[CI] Increase the MTEB_EMBED_TOL threshold to 5e-4. (#31797)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>

tests/models/language/pooling_mteb_test/mteb_embed_utils.py

@@ -19,9 +19,9 @@ from tests.models.utils import (
 # - Model implementation and minor changes in tensor dtype
 #   results in differences less than 1e-4
 # - Different model results in differences more than 1e-3
-# 1e-4 is a good tolerance threshold
+# 5e-4 is a good tolerance threshold
 MTEB_EMBED_TASKS = ["STS12"]
-MTEB_EMBED_TOL = 1e-4
+MTEB_EMBED_TOL = 5e-4
 
 
 _empty_model_meta = ModelMeta(

tests/models/language/pooling_mteb_test/test_baai.py

@@ -65,7 +65,6 @@ MODELS = [
         "BAAI/bge-code-v1",
         architecture="Qwen2Model",
         mteb_score=0.75724465,
-        dtype="float32",
         pooling_type="LAST",
         attn_type="decoder",
         is_prefix_caching_supported=True,

tests/models/language/pooling_mteb_test/test_gte.py

@@ -89,7 +89,6 @@ MODELS = [
         "Qwen/Qwen3-Embedding-0.6B",
         mteb_score=0.771163695,
         architecture="Qwen3ForCausalLM",
-        dtype="float32",
         pooling_type="LAST",
         attn_type="decoder",
         is_prefix_caching_supported=True,
@@ -99,7 +98,6 @@ MODELS = [
     EmbedModelInfo(
         "Qwen/Qwen3-Embedding-4B",
         architecture="Qwen3ForCausalLM",
-        dtype="float32",
         enable_test=False,
     ),
 ]

tests/models/language/pooling_mteb_test/test_jina.py

@@ -28,7 +28,6 @@ EMBEDDING_MODELS = [
         attn_type="encoder_only",
         is_prefix_caching_supported=False,
         is_chunked_prefill_supported=False,
-        dtype="float32",
     )
 ]
 

tests/models/language/pooling_mteb_test/test_st_projector.py

@@ -29,7 +29,6 @@ ST_PROJECTOR_MODELS = [
         is_prefix_caching_supported=False,
         is_chunked_prefill_supported=False,
         enable_test=True,
-        dtype="float32",
     ),
 ]