A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 02:00 UTC
[Bugfix] Fix GPT-OSS on AMD after #28603 (#28816)
Signed-off-by: zhewenli <zhewenli@meta.com>
Zhewen Li committed
Nov 17, 2025 at 10:15 UTC
f8b19c0ffd65f7f6f01a0da4a39b6890f5db40cb
2 files changed
+7
-6
.buildkite/test-amd.yaml
+5
-4
@@ -1068,7 +1068,7 @@ steps:
1068
# this runner has 2 GPUs available even though num_gpus=2 is not set
1069
- pytest -v -s tests/compile/test_fusion_all_reduce.py
1070
# Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
1071
- # Wrap with quotes to escape yaml
1071
+ # Wrap with quotes to escape yaml
1072
- "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'"
1073
1074
- label: Blackwell Fusion E2E Tests # 30 min
@@ -1095,10 +1095,11 @@ steps:
1095
# test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
1096
- pytest -v -s tests/compile/test_full_graph.py::test_fp8_kv_scale_compile
1097
1098
-- label: Blackwell GPT-OSS Eval
1098
+- label: ROCm GPT-OSS Eval
1099
timeout_in_minutes: 60
1100
working_dir: "/vllm-workspace/"
1101
- gpu: b200
1101
+ agent_pool: mi325_1
1102
+ mirror_hardwares: [amdproduction]
1103
optional: true # run on nightlies
1104
source_file_dependencies:
1105
- tests/evals/gpt_oss
@@ -1107,7 +1108,7 @@ steps:
1108
- vllm/v1/attention/backends/flashinfer.py
1109
commands:
1110
- uv pip install --system 'gpt-oss[eval]==0.0.5'
1110
- - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
1111
+ - 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
1112
1113
- label: Blackwell Quantized MoE Test
1114
timeout_in_minutes: 60
vllm/model_executor/layers/quantization/mxfp4.py
+2
-2
@@ -755,8 +755,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
755
756
self.w13_weight = w13_weight
757
self.w2_weight = w2_weight
758
- layer.w13_weight = Parameter(w13_weight.data, requires_grad=False)
759
- layer.w2_weight = Parameter(w2_weight.data, requires_grad=False)
758
+ layer.w13_weight = Parameter(w13_weight.storage.data, requires_grad=False)
759
+ layer.w2_weight = Parameter(w2_weight.storage.data, requires_grad=False)
760
else:
761
raise ValueError(f"Unsupported backend: {self.mxfp4_backend}")
762