minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[Fix] Fix GLM5 mHC PP forward (#39720)

Browse Source
This commit is contained in:
Void
2026-09-16 00:32:53 -07:00
committed by GitHub
parent 54f0d72adf
commit 9c8d4641ff
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/glm5_next.py
+4 -1
View File
@@ -981,7 +981,8 @@ class Glm5NextModel(nn.Module):
else: else:
assert pp_proxy_tensors is not None assert pp_proxy_tensors is not None
hidden_states = pp_proxy_tensors["hidden_states"] hidden_states = pp_proxy_tensors["hidden_states"]
residual = pp_proxy_tensors["residual"] # mHC carries its residual streams in hidden_states across PP stages.
residual = None if self.config.mhc else pp_proxy_tensors["residual"]
device = hidden_states.device device = hidden_states.device
zero_allocator = BumpAllocator( zero_allocator = BumpAllocator(
buffer_size=total_num_layers * 2 * (2 if forward_batch.can_run_tbo else 1), buffer_size=total_num_layers * 2 * (2 if forward_batch.can_run_tbo else 1),
@@ -1059,6 +1060,8 @@ class Glm5NextModel(nn.Module):
) )
if not self.pp_group.is_last_rank: if not self.pp_group.is_last_rank:
if self.config.mhc:
return PPProxyTensors({"hidden_states": hidden_states})
return PPProxyTensors( return PPProxyTensors(
{ {
"hidden_states": hidden_states, "hidden_states": hidden_states,