iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm][Bugfix] Fix HIP fork re-init in multimodal offline examples (#46741)

Browse Source 
Signed-off-by: pei.zhang <pei.zhang@amd.com>
This commit is contained in:
peizhang56
2026-06-26 00:37:47 -05:00
committed by GitHub
parent 5b33041746
commit 915e99ec67
1 changed files with 14 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.buildkite/test-amd.yaml
+14 -5
View File
@@ -1374,8 +1374,11 @@ steps:
- python3 basic/offline_inference/score.py
# Multi-modal models
- python3 generate/multimodal/audio_language_offline.py --seed 0
- python3 generate/multimodal/vision_language_offline.py --seed 0
- python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
# These two examples import transformers before vllm, so on ROCm the HIP context
# is initialized in the parent before vllm sets this guard, poisoning fork. Set it
# inline to keep torch.cuda.is_available() fork-safe (see vllm/env_override.py).
- PYTORCH_NVML_BASED_CUDA_CHECK=1 python3 generate/multimodal/vision_language_offline.py --seed 0
- PYTORCH_NVML_BASED_CUDA_CHECK=1 python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
- python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
# Pooling models
- python3 pooling/embed/vision_embedding_offline.py --seed 0
@@ -1817,7 +1820,10 @@ steps:
- pytest -v -s tests/models/test_transformers.py
- pytest -v -s tests/models/multimodal/test_mapping.py
- python3 examples/basic/offline_inference/chat.py
- python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
# This example imports transformers before vllm, so on ROCm the HIP context is
# initialized in the parent before vllm sets this guard, poisoning fork. Set it
# inline to keep torch.cuda.is_available() fork-safe (see vllm/env_override.py).
- PYTORCH_NVML_BASED_CUDA_CHECK=1 python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
- VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper
#------------------------------------------------------- mi300 · quantization --------------------------------------------------------#
@@ -2893,8 +2899,11 @@ steps:
- python3 basic/offline_inference/score.py
# Multi-modal models
- python3 generate/multimodal/audio_language_offline.py --seed 0
- python3 generate/multimodal/vision_language_offline.py --seed 0
- python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
# These two examples import transformers before vllm, so on ROCm the HIP context
# is initialized in the parent before vllm sets this guard, poisoning fork. Set it
# inline to keep torch.cuda.is_available() fork-safe (see vllm/env_override.py).
- PYTORCH_NVML_BASED_CUDA_CHECK=1 python3 generate/multimodal/vision_language_offline.py --seed 0
- PYTORCH_NVML_BASED_CUDA_CHECK=1 python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
- python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
# Pooling models
- python3 pooling/embed/vision_embedding_offline.py --seed 0