EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[BugFix][Cherry-pick] Cherry-pick PR 3675 to v0.11.0-dev (#3732)

Browse Source 
This PR cherry-picks the bugfix related with running multi-modal models
with AscendScheduler to v0.11.0-dev

Signed-off-by: hw_whx <wanghexiang7@huawei.com>
Co-authored-by: hw_whx <wanghexiang7@huawei.com>
This commit is contained in:
whx
2025-10-25 09:41:51 +08:00
committed by GitHub
parent 12bc78d252
commit 5a2c5be229
2 changed files with 35 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
tests/e2e/singlecard/test_vlm.py
View File

@@ -55,6 +55,38 @@ def test_multimodal_vl(prompt_template):
assert output_str, "Generated output should not be empty."
def test_multimodal_ascend_scheduler(prompt_template):
image = ImageAsset("cherry_blossom") \
.pil_image.convert("RGB")
img_questions = [
"What is the content of this image?",
"Describe the content of this image in detail.",
"What's in the image?",
"Where is this image taken?",
]
images = [image] * len(img_questions)
prompts = prompt_template(img_questions)
with VllmRunner("Qwen/Qwen2.5-VL-3B-Instruct",
max_model_len=4096,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
},
},
mm_processor_kwargs={
"min_pixels": 28 * 28,
"max_pixels": 1280 * 28 * 28,
"fps": 1,
},
enforce_eager=True) as vllm_model:
outputs = vllm_model.generate_greedy(prompts=prompts,
images=images,
max_tokens=64)
assert len(outputs) == len(prompts)
for _, output_str in outputs:
assert output_str, "Generated output should not be empty."
def test_multimodal_audio():
audio_prompt = "".join([
f"Audio {idx+1}: <|audio_bos|><|AUDIO|><|audio_eos|>\n"

6
vllm_ascend/core/schedule_config.py
View File

@@ -26,7 +26,7 @@ MAX_INT = 2147483647
@dataclass
class AscendSchedulerConfig(SchedulerConfig):
enable_chunked_prefill: bool = False
max_long_partial_prefills: int = MAX_INT
max_long_partial_prefills: int = 1
long_prefill_token_threshold: int = MAX_INT
policy: str = "fcfs"
scheduler_cls: Union[str, Type[object]] = (
@@ -73,9 +73,9 @@ class AscendSchedulerConfig(SchedulerConfig):
"max_num_batched_tokens and makes vLLM reject longer "
"sequences. Please increase max_num_batched_tokens or "
"decrease max_model_len.")
# concurrent partial prefills. Default is inf
# concurrent partial prefills. Default is 1 meaning not enabled.
if self.max_long_partial_prefills is None:
self.max_long_partial_prefills = MAX_INT
self.max_long_partial_prefills = 1
self.long_prefill_token_threshold = MAX_INT
if self.long_prefill_token_threshold is None or \