From 7e0ef4084affa9de84904ba7726c46f53f4f6379 Mon Sep 17 00:00:00 2001 From: Michael Goin Date: Tue, 14 Oct 2025 19:41:43 -0400 Subject: [PATCH] [CI Failure] Fix torchao dep failure for Quantization Test (#26824) Signed-off-by: mgoin --- .buildkite/test-amd.yaml | 3 ++- .buildkite/test-pipeline.yaml | 3 ++- tests/quantization/test_compressed_tensors.py | 3 ++- vllm/model_executor/layers/quantization/rtn.py | 3 ++- 4 files changed, 8 insertions(+), 4 deletions(-) diff --git a/.buildkite/test-amd.yaml b/.buildkite/test-amd.yaml index b2a3a0a775ba..91f0b850575c 100644 --- a/.buildkite/test-amd.yaml +++ b/.buildkite/test-amd.yaml @@ -603,7 +603,8 @@ steps: # since torchao nightly is only compatible with torch nightly currently # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now # we can only upgrade after this is resolved - - pip install --pre torchao==0.13.0.dev20250814 --index-url https://download.pytorch.org/whl/nightly/cu128 + # TODO(jerryzh168): resolve the above comment + - uv pip install --system torchao==0.13.0 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ - label: LM Eval Small Models # 53min diff --git a/.buildkite/test-pipeline.yaml b/.buildkite/test-pipeline.yaml index ebe0602a1b5d..94c0944c838c 100644 --- a/.buildkite/test-pipeline.yaml +++ b/.buildkite/test-pipeline.yaml @@ -527,7 +527,8 @@ steps: # since torchao nightly is only compatible with torch nightly currently # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now # we can only upgrade after this is resolved - - pip install --pre torchao==0.13.0.dev20250814 --index-url https://download.pytorch.org/whl/nightly/cu128 + # TODO(jerryzh168): resolve the above comment + - uv pip install --system torchao==0.13.0 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ - label: LM Eval Small Models # 53min diff --git a/tests/quantization/test_compressed_tensors.py b/tests/quantization/test_compressed_tensors.py index ef7164c8813d..5aeb002238cf 100644 --- a/tests/quantization/test_compressed_tensors.py +++ b/tests/quantization/test_compressed_tensors.py @@ -697,7 +697,8 @@ def test_compressed_tensors_2of4_sparse_compressed(vllm_runner, args_2of4): @pytest.mark.parametrize( "args", [ - ("nm-testing/TinyLlama-1.1B-Chat-v1.0-NVFP4A16", CompressedTensorsW4A16Fp4), + # TODO: Enable once model is available again + # ("nm-testing/TinyLlama-1.1B-Chat-v1.0-NVFP4A16", CompressedTensorsW4A16Fp4), ("nm-testing/TinyLlama-1.1B-Chat-v1.0-NVFP4", CompressedTensorsW4A4Fp4), ], ) diff --git a/vllm/model_executor/layers/quantization/rtn.py b/vllm/model_executor/layers/quantization/rtn.py index c041d2fd0ba4..e4f7ff833956 100644 --- a/vllm/model_executor/layers/quantization/rtn.py +++ b/vllm/model_executor/layers/quantization/rtn.py @@ -15,6 +15,7 @@ from vllm.model_executor.layers.fused_moe.config import ( FusedMoEConfig, FusedMoEQuantConfig, ) +from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe from vllm.model_executor.layers.fused_moe.layer import FusedMoE, FusedMoEMethodBase from vllm.model_executor.layers.linear import ( LinearBase, @@ -396,7 +397,7 @@ class RTNMoEMethod(FusedMoEMethodBase): indices_type=self.topk_indices_dtype, ) - return torch.ops.vllm.fused_marlin_moe( + return fused_marlin_moe( x, layer.w13_weight, layer.w2_weight,