From 045b396d090f4a16fbba760bef86e9a24a7ba9ce Mon Sep 17 00:00:00 2001
From: Isotr0py <mozf@mail2.sysu.edu.cn>
Date: Sun, 12 Oct 2025 17:42:42 +0800
Subject: [PATCH] [Bugfix][CI/Build] Fix failing Mteb CI (#26638)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
---
 tests/models/language/pooling_mteb_test/mteb_utils.py       | 2 +-
 tests/models/language/pooling_mteb_test/test_jina.py        | 5 +++++
 .../models/language/pooling_mteb_test/test_st_projector.py  | 1 +
 tests/models/utils.py                                       | 1 +
 vllm/model_executor/layers/layernorm.py                     | 6 +++++-
 5 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/tests/models/language/pooling_mteb_test/mteb_utils.py b/tests/models/language/pooling_mteb_test/mteb_utils.py
index d96dc90416855..65ad49fad3653 100644
--- a/tests/models/language/pooling_mteb_test/mteb_utils.py
+++ b/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
diff --git a/tests/models/language/pooling_mteb_test/test_jina.py b/tests/models/language/pooling_mteb_test/test_jina.py
index 0a712b2542f3c..dbdf82af33c72 100644
--- a/tests/models/language/pooling_mteb_test/test_jina.py
+++ b/tests/models/language/pooling_mteb_test/test_jina.py
@@ -25,6 +25,11 @@ EMBEDDING_MODELS = [
         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",
     )
 ]
 
diff --git a/tests/models/language/pooling_mteb_test/test_st_projector.py b/tests/models/language/pooling_mteb_test/test_st_projector.py
index 91b1ef828d0df..74fe4b9bcc03f 100644
--- a/tests/models/language/pooling_mteb_test/test_st_projector.py
+++ b/tests/models/language/pooling_mteb_test/test_st_projector.py
@@ -23,6 +23,7 @@ ST_PROJECTOR_MODELS = [
         architecture="Gemma3TextModel",
         mteb_score=0.7473819294684156,
         enable_test=True,
+        dtype="float32",
     ),
 ]
 
diff --git a/tests/models/utils.py b/tests/models/utils.py
index 84697ad68d441..3d6e6cb89d62a 100644
--- a/tests/models/utils.py
+++ b/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 = ""
diff --git a/vllm/model_executor/layers/layernorm.py b/vllm/model_executor/layers/layernorm.py
index 6a49ae42ca895..910f145b1f8c2 100644
--- a/vllm/model_executor/layers/layernorm.py
+++ b/vllm/model_executor/layers/layernorm.py
@@ -318,7 +318,11 @@ class GemmaRMSNorm(CustomOp):
         """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()