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

Eagle3 mm support, enablement on qwen3vl (#4848)

Browse Source 
### What this PR does / why we need it?
follow pr
[https://github.com/vllm-project/vllm/pull/20788](https://github.com/vllm-project/vllm/pull/20788)
, Eagle3 mm support, enablement on qwen3vl
target model
[Qwen/Qwen3-VL-8B-Instruct]([https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct])
eagle3
[MNN/Qwen3-VL-8B-Instruct-Eagle3](https://www.modelscope.cn/models/MNN/Qwen3-VL-8B-Instruct-Eagle3)
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?

pytest ./tests/e2e/singlecard/test_completion_with_prompt_embeds.py -vv

vLLM with eagle3 :
```bash
vllm serve /model/Qwen3-VL-8B-Instruct   --enforce-eager   --port 9100    --max-model-len 32768   --max-num-seqs 32   --tensor-parallel-size 2   --allowed-local-media-path /model/gx/images  --speculative-config '{
    "method": "eagle3",
    "model": "/model/hf/Qwen3-VL-8B-Instruct-Eagle3",
    "num_speculative_tokens": 3
  }'
```
vLLM without eagle3 :
```bash
vllm serve /model/Qwen3-VL-8B-Instruct   --enforce-eager   --port 9100    --max-model-len 32768   --max-num-seqs 32   --tensor-parallel-size 2   --allowed-local-media-path /model/gx/images 
```

bench:
```
vllm bench serve   --backend openai-chat   --base-url http://127.0.0.1:9100   --tokenizer /model/Qwen3-VL-8B-Instruct   --endpoint /v1/chat/completions   --model /model/Qwen3-VL-8B-Instruct   --dataset-name random  --num-prompts 50   --max-concurrency 5   --temperature 0   --top-p 1.0   --seed 123
```

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

---------

Signed-off-by: jesse <szxfml@gmail.com>
This commit is contained in:
Song Zhixin
2026-01-19 08:58:07 +08:00
committed by GitHub
parent 05e69b99e5
commit 2b6dc100b5
5 changed files with 145 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
tests/e2e/singlecard/spec_decode/test_v1_spec_decode.py
View File

@@ -85,6 +85,14 @@ def eagle3_model_name():
return "vllm-ascend/EAGLE3-LLaMA3.1-Instruct-8B"
@pytest.fixture
def vl_model_name():
return "Qwen/Qwen3-VL-8B-Instruct"
def vl_eagle3_model_name():
return "MNN/Qwen3-VL-8B-Instruct-Eagle3"
def test_ngram_correctness(
test_prompts: list[list[dict[str, Any]]],
sampling_config: SamplingParams,
@@ -129,6 +137,48 @@ def test_ngram_correctness(
assert matches > int(0.66 * len(ref_outputs))
def test_qwen3_vl_eagle_correctness(
test_prompts: list[list[dict[str, Any]]],
sampling_config: SamplingParams,
vl_model_name: str,
):
'''
Compare the outputs of a original LLM and a speculative LLM
should be the same when using eagle speculative decoding.
'''
with VllmRunner(
vl_model_name,
max_model_len=1024,
cudagraph_capture_sizes=[1, 2, 4, 8],
) as ref_llm:
ref_outputs = ref_llm.model.chat(test_prompts, sampling_config)
spec_model_name = vl_eagle3_model_name()
with VllmRunner(
vl_model_name,
speculative_config={
"method": "eagle3",
"model": spec_model_name,
"num_speculative_tokens": 2,
},
max_model_len=1024,
cudagraph_capture_sizes=[1, 2, 4, 8],
) as runner:
spec_outputs = runner.model.chat(test_prompts, sampling_config)
matches = 0
misses = 0
for ref_output, spec_output in zip(ref_outputs, spec_outputs):
if ref_output.outputs[0].text == spec_output.outputs[0].text:
matches += 1
else:
misses += 1
print(f"ref_output: {ref_output.outputs[0].text}")
print(f"spec_output: {spec_output.outputs[0].text}")
# Heuristic: expect at least 70% of the prompts to match exactly
# Upon failure, inspect the outputs to check for inaccuracy.
assert matches > int(0.66 * len(ref_outputs))
def test_suffix_correctness(
test_prompts: list[list[dict[str, Any]]],
sampling_config: SamplingParams,