minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[fix typo] expert_indicies -> expert_indices (#19627)

Browse Source 
Co-authored-by: ishandhanani <82981111+ishandhanani@users.noreply.github.com>
This commit is contained in:
SoluMilken
2026-03-01 17:37:34 -08:00
committed by GitHub
co-authored by ishandhanani
parent f51ddba131
commit 20282f5664
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/mixtral_quant.py
+4 -4
View File
@@ -115,10 +115,10 @@ class MixtralMoE(nn.Module):
f"the number of experts {self.num_total_experts}." f"the number of experts {self.num_total_experts}."
) )
# Split experts equally between ranks # Split experts equally between ranks
self.expert_indicies = np.array_split( self.expert_indices = np.array_split(
range(self.num_total_experts), self.tp_size range(self.num_total_experts), self.tp_size
)[self.rank].tolist() )[self.rank].tolist()
if not self.expert_indicies: if not self.expert_indices:
raise ValueError(f"Rank {self.rank} has no experts assigned to it.") raise ValueError(f"Rank {self.rank} has no experts assigned to it.")
self.experts = nn.ModuleList( self.experts = nn.ModuleList(
@@ -131,7 +131,7 @@ class MixtralMoE(nn.Module):
quant_config=quant_config, quant_config=quant_config,
prefix=add_prefix(f"experts.{idx}", prefix), prefix=add_prefix(f"experts.{idx}", prefix),
) )
if idx in self.expert_indicies if idx in self.expert_indices
else None else None
) )
for idx in range(self.num_total_experts) for idx in range(self.num_total_experts)
@@ -155,7 +155,7 @@ class MixtralMoE(nn.Module):
routing_weights /= routing_weights.sum(dim=-1, keepdim=True) routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
final_hidden_states = None final_hidden_states = None
for expert_idx in self.expert_indicies: for expert_idx in self.expert_indices:
expert_layer = self.experts[expert_idx] expert_layer = self.experts[expert_idx]
expert_mask = selected_experts == expert_idx expert_mask = selected_experts == expert_idx
expert_weights = (routing_weights * expert_mask).sum(dim=-1, keepdim=True) expert_weights = (routing_weights * expert_mask).sum(dim=-1, keepdim=True)