karylab_agents/vllm
1
0
Fork 0
forked from Karylab-cklius/vllm
Code Pull Requests 2 Actions 1 Packages Activity

[DSV4] Guard megamoe flag with Pure TP (#41522)

Browse Source 
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-05-02 16:44:33 -07:00
committed by Roger Wang
parent a4debbd5bd
commit 228d225c4e
1 changed files with 17 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/models/deepseek_v4.py
+17 -6
View File
@@ -715,12 +715,15 @@ class DeepseekV4MoE(nn.Module):
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.prefix = prefix
if vllm_config.parallel_config.enable_expert_parallel:
self.use_mega_moe = (
vllm_config.kernel_config.moe_backend == "deep_gemm_mega_moe"
self.use_mega_moe = (
vllm_config.kernel_config.moe_backend == "deep_gemm_mega_moe"
)
if self.use_mega_moe and not vllm_config.parallel_config.enable_expert_parallel:
raise NotImplementedError(
"DeepSeek V4 MegaMoE currently requires expert parallel. "
"Enable it with --enable-expert-parallel, or pick a different "
"moe backend."
)
else:
self.use_mega_moe = False
self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
self.hidden_size = config.hidden_size
@@ -1224,7 +1227,15 @@ class DeepseekV4Model(nn.Module):
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.use_mega_moe = (
vllm_config.kernel_config.moe_backend == "deep_gemm_mega_moe"
)
if self.use_mega_moe and not vllm_config.parallel_config.enable_expert_parallel:
raise NotImplementedError(
"DeepSeek V4 MegaMoE currently requires expert parallel. "
"Enable it with --enable-expert-parallel, or pick a different "
"moe backend."
)
self.vocab_size = config.vocab_size
self.hc_eps = config.hc_eps
self.hc_mult = config.hc_mult