@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 5406ebf5c9
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 14, 05:30 UTC

[CI] Pooling models mteb test uses enforce_eager (#22878)

Signed-off-by: wang.yuqi <noooop@126.com>

wang.yuqi committed Aug 15, 2025 at 16:16 UTC 5406ebf5c9606b717ee11a40db8ba56d8c8dde17
1 file changed +5 -1
tests/models/language/pooling/mteb_utils.py
+5 -1
@@ -18,7 +18,7 @@ from tests.models.utils import EmbedModelInfo, RerankModelInfo
18 # - Different model results in differences more than 1e-3
19 # 1e-4 is a good tolerance threshold
20 MTEB_EMBED_TASKS = ["STS12"]
21 -MTEB_EMBED_TOL = 1e-4
21 +MTEB_EMBED_TOL = 0.02
22
23 # See #19344
24 MTEB_RERANK_TASKS = ["NFCorpus"]
@@ -175,6 +175,7 @@ def mteb_test_embed_models(hf_runner,
175 with vllm_runner(model_info.name,
176 runner="pooling",
177 max_model_len=None,
178 + enforce_eager=True,
179 **vllm_extra_kwargs) as vllm_model:
180
181 model_config = vllm_model.llm.llm_engine.model_config
@@ -198,6 +199,7 @@ def mteb_test_embed_models(hf_runner,
199 st_main_score = run_mteb_embed_task(hf_model, MTEB_EMBED_TASKS)
200 st_dtype = next(hf_model.model.parameters()).dtype
201
202 + print("Model:", model_info.name)
203 print("VLLM:", vllm_dtype, vllm_main_score)
204 print("SentenceTransformers:", st_dtype, st_main_score)
205 print("Difference:", st_main_score - vllm_main_score)
@@ -286,6 +288,7 @@ def mteb_test_rerank_models(hf_runner,
288 runner="pooling",
289 max_model_len=None,
290 max_num_seqs=8,
291 + enforce_eager=True,
292 **vllm_extra_kwargs) as vllm_model:
293
294 model_config = vllm_model.llm.llm_engine.model_config
@@ -304,6 +307,7 @@ def mteb_test_rerank_models(hf_runner,
307 st_main_score, st_dtype = mteb_test_rerank_models_hf(
308 hf_runner, model_info.name, hf_model_callback)
309
310 + print("Model:", model_info.name)
311 print("VLLM:", vllm_dtype, vllm_main_score)
312 print("SentenceTransformers:", st_dtype, st_main_score)
313 print("Difference:", st_main_score - vllm_main_score)