[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
@@ -66,6 +66,12 @@ def _make_router(eplb_state: EplbLayerState | None = None) -> DummyRouter:
)
def _make_modular_routed_experts():
return types.SimpleNamespace(
quant_method=types.SimpleNamespace(is_monolithic=False),
)
def test_base_router_capture_pre_eplb_mapping():
router = _make_router()
captured = []
@@ -122,6 +128,8 @@ def test_gpu_model_runner_binds_router_capture(monkeypatch):
def __init__(self):
self.layer_id = 7
self.router = _make_router()
self.routed_experts = _make_modular_routed_experts()
self._quant_method = self.routed_experts.quant_method
class DummyCapturer:
def __init__(self):
@@ -160,6 +168,8 @@ def test_gpu_model_runner_binding_stage(monkeypatch):
def __init__(self):
self.layer_id = 11
self.router = _make_router()
self.routed_experts = _make_modular_routed_experts()
self._quant_method = self.routed_experts.quant_method
class DummyCapturer:
def __init__(self):
@@ -197,6 +207,8 @@ def test_gpu_model_runner_does_not_bind_draft_router_capture(monkeypatch):
def __init__(self, layer_id):
self.layer_id = layer_id
self.router = _make_router()
self.routed_experts = _make_modular_routed_experts()
self._quant_method = self.routed_experts.quant_method
target_module = DummyFusedMoE(layer_id=7)
draft_module = DummyFusedMoE(layer_id=0)
@@ -222,6 +234,49 @@ def test_gpu_model_runner_does_not_bind_draft_router_capture(monkeypatch):
assert draft_module.router.capture_fn is None
def test_gpu_model_runner_rejects_monolithic_without_replay_support(monkeypatch):
from vllm.v1.worker import gpu_model_runner as gmr
class DummyFusedMoE:
def __init__(self):
self.layer_id = 3
self.router = _make_router()
# Use a concrete monolithic expert and override its capability
# instead of instantiating the abstract base class directly.
from vllm.model_executor.layers.fused_moe.experts.cpu_moe import (
CPUExpertsFp8,
)
fused_experts = CPUExpertsFp8.__new__(CPUExpertsFp8)
self.routed_experts = types.SimpleNamespace(
quant_method=types.SimpleNamespace(
is_monolithic=True,
moe_kernel=types.SimpleNamespace(
impl=types.SimpleNamespace(fused_experts=fused_experts)
),
)
)
self._quant_method = self.routed_experts.quant_method
self._quant_method.moe_kernel.impl.fused_experts = fused_experts
fused_experts.supports_routing_replay_capture = lambda: False
class DummyCapturer:
def capture(self, layer_id, topk_ids):
pass
dummy_module = DummyFusedMoE()
import vllm.model_executor.layers.fused_moe.layer as fused_moe_layer
monkeypatch.setattr(fused_moe_layer, "MoERunner", DummyFusedMoE)
dummy_self = types.SimpleNamespace(
model=types.SimpleNamespace(modules=lambda: [dummy_module])
)
with pytest.raises(ValueError, match="monolithic MoE kernel"):
gmr.GPUModelRunner._bind_routed_experts_capturer(dummy_self, DummyCapturer())
def test_routed_experts_capturer_single_dp_no_metadata():
"""dp_metadata is None: capture writes the full topk_ids rows."""
capturer = _capturer_with_buffer(dp_rank=0)