mirror of
https://git.datalinker.icu/vllm-project/vllm.git
synced 2026-06-08 05:42:16 +08:00
[perf] Enable concurrent execution of "shared_experts" and "selected_experts" in qwen3-next (#27578)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
parent
4fb8771cc0
commit
8df98c2161
@ -159,6 +159,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
|
|||||||
|
|
||||||
self.experts = SharedFusedMoE(
|
self.experts = SharedFusedMoE(
|
||||||
shared_experts=self.shared_expert,
|
shared_experts=self.shared_expert,
|
||||||
|
gate=self.gate,
|
||||||
num_experts=self.n_routed_experts,
|
num_experts=self.n_routed_experts,
|
||||||
top_k=config.num_experts_per_tok,
|
top_k=config.num_experts_per_tok,
|
||||||
hidden_size=config.hidden_size,
|
hidden_size=config.hidden_size,
|
||||||
@ -181,11 +182,17 @@ class Qwen3NextSparseMoeBlock(nn.Module):
|
|||||||
if self.is_sequence_parallel:
|
if self.is_sequence_parallel:
|
||||||
hidden_states = sequence_parallel_chunk(hidden_states)
|
hidden_states = sequence_parallel_chunk(hidden_states)
|
||||||
|
|
||||||
# router_logits: (num_tokens, n_experts)
|
if self.experts.is_internal_router:
|
||||||
router_logits, _ = self.gate(hidden_states)
|
# In this case, the gate/router runs inside the FusedMoE class
|
||||||
final_hidden_states = self.experts(
|
final_hidden_states = self.experts(
|
||||||
hidden_states=hidden_states, router_logits=router_logits
|
hidden_states=hidden_states, router_logits=hidden_states
|
||||||
)
|
)
|
||||||
|
else:
|
||||||
|
# router_logits: (num_tokens, n_experts)
|
||||||
|
router_logits, _ = self.gate(hidden_states)
|
||||||
|
final_hidden_states = self.experts(
|
||||||
|
hidden_states=hidden_states, router_logits=router_logits
|
||||||
|
)
|
||||||
|
|
||||||
if self.shared_expert is not None:
|
if self.shared_expert is not None:
|
||||||
final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
|
final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user