EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Small refactor DeepEPDispatcher into subclasses (#4994)

Browse Source
This commit is contained in:
fzyzcjy
2025-04-04 15:24:18 +08:00
committed by GitHub
parent a995a773a0
commit febe21ce03
2 changed files with 315 additions and 197 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
python/sglang/srt/models/deepseek_v2.py
View File

@@ -188,35 +188,24 @@ class DeepseekV2MoE(nn.Module):
if global_server_args_dict["enable_deepep_moe"]
else (EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE)
)
if not global_server_args_dict["enable_deepep_moe"]:
self.experts = MoEImpl(
num_experts=config.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
num_expert_group=config.n_group,
topk_group=config.topk_group,
correction_bias=self.gate.e_score_correction_bias,
prefix=add_prefix("experts", prefix),
)
else:
self.experts = MoEImpl(
num_experts=config.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
num_expert_group=config.n_group,
topk_group=config.topk_group,
correction_bias=self.gate.e_score_correction_bias,
prefix=add_prefix("experts", prefix),
deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]],
)
self.experts = MoEImpl(
num_experts=config.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
num_expert_group=config.n_group,
topk_group=config.topk_group,
correction_bias=self.gate.e_score_correction_bias,
prefix=add_prefix("experts", prefix),
**(
dict(deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]])
if global_server_args_dict["enable_deepep_moe"]
else {}
),
)
if config.n_shared_experts is not None:
intermediate_size = config.moe_intermediate_size * config.n_shared_experts