[Kernel] Delegate construction of FusedMoEQuantConfig to FusedMoEMethodBase subclasses (#22537)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-09-17 17:43:31 -06:00
committed by GitHub
parent e6585ddb45
commit 5963b98b46
68 changed files with 2698 additions and 2526 deletions
+14 -8
View File
@@ -37,7 +37,7 @@ from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_reduce)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
QKVParallelLinear,
@@ -163,13 +163,19 @@ class DeepseekMoE(nn.Module):
shared_output = self.shared_experts(hidden_states)
# router_logits: (num_tokens, n_experts)
router_logits, _ = self.gate(hidden_states)
final_hidden_states = fused_moe(hidden_states,
self.w1,
self.w2,
router_logits,
self.top_k,
renormalize=self.config.norm_topk_prob,
inplace=True)
topk_weights, topk_ids, _ = fused_topk(
hidden_states,
router_logits,
self.top_k,
renormalize=self.config.norm_topk_prob)
final_hidden_states = fused_experts(hidden_states,
self.w1,
self.w2,
topk_weights,
topk_ids,
inplace=True)
if self.config.n_shared_experts is not None:
final_hidden_states = final_hidden_states + shared_output