cleanup + another way

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

.github/pins/pytorch.txt

@@ -1 +1 @@
-3560b8ebe9277e8c25335e35d8c9e0872052b2dc
+29e6033ff38f0deb8dde1146a1612e64ed00e3b7

.github/workflows/e2e-reusable.yml

@@ -203,22 +203,9 @@ jobs:
         if: ${{ inputs.suite == 'torchbench' }}
         run: |
           cd benchmark
-          if [[ "${{ inputs.only_one_model }}" ]]; then
-            python install.py "${{ inputs.only_one_model }}"
-          else
-            # install all models
-            python install.py
-          fi
+          python install.py
           pip install -e .
 
-      - name: Install torchvision package from PyTorch wheel
-        if: ${{ inputs.suite == 'timm_models' || inputs.suite == 'torchbench' }}
-        shell: bash
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          pip install --no-deps torchvision==0.22.0.dev20250525+xpu --pre --index-url https://download.pytorch.org/whl/nightly/xpu
-          python -c "import torchvision; print(torchvision.__version__)"
-
       - name: Run e2e ${{ inputs.test_mode }} tests
         env:
           HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}

scripts/inductor_xpu_test.sh

@@ -51,4 +51,11 @@ if (( $EUID == 0 )); then
     ulimit -n 1048576
 fi
 
-python benchmarks/dynamo/torchbench.py --accuracy --float32 -d xpu -n10 --inference --only detectron2_fasterrcnn_r_50_fpn --backend=inductor --cold-start-latency --disable-cudagraphs
+if [[ $DT == "amp_bf16" ]]; then
+    ZE_AFFINITY_MASK=${CARD} python benchmarks/dynamo/${SUITE}.py --${SCENARIO} --amp -d${DEVICE} -n10 --no-skip --dashboard ${Mode_extra} ${Shape_extra} ${partition_flags} ${Model_only_extra} --backend=inductor --timeout=4800 --output=${LOG_DIR}/${LOG_NAME}.csv 2>&1 | tee ${LOG_DIR}/${LOG_NAME}.log
+elif [[ $DT == "amp_fp16" ]]; then
+    export INDUCTOR_AMP_DT=float16
+    ZE_AFFINITY_MASK=${CARD} python benchmarks/dynamo/${SUITE}.py --${SCENARIO} --amp -d${DEVICE} -n10 --no-skip --dashboard ${Mode_extra} ${Shape_extra} ${partition_flags} ${Model_only_extra}  --backend=inductor --timeout=4800 --output=${LOG_DIR}/${LOG_NAME}.csv 2>&1 | tee ${LOG_DIR}/${LOG_NAME}.log
+else
+    ZE_AFFINITY_MASK=${CARD} python benchmarks/dynamo/${SUITE}.py --${SCENARIO} --${DT} -d${DEVICE} -n10 --no-skip --dashboard ${Mode_extra} ${Shape_extra} ${partition_flags} ${Model_only_extra}  --backend=inductor --timeout=4800 --output=${LOG_DIR}/${LOG_NAME}.csv 2>&1 | tee ${LOG_DIR}/${LOG_NAME}.log
+fi

scripts/patch-pytorch.sh

@@ -35,3 +35,4 @@ apply_patch() {
 echo "Applying PyTorch patches in $REPO_ROOT"
 
 # put your patch applies here
+apply_patch https://github.com/pytorch/pytorch/pull/143553.diff

third_party/intel/lib/Target/SPIRV/SPIRVTranslation.cpp

@@ -107,13 +107,14 @@ class SmallVectorBuffer : public std::streambuf {
 
 static SPIRV::TranslatorOpts getSPIRVOpts() {
   SPIRV::TranslatorOpts SPIRVOpts{SPIRV::VersionNumber::SPIRV_1_4};
-  static constexpr std::array<SPIRV::ExtensionID, 17> AllowedExtensions{
+  static constexpr std::array<SPIRV::ExtensionID, 18> AllowedExtensions{
       SPIRV::ExtensionID::SPV_EXT_shader_atomic_float_add,
       SPIRV::ExtensionID::SPV_INTEL_2d_block_io,
       SPIRV::ExtensionID::SPV_INTEL_arbitrary_precision_integers,
       SPIRV::ExtensionID::SPV_INTEL_arithmetic_fence,
       SPIRV::ExtensionID::SPV_INTEL_bfloat16_conversion,
       SPIRV::ExtensionID::SPV_INTEL_cache_controls,
+      SPIRV::ExtensionID::SPV_INTEL_fp_fast_math_mode,
       SPIRV::ExtensionID::SPV_INTEL_inline_assembly,
       SPIRV::ExtensionID::SPV_INTEL_kernel_attributes,
       SPIRV::ExtensionID::SPV_INTEL_memory_access_aliasing,

third_party/intel/triton_xpu.cc

@@ -293,7 +293,7 @@ void init_triton_intel(py::module &&m) {
         if (auto *op = dyn_cast<FPMathOperator>(&inst)) {
           FastMathFlags FMF;
           // Default to allow contract when default fp fusion is not disabled.
-          if ((!enableFpFusion.has_value() || enableFpFusion.value()) &&
+          if ((enableFpFusion.has_value() && enableFpFusion.value()) &&
               !fastMath.has_value()) {
             if (op->getOpcode() == Instruction::FAdd ||
                 op->getOpcode() == Instruction::FMul)