biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-09-17 19:43:31 -04:00
committed by GitHub
parent e6585ddb45
commit 5963b98b46
68 changed files with 2698 additions and 2526 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm/model_executor/models/minicpm.py
View File

@@ -39,7 +39,7 @@ 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 FatreluAndMul, 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,
@@ -136,13 +136,18 @@ class MiniCPMMoE(nn.Module):
hidden_states = hidden_states.view(-1, self.hidden_size)
# router_logits: (num_tokens, n_experts)
router_logits, _ = self.gate(hidden_states)
final_hidden_states = fused_moe(hidden_states,
self.ws,
self.w2s,
router_logits,
self.top_k,
renormalize=True,
inplace=True)
topk_weights, topk_ids, _ = fused_topk(hidden_states,
router_logits,
self.top_k,
renormalize=True)
final_hidden_states = fused_experts(hidden_states,
self.ws,
self.w2s,
topk_weights,
topk_ids,
inplace=True)
if self.tp_size > 1:
final_hidden_states = tensor_model_parallel_all_reduce(