[RL Infra][FlashInfer] Enable router replay output from FlashInfer monolithic MoE kernel (#44214)

Signed-off-by: Xuanyu Zhang <xuanyu.zhang@mistral.ai>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Xuanyu Zhang
2026-07-20 16:45:10 -07:00
committed by GitHub
co-authored by Claude aoshen02 OpenAI Codex
parent 58b2012aa2
commit 97a668152b
11 changed files with 1097 additions and 10 deletions
+24 -4
View File
@@ -7676,17 +7676,37 @@ class GPUModelRunner(
def _bind_routed_experts_capturer(self, capturer: RoutedExpertsCapturer) -> None:
from vllm.model_executor.layers.fused_moe.layer import MoERunner
from vllm.model_executor.layers.fused_moe.modular_kernel import (
FusedMoEExpertsMonolithic,
)
from vllm.model_executor.layers.fused_moe.router.base_router import (
BaseRouter,
)
for module in self.model.modules():
if isinstance(module, MoERunner) and isinstance(module.router, BaseRouter):
layer_id = module.layer_id
if not isinstance(module, MoERunner):
continue
layer_id = module.layer_id
def _capture_fn(topk_ids, _layer_id=layer_id, _capturer=capturer):
_capturer.capture(_layer_id, topk_ids)
def _capture_fn(topk_ids, _layer_id=layer_id, _capturer=capturer):
_capturer.capture(_layer_id, topk_ids)
quant_method = module._quant_method
moe_kernel = getattr(quant_method, "moe_kernel", None)
impl = getattr(moe_kernel, "impl", None)
fused_experts = getattr(impl, "fused_experts", None)
if quant_method.is_monolithic:
if not (
isinstance(fused_experts, FusedMoEExpertsMonolithic)
and fused_experts.supports_routing_replay_capture()
):
raise ValueError(
"--enable-return-routed-experts is not supported with "
f"monolithic MoE kernel {type(fused_experts).__name__}; "
"routed expert IDs would be silently all-zero."
)
fused_experts.set_capture_fn(_capture_fn)
elif isinstance(module.router, BaseRouter):
module.router.set_capture_fn(_capture_fn)
def may_add_encoder_only_layers_to_kv_cache_config(self) -> None: