[Model Runner V2] Use MRV2 for all models by default (#53183)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-08-27 11:59:34 -07:00
committed by GitHub
parent a18dbe49ad
commit 4aab2b0ebe
5 changed files with 122 additions and 117 deletions
@@ -487,6 +487,7 @@ def test_models_fse_init(
quantization_config=None,
runner_type="generate",
is_moe=True,
logits_processors=None,
)
vllm_config.parallel_config.enable_expert_parallel = False
if model_type == "deepseek_v4":
@@ -545,6 +546,8 @@ def test_models_fse_init(
vllm_config.speculative_config = SimpleNamespace(
draft_model_config=SimpleNamespace(hf_config=config),
method="mtp",
parallel_drafting=False,
enable_adaptive_verification=False,
)
mtp = DeepSeekV4MTP(vllm_config=vllm_config)
assert model.is_fused_shared_expert_enabled is (fse_enabled and not exclude)