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

qwen3_moe/qwen25 support torchair graph (#2403)

Browse Source 
### What this PR does / why we need it?
Added support for the TorchAir graph mode in qwen3_moe and qwen2.5
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
```bash
llm = LLM(
    model=model,
    tensor_parallel_size=GPUs_per_dp_rank,
    enforce_eager=False,
    enable_expert_parallel=True,
    max_model_len=4096,
    max_num_seqs=16,
    trust_remote_code=trust_remote_code,
    gpu_memory_utilization=0.4,
    additional_config={
             "torchair_graph_config": {
                 "enabled": True,
                 "use_cached_graph": False,
                 "graph_batch_sizes_init": False,
                 "graph_batch_sizes": [16]
             },
             "ascend_scheduler_config": {
                 "enabled": True,
                 "chunked_prefill_enabled":True,
             },
             "refresh": True,
    },
)
```

- vLLM version: v0.10.0
- vLLM main:
b87cb97a53

Signed-off-by: taoyuxiang <oui.nicholas.tao@gmail.com>
This commit is contained in:
Nicholas Tao
2025-08-20 11:23:50 +08:00
committed by GitHub
parent 31ae249742
commit 7bec1a9b9c
9 changed files with 1123 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/ut/test_ascend_config.py
View File

@@ -232,7 +232,7 @@ class TestAscendConfig(TestBase):
def test_check_torchair_supported(self):
test_cases = [('deepseek_v3', True), ('PanguProMoE', True),
('qwen', False), ('llama', False)]
('qwen', True), ('llama', False)]
for model_type, expected_output in test_cases:
self.assertEqual(_check_torchair_supported(model_type),
expected_output)