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

fix profile run for vl model (#5136)

Browse Source 
### What this PR does / why we need it?

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>
This commit is contained in:
shaopeng-666
2025-12-17 23:51:31 +08:00
committed by GitHub
parent 43d974c6f7
commit 39bdd4cfaa
4 changed files with 1 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tests/e2e/models/configs/Qwen3-VL-30B-A3B-Instruct.yaml
View File

@@ -6,7 +6,6 @@ tasks:
metrics: metrics:
- name: "acc,none" - name: "acc,none"
value: 0.58 value: 0.58
max_model_len: 8192
tensor_parallel_size: 2 tensor_parallel_size: 2
gpu_memory_utilization: 0.7 gpu_memory_utilization: 0.7
enable_expert_parallel: True enable_expert_parallel: True

1
tests/e2e/models/configs/Qwen3-VL-8B-Instruct.yaml
View File

@@ -6,6 +6,5 @@ tasks:
metrics: metrics:
- name: "acc,none" - name: "acc,none"
value: 0.55 value: 0.55
max_model_len: 8192
batch_size: 32 batch_size: 32
gpu_memory_utilization: 0.7 gpu_memory_utilization: 0.7

1
tests/e2e/singlecard/test_vlm.py
View File

@@ -39,7 +39,6 @@ def test_multimodal_vl(prompt_template):
images = [image] * len(img_questions) images = [image] * len(img_questions)
prompts = prompt_template(img_questions) prompts = prompt_template(img_questions)
with VllmRunner("Qwen/Qwen3-VL-8B-Instruct", with VllmRunner("Qwen/Qwen3-VL-8B-Instruct",
max_model_len=4096,
mm_processor_kwargs={ mm_processor_kwargs={
"min_pixels": 28 * 28, "min_pixels": 28 * 28,
"max_pixels": 1280 * 28 * 28, "max_pixels": 1280 * 28 * 28,

1
vllm_ascend/worker/worker_v1.py
View File

@@ -232,6 +232,7 @@ class NPUWorker(WorkerBase):
# Init ModelRunner here, so that we have access to self.device. # Init ModelRunner here, so that we have access to self.device.
self.model_runner = NPUModelRunner(self.vllm_config, self.device) self.model_runner = NPUModelRunner(self.vllm_config, self.device)
@torch.inference_mode()
def determine_available_memory(self) -> int: def determine_available_memory(self) -> int:
# Profile the memory usage of the model and get the maximum number of # Profile the memory usage of the model and get the maximum number of
# cache blocks that can be allocated with the remaining free memory. # cache blocks that can be allocated with the remaining free memory.