[MoE] Defer the TRTLLM-Gen top-k finalize on the modular path (#58635)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Yongye Zhu
2026-09-25 15:38:29 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 24928a4d62
commit 378504a544
26 changed files with 652 additions and 117 deletions
+6 -6
View File
@@ -15,6 +15,7 @@ from tests.utils import (
multi_process_parallel,
)
from vllm.distributed import get_tp_group
from vllm.model_executor.layers.fused_moe.config import FusedMoEConfig
from vllm.model_executor.layers.fused_moe.experts.trtllm_mxfp4_moe import (
TrtLlmMxfp4ExpertsMonolithic,
)
@@ -43,12 +44,11 @@ def test_deferred_finalize_enabled_before_moe_kernel_setup(
hidden_dim = LATENT_SIZE
hidden_dim_unpadded = LATENT_SIZE
experts_per_token = 16
defer_moe_finalize = False
_defer_moe_finalize = False
defer_moe_finalize_max_num_tokens = -1
@property
def use_deferred_moe_finalize(self) -> bool:
return self.defer_moe_finalize
defer_moe_finalize = FusedMoEConfig.defer_moe_finalize
limit_deferred_moe_finalize = FusedMoEConfig.limit_deferred_moe_finalize
use_deferred_moe_finalize = FusedMoEConfig.use_deferred_moe_finalize
moe_config = FakeMoEConfig()
quant_method = SimpleNamespace(
@@ -99,7 +99,7 @@ def test_deferred_finalize_enabled_before_moe_kernel_setup(
latent_moe_runner.LatentMoERunner()
assert moe_config.defer_moe_finalize
assert moe_config.use_deferred_moe_finalize
assert moe_config.defer_moe_finalize_max_num_tokens == 128
assert initialized_with["experts_per_token"] == 16