[Model Runner V2] Use MRV2 for all models by default (#53183)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -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)
|
||||
|
||||
Reference in New Issue
Block a user