[Bugfix] Fix num_hidden_layers when Qwen2-Audio 7B

Signed-off-by: hfadzxy <[email protected]>

Author: zhangxinyuehfad

docs/source/tutorials/single_npu_audio.md

@@ -90,8 +90,7 @@ def main(audio_count: int):
     llm = LLM(model="Qwen/Qwen2-Audio-7B-Instruct",
               max_model_len=4096,
               max_num_seqs=5,
-              limit_mm_per_prompt={"audio": audio_count},
-              enforce_eager=True)
+              limit_mm_per_prompt={"audio": audio_count})
 
     inputs = prepare_inputs(audio_count)
 

docs/source/tutorials/single_npu_multimodal.md

@@ -57,7 +57,6 @@ llm = LLM(
     model=MODEL_PATH,
     max_model_len=16384,
     limit_mm_per_prompt={"image": 10},
-    enforce_eager=True,
 )
 
 sampling_params = SamplingParams(
@@ -146,8 +145,7 @@ docker run --rm \
 vllm serve Qwen/Qwen2.5-VL-7B-Instruct \
 --dtype bfloat16 \
 --max_model_len 16384 \
---max-num-batched-tokens 16384 \
---enforce-eager
+--max-num-batched-tokens 16384 
 ```
 
 :::{note}

tests/e2e/singlecard/test_offline_inference.py

@@ -28,6 +28,7 @@
 from modelscope import snapshot_download  # type: ignore[import-untyped]
 from vllm import SamplingParams
 from vllm.assets.image import ImageAsset
+from vllm.assets.audio import AudioAsset
 
 import vllm_ascend  # noqa: F401
 from tests.e2e.conftest import VllmRunner
@@ -36,12 +37,18 @@
     "Qwen/Qwen2.5-0.5B-Instruct",
     "Qwen/Qwen3-0.6B-Base",
 ]
-MULTIMODALITY_MODELS = ["Qwen/Qwen2.5-VL-3B-Instruct"]
+MULTIMODALITY_VL_MODELS = ["Qwen/Qwen2.5-VL-3B-Instruct"]
+MULTIMODALITY_AUDIO_MODELS = ["Qwen/Qwen2-Audio-7B-Instruct"]
 
 QUANTIZATION_MODELS = [
     "vllm-ascend/Qwen2.5-0.5B-Instruct-W8A8",
 ]
 os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
+AUDIO_ASSETS = [AudioAsset("mary_had_lamb"), AudioAsset("winning_call")]
+AUDIO_PROMPT_TEMPLATES = {
+    1: "What is recited in the audio?",
+    2: "What sport and what nursery rhyme are referenced?"
+}
 
 
 @pytest.mark.parametrize("model", MODELS)
@@ -84,8 +91,8 @@ def test_quantization_models(model: str, max_tokens: int) -> None:
         vllm_model.generate_greedy(example_prompts, max_tokens)
 
 
-@pytest.mark.parametrize("model", MULTIMODALITY_MODELS)
-def test_multimodal(model, prompt_template, vllm_runner):
+@pytest.mark.parametrize("model", MULTIMODALITY_VL_MODELS)
+def test_multimodal_vl(model, prompt_template, vllm_runner):
     image = ImageAsset("cherry_blossom") \
         .pil_image.convert("RGB")
     img_questions = [
@@ -108,6 +115,47 @@ def test_multimodal(model, prompt_template, vllm_runner):
                                    max_tokens=64)
 
 
+def prepare_audio_inputs(audio_count: int):
+    audio_prompt = "".join([
+        f"Audio {idx+1}: <|audio_bos|><|AUDIO|><|audio_eos|>\n"
+        for idx in range(audio_count)
+    ])
+    question = AUDIO_PROMPT_TEMPLATES[audio_count]
+    prompt = ("<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
+              "<|im_start|>user\n"
+              f"{audio_prompt}{question}<|im_end|>\n"
+              "<|im_start|>assistant\n")
+    mm_data = {
+        "audio":
+        [asset.audio_and_sample_rate for asset in AUDIO_ASSETS[:audio_count]]
+    }
+    inputs = {"prompt": prompt, "multi_modal_data": mm_data}
+    return inputs
+
+
+@pytest.mark.parametrize("model", MULTIMODALITY_AUDIO_MODELS)
+@pytest.mark.parametrize("audio_count", [2])
+@pytest.mark.parametrize("max_tokens", [10])
+def test_multimodal_audio(model: str, audio_count: int,
+                          max_tokens: int) -> None:
+    inputs = prepare_audio_inputs(audio_count)
+
+    llm_config = {
+        "max_model_len": 4096,
+        "max_num_seqs": 5,
+        "limit_mm_per_prompt": {
+            "audio": audio_count
+        },
+        "gpu_memory_utilization": 0.9
+    }
+
+    with VllmRunner(model, **llm_config) as vllm_model:
+        sampling_params = SamplingParams(temperature=0.2,
+                                         max_tokens=max_tokens,
+                                         stop_token_ids=None)
+        vllm_model.generate(inputs, sampling_params=sampling_params)
+
+
 @patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION": "1"})
 def test_models_topk() -> None:
     example_prompts = [

vllm_ascend/utils.py

@@ -296,6 +296,24 @@ def vllm_version_is(target_vllm_version: str):
             "format of x.y.z.")
 
 
+def get_max_hidden_layers(hf_config) -> int:
+    cfg_dict = hf_config.to_dict()
+    layer_counts = []
+
+    def _rec_find(d):
+        if isinstance(d, dict):
+            for k, v in d.items():
+                if k == "num_hidden_layers" and isinstance(v, int):
+                    layer_counts.append(v)
+                else:
+                    _rec_find(v)
+
+    _rec_find(cfg_dict)
+    if not layer_counts:
+        raise ValueError("Not found num_hidden_layers in model config.")
+    return max(layer_counts)
+
+
 def update_aclgraph_sizes(vllm_config: VllmConfig) -> None:
     """Update ACL graph capture sizes based on hardware limitations"""
     # Store original configuration and temporarily clear it
@@ -304,7 +322,11 @@ def update_aclgraph_sizes(vllm_config: VllmConfig) -> None:
         compilation_config.cudagraph_capture_sizes, None
 
     # Calculate parallel configuration factor
-    num_hidden_layers = vllm_config.model_config.hf_config.num_hidden_layers
+    hf_config = vllm_config.model_config.hf_config
+    if hasattr(hf_config, 'num_hidden_layers'):
+        num_hidden_layers = hf_config.num_hidden_layers
+    else:
+        num_hidden_layers = get_max_hidden_layers(hf_config)
     parallel_config = vllm_config.parallel_config
 
     # TODO: Find out whether we need to take into account the pp_size