ducks::sb::identifier;
+//     Args... args;
+// };
+
+// namespace ducks {
+// namespace sb {
+// template<typename T> concept all = requires {
+//     typename T::identifier;
+// } && std::is_same_v<T::identifier, identifier>;
+// }
+// }
+
+// Joyously stolen from https://github.com/NVIDIA/cutlass/blob/5c447dd84f8ae0e1d48ff9a2eae26ce8c4958101/include/cute/container/alignment.hpp#L51
+#if defined(__CUDACC__)
+#define KITTENS_ALIGN_AS(n) __align__(n)
+#else
+#define KITTENS_ALIGN_AS(n) alignas(n)
+#endif
+
+#ifdef KITTENS_HOPPER
+#define KITTENS_DEFAULT_ALIGN KITTENS_ALIGN_AS(128)
+#else
+#define KITTENS_DEFAULT_ALIGN KITTENS_ALIGN_AS(16)
+#endif
+
+/**
+ * @brief Dummy structure for alignment purposes. Needed for WGMMA and TMA calls.
+ */
+struct KITTENS_DEFAULT_ALIGN alignment_dummy { int dummy; };
+/**
+ * @brief Very simple allocator for dynamic shared memory. Advances pointer and tracks alignments.
+ * @tparam default_alignment The default alignment this allocator will enforce. If <=0 (default -1) it will not align.
+ */
+#ifdef KITTENS_HOPPER
+template<int default_alignment=1024> 
+#else
+template<int default_alignment=16> 
+#endif
+struct shared_allocator {
+    int *ptr;
+
+    private:
+        // Recursive template to generate N-dimensional array type
+        template<typename A, size_t... dims>
+        struct variadic_array;
+        template<typename A, size_t first_dim, size_t... rest_dims>
+        struct variadic_array<A, first_dim, rest_dims...> {
+            using type = typename variadic_array<A, rest_dims...>::type[first_dim];
+        };
+        template<typename A>
+        struct variadic_array<A> {
+            using type = A;
+        };
+        template<typename A, size_t... dims> 
+        using variadic_array_t = typename variadic_array<A, dims...>::type;
+
+        template<int alignment>
+        __device__ inline void align_ptr() {
+            if constexpr (alignment > 0) {
+                uint64_t p = reinterpret_cast<uint64_t>(ptr);
+                if(p % alignment != 0) {
+                    ptr = (int*)(p + (alignment-(p%alignment)));
+                }
+            }
+        }
+
+    public:
+        /**
+        * @brief Construct a new shared allocator using a pointer to extern shared memory.
+        * @param[in] _ptr Pointer to the start of the extern shared memory.
+        */
+        __device__ shared_allocator(int *_ptr): ptr(_ptr) {}
+        /**
+        * @brief Allocate shared memory for a single instance or N-dimensional array of type A.
+        * @tparam A The type of the object to allocate.
+        * @tparam dims... A list of dimensions for the N-dimensional array.
+        * @return Reference to the allocated object.
+        */
+        template<typename A, size_t... dims> 
+        __device__ inline variadic_array_t<A, dims...>& allocate() {
+            // static_assert(sizeof(A) % default_alignment == 0, "Type is not aligned properly for array allocation");
+            align_ptr<default_alignment>();
+            using at = variadic_array_t<A, dims...>;
+            at*p = reinterpret_cast<at*>(ptr);
+            ptr += sizeof(at)/sizeof(int);
+            return *p;
+        }
+        /**
+        * @brief Allocate shared memory for a single instance or N-dimensional array of type A.
+        * @tparam alignment An alignment to enforce for this particular object.
+        * @tparam A The type of the object to allocate.
+        * @tparam dims... A list of dimensions for the N-dimensional array.
+        * @return Reference to the allocated object.
+        */
+        template<int alignment, typename A, size_t... dims> 
+        __device__ inline variadic_array_t<A, dims...>& allocate() {
+            // static_assert(sizeof(A) % alignment == 0, "Type is not aligned properly for array allocation");
+            align_ptr<alignment>();
+            using at = variadic_array_t<A, dims...>;
+            at*p = reinterpret_cast<at*>(ptr);
+            ptr += sizeof(at)/sizeof(int);
+            return *p;
+        }
+};
+#if (defined(KITTENS_HOPPER) || defined(KITTENS_BLACKWELL))
+/**
+ * @brief A wrapper for an allocator that enforces sufficient alignment to be used for TMA loads and stores.
+ */
+using tma_allocator = shared_allocator<1024>;
+using tma_swizzle_allocator = tma_allocator; // swizzled TMA modes require up to 1024 byte alignments :/
+
+/* Get CTA ID within a cluster */
+__device__ static inline int3 clusterIdx() {
+    int3 cluster_idx;
+    asm volatile("mov.u32 %0, %clusterid.x;\n" : "=r"(cluster_idx.x));
+    asm volatile("mov.u32 %0, %clusterid.y;\n" : "=r"(cluster_idx.y));
+    asm volatile("mov.u32 %0, %clusterid.z;\n" : "=r"(cluster_idx.z));
+    return cluster_idx;
+}
+__device__ static inline int cluster_ctarank() {
+    uint32_t ctarank;
+    asm volatile("mov.u32 %0, %cluster_ctarank;\n" : "=r"(ctarank));
+    return ctarank;
+}
+#endif
+
+} // namespace kittens
@@ -0,0 +1,12 @@
+/**
+ * @file
+ * @brief The master header file of ThunderKittens. This file includes everything you need!
+ */
+
+#pragma once
+
+#include "common/common.cuh"
+#include "types/types.cuh"
+#include "ops/ops.cuh"
+#include "pyutils/util.cuh"
+// #include "pyutils/pyutils.cuh" // for simple binding without including torch
@@ -0,0 +1,51 @@
+/**
+ * @file
+ * @brief An aggregate header of all device (multi-GPU) operations defined by ThunderKittens
+ */
+
+#pragma once
+
+#include "../../types/types.cuh"
+
+namespace kittens {
+
+template<int _NUM_DEVICES>
+struct device {
+
+static_assert(_NUM_DEVICES >= 0 && _NUM_DEVICES <= 72, "Invalid number of devices");
+static constexpr int NUM_DEVICES = _NUM_DEVICES;
+
+#ifdef KITTENS_HOPPER
+
+using barrier_t = pgl<gl<int, 1, 1, 1, -1>, NUM_DEVICES, true>;
+
+/**
+ * @brief Multi-GPU synchronization barrier for coordinated kernel exit
+ * 
+ * Performs a synchronization across all devices to ensure all GPUs complete 
+ * their work before any kernel exits. Does not synchronize intra-node threads
+ * or threadblocks.
+ * 
+ * @param barrier Pre-allocated barrier structure, must be initialized to 0
+ * @param dev_idx Current device index (0 to NUM_DEVICES - 1)
+ * @param id Synchronization point identifier (default: 0). 0 is fine for most cases
+ * 
+ */
+__device__ static inline void sync_on_exit(const barrier_t &barrier, const int dev_idx, const int id = 0) {
+    if (blockIdx.x == 0 && blockIdx.y == 0 && blockIdx.z == 0 &&
+        threadIdx.x == 0 && threadIdx.y == 0 && threadIdx.z == 0) {
+        cuda::atomic_ref<int, cuda::thread_scope_system> barrier_uc(barrier[dev_idx][{id}]);
+
+        // Inter-note check-in
+        multimem<int>::red<reduce_op::ADD>(barrier.mc_ptr_at({id}), 1);
+        asm volatile ("{fence.proxy.alias;}" ::: "memory");
+        while (barrier_uc.load(cuda::memory_order_acquire) < NUM_DEVICES);
+        barrier_uc.fetch_sub(NUM_DEVICES, cuda::memory_order_release);
+    }
+}
+
+#endif
+
+};
+
+} // namespace kittens
@@ -0,0 +1,96 @@
+/**
+ * @file
+ * @brief An aggregate header of all group (multi-warp) operations defined by ThunderKittens
+ */
+
+#pragma once
+
+#include <cuda/pipeline>
+
+#include "../../common/common.cuh"
+#include "../../types/types.cuh"
+#include "../thread/thread.cuh" // several group memory ops rely on underlying warp-scope ops
+
+#define KITTENS_CHECK_WARP static_assert(GROUP_WARPS==1, "Warp (GROUP_WARPS=1) function called from a non-warp group.");
+// A "warpgroup" is a special group of 4 consecutive warps defined by NVIDIA for certain SM_90+ operations.
+#define KITTENS_CHECK_WARPGROUP static_assert(GROUP_WARPS==4, "Warpgroup (GROUP_WARPS=4) function called from a non-warpgroup group.");
+
+// WGMMA relies on some template structures that cannot be specialized within the group struct, so we declare them in advance.
+#ifdef KITTENS_HOPPER
+#include "mma/warpgroup/base/base.cuh"
+#endif
+
+namespace kittens {
+/*
+This is meant to be used with a `using group_N = kittens::group<NUM_WORKERS>;` at the start of every kernel.
+*/
+template<int _GROUP_WARPS>
+struct group {
+static constexpr int GROUP_WARPS = _GROUP_WARPS; // This alias produces nice parallelism.
+static constexpr int GROUP_THREADS = GROUP_WARPS * kittens::WARP_THREADS; // This alias produces nice parallelism.
+__device__ static inline int laneid() { return threadIdx.x % GROUP_THREADS; }
+__device__ static inline int warpid() { return laneid() / kittens::WARP_THREADS; }
+__device__ static inline int groupid() { return threadIdx.x / GROUP_THREADS; }
+
+__device__ static inline void sync(int id) {
+    asm volatile("bar.sync %0, %1;\n" :: "r"(id), "n"(GROUP_THREADS));
+}
+template<uint32_t MASK=0xFFFFFFFF> __device__ static inline void sync() {
+    static_assert(GROUP_WARPS==1, "barrier-less sync() can only be called by a single warp!");
+    asm volatile("bar.warp.sync %0;\n" :: "n"(MASK));
+}
+__device__ static inline void arrive(int id) {
+    asm volatile("bar.arrive %0, %1;\n" :: "r"(id), "n"(GROUP_THREADS));
+}
+
+#include "memory/memory.cuh"
+#include "shared/shared.cuh"
+#include "register/register.cuh"
+
+#ifdef KITTENS_HOPPER
+#include "mma/mma.cuh"
+
+template<int n_reg> __device__ static inline void increase_registers() {
+    static_assert(n_reg % 8 == 0, "n_reg must be a multiple of 8");
+    asm volatile("setmaxnreg.inc.sync.aligned.u32 %0;\n" :: "n"(n_reg));
+}
+template<int n_reg> __device__ static inline void decrease_registers() {
+    static_assert(n_reg % 8 == 0, "n_reg must be a multiple of 8");
+    asm volatile("setmaxnreg.dec.sync.aligned.u32 %0;\n" :: "n"(n_reg));
+}
+__device__ static inline void producer_registers() { decrease_registers<24>(); }
+template<int NCWG> __device__ static inline void consumer_registers() { increase_registers<480/NCWG - 8*(NCWG>3) - 224*(NCWG==1)>(); }
+
+#endif
+
+};
+
+namespace everyone {
+
+// Block-level synchronization
+__device__ static inline void sync(int id) {
+    asm volatile("bar.sync %0;\n" :: "r"(id));
+}
+
+// Cluster-level synchronization functions
+namespace tma {
+namespace cluster {
+__device__ static inline void arrive_aligned() { // All threads in the cluster must call this
+    asm volatile ("barrier.cluster.arrive.release.aligned;\n");
+}
+__device__ static inline void wait_aligned() {
+    asm volatile ("barrier.cluster.wait.acquire.aligned;\n");
+}
+__device__ static inline void sync() {
+    arrive_aligned();
+    wait_aligned();
+}
+}
+}
+
+};
+
+using warp = group<1>;      // scope used by most pre-Hopper GPUs, and also for most register operations.
+using warpgroup = group<4>; // special scope commonly used by Hopper and later.
+
+}
@@ -0,0 +1,21 @@
+/**
+ * @file
+ * @brief An aggregate header of colaborative group memory movement operations
+ */
+
+#include "util/util.cuh"
+#include "tile/tile.cuh"
+#include "vec/vec.cuh"
+
+#ifdef KITTENS_HOPPER
+struct tma {
+#include "util/tma.cuh"
+#include "tile/tma.cuh"
+#include "vec/tma.cuh"
+struct cluster {
+#include "util/tma_cluster.cuh"
+#include "tile/tma_cluster.cuh"
+#include "vec/tma_cluster.cuh"
+};
+};
+#endif
@@ -0,0 +1,42 @@
+/**
+ * @file
+ * @brief Functions for a group to collaboratively transfer data directly between global memory and registers and back.
+ */
+
+/**
+ * @brief Collaboratively loads data from a source array into register tiles.
+ *
+ * @tparam RT The register tile type.
+ * @tparam U The data type of the source array.
+ * @param dst[out] The destination tile to load data into.
+ * @param src[in] The source array to load data from.
+ * @param row_stride[in] The stride in elements between rows in the source array.
+ */
+template<int axis, ducks::crt::all CRT, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<crt<typename CRT::T, GROUP_WARPS*CRT::rows, CRT::cols, typename CRT::layout>>>
+__device__ inline static void load(CRT &dst, const CGL &src, const COORD &idx) {
+    load<axis, CRT::component, CGL::component, COORD>(dst.real, src.real, idx);
+    load<axis, CRT::component, CGL::component, COORD>(dst.imag, src.imag, idx);
+}
+template<ducks::crt::all CRT, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<crt<typename CRT::T, GROUP_WARPS*CRT::rows, CRT::cols, typename CRT::layout>>>
+__device__ inline static void load(CRT &dst, const CGL &src, const COORD &idx) {
+    load<2, CRT, CGL>(dst, src, idx);
+}
+
+/**
+ * @brief Collaboratively stores data from register tiles to a destination array in global memory.
+ *
+ * @tparam RT The register tile type.
+ * @tparam U The data type of the destination array.
+ * @param[out] dst The destination array in global memory to store data into.
+ * @param[in] src The source register tile to store data from.
+ * @param row_stride[in] The stride in elements between rows in the destination array.
+ */
+template<int axis, ducks::crt::all CRT, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<crt<typename CRT::T, GROUP_WARPS*CRT::rows, CRT::cols, typename CRT::layout>>>
+__device__ inline static void store(CGL &dst, const CRT &src, const COORD &idx) {
+    store<axis, typename CRT::component, typename CGL::component>(dst.real, src.real, idx);
+    store<axis, typename CRT::component, typename CGL::component>(dst.imag, src.imag, idx);
+}
+template<ducks::crt::all CRT, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<crt<typename CRT::T, GROUP_WARPS*CRT::rows, CRT::cols, typename CRT::layout>>>
+__device__ inline static void store(CGL &dst, const CRT &src, const COORD &idx) {
+    store<2, CRT, CGL>(dst, src, idx);
+}
@@ -0,0 +1,37 @@
+/**
+ * @file
+ * @brief Group (collaborative warp) ops for loading shared tiles from and storing to global memory. 
+ */
+
+template<int axis, bool assume_aligned, ducks::cst::all CST, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<CST>>
+__device__ static inline void load(CST &dst, const CGL &src, const COORD &idx) {
+    load<axis, assume_aligned, typename CST::component, typename CGL::component, COORD>(dst.real, src.real, idx);
+    load<axis, assume_aligned, typename CST::component, typename CGL::component, COORD>(dst.imag, src.imag, idx);
+}
+template<ducks::cst::all CST, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<CST>>
+__device__ static inline void load(CST &dst, const CGL &src, const COORD &idx) {
+    load<2, false, typename CST::component, typename CGL::component, COORD>(dst.real, src.real, idx);
+    load<2, false, typename CST::component, typename CGL::component, COORD>(dst.imag, src.imag, idx);
+}
+
+template<int axis, bool assume_aligned, ducks::cst::all CST, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<CST>>
+__device__ static inline void store(CGL &dst, const CST &src, const COORD &idx) {
+    store<axis, assume_aligned, typename CST::component, typename CGL::component, COORD>(dst.real, src.real, idx);
+    store<axis, assume_aligned, typename CST::component, typename CGL::component, COORD>(dst.imag, src.imag, idx);
+}
+template<ducks::cst::all CST, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<CST>>
+__device__ static inline void store(CGL &dst, const CST &src, const COORD &idx) {
+    store<2, false, typename CST::component, typename CGL::component, COORD>(dst.real, src.real, idx);
+    store<2, false, typename CST::component, typename CGL::component, COORD>(dst.imag, src.imag, idx);
+}
+
+template<int axis, bool assume_aligned, ducks::cst::all CST, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<CST>>
+__device__ static inline void load_async(CST &dst, const CGL &src, const COORD &idx) {
+    load_async<axis, assume_aligned, typename CST::component, typename CGL::component, COORD>(dst.real, src.real, idx);
+    load_async<axis, assume_aligned, typename CST::component, typename CGL::component, COORD>(dst.imag, src.imag, idx);
+}
+template<ducks::cst::all CST, ducks::cgl::all CGL, ducks::coord::tile COORD=coord<CST>>
+__device__ static inline void load_async(CST &dst, const CGL &src, const COORD &idx) {
+    load_async<2, false, typename CST::component, typename CGL::component, COORD>(dst.real, src.real, idx);
+    load_async<2, false, typename CST::component, typename CGL::component, COORD>(dst.imag, src.imag, idx);
+}
@@ -0,0 +1,34 @@
+/**
+ * @file
+ * @brief Functions for a warpgroup to collaboratively transfer data directly between shared memory and registers and back.
+ */
+
+/**
+ * @brief Collaboratively load data from a shared tile into register tiles split across a warpgroup.
+ *
+ * @tparam RT The register tile type
+ * @tparam ST The shared tile type
+ * @param dst[out] The destination register tile.
+ * @param src[in]  The source shared tile.
+ */
+template<ducks::crt::all RT, ducks::cst::all ST>
+__device__ inline static void load(RT &dst, const ST &src) {
+    load(dst.real, src.real);
+    load(dst.imag, src.imag);
+}
+
+
+/**
+ * @brief Collaboratively store data into a shared tile from register tiles split across a warpgroup.
+ *
+ * @tparam RT The register tile type
+ * @tparam ST The shared tile type
+ * @param dst[out] The destination shared tile.
+ * @param src[in]  The source register tile.
+ */
+template<ducks::cst::all ST, ducks::crt::all RT>
+__device__ inline static void store(ST &dst, const RT &src) {
+    store(dst.real, src.real);
+    store(dst.imag, src.imag);
+}
+
@@ -0,0 +1,207 @@
+/**
+ * @file
+ * @brief Functions for a group to collaboratively transfer data directly between global memory and registers and back.
+ */
+
+/**
+ * @brief Collaboratively loads data from a source array into row-major layout tiles.
+ *
+ * @tparam RT The row-major layout tile type.
+ * @tparam U The data type of the source array.
+ * @param dst[out] The destination tile to load data into.
+ * @param src[in] The source array to load data from.
+ * @param row_stride[in] The stride in elements between rows in the source array.
+ */
+template<int axis, ducks::rt::row_layout RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<rt<typename RT::T, GROUP_WARPS*RT::rows, RT::cols, typename RT::layout>>>
+__device__ inline static void load(RT &dst, const GL &src, const COORD &idx) {
+    using T2 = RT::dtype;
+    using U = typename GL::dtype;
+
+    #ifdef KITTENS_HOPPER
+    static_assert(!std::is_same_v<T2, fp8e4m3_4> && !std::is_same_v<T2, fp8e5m2_4>, "Unsupported type for load/store");
+    #endif
+
+    U *src_ptr = (U*)&src[(idx.template unit_coord<axis, 3>())];
+    const int row_stride = src.template stride<axis>();
+    using U2 = base_types::packing<U>::packed_type;
+    int warp_laneid = threadIdx.x % WARP_THREADS;
+    int local_warpid;
+    if constexpr(GROUP_WARPS % 4 == 0) local_warpid = (warpid()/4+(warpid()%4)*(GROUP_WARPS/4));
+    else local_warpid = warpid();
+    const int row_offset = dst.rows*local_warpid;
+    #pragma unroll
+    for(int i = 0; i < dst.height; i++) {
+        int row = row_offset + i*dst.tile_size_row + (warp_laneid / 4);
+        #pragma unroll
+        for(int j = 0; j < dst.width; j++) {
+            int col = j*dst.tile_size_col + 2*(warp_laneid % 4);
+            dst.tiles[i][j].data[0] = base_types::convertor<T2, U2>::convert(*(U2*)(&src_ptr[(row+0)*row_stride + (col+0)]));
+            dst.tiles[i][j].data[2] = base_types::convertor<T2, U2>::convert(*(U2*)(&src_ptr[(row+0)*row_stride + (col+8)]));
+        }
+        #pragma unroll
+        for(int j = 0; j < dst.width; j++) {
+            int col = j*dst.tile_size_col + 2*(warp_laneid % 4);
+            dst.tiles[i][j].data[1] = base_types::convertor<T2, U2>::convert(*(U2*)(&src_ptr[(row+8)*row_stride + (col+0)]));
+            dst.tiles[i][j].data[3] = base_types::convertor<T2, U2>::convert(*(U2*)(&src_ptr[(row+8)*row_stride + (col+8)]));
+        }
+    }
+}
+/**
+ * @brief Collaboratively loads data from a source array into column-major layout tiles.
+ *
+ * @tparam RT The column-major layout tile type.
+ * @tparam U The data type of the source array.
+ * @param dst[out] The destination tile to load data into.
+ * @param src[in] The source array to load data from.
+ * @param row_stride[in] The stride in elements between rows in the source array.
+ */
+template<int axis, ducks::rt::col_layout RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<rt<typename RT::T, GROUP_WARPS*RT::rows, RT::cols, typename RT::layout>>>
+__device__ inline static void load(RT &dst, const GL &src, const COORD &idx) {
+    using T = typename RT::T;
+    using U = typename GL::dtype;
+
+    #ifdef KITTENS_HOPPER
+    static_assert(!std::is_same_v<T, fp8e4m3> && !std::is_same_v<T, fp8e5m2>, "Unsupported type for load/store");
+    #endif
+
+    U *src_ptr = (U*)&src[(idx.template unit_coord<axis, 3>())];
+    const int row_stride = src.template stride<axis>();
+    int warp_laneid = threadIdx.x % WARP_THREADS;
+    int local_warpid;
+    if constexpr(GROUP_WARPS % 4 == 0) local_warpid = (warpid()/4+(warpid()%4)*(GROUP_WARPS/4));
+    else local_warpid = warpid();
+    const int row_offset = dst.rows*local_warpid;
+    #pragma unroll
+    for(int i = 0; i < dst.height; i++) {
+        int row = row_offset + i*dst.tile_size_row + 2*(warp_laneid % 4);
+        #pragma unroll
+        for(int j = 0; j < dst.width; j++) {
+            int col = j*dst.tile_size_col + (warp_laneid / 4);
+            dst.tiles[i][j].data[0].x = base_types::convertor<T, U>::convert(src_ptr[(row+0)*row_stride + (col+0)]);
+            dst.tiles[i][j].data[1].x = base_types::convertor<T, U>::convert(src_ptr[(row+0)*row_stride + (col+8)]);
+        }
+        #pragma unroll
+        for(int j = 0; j < dst.width; j++) {
+            int col = j*dst.tile_size_col + (warp_laneid / 4);
+            dst.tiles[i][j].data[0].y = base_types::convertor<T, U>::convert(src_ptr[(row+1)*row_stride + (col+0)]);
+            dst.tiles[i][j].data[1].y = base_types::convertor<T, U>::convert(src_ptr[(row+1)*row_stride + (col+8)]);
+        }
+        #pragma unroll
+        for(int j = 0; j < dst.width; j++) {
+            int col = j*dst.tile_size_col + (warp_laneid / 4);
+            dst.tiles[i][j].data[2].x = base_types::convertor<T, U>::convert(src_ptr[(row+8)*row_stride + (col+0)]);
+            dst.tiles[i][j].data[3].x = base_types::convertor<T, U>::convert(src_ptr[(row+8)*row_stride + (col+8)]);
+        }
+        #pragma unroll
+        for(int j = 0; j < dst.width; j++) {
+            int col = j*dst.tile_size_col + (warp_laneid / 4);
+            dst.tiles[i][j].data[2].y = base_types::convertor<T, U>::convert(src_ptr[(row+9)*row_stride + (col+0)]);
+            dst.tiles[i][j].data[3].y = base_types::convertor<T, U>::convert(src_ptr[(row+9)*row_stride + (col+8)]);
+        }
+    }
+}
+template<ducks::rt::all RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<rt<typename RT::T, GROUP_WARPS*RT::rows, RT::cols, typename RT::layout>>>
+__device__ inline static void load(RT &dst, const GL &src, const COORD &idx) {
+    load<2>(dst, src, idx);
+}
+/**
+ * @brief Collaboratively stores data from register tiles to a destination array in global memory with a row-major layout.
+ *
+ * @tparam RT The register tile type with a row-major layout.
+ * @tparam U The data type of the destination array.
+ * @param[out] dst The destination array in global memory to store data into.
+ * @param[in] src The source register tile to store data from.
+ * @param row_stride[in] The stride in elements between rows in the destination array.
+ */
+template<int axis, ducks::rt::row_layout RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<rt<typename RT::T, GROUP_WARPS*RT::rows, RT::cols, typename RT::layout>>>
+__device__ inline static void store(const GL &dst, const RT &src, const COORD &idx) {
+    using T2 = RT::dtype;
+    using U = typename GL::dtype;
+
+    #ifdef KITTENS_HOPPER
+    static_assert(!std::is_same_v<T2, fp8e4m3_4> && !std::is_same_v<T2, fp8e5m2_4>, "Unsupported type for load/store");
+    #endif
+
+    U *dst_ptr = (U*)&dst[(idx.template unit_coord<axis, 3>())];
+    const int row_stride = dst.template stride<axis>();
+    using U2 = base_types::packing<U>::packed_type;
+    int warp_laneid = threadIdx.x % WARP_THREADS;
+    int local_warpid;
+    if constexpr(GROUP_WARPS % 4 == 0) local_warpid = (warpid()/4+(warpid()%4)*(GROUP_WARPS/4));
+    else local_warpid = warpid();
+    const int row_offset = src.rows*local_warpid;
+    #pragma unroll
+    for(int i = 0; i < src.height; i++) {
+        int row = row_offset + i*src.tile_size_row + (warp_laneid / 4);
+        #pragma unroll
+        for(int j = 0; j < src.width; j++) {
+            int col = j*src.tile_size_col + 2*(warp_laneid % 4);
+            *(U2*)(&dst_ptr[(row+0)*row_stride + (col+0)]) = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[0]);
+            *(U2*)(&dst_ptr[(row+0)*row_stride + (col+8)]) = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[2]);
+        }
+        #pragma unroll
+        for(int j = 0; j < src.width; j++) {
+            int col = j*src.tile_size_col + 2*(warp_laneid % 4);
+            *(U2*)(&dst_ptr[(row+8)*row_stride + (col+0)]) = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[1]);
+            *(U2*)(&dst_ptr[(row+8)*row_stride + (col+8)]) = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[3]);
+        }
+    }
+}
+/**
+ * @brief Collaboratively stores data from register tiles to a destination array in global memory with a column-major layout.
+ *
+ * @tparam RT The register tile type with a column-major layout.
+ * @tparam U The data type of the destination array.
+ * @param[out] dst The destination array in global memory to store data into.
+ * @param[in] src The source register tile to store data from.
+ * @param row_stride[in] The stride in elements between rows in the destination array.
+ */
+template<int axis, ducks::rt::col_layout RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<rt<typename RT::T, GROUP_WARPS*RT::rows, RT::cols, typename RT::layout>>>
+__device__ inline static void store(const GL &dst, const RT &src, const COORD &idx) {
+    using T = base_types::packing<typename RT::dtype>::unpacked_type;
+    using U = typename GL::dtype;
+
+    #ifdef KITTENS_HOPPER
+    static_assert(!std::is_same_v<T, fp8e4m3_4> && !std::is_same_v<T, fp8e5m2_4>, "Unsupported type for load/store");
+    #endif
+    
+    U *dst_ptr = (U*)&dst[(idx.template unit_coord<axis, 3>())];
+    const int row_stride = dst.template stride<axis>();
+    int warp_laneid = threadIdx.x % WARP_THREADS;
+    int local_warpid;
+    if constexpr(GROUP_WARPS % 4 == 0) local_warpid = (warpid()/4+(warpid()%4)*(GROUP_WARPS/4));
+    else local_warpid = warpid();
+    const int row_offset = src.rows*local_warpid;
+    #pragma unroll
+    for(int i = 0; i < src.height; i++) {
+        int row = row_offset + i*src.tile_size_row + 2*(warp_laneid % 4);
+        #pragma unroll
+        for(int j = 0; j < src.width; j++) {
+            int col = j*src.tile_size_col + (warp_laneid / 4);
+            dst_ptr[(row+0)*row_stride + (col+0)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[0].x);
+            dst_ptr[(row+0)*row_stride + (col+8)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[1].x);
+        }
+        #pragma unroll
+        for(int j = 0; j < src.width; j++) {
+            int col = j*src.tile_size_col + (warp_laneid / 4);
+            dst_ptr[(row+1)*row_stride + (col+0)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[0].y);
+            dst_ptr[(row+1)*row_stride + (col+8)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[1].y);
+        }
+        #pragma unroll
+        for(int j = 0; j < src.width; j++) {
+            int col = j*src.tile_size_col + (warp_laneid / 4);
+            dst_ptr[(row+8)*row_stride + (col+0)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[2].x);
+            dst_ptr[(row+8)*row_stride + (col+8)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[3].x);
+        }
+        #pragma unroll
+        for(int j = 0; j < src.width; j++) {
+            int col = j*src.tile_size_col + (warp_laneid / 4);
+            dst_ptr[(row+9)*row_stride + (col+0)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[2].y);
+            dst_ptr[(row+9)*row_stride + (col+8)] = base_types::convertor<U, T>::convert(src.tiles[i][j].data[3].y);
+        }
+    }
+}
+template<ducks::rt::all RT, ducks::gl::all GL, ducks::coord::tile COORD=coord<rt<typename RT::T, GROUP_WARPS*RT::rows, RT::cols, typename RT::layout>>>
+__device__ inline static void store(const GL &dst, const RT &src, const COORD &idx) {
+    store<2>(dst, src, idx);
+}
@@ -0,0 +1,168 @@
+/**
+ * @file
+ * @brief Group (collaborative warp) ops for loading shared tiles from and storing to global memory. 
+ */
+ 
+
+/**
+ * @brief Loads data from global memory into a shared memory tile.
+ *
+ * @tparam ST The type of the shared tile.
+ * @param[out] dst The destination shared memory tile.
+ * @param[in] src The source global memory array.
+ * @param[in] idx The coordinate of the tile in the global memory array.
+ */
+template<int axis, bool assume_aligned, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load(ST &dst, const GL &src, const COORD &idx) {
+    using T = typename ST::dtype;
+    const int row_stride = src.template stride<axis>();
+    // we can handle this many rows each time we run a memcpy_async
+    constexpr int elem_per_memcpy = sizeof(float4)/sizeof(typename ST::dtype);
+    constexpr int memcpy_per_row = dst.cols / elem_per_memcpy;
+    constexpr int total_calls = (dst.height*dst.width * kittens::TILE_ROW_DIM<T>*kittens::TILE_COL_DIM<T> + GROUP_THREADS*elem_per_memcpy-1) / (GROUP_THREADS*elem_per_memcpy); // round up
+    constexpr int total_rows = dst.height*dst.width;
+
+    coord<> unit_coord = idx.template unit_coord<axis, 3>();
+    typename GL::dtype *src_ptr = (typename GL::dtype*)&src[unit_coord];
+    uint32_t dst_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(&dst.data[0]));
+    int laneid = threadIdx.x % GROUP_THREADS;
+
+    #pragma unroll
+    for(int i = 0; i < total_calls; i++) {
+
+        int load_idx = i * GROUP_THREADS + laneid;
+        
+        int row = load_idx / memcpy_per_row;
+        int col = (load_idx*elem_per_memcpy) % dst.cols;
+
+        if constexpr (assume_aligned) {
+            float4 tmp;
+            move<float4>::ldg(tmp, (float4*)&src_ptr[row*row_stride + col]);
+            move<float4>::sts(dst.idx(dst_ptr, {row, col}), tmp);
+        }
+        else {
+            if (row + unit_coord.template dim<axis>() < src.template shape<axis>()) {
+                float4 tmp;
+                move<float4>::ldg(tmp, (float4*)&src_ptr[row*row_stride + col]);
+                move<float4>::sts(dst.idx(dst_ptr, {row, col}), tmp);
+            }
+            else {
+                float4 zeros = {0.f,0.f,0.f,0.f};
+                move<float4>::sts(dst.idx(dst_ptr, {row, col}), zeros); // use the default value
+            }
+        }
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load(ST &dst, const GL &src, const COORD &idx) {
+    load<2, false, ST, GL, COORD>(dst, src, idx);
+}
+
+/**
+ * @brief Stores data from a shared memory tile into global memory.
+ *
+ * @tparam ST The type of the shared tile.
+ * @param[out] dst The destination global memory array.
+ * @param[in] src The source shared memory tile.
+ * @param row_stride[in] The stride between rows in the destination array.
+ */
+template<int axis, bool assume_aligned, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store(const GL &dst, const ST &src, const COORD &idx) {
+    using T = typename ST::dtype;
+    const int row_stride = dst.template stride<axis>();
+    // we can handle this many rows each time we run a memcpy_async
+    constexpr int elem_per_memcpy = sizeof(float4)/sizeof(typename ST::dtype);
+    constexpr int memcpy_per_row = src.cols / elem_per_memcpy;
+    constexpr int total_calls = (src.height*src.width * kittens::TILE_ROW_DIM<T>*kittens::TILE_COL_DIM<T> + GROUP_THREADS*elem_per_memcpy-1) / (GROUP_THREADS*elem_per_memcpy); // round up
+
+    coord<> unit_coord = idx.template unit_coord<axis, 3>();
+    typename GL::dtype *dst_ptr = (typename GL::dtype*)&dst[unit_coord];
+    uint32_t src_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(&src.data[0]));
+    int laneid = threadIdx.x % GROUP_THREADS;
+
+    #pragma unroll
+    for(int i = 0; i < total_calls; i++) {
+
+        int load_idx = i * GROUP_THREADS + laneid;
+        
+        int row = load_idx / memcpy_per_row;
+        int col = (load_idx*elem_per_memcpy) % src.cols;
+
+        if constexpr (assume_aligned) {
+            float4 tmp;
+            move<float4>::lds(tmp, src.idx(src_ptr, {row, col}));
+            move<float4>::stg((float4*)&dst_ptr[row*row_stride + col], tmp);
+        }
+        else {
+            if (row + unit_coord.template dim<axis>() < dst.template shape<axis>()) {
+                float4 tmp;
+                move<float4>::lds(tmp, src.idx(src_ptr, {row, col}));
+                move<float4>::stg((float4*)&dst_ptr[row*row_stride + col], tmp);
+            }
+        }
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store(const GL &dst, const ST &src, const COORD &idx) {
+    store<2, false, ST, GL, COORD>(dst, src, idx);
+}
+
+/**
+ * @brief Asynchronously loads data from global memory into a shared memory tile.
+ *
+ * @tparam ST The type of the shared tile.
+ * @param[out] dst The destination shared memory tile.
+ * @param[in] src The source global memory array.
+ *
+ * @note This function expects 16-byte alignments. Otherwise, behavior is undefined.
+ */
+template<int axis, bool assume_aligned, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx) {
+    using T = typename ST::dtype;
+    const int row_stride = src.template stride<axis>();
+    // we can handle this many rows each time we run a memcpy_async
+    constexpr int elem_per_memcpy = sizeof(float4)/sizeof(typename ST::dtype);
+    constexpr int memcpy_per_row = dst.cols / elem_per_memcpy;
+    constexpr int total_calls = (dst.height*dst.width * kittens::TILE_ROW_DIM<T>*kittens::TILE_COL_DIM<T> + GROUP_THREADS*elem_per_memcpy-1) / (GROUP_THREADS*elem_per_memcpy); // round up
+
+    coord<> unit_coord = idx.template unit_coord<axis, 3>();
+    typename GL::dtype *src_ptr = (typename GL::dtype*)&src[unit_coord];
+    uint32_t dst_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(&dst.data[0]));
+    int laneid = threadIdx.x % GROUP_THREADS;
+
+    #pragma unroll
+    for(int i = 0; i < total_calls; i++) {
+
+        int load_idx = i * GROUP_THREADS + laneid;
+        
+        int row = load_idx / memcpy_per_row;
+        int col = (load_idx*elem_per_memcpy) % dst.cols;
+
+        if constexpr (assume_aligned) {
+            asm volatile(
+                "cp.async.cg.shared.global.L2::128B [%0], [%1], 16;\n"
+                :: "r"(dst.idx(dst_ptr, {row, col})), "l"(&src_ptr[row*row_stride + col])
+                : "memory"
+            );
+        }
+        else {
+            if (row + unit_coord.template dim<axis>() < src.template shape<axis>()) {
+                asm volatile(
+                    "cp.async.cg.shared.global.L2::128B [%0], [%1], 16;\n"
+                    :: "r"(dst.idx(dst_ptr, {row, col})), "l"(&src_ptr[row*row_stride + col])
+                    : "memory"
+                );
+            }
+            else {
+                // printf("thread %d skipping async load on row %d, col %d\n", threadIdx.x, row + unit_coord.template dim<axis>(), col);
+                float4 zeros = {0.f,0.f,0.f,0.f};
+                move<float4>::sts(dst.idx(dst_ptr, {row, col}), zeros); // use the default value
+            }
+        }
+    }
+    asm volatile("cp.async.commit_group;\n" ::: "memory");
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx) {
+    load_async<2, false, ST, GL, COORD>(dst, src, idx);
+}
@@ -0,0 +1,323 @@
+/**
+ * @file
+ * @brief Functions for a warpgroup to collaboratively transfer data directly between shared memory and registers and back.
+ */
+
+/**
+ * @brief Collaboratively load data from a shared tile into register tiles split across a warpgroup.
+ *
+ * @tparam RT The register tile type
+ * @tparam ST The shared tile type
+ * @param dst[out] The destination register tile.
+ * @param src[in]  The source shared tile.
+ */
+template<ducks::rt::all RT, ducks::st::all ST>
+__device__ inline static void load(RT &dst, const ST &src) {
+    constexpr int height = ST::height;
+    constexpr int warp_height = RT::height;
+    static_assert(height%GROUP_WARPS == 0, "Group load / store requires tile height to be a multiple of GROUP_WARPS.");
+    static_assert(height%warp_height == 0, "Group load / store requires tile height to be a multiple of the RT height.");
+    static_assert(ST::width==RT::width, "Group load / store requires tile widths to match.");
+    int local_warpid;
+    if constexpr(GROUP_WARPS % 4 == 0) local_warpid = (warpid()/4+(warpid()%4)*(GROUP_WARPS/4));
+    else local_warpid = warpid();
+    using T2 = RT::dtype;
+    using U  = ST::dtype;
+    using T  = base_types::packing<T2>::unpacked_type;
+    using U2 = base_types::packing<U>::packed_type;
+    int warp_laneid = ::kittens::laneid();
+
+    // convert to shared state space
+    uint32_t shared_addr = static_cast<uint32_t>(__cvta_generic_to_shared(&src.data[0]));
+
+    #pragma unroll
+    for(int i = 0; i < dst.height; i++) {
+        #pragma unroll
+        for(int j = 0; j < dst.width; j++) {
+            if constexpr (sizeof(typename ST::dtype) == 2) {
+                // handle the row-major layout for 16-bit types
+                U2 tmp[4];
+                int row = (local_warpid*warp_height + i)*dst.tile_size_row + (warp_laneid % 16);
+                int col = j*dst.tile_size_col + (warp_laneid / 16) * 8;
+                if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row>) {
+                    move<U2>::ldsm4(tmp[0], tmp[1], tmp[2], tmp[3], src.idx(shared_addr, {row, col}));
+                }
+                else {
+                    move<U2>::ldsm4t(tmp[0], tmp[2], tmp[1], tmp[3], src.idx(shared_addr, {row, col}));
+                }
+                dst.tiles[i][j].data[0] = base_types::convertor<T2, U2>::convert(tmp[0]);
+                dst.tiles[i][j].data[1] = base_types::convertor<T2, U2>::convert(tmp[1]);
+                dst.tiles[i][j].data[2] = base_types::convertor<T2, U2>::convert(tmp[2]);
+                dst.tiles[i][j].data[3] = base_types::convertor<T2, U2>::convert(tmp[3]);
+            }
+            else if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row> && sizeof(typename ST::dtype) == 1) {
+                // handle the row-major layout for 8-bit types
+                int warp_group_16 = (warp_laneid / 16);  // divide each warp into two groups of 16 threads
+                int lane_in_16 = warp_laneid % 16;       // position in group of 16 threads
+                int row = (local_warpid*warp_height + i)*dst.tile_size_row + (lane_in_16 % 16); // find base row for warp in warpgroup and then distribute the 16 threads in the warp across the rows
+                int col = j*dst.tile_size_col + warp_group_16 * 16; // find base column and then *16 for second half of the warp
+
+                U2 tmp[4];
+                if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row>) {
+                    move<U2>::ldsm4(tmp[0], tmp[1], tmp[2], tmp[3], src.idx(shared_addr, {row, col}));
+                }
+                else {
+                    move<U2>::ldsm4t(tmp[0], tmp[2], tmp[1], tmp[3], src.idx(shared_addr, {row, col}));
+                }
+                dst.tiles[i][j].data[0] = base_types::convertor<T2, U2>::convert(tmp[0]);
+                dst.tiles[i][j].data[1] = base_types::convertor<T2, U2>::convert(tmp[1]);
+                dst.tiles[i][j].data[2] = base_types::convertor<T2, U2>::convert(tmp[2]);
+                dst.tiles[i][j].data[3] = base_types::convertor<T2, U2>::convert(tmp[3]);
+            }
+            else if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row> && sizeof(typename ST::dtype) == 4) {
+                // handle the row-major layout for 32-bit types
+                int row = (local_warpid*warp_height + i)*dst.tile_size_row + (warp_laneid / 4);
+                int col = j*dst.tile_size_col + 2*(warp_laneid % 4);
+                if constexpr (ST::rows != ST::underlying_rows || ST::cols != ST::underlying_cols) { // subtile case
+                    row += src.row_offset;
+                    col += src.col_offset;
+                }
+                int blit = sizeof(typename ST::dtype) * ((warp_laneid%4) / 2);
+                U2 tmp[4];
+                static constexpr int swizzle_repeat = ST::swizzle_bytes * 8;
+                static constexpr int subtile_cols   = ST::swizzle_bytes / sizeof(U);
+                const int outer_idx = col/subtile_cols;
+                const uint32_t addr_1 = shared_addr + sizeof(U)*(outer_idx*ST::underlying_rows*subtile_cols + (row+0)*subtile_cols + col%subtile_cols);
+                const uint32_t addr_2 = shared_addr + sizeof(U)*(outer_idx*ST::underlying_rows*subtile_cols + (row+8)*subtile_cols + col%subtile_cols);
+                const int swizzle_1 = blit ^ ((addr_1 % swizzle_repeat) >> 7) << 4;
+                const int swizzle_2 = blit ^ ((addr_2 % swizzle_repeat) >> 7) << 4;
+                move<U>::lds(tmp[0].x, (addr_1+ 0)^swizzle_1);
+                move<U>::lds(tmp[0].y, (addr_1+ 4)^swizzle_1);
+                move<U>::lds(tmp[2].x, (addr_1+32)^swizzle_1);
+                move<U>::lds(tmp[2].y, (addr_1+36)^swizzle_1);
+                move<U>::lds(tmp[1].x, (addr_2+ 0)^swizzle_2);
+                move<U>::lds(tmp[1].y, (addr_2+ 4)^swizzle_2);
+                move<U>::lds(tmp[3].x, (addr_2+32)^swizzle_2);
+                move<U>::lds(tmp[3].y, (addr_2+36)^swizzle_2);
+                dst.tiles[i][j].data[0] = base_types::convertor<T2, U2>::convert(tmp[0]);
+                dst.tiles[i][j].data[1] = base_types::convertor<T2, U2>::convert(tmp[1]);
+                dst.tiles[i][j].data[2] = base_types::convertor<T2, U2>::convert(tmp[2]);
+                dst.tiles[i][j].data[3] = base_types::convertor<T2, U2>::convert(tmp[3]);
+                if(blit) {
+                    #pragma unroll
+                    for(int k = 0; k < 4; k++) {
+                        dst.tiles[i][j].data[k] = T2{dst.tiles[i][j].data[k].y, dst.tiles[i][j].data[k].x};
+                    }
+                }
+            }
+            else {
+                // handle the column-major layout
+                int row = (local_warpid*warp_height + i)*dst.tile_size_row + 2*(warp_laneid % 4);
+                int col = j*dst.tile_size_col + (warp_laneid / 4);
+                U2 tmp[4];
+                move<U>::lds(tmp[0].x, src.idx(shared_addr, {row+0, col+0}));
+                move<U>::lds(tmp[0].y, src.idx(shared_addr, {row+1, col+0}));
+                move<U>::lds(tmp[1].x, src.idx(shared_addr, {row+0, col+8}));
+                move<U>::lds(tmp[1].y, src.idx(shared_addr, {row+1, col+8}));
+                move<U>::lds(tmp[2].x, src.idx(shared_addr, {row+8, col+0}));
+                move<U>::lds(tmp[2].y, src.idx(shared_addr, {row+9, col+0}));
+                move<U>::lds(tmp[3].x, src.idx(shared_addr, {row+8, col+8}));
+                move<U>::lds(tmp[3].y, src.idx(shared_addr, {row+9, col+8}));
+                dst.tiles[i][j].data[0] = base_types::convertor<T2, U2>::convert(tmp[0]);
+                dst.tiles[i][j].data[1] = base_types::convertor<T2, U2>::convert(tmp[1]);
+                dst.tiles[i][j].data[2] = base_types::convertor<T2, U2>::convert(tmp[2]);
+                dst.tiles[i][j].data[3] = base_types::convertor<T2, U2>::convert(tmp[3]);
+            }
+        }
+    }
+}
+
+
+/**
+ * @brief Collaboratively store data into a shared tile from register tiles split across a warpgroup.
+ *
+ * @tparam RT The register tile type
+ * @tparam ST The shared tile type
+ * @param dst[out] The destination shared tile.
+ * @param src[in]  The source register tile.
+ */
+template<ducks::st::all ST, ducks::rt::all RT>
+__device__ inline static void store(ST &dst, const RT &src) {
+    constexpr int height = ST::height;
+    constexpr int warp_height = RT::height;
+    static_assert(height%GROUP_WARPS == 0, "Group load / store requires tile height to be a multiple of GROUP_WARPS.");
+    static_assert(height%warp_height == 0, "Group load / store requires tile height to be a multiple of the RT height.");
+    static_assert(ST::width==RT::width, "Group load / store requires tile widths to match.");
+    int local_warpid;
+    if constexpr(GROUP_WARPS % 4 == 0) local_warpid = (warpid()/4+(warpid()%4)*(GROUP_WARPS/4));
+    else local_warpid = warpid();
+    using T2 = RT::dtype;
+    using U  = ST::dtype;
+    using T  = base_types::packing<T2>::unpacked_type;
+    using U2 = base_types::packing<U>::packed_type;
+    int warp_laneid = ::kittens::laneid();
+
+    // convert to shared state space
+    uint32_t shared_addr = static_cast<uint32_t>(__cvta_generic_to_shared(&dst.data[0]));
+
+    #pragma unroll
+    for(int i = 0; i < warp_height; i++) {
+        #pragma unroll
+        for(int j = 0; j < src.width; j++) {
+            if constexpr (sizeof(typename ST::dtype) == 2) {
+                // handle the row-major layout
+                U2 tmp[4];
+                tmp[0] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[0]);
+                tmp[1] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[1]);
+                tmp[2] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[2]);
+                tmp[3] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[3]);
+#ifdef KITTENS_HOPPER
+                int row = (local_warpid*warp_height + i)*src.tile_size_row + (warp_laneid % 16);
+                int col = j*src.tile_size_col + (warp_laneid / 16) * 8;
+                if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row>) {
+                    move<U2>::stsm4(dst.idx(shared_addr, {row, col}), tmp[0], tmp[1], tmp[2], tmp[3]);
+                }
+                else {
+                    move<U2>::stsm4t(dst.idx(shared_addr, {row, col}), tmp[0], tmp[2], tmp[1], tmp[3]);
+                }
+#else
+                if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row>) {
+                    int row = (local_warpid*warp_height + i)*src.tile_size_row + (warp_laneid / 4);
+                    int col = j*src.tile_size_col + 2*(warp_laneid % 4);
+                    move<U2>::sts(dst.idx(shared_addr, {row+0, col+0}), tmp[0]);
+                    move<U2>::sts(dst.idx(shared_addr, {row+8, col+0}), tmp[1]);
+                    move<U2>::sts(dst.idx(shared_addr, {row+0, col+8}), tmp[2]);
+                    move<U2>::sts(dst.idx(shared_addr, {row+8, col+8}), tmp[3]);
+                }
+                else {
+                    int row = (local_warpid*warp_height + i)*src.tile_size_row + 2*(warp_laneid % 4);
+                    int col = j*src.tile_size_col + (warp_laneid / 4);
+                    move<U>::sts(dst.idx(shared_addr, {row+0, col+0}), tmp[0].x);
+                    move<U>::sts(dst.idx(shared_addr, {row+1, col+0}), tmp[0].y);
+                    move<U>::sts(dst.idx(shared_addr, {row+0, col+8}), tmp[1].x);
+                    move<U>::sts(dst.idx(shared_addr, {row+1, col+8}), tmp[1].y);
+                    move<U>::sts(dst.idx(shared_addr, {row+8, col+0}), tmp[2].x);
+                    move<U>::sts(dst.idx(shared_addr, {row+9, col+0}), tmp[2].y);
+                    move<U>::sts(dst.idx(shared_addr, {row+8, col+8}), tmp[3].x);
+                    move<U>::sts(dst.idx(shared_addr, {row+9, col+8}), tmp[3].y);
+                }
+#endif
+            }
+            else if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row> && sizeof(typename ST::dtype) == 1) { 
+                // handle the row-major layout for 8-bit types
+
+                int warp_group_16 = (warp_laneid / 16);  // divide each warp into two groups of 16 threads
+                int lane_in_16 = warp_laneid % 16;       // position in group of 16 threads
+                int row = (local_warpid*warp_height + i)*src.tile_size_row + (lane_in_16 % 16); // find base row for warp in warpgroup and then distribute the 16 threads in the warp across the rows
+                int col = j*src.tile_size_col + warp_group_16 * 16; // find base column and then *16 for second half of the warp
+
+                U2 tmp[4];
+                tmp[0] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[0]);
+                tmp[1] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[1]);
+                tmp[2] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[2]);
+                tmp[3] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[3]);
+                if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row>) {
+                    move<U2>::stsm4(dst.idx(shared_addr, {row, col}), tmp[0], tmp[1], tmp[2], tmp[3]);
+                }
+                else {
+                    move<U2>::stsm4t(dst.idx(shared_addr, {row, col}), tmp[0], tmp[2], tmp[1], tmp[3]);
+                }
+            }
+            else if constexpr (std::is_same_v<typename RT::layout, ducks::rt_layout::row> && sizeof(typename ST::dtype) == 4) {
+                // handle the row-major layout for 32-bit types
+                int row = (local_warpid*warp_height + i)*src.tile_size_row + (warp_laneid / 4);
+                int col = j*src.tile_size_col + 2*(warp_laneid % 4);
+                if constexpr (ST::rows != ST::underlying_rows || ST::cols != ST::underlying_cols) { // subtile case
+                    row += dst.row_offset;
+                    col += dst.col_offset;
+                }
+                int blit = sizeof(typename ST::dtype) * ((warp_laneid%4) / 2);
+                T2 reg_tmp[4];
+                if(blit) {
+                    #pragma unroll
+                    for(int k = 0; k < 4; k++) {
+                        reg_tmp[k] = T2{src.tiles[i][j].data[k].y, src.tiles[i][j].data[k].x};
+                    }
+                }
+                else {
+                    #pragma unroll
+                    for(int k = 0; k < 4; k++) {
+                        reg_tmp[k] = src.tiles[i][j].data[k];
+                    }
+                }
+                U2 tmp[4];
+                tmp[0] = base_types::convertor<U2, T2>::convert(reg_tmp[0]);
+                tmp[1] = base_types::convertor<U2, T2>::convert(reg_tmp[1]);
+                tmp[2] = base_types::convertor<U2, T2>::convert(reg_tmp[2]);
+                tmp[3] = base_types::convertor<U2, T2>::convert(reg_tmp[3]);
+                static constexpr int swizzle_repeat = ST::swizzle_bytes * 8;
+                static constexpr int subtile_cols   = ST::swizzle_bytes / sizeof(U);
+                const int outer_idx = col/subtile_cols;
+                const uint32_t addr_1 = shared_addr + sizeof(U)*(outer_idx*ST::underlying_rows*subtile_cols + (row+0)*subtile_cols + col%subtile_cols);
+                const uint32_t addr_2 = shared_addr + sizeof(U)*(outer_idx*ST::underlying_rows*subtile_cols + (row+8)*subtile_cols + col%subtile_cols);
+                const int swizzle_1 = blit ^ ((addr_1 % swizzle_repeat) >> 7) << 4;
+                const int swizzle_2 = blit ^ ((addr_2 % swizzle_repeat) >> 7) << 4;
+                move<U>::sts((addr_1+ 0)^swizzle_1, tmp[0].x);
+                move<U>::sts((addr_1+ 4)^swizzle_1, tmp[0].y);
+                move<U>::sts((addr_1+32)^swizzle_1, tmp[2].x);
+                move<U>::sts((addr_1+36)^swizzle_1, tmp[2].y);
+                move<U>::sts((addr_2+ 0)^swizzle_2, tmp[1].x);
+                move<U>::sts((addr_2+ 4)^swizzle_2, tmp[1].y);
+                move<U>::sts((addr_2+32)^swizzle_2, tmp[3].x);
+                move<U>::sts((addr_2+36)^swizzle_2, tmp[3].y);
+            }
+            else {
+                // handle the column-major layout
+                int row = (local_warpid*warp_height + i)*src.tile_size_row + 2*(warp_laneid % 4);
+                int col = j*src.tile_size_col + (warp_laneid / 4);
+                U2 tmp[4];
+                tmp[0] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[0]);
+                tmp[1] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[1]);
+                tmp[2] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[2]);
+                tmp[3] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[3]);
+                move<U>::sts(dst.idx(shared_addr, {row+0, col+0}), tmp[0].x);
+                move<U>::sts(dst.idx(shared_addr, {row+1, col+0}), tmp[0].y);
+                move<U>::sts(dst.idx(shared_addr, {row+0, col+8}), tmp[1].x);
+                move<U>::sts(dst.idx(shared_addr, {row+1, col+8}), tmp[1].y);
+                move<U>::sts(dst.idx(shared_addr, {row+8, col+0}), tmp[2].x);
+                move<U>::sts(dst.idx(shared_addr, {row+9, col+0}), tmp[2].y);
+                move<U>::sts(dst.idx(shared_addr, {row+8, col+8}), tmp[3].x);
+                move<U>::sts(dst.idx(shared_addr, {row+9, col+8}), tmp[3].y);
+            }
+        }
+    }
+}
+
+// Load and store of vectors from/to shared tiles.
+
+template<ducks::rv::naive_layout RV, ducks::st::all ST>
+__device__ inline static auto load(RV &dst, const ST &src, int2 row_col) {
+    KITTENS_CHECK_WARP;
+    static_assert(ST::cols>=RV::length, "Shared tile must be at least as wide as the vector.");
+    using T = RV::T;
+    using U = ST::T;
+    int warp_laneid = ::kittens::laneid();
+
+    // convert to shared state space
+    uint32_t shared_addr = static_cast<uint32_t>(__cvta_generic_to_shared(&src.data[0]));
+
+    #pragma unroll
+    for(int col = warp_laneid; col < dst.length; col+=WARP_THREADS) {
+        U tmp;
+        move<U>::lds(tmp, src.idx(shared_addr, {row_col.x, row_col.y + col}));
+        dst.data[col/WARP_THREADS][0] = base_types::convertor<T, U>::convert(tmp);
+    }
+}
+
+template<ducks::rv::naive_layout RV, ducks::st::all ST>
+__device__ inline static auto store(ST &dst, const RV &src, int2 row_col) {
+    KITTENS_CHECK_WARP;
+    static_assert(ST::cols>=RV::length, "Shared tile must be at least as wide as the vector.");
+    using T = RV::T;
+    using U = ST::T;
+    int warp_laneid = ::kittens::laneid();
+
+    // convert to shared state space
+    uint32_t shared_addr = static_cast<uint32_t>(__cvta_generic_to_shared(&dst.data[0]));
+
+    #pragma unroll
+    for(int col = warp_laneid; col < src.length; col+=WARP_THREADS) {
+        U tmp = base_types::convertor<U, T>::convert(src.data[col/WARP_THREADS][0]);
+        move<U>::sts(dst.idx(shared_addr, {row_col.x, row_col.y + col}), tmp);
+    }
+}
@@ -0,0 +1,325 @@
+/**
+ * @file
+ * @brief Group (collaborative warp) ops for loading tensor tiles into register tiles.
+ */
+
+/**
+ * @brief Load data from a tensor tile into a register tile.
+ *
+ * @tparam RT The register tile type
+ * @tparam TM The tensor memory tile type
+ * @param dst[out] The destination register tile.
+ * @param src[in]  The source tensor tile.
+ */
+template<ducks::rt::row_layout RT, ducks::tt::all TM>
+__device__ inline static void load_async(RT &dst, const TM &src) {
+    if constexpr (GROUP_WARPS == 1) {
+        static_assert(RT::height == TM::height, "register tile and tensor tile must match height");
+        static_assert(RT::width == TM::width, "register tile and tensor tile must match width");
+
+        using T2 = RT::dtype;
+        using U  = typename TM::dtype;
+        using U2 = base_types::packing<typename TM::dtype>::packed_type;
+
+        if constexpr (sizeof(typename TM::dtype) == 1) {
+            #pragma unroll
+            for(int i = 0; i < dst.height; i++) {
+                #pragma unroll
+                for(int j = 0; j < dst.width; j++) {
+                    asm volatile(
+                        "tcgen05.ld.sync.aligned.16x128b.x2.pack::16b.b32 {%0, %1, %2, %3}, [%4];\n"
+                        : "=r"(*(uint32_t*) &dst.tiles[i][j].data[0]),
+                            "=r"(*(uint32_t*) &dst.tiles[i][j].data[1]),
+                            "=r"(*(uint32_t*) &dst.tiles[i][j].data[2]),
+                            "=r"(*(uint32_t*) &dst.tiles[i][j].data[3])
+                        : "r"(src.addr + ((i * dst.tile_size_row) << 16) + (j * dst.tile_size_col)/(4/(uint32_t)sizeof(U)))
+                    );
+                }
+            }
+        } else if constexpr (sizeof(typename TM::dtype) == 2) {
+            #pragma unroll
+            for(int i = 0; i < dst.height; i++) {
+                #pragma unroll
+                for(int j = 0; j < dst.width; j++) {
+                    asm volatile(
+                        "tcgen05.ld.sync.aligned.16x128b.x2.pack::16b.b32 {%0, %1, %2, %3}, [%4];\n"
+                        : "=r"(*(uint32_t*) &dst.tiles[i][j].data[0]),
+                            "=r"(*(uint32_t*) &dst.tiles[i][j].data[1]),
+                            "=r"(*(uint32_t*) &dst.tiles[i][j].data[2]),
+                            "=r"(*(uint32_t*) &dst.tiles[i][j].data[3])
+                        : "r"(src.addr + ((i * dst.tile_size_row) << 16) + (j * dst.tile_size_col))
+                    );
+                }
+            }
+        }
+        else if constexpr (sizeof(typename TM::dtype) == 4) {
+            #pragma unroll
+            for(int i = 0; i < dst.height; i++) {
+                if constexpr (dst.width%4 == 0) {
+                    #pragma unroll
+                    for(int j = 0; j < dst.width; j+=4) {
+                        U2 data[16];
+                        asm volatile(
+                            "tcgen05.ld.sync.aligned.16x256b.x8.b32 {%0, %1, %2, %3, %4, %5, %6, %7, %8, %9, %10, %11, %12, %13, %14, %15, %16, %17, %18, %19, %20, %21, %22, %23, %24, %25, %26, %27, %28, %29, %30, %31}, [%32];\n"
+                            : "=f"(data[0].x), "=f"(data[0].y),
+                            "=f"(data[1].x), "=f"(data[1].y),
+                            "=f"(data[2].x), "=f"(data[2].y),
+                            "=f"(data[3].x), "=f"(data[3].y),
+                            "=f"(data[4].x), "=f"(data[4].y),
+                            "=f"(data[5].x), "=f"(data[5].y),
+                            "=f"(data[6].x), "=f"(data[6].y),
+                            "=f"(data[7].x), "=f"(data[7].y),
+                            "=f"(data[8].x), "=f"(data[8].y),
+                            "=f"(data[9].x), "=f"(data[9].y),
+                            "=f"(data[10].x), "=f"(data[10].y),
+                            "=f"(data[11].x), "=f"(data[11].y),
+                            "=f"(data[12].x), "=f"(data[12].y),
+                            "=f"(data[13].x), "=f"(data[13].y),
+                            "=f"(data[14].x), "=f"(data[14].y),
+                            "=f"(data[15].x), "=f"(data[15].y)
+                            : "r"(src.addr + ((i * dst.tile_size_row) << 16) + (j * dst.tile_size_col)/(4/(uint32_t)sizeof(U)))
+                        );
+                        #pragma unroll
+                        for(int k = 0; k < 4; k++) {
+                            dst.tiles[i][j+0].data[k] = base_types::convertor<T2, U2>::convert(data[k]);
+                            dst.tiles[i][j+1].data[k] = base_types::convertor<T2, U2>::convert(data[k+4]);
+                            dst.tiles[i][j+2].data[k] = base_types::convertor<T2, U2>::convert(data[k+8]);
+                            dst.tiles[i][j+3].data[k] = base_types::convertor<T2, U2>::convert(data[k+12]);
+                        }
+                    }
+                }
+                else if constexpr (dst.width%2 == 0) {
+                    #pragma unroll
+                    for(int j = 0; j < dst.width; j+=2) {
+                        U2 data[8];
+                        asm volatile(
+                            "tcgen05.ld.sync.aligned.16x256b.x4.b32 {%0, %1, %2, %3, %4, %5, %6, %7, %8, %9, %10, %11, %12, %13, %14, %15}, [%16];\n"
+                            : "=f"(data[0].x), "=f"(data[0].y),
+                            "=f"(data[1].x), "=f"(data[1].y),
+                            "=f"(data[2].x), "=f"(data[2].y),
+                            "=f"(data[3].x), "=f"(data[3].y),
+                            "=f"(data[4].x), "=f"(data[4].y),
+                            "=f"(data[5].x), "=f"(data[5].y),
+                            "=f"(data[6].x), "=f"(data[6].y),
+                            "=f"(data[7].x), "=f"(data[7].y)
+                            : "r"(src.addr + ((i * dst.tile_size_row) << 16) + (j * dst.tile_size_col)/(4/(uint32_t)sizeof(U)))
+                        );
+                        #pragma unroll
+                        for(int k = 0; k < 4; k++) {
+                            dst.tiles[i][j+0].data[k] = base_types::convertor<T2, U2>::convert(data[k]);
+                            dst.tiles[i][j+1].data[k] = base_types::convertor<T2, U2>::convert(data[k+4]);
+                        }
+                    }
+                }
+                else {
+                    #pragma unroll
+                    for(int j = 0; j < dst.width; j++) {
+                        U2 data[4];
+                        asm volatile(
+                            "tcgen05.ld.sync.aligned.16x256b.x2.b32 {%0, %1, %2, %3, %4, %5, %6, %7}, [%8];\n"
+                            : "=f"(data[0].x), "=f"(data[0].y),
+                            "=f"(data[1].x), "=f"(data[1].y),
+                            "=f"(data[2].x), "=f"(data[2].y),
+                            "=f"(data[3].x), "=f"(data[3].y)
+                            : "r"(src.addr + ((i * dst.tile_size_row) << 16) + (j * dst.tile_size_col)/(4/(uint32_t)sizeof(U)))
+                        );
+                        #pragma unroll
+                        for(int k = 0; k < 4; k++) {
+                            dst.tiles[i][j].data[k] = base_types::convertor<T2, U2>::convert(data[k]);
+                        }
+                    }
+                }
+            }
+        }
+    }
+    else {
+        static_assert(GROUP_WARPS==4 || GROUP_WARPS==8);
+        constexpr int warp_rows = TM::rows/GROUP_WARPS;
+        static_assert(TM::cols==RT::cols);
+        static_assert(warp_rows==RT::rows);
+        if constexpr (GROUP_WARPS == 4) {
+            auto src_subtile = src.template subtile<tt<typename TM::dtype, warp_rows, TM::cols>>(32*warpid(), 0);
+            ::kittens::group<1>::load_async(dst, src_subtile);
+        }
+        else {
+            auto src_subtile = src.template subtile<tt<typename TM::dtype, warp_rows, TM::cols>>(32*(warpid()%4)+16*(warpid()/4), 0);
+            ::kittens::group<1>::load_async(dst, src_subtile);
+        }
+    }
+}
+
+
+/**
+ * @brief Store data into a tensor tile from a register tile.
+ *
+ * @tparam RT The register tile type
+ * @tparam TM The tensor memory tile type
+ * @param dst[out] The destination tensor tile.
+ * @param src[in]  The source register tile.
+ */
+template<ducks::rt::all RT, ducks::tt::all TM>
+__device__ inline static void store_async(TM &dst, const RT &src) {
+    if constexpr (GROUP_WARPS == 1) {
+        static_assert(RT::height == TM::height, "register tile and tensor tile must match height");
+        static_assert(RT::width == TM::width, "register tile and tensor tile must match width");
+
+        using T2 = RT::dtype;
+        using T = base_types::packing<T2>::unpacked_type;
+        using U = TM::dtype;
+        using U2 = base_types::packing<U>::packed_type;
+
+        if constexpr (sizeof(typename TM::dtype) == 2) {
+            #pragma unroll
+            for(int i = 0; i < src.height; i++) {
+                if constexpr (src.width%4 == 0) {
+                    #pragma unroll
+                    for(int j = 0; j < src.width; j+=4) {
+                        asm volatile(
+                            "tcgen05.st.sync.aligned.16x128b.x8.b32 [%0], {%1, %2, %3, %4, %5, %6, %7, %8, %9, %10, %11, %12, %13, %14, %15, %16};\n"
+                            :: "r"(dst.addr + ((i * src.tile_size_row) << 16) + (j * src.tile_size_col)/(4/(uint32_t)sizeof(U))),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[0]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[1]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[2]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[3]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[0]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[1]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[2]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[3]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+2].data[0]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+2].data[1]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+2].data[2]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+2].data[3]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+3].data[0]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+3].data[1]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+3].data[2]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+3].data[3])
+                        );
+                    }
+                }
+                else if constexpr (src.width%2 == 0) {
+                    #pragma unroll
+                    for(int j = 0; j < src.width; j+=2) {
+                        asm volatile(
+                            "tcgen05.st.sync.aligned.16x128b.x4.b32 [%0], {%1, %2, %3, %4, %5, %6, %7, %8};\n"
+                            :: "r"(dst.addr + ((i * src.tile_size_row) << 16) + (j * src.tile_size_col)/(4/(uint32_t)sizeof(U))),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[0]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[1]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[2]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+0].data[3]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[0]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[1]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[2]),
+                            "r"(*(uint32_t*)&src.tiles[i][j+1].data[3])
+                        );
+                    }
+                }
+                else {
+                    #pragma unroll
+                    for(int j = 0; j < src.width; j++) {
+                        asm volatile(
+                            "tcgen05.st.sync.aligned.16x128b.x2.b32 [%0], {%1, %2, %3, %4};\n"
+                            :: "r"(dst.addr + ((i * src.tile_size_row) << 16) + (j * src.tile_size_col)/(4/(uint32_t)sizeof(U))),
+                            "r"(*(uint32_t*)&src.tiles[i][j].data[0]),
+                            "r"(*(uint32_t*)&src.tiles[i][j].data[1]),
+                            "r"(*(uint32_t*)&src.tiles[i][j].data[2]),
+                            "r"(*(uint32_t*)&src.tiles[i][j].data[3])
+                        );
+                    }
+                }
+            }
+        }
+        else if constexpr (sizeof(typename TM::dtype) == 4) {
+            #pragma unroll
+            for(int i = 0; i < src.height; i++) {
+                if constexpr(src.width%4 == 0) {
+                    #pragma unroll
+                    for(int j = 0; j < src.width; j+=4) {
+                        U2 data[16];
+                        #pragma unroll
+                        for(int k = 0; k < 4; k++) {
+                            data[k] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[k]);
+                            data[k+4] = base_types::convertor<U2, T2>::convert(src.tiles[i][j+1].data[k]);
+                            data[k+8] = base_types::convertor<U2, T2>::convert(src.tiles[i][j+2].data[k]);
+                            data[k+12] = base_types::convertor<U2, T2>::convert(src.tiles[i][j+3].data[k]);
+                        }
+                        asm volatile(
+                            "tcgen05.st.sync.aligned.16x256b.x8.b32 [%0], {%1, %2, %3, %4, %5, %6, %7, %8, %9, %10, %11, %12, %13, %14, %15, %16, %17, %18, %19, %20, %21, %22, %23, %24, %25, %26, %27, %28, %29, %30, %31, %32};\n"
+                            :: "r"(dst.addr + ((i * src.tile_size_row) << 16) + (j * src.tile_size_col)/(4/(uint32_t)sizeof(U))),
+                            "f"(data[0].x), "f"(data[0].y),
+                            "f"(data[1].x), "f"(data[1].y),
+                            "f"(data[2].x), "f"(data[2].y),
+                            "f"(data[3].x), "f"(data[3].y),
+                            "f"(data[4].x), "f"(data[4].y),
+                            "f"(data[5].x), "f"(data[5].y),
+                            "f"(data[6].x), "f"(data[6].y),
+                            "f"(data[7].x), "f"(data[7].y),
+                            "f"(data[8].x), "f"(data[8].y),
+                            "f"(data[9].x), "f"(data[9].y),
+                            "f"(data[10].x), "f"(data[10].y),
+                            "f"(data[11].x), "f"(data[11].y),
+                            "f"(data[12].x), "f"(data[12].y),
+                            "f"(data[13].x), "f"(data[13].y),
+                            "f"(data[14].x), "f"(data[14].y),
+                            "f"(data[15].x), "f"(data[15].y)
+                        );
+                    }
+                }
+                else if constexpr(src.width%2 == 0) {
+                    #pragma unroll
+                    for(int j = 0; j < src.width; j+=2) {
+                        U2 data[8];
+                        #pragma unroll
+                        for(int k = 0; k < 4; k++) {
+                            data[k] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[k]);
+                            data[k+4] = base_types::convertor<U2, T2>::convert(src.tiles[i][j+1].data[k]);
+                        }
+                        asm volatile(
+                            "tcgen05.st.sync.aligned.16x256b.x4.b32 [%0], {%1, %2, %3, %4, %5, %6, %7, %8, %9, %10, %11, %12, %13, %14, %15, %16};\n"
+                            :: "r"(dst.addr + ((i * src.tile_size_row) << 16) + (j * src.tile_size_col)/(4/(uint32_t)sizeof(U))),
+                            "f"(data[0].x), "f"(data[0].y),
+                            "f"(data[1].x), "f"(data[1].y),
+                            "f"(data[2].x), "f"(data[2].y),
+                            "f"(data[3].x), "f"(data[3].y),
+                            "f"(data[4].x), "f"(data[4].y),
+                            "f"(data[5].x), "f"(data[5].y),
+                            "f"(data[6].x), "f"(data[6].y),
+                            "f"(data[7].x), "f"(data[7].y)
+                        );
+                    }
+                }
+                else {
+                    #pragma unroll
+                    for(int j = 0; j < src.width; j++) {
+                        U2 data[4];
+                        #pragma unroll
+                        for(int k = 0; k < 4; k++) {
+                            data[k] = base_types::convertor<U2, T2>::convert(src.tiles[i][j].data[k]);
+                        }
+                        asm volatile(
+                            "tcgen05.st.sync.aligned.16x256b.x2.b32 [%0], {%1, %2, %3, %4, %5, %6, %7, %8};\n"
+                            :: "r"(dst.addr + ((i * src.tile_size_row) << 16) + (j * src.tile_size_col)/(4/(uint32_t)sizeof(U))),
+                            "f"(data[0].x), "f"(data[0].y),
+                            "f"(data[1].x), "f"(data[1].y),
+                            "f"(data[2].x), "f"(data[2].y),
+                            "f"(data[3].x), "f"(data[3].y)
+                        );
+                    }
+                }
+            }
+        }
+    }
+    else {
+        static_assert(GROUP_WARPS==4 || GROUP_WARPS==8);
+        constexpr int warp_rows = TM::rows/GROUP_WARPS;
+        static_assert(TM::cols==RT::cols);
+        static_assert(warp_rows==RT::rows);
+        if constexpr (GROUP_WARPS == 4) {
+            auto dst_subtile = dst.template subtile<tt<typename TM::dtype, warp_rows, TM::cols>>(32*warpid(), 0);
+            ::kittens::group<1>::store_async(dst_subtile, src);
+        }
+        else {
+            auto dst_subtile = dst.template subtile<tt<typename TM::dtype, warp_rows, TM::cols>>(32*(warpid()%4)+16*(warpid()/4), 0);
+            ::kittens::group<1>::store_async(dst_subtile, src);
+        }
+    }
+}
@@ -0,0 +1,16 @@
+/**
+ * @file
+ * @brief An aggregate header of group memory operations on tiles.
+ */
+
+#include "shared_to_register.cuh"
+#include "global_to_register.cuh"
+#include "global_to_shared.cuh"
+#ifdef KITTENS_BLACKWELL
+#include "tensor_to_register.cuh"
+#endif
+
+#include "complex/complex_shared_to_register.cuh"
+#include "complex/complex_global_to_register.cuh"
+#include "complex/complex_global_to_shared.cuh"
+
@@ -0,0 +1,134 @@
+/**
+ * @file
+ * @brief Functions for a group scope to call tile TMA functions.
+ */
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void prefetch(ST &dst, const GL &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::prefetch<axis, policy, ST, GL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void prefetch(ST &dst, const GL &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::prefetch<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_async<axis, policy, ST, GL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_async<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_async<axis, policy, ST, PGL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_async<dim::ROW, cache_policy::NORMAL, ST, PGL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_add_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_add_async<axis, policy, ST, GL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_add_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_add_async<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_add_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_add_async<axis, policy, ST, PGL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_add_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_add_async<dim::ROW, cache_policy::NORMAL, ST, PGL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_min_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_min_async<axis, policy, ST, GL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_min_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_min_async<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_min_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_min_async<axis, policy, ST, PGL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_min_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_min_async<dim::ROW, cache_policy::NORMAL, ST, PGL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_max_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_max_async<axis, policy, ST, GL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_max_async(const GL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_max_async<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_max_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_max_async<axis, policy, ST, PGL, COORD>(dst, src, idx); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::pgl::all PGL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void store_max_async(const PGL &dst, const ST &src, const COORD &idx) {
+    if(laneid() == 0) {
+        ::kittens::tma::store_max_async<dim::ROW, cache_policy::NORMAL, ST, PGL, COORD>(dst, src, idx);
+    }
+}
+
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx, semaphore& bar) {
+    if(laneid() == 0) {
+        ::kittens::tma::load_async<axis, policy, ST, GL, COORD>(dst, src, idx, bar); // Don't do the mask
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx, semaphore& bar) {
+    if(laneid() == 0) {
+        ::kittens::tma::load_async<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx, bar);
+    }
+}
@@ -0,0 +1,33 @@
+/**
+ * @file
+ * @brief Functions for a group scope to call tile TMA cluster functions.
+ */
+
+
+#ifdef KITTENS_BLACKWELL
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx, semaphore& bar, uint16_t cluster_mask, int dst_mbar_cta=-1) {
+    if(laneid() == 0) {
+        ::kittens::tma::cluster::load_async<axis, policy, ST, GL, COORD>(dst, src, idx, bar, cluster_mask, dst_mbar_cta);
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx, semaphore& bar, uint16_t cluster_mask, int dst_mbar_cta=-1) {
+    if(laneid() == 0) {
+        ::kittens::tma::cluster::load_async<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx, bar, cluster_mask, dst_mbar_cta);
+    }
+}
+#else
+template<int axis, cache_policy policy, ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx, semaphore& bar, uint16_t cluster_mask) {
+    if(laneid() == 0) {
+        ::kittens::tma::cluster::load_async<axis, policy, ST, GL, COORD>(dst, src, idx, bar, cluster_mask);
+    }
+}
+template<ducks::st::all ST, ducks::gl::all GL, ducks::coord::tile COORD=coord<ST>>
+__device__ static inline void load_async(ST &dst, const GL &src, const COORD &idx, semaphore& bar, uint16_t cluster_mask) {
+    if(laneid() == 0) {
+        ::kittens::tma::cluster::load_async<dim::ROW, cache_policy::NORMAL, ST, GL, COORD>(dst, src, idx, bar, cluster_mask);
+    }
+}
+#endif
@@ -0,0 +1,68 @@
+/**
+ * @file
+ * @brief Various utilities for group TMA memory operations.
+ */
+
+/* ----------   Barrier functions for async load  ---------- */
+
+/**
+* @brief Sets the number of bytes expected at the semaphore.
+*
+* This function sets the number of bytes expected at the semaphore for the first thread in the warp.
+* It converts the semaphore pointer to a generic shared memory pointer and uses an inline assembly
+* instruction to set the expected number of bytes.
+*
+* @param semaphore Reference to the semaphore variable.
+* @param bytes The number of bytes expected at the semaphore.
+*/
+__device__ static inline void expect_bytes(semaphore& bar, uint32_t bytes) {
+    if(laneid() == 0) {
+        ::kittens::tma::expect_bytes(bar, bytes);
+    }
+}
+/**
+* @brief Sets the number of bytes expected at the semaphore.
+*
+* This function sets the number of bytes expected at the mbarrier before the transaction arrives.
+*/
+template<typename T, typename... args>
+__device__ static inline void expect(semaphore& bar, const T& _1, const args&... _2) {
+    expect_bytes(bar, size_bytes<T, args...>);
+}
+
+/* ----------   Synchronization functions for async store  ---------- */
+
+/**
+ * @brief Commits previous asynchronous TMA stores to a group and performs them.
+*/
+__device__ static inline void store_commit_group() {
+    asm volatile("cp.async.bulk.commit_group;");
+}
+/**
+ * @brief Waits for previous committed TMA store groups to complete.
+ *
+ * @tparam N The maximum number of remaining TMA store groups. Defaults to 0.
+*/
+template <int N=0>
+__device__ static inline void store_async_wait() {
+    asm volatile (
+        "cp.async.bulk.wait_group %0;"
+        :
+        : "n"(N)
+        : "memory"
+    );
+}
+/**
+ * @brief Waits for previous committed TMA store groups to finish reading from shared memory.
+ *
+ * @tparam N The maximum number of remaining TMA store groups. Defaults to 0.
+*/
+template <int N=0>
+__device__ static inline void store_async_read_wait() {
+    asm volatile (
+        "cp.async.bulk.wait_group.read %0;"
+        :
+        : "n"(N)
+        : "memory"
+    );
+}
@@ -0,0 +1,90 @@
+
+/**
+* @brief Waits for the requested semaphore phase, at cluster scope
+*
+* @param semaphore Reference to the semaphore variable.
+* @param kPhaseBit The phase bit used for the semaphore.
+*/
+__device__ static inline void wait(semaphore& bar, int kPhaseBit) {
+    void const* const ptr = &bar;
+    uint32_t mbar_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(ptr)); 
+
+    asm volatile (
+        "{\n"
+        ".reg .pred                P1;\n"
+        "LAB_WAIT:\n"
+        "mbarrier.try_wait.parity.acquire.cluster.shared::cta.b64 P1, [%0], %1;\n"
+        "@P1                       bra.uni DONE;\n"
+        "bra.uni                   LAB_WAIT;\n"
+        "DONE:\n"
+        "}\n"
+        :: "r"(mbar_ptr),
+        "r"(kPhaseBit)
+    );
+}
+
+/**
+* @brief Sets the number of bytes expected at the semaphore, assuming a multicast instruction.
+*
+* This function sets the number of bytes expected at the semaphore for the first thread in the warp.
+* It converts the semaphore pointer to a generic shared memory pointer and uses an inline assembly
+* instruction to set the expected number of bytes.
+* 
+* It's worth being aware that this function is particularly necessary for multicast loads, and
+* distributed shared memory can actually be done with a normal tma::expect followed by wait. See
+* the unit tests of dsmem for an example.
+*
+* @param semaphore Reference to the semaphore variable.
+* @param bytes The number of bytes expected at the semaphore.
+*/
+__device__ static inline void expect_bytes(semaphore& bar, uint32_t bytes, int dst_cta) {
+    if(laneid() == 0) {
+        ::kittens::tma::cluster::expect_bytes(bar, bytes, dst_cta);
+    }
+}
+/**
+* @brief Sets the number of bytes expected at the semaphore.
+*
+* This function sets the number of bytes expected at the semaphore for the first thread in the warp.
+* It converts the semaphore pointer to a generic shared memory pointer and uses an inline assembly
+* instruction to set the expected number of bytes.
+*
+* @tparam T The type of the data to be stored at the semaphore.
+* @param semaphore Reference to the semaphore variable.
+*/
+/**
+* @brief Sets the number of bytes expected at the semaphore.
+*
+* This function sets the number of bytes expected at the mbarrier before the transaction arrives.
+*/
+template<typename T, typename... args>
+__device__ static inline void expect(semaphore& bar, int dst_cta, const T& _1, const args&... _2) {
+    expect_bytes(bar, size_bytes<T, args...>, dst_cta);
+}
+
+/**
+* @brief Arrives at a semaphore in cluster scope.
+*
+* Marks a thread arrival at an mbarrier
+*
+* @param semaphore Reference to the semaphore variable.
+* @param kPhaseBit The phase bit used for the semaphore.
+*/
+__device__ static inline void arrive(semaphore& bar, int dst_cta, uint32_t count=1) {
+    if(laneid() == 0) {
+        ::kittens::tma::cluster::arrive(bar, dst_cta, count);
+    }
+}
+
+// Generic transfer
+__device__ static inline void store_async(void *dst, void *src, int dst_cta, uint32_t size_bytes, semaphore& bar) {
+    if(laneid() == 0) {
+        ::kittens::tma::cluster::store_async(dst, src, dst_cta, size_bytes, bar);
+    }
+}
+
+// Templated transfer for convenience
+template<typename T>
+__device__ static inline void store_async(T &dst_, T &src_, int dst_cta, semaphore& bar) {
+    store_async((void*)&dst_, (void*)&src_, dst_cta, size_bytes<T>, bar);
+}
@@ -0,0 +1,168 @@
+/**
+ * @file
+ * @brief Various utilities for group memory operations.
+ */
+
+
+template<int N=0> __device__ static inline void load_async_wait(int bar_id) { // for completing (non-TMA) async loads
+    asm volatile("cp.async.wait_group %0;\n" : : "n"(N) : "memory");
+    sync(bar_id);
+}
+template<int N=0> __device__ static inline void load_async_wait() { // for completing (non-TMA) async loads
+    KITTENS_CHECK_WARP
+    asm volatile("cp.async.wait_group %0;\n" : : "n"(N) : "memory");
+    __syncwarp();
+}
+
+__device__ static inline void arrive(barrier<GROUP_WARPS> bar) {
+    asm volatile("bar.arrive %0, %1;\n" :: "r"(bar.barrier_id), "n"(GROUP_WARPS*WARP_THREADS) : "memory");
+}
+__device__ static inline void arrive_and_wait(barrier<GROUP_WARPS> bar) {
+    asm volatile("bar.sync %0, %1;\n" :: "r"(bar.barrier_id), "n"(GROUP_WARPS*WARP_THREADS) : "memory");
+}
+
+/**
+ * @brief Initializes a synchronization semaphore with a transaction count and sets the expected number of bytes.
+ *
+ * This function sets up a semaphore that is used to synchronize threads within a block during asynchronous operations.
+ * It initializes the semaphore with a thread count semaphore.
+ *
+ * Additionally, if it is given a shared tile type, it will also call `set_bytes` to prepare for the memory transaction.
+ *
+ * @param[out] semaphore The semaphore variable to initialize.
+ * @param[in] tc The thread counter for the semaphore.
+ */
+__device__ static inline void init_semaphore(semaphore& bar, int thread_count, int transaction_count=0) {
+    if (laneid() == 0) {
+        void const* const ptr = &bar;
+        uint32_t bar_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(ptr)); 
+
+        asm volatile (
+            "mbarrier.init.shared::cta.b64 [%0], %1;\n"
+            :: "r"(bar_ptr), "r"(thread_count+transaction_count)
+        );
+    }
+}
+/**
+ * @brief Invalidate an mbarrier
+ *
+ * @param[out] semaphore The semaphore variable to initialize.
+ * @param[in] tc The thread counter for the semaphore.
+ */
+__device__ static inline void invalidate_semaphore(semaphore& bar) {
+    if (laneid() == 0) {
+        void const* const ptr = &bar;
+        uint32_t bar_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(ptr)); 
+        asm volatile (
+            "mbarrier.inval.shared::cta.b64 [%0];\n"
+            :: "r"(bar_ptr)
+        );
+    }
+}
+
+/**
+* @brief Arrives at a semaphore.
+*
+* Marks a warp arrival at an mbarrier
+*
+* @param semaphore Reference to the semaphore variable.
+* @param kPhaseBit The phase bit used for the semaphore.
+*/
+__device__ static inline void arrive(semaphore& sem) {
+    if(laneid() == 0) {
+            uint32_t mbar_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(&sem)); 
+            asm volatile (
+                "mbarrier.arrive.release.cta.shared::cta.b64 _, [%0];\n"
+            :
+            : "r"(mbar_ptr)
+            : "memory"
+        );
+    }
+}
+template<int num_warps> __device__ static inline void arrive(barrier<num_warps> bar) {
+    asm volatile("bar.arrive %0, %1;\n" :: "r"(bar.barrier_id), "n"(num_warps*WARP_THREADS) : "memory");
+}
+
+#ifdef KITTENS_HOPPER
+/**
+* @brief Arrives at a semaphore.
+*
+* Marks a warp arrival at an mbarrier
+*
+* @param semaphore Reference to the semaphore variable.
+* @param kPhaseBit The phase bit used for the semaphore.
+*/
+__device__ static inline void arrive(semaphore& sem, uint32_t count) {
+    if(laneid() == 0) {
+        uint32_t mbar_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(&sem));
+        asm volatile (
+            "mbarrier.arrive.release.cta.shared::cta.b64 _, [%0], %1;\n"
+            :
+            : "r"(mbar_ptr), "r"(count)
+            : "memory"
+        );
+    }
+}
+#endif
+
+/**
+* @brief Waits for the requested semaphore phase.
+*
+* @param semaphore Reference to the semaphore variable.
+* @param kPhaseBit The phase bit used for the semaphore.
+*/
+__device__ static inline void wait(semaphore& sem, int kPhaseBit) {
+    void const* const ptr = &sem;
+    uint32_t mbar_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(ptr)); 
+
+#ifdef KITTENS_HOPPER
+    asm volatile (
+        "{\n"
+        ".reg .pred                P1;\n"
+        "LAB_WAIT:\n"
+        "mbarrier.try_wait.parity.shared::cta.b64 P1, [%0], %1;\n"
+        "@P1                       bra.uni DONE;\n"
+        "bra.uni                   LAB_WAIT;\n"
+        "DONE:\n"
+        "}\n"
+        :: "r"(mbar_ptr),
+        "r"(kPhaseBit)
+    );
+#else
+    asm volatile (
+        "{\n"
+        ".reg .pred                P1;\n"
+        "LAB_WAIT:\n"
+        "mbarrier.test_wait.parity.shared::cta.b64 P1, [%0], %1;\n"
+        "@P1                       bra.uni DONE;\n"
+        "nanosleep.u32 5;\n" // wait a few nanoseconds on pre-Hopper architectures to save instruction issue slots
+        "bra.uni                   LAB_WAIT;\n"
+        "DONE:\n"
+        "}\n"
+        :: "r"(mbar_ptr),
+        "r"(kPhaseBit)
+    );
+#endif
+}
+
+/**
+* @brief Checks if the requested semaphore phase is ready.
+*
+* @param semaphore Reference to the semaphore variable.
+* @param kPhaseBit The phase bit used for the semaphore.
+*/
+__device__ static inline int test_wait(semaphore& sem, int kPhaseBit) {
+    void const* const ptr = &sem;
+    uint32_t mbar_ptr = static_cast<uint32_t>(__cvta_generic_to_shared(ptr));
+    int result;
+    asm volatile (
+        "{\n"
+        ".reg .pred P1;\n"
+        "mbarrier.test_wait.parity.shared::cta.b64 P1, [%1], %2;\n"
+        "selp.u32 %0,1,0,P1;"
+        "}\n"
+        : "=r"(result)
+        : "r"(mbar_ptr), "r"(kPhaseBit)
+    );
+    return result;
+}
--- a/Show More
+++ b/Show More