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 12, 06:00 UTC
[CI][AMD] spec_decode:eagle skip FLASH_ATTN for deepseek on ROCm (#29827)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Divakar Verma committed
Dec 1, 2025 at 23:27 UTC
e2fbfc955e1bc9f67c349c5a2bde63a3edd86b84
1 file changed
+4
-1
tests/v1/e2e/test_spec_decode.py
+4
-1
@@ -414,7 +414,10 @@ def test_eagle_correctness(
414
)
415
416
if attn_backend == "FLASH_ATTN" and current_platform.is_rocm():
417
- m.setenv("VLLM_ROCM_USE_AITER", "1")
417
+ if "deepseek" in model_setup[1].lower():
418
+ pytest.skip("FLASH_ATTN for deepseek not supported on ROCm platform")
419
+ else:
420
+ m.setenv("VLLM_ROCM_USE_AITER", "1")
421
422
method, model_name, spec_model_name, tp_size = model_setup
423
max_model_len = 2048