[Bugfix][CI/Build] Fix failing Mteb CI (vllm-project#26638)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: xuebwang-amd <xuebwang@amd.com>

Author: Isotr0py

tests/models/language/pooling_mteb_test/mteb_utils.py

@@ -191,7 +191,7 @@ def mteb_test_embed_models(
     with vllm_runner(
         model_info.name,
         runner="pooling",
-        max_model_len=None,
+        max_model_len=model_info.max_model_len,
         **vllm_extra_kwargs,
     ) as vllm_model:
         model_config = vllm_model.llm.llm_engine.model_config

tests/models/language/pooling_mteb_test/test_jina.py

@@ -25,6 +25,11 @@
         mteb_score=0.824413164,
         architecture="XLMRobertaModel",
         is_matryoshka=True,
+        # The default max length of the model is 8194, which will crash
+        # CUDAGraph due to odd length for Gemm. We set it to 8192 to avoid
+        # avoid this issue.
+        max_model_len=8192,
+        dtype="float32",
     )
 ]
 

tests/models/language/pooling_mteb_test/test_st_projector.py

@@ -23,6 +23,7 @@
         architecture="Gemma3TextModel",
         mteb_score=0.7473819294684156,
         enable_test=True,
+        dtype="float32",
     ),
 ]
 

tests/models/utils.py

@@ -369,6 +369,7 @@ class ModelInfo:
     name: str
     architecture: str = ""
     dtype: str = "auto"
+    max_model_len: Optional[int] = None
     hf_dtype: str = "float32"
     hf_overrides: Optional[dict[str, Any]] = None
     default_pooling_type: str = ""

vllm/model_executor/layers/layernorm.py

@@ -318,7 +318,11 @@ def forward_static(
         """PyTorch-native implementation equivalent to forward()."""
         orig_dtype = x.dtype
         if residual is not None:
-            x = x + residual.float() if orig_dtype == torch.float16 else x + residual
+            x = (
+                x.float() + residual.float()
+                if orig_dtype == torch.float16
+                else x + residual
+            )
             residual = x
 
         x = x.float()