[ModelRunner v2] Enable MRV2 for pooling models by default (#48290)
Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
Taneem Ibrahim committed
Aug 18, 2026 at 23:36 UTC
08afae27863e884d68937f76e5bb87bba171c5de
3 files changed
+26
-6
tests/models/language/pooling/test_colbert.py
+2
-2
@@ -323,8 +323,8 @@ class TestColbertSharedEngine:
323
[
324
pytest.param("bert", True, id="bert-v2"),
325
pytest.param("modernbert", True, id="modernbert-v2"),
326
- pytest.param("jina", False, id="jina-v1"),
327
- pytest.param("lfm2", False, id="lfm2-v1"),
326
+ pytest.param("jina", True, id="jina-v2"),
327
+ pytest.param("lfm2", True, id="lfm2-v2"),
328
],
329
)
330
def test_colbert_hf_comparison(vllm_runner, monkeypatch, backend, use_v2):
tests/test_config.py
+24
-1
@@ -342,15 +342,38 @@ def test_resolve_cudagraph_mode_adjusts_spec_decode_sizes_only_for_v1(
342
),
343
False,
344
),
345
+ (
346
+ SimpleNamespace(
347
+ model="sentence-transformers/all-MiniLM-L6-v2",
348
+ architectures=["BertModel"],
349
+ runner_type="pooling",
350
+ is_multimodal_model=False,
351
+ is_moe=False,
352
+ is_quantized=False,
353
+ ),
354
+ True,
355
+ ),
356
(
357
SimpleNamespace(
358
model="Qwen/Qwen3-Embedding-0.6B",
359
architectures=["Qwen3ForCausalLM"],
360
runner_type="pooling",
361
+ is_multimodal_model=False,
362
is_moe=False,
363
is_quantized=False,
364
),
353
- False,
365
+ True,
366
+ ),
367
+ (
368
+ SimpleNamespace(
369
+ model="TomoroAI/tomoro-colqwen3-embed-4b",
370
+ architectures=["ColQwen3"],
371
+ runner_type="pooling",
372
+ is_multimodal_model=True,
373
+ is_moe=False,
374
+ is_quantized=False,
375
+ ),
376
+ True,
377
),
378
],
379
)
vllm/config/vllm.py
-3
@@ -676,9 +676,6 @@ class VllmConfig:
676
if model_config is None:
677
return False
678
679
- if model_config.runner_type != "generate":
680
- return False
681
-
679
architectures = getattr(model_config, "architectures", [])
680
default_architectures = default_v2_model_runner_architectures()
681
is_default_v2_architecture = any(