xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-16 14:17:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

[CI][AMD] spec_decode:eagle skip FLASH_ATTN for deepseek on ROCm (#29827)

Browse Source 
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
This commit is contained in:
Divakar Verma 2025-12-01 23:27:46 -06:00 committed by GitHub
parent a690fb5bd6
commit e2fbfc955e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tests/v1/e2e/test_spec_decode.py
View File

@ -414,7 +414,10 @@ def test_eagle_correctness(
) )
if attn_backend == "FLASH_ATTN" and current_platform.is_rocm(): if attn_backend == "FLASH_ATTN" and current_platform.is_rocm():
m.setenv("VLLM_ROCM_USE_AITER", "1") if "deepseek" in model_setup[1].lower():
pytest.skip("FLASH_ATTN for deepseek not supported on ROCm platform")
else:
m.setenv("VLLM_ROCM_USE_AITER", "1")
method, model_name, spec_model_name, tp_size = model_setup method, model_name, spec_model_name, tp_size = model_setup
max_model_len = 2048 max_model_len = 2048