[CI/Build][AMD] Fix ROCm OOM in eagle_correctness_heavy by reserving CUDA graph memory (#47366)

Signed-off-by: pei.zhang <pei.zhang@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
peizhang56
2026-07-10 02:14:38 +00:00
committed by GitHub
co-authored by Claude Andreas Karatzas
parent feb384ada2
commit 88e5e2c57b
2 changed files with 14 additions and 5 deletions
@@ -17,6 +17,7 @@ from tests.utils import (
multi_gpu_marks,
multi_gpu_only,
single_gpu_only,
wait_for_rocm_memory_to_settle,
)
from vllm import LLM, SamplingParams
from vllm.assets.base import VLLM_S3_BUCKET_URL
@@ -448,6 +449,9 @@ def _run_eagle_correctness(
del ref_llm
torch.accelerator.empty_cache()
cleanup_dist_env_and_memory()
# ROCm frees VRAM lazily; wait so the spec engine started right after
# does not OOM on its startup memory guard.
wait_for_rocm_memory_to_settle()
spec_llm = LLM(
model=model_name,
@@ -485,6 +489,9 @@ def _run_eagle_correctness(
del spec_llm
torch.accelerator.empty_cache()
cleanup_dist_env_and_memory()
# ROCm frees VRAM lazily; wait so the next parametrization's engine does
# not OOM on its startup memory guard.
wait_for_rocm_memory_to_settle()
@single_gpu_only