xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-11 00:15:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix GPT-OSS on AMD after #28603 (#28816)

Browse Source 
Signed-off-by: zhewenli <zhewenli@meta.com>
This commit is contained in:
Zhewen Li 2025-11-17 10:15:26 -08:00 committed by GitHub
parent e42bd8c2e3
commit f8b19c0ffd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
.buildkite/test-amd.yaml
View File

@ -1068,7 +1068,7 @@ steps:
# this runner has 2 GPUs available even though num_gpus=2 is not set # this runner has 2 GPUs available even though num_gpus=2 is not set
- pytest -v -s tests/compile/test_fusion_all_reduce.py - pytest -v -s tests/compile/test_fusion_all_reduce.py
# Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
# Wrap with quotes to escape yaml # Wrap with quotes to escape yaml
- "pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and Llama-3.1 and -quant_fp8 and -rms_norm'" - "pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and Llama-3.1 and -quant_fp8 and -rms_norm'"
- label: Blackwell Fusion E2E Tests # 30 min - label: Blackwell Fusion E2E Tests # 30 min
@ -1095,10 +1095,11 @@ steps:
# test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40) # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
- pytest -v -s tests/compile/test_full_graph.py::test_fp8_kv_scale_compile - pytest -v -s tests/compile/test_full_graph.py::test_fp8_kv_scale_compile
- label: Blackwell GPT-OSS Eval - label: ROCm GPT-OSS Eval
timeout_in_minutes: 60 timeout_in_minutes: 60
working_dir: "/vllm-workspace/" working_dir: "/vllm-workspace/"
gpu: b200 agent_pool: mi325_1
mirror_hardwares: [amdproduction]
optional: true # run on nightlies optional: true # run on nightlies
source_file_dependencies: source_file_dependencies:
- tests/evals/gpt_oss - tests/evals/gpt_oss
@ -1107,7 +1108,7 @@ steps:
- vllm/v1/attention/backends/flashinfer.py - vllm/v1/attention/backends/flashinfer.py
commands: commands:
- uv pip install --system 'gpt-oss[eval]==0.0.5' - uv pip install --system 'gpt-oss[eval]==0.0.5'
- pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58 - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
- label: Blackwell Quantized MoE Test - label: Blackwell Quantized MoE Test
timeout_in_minutes: 60 timeout_in_minutes: 60

4
vllm/model_executor/layers/quantization/mxfp4.py
View File

@ -755,8 +755,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
self.w13_weight = w13_weight self.w13_weight = w13_weight
self.w2_weight = w2_weight self.w2_weight = w2_weight
layer.w13_weight = Parameter(w13_weight.data, requires_grad=False) layer.w13_weight = Parameter(w13_weight.storage.data, requires_grad=False)
layer.w2_weight = Parameter(w2_weight.data, requires_grad=False) layer.w2_weight = Parameter(w2_weight.storage.data, requires_grad=False)
else: else:
raise ValueError(f"Unsupported backend: {self.mxfp4_backend}") raise ValueError(f"Unsupported backend: {self.mxfp4_backend}")