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

[sgl] Fix kimi-k2.5 EAGLE3 MLA draft embeds for batched MM prefill (#27647)

Browse Source
This commit is contained in:
Bi Xue
2026-06-09 11:26:48 -07:00
committed by GitHub
parent 5babb902a9
commit 8ae328e5f0
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/kimi_k25_eagle3.py
+4 -3
View File
@@ -261,9 +261,10 @@ class Eagle3MLAModel(nn.Module):
and not forward_batch.forward_mode.is_draft_extend(include_v2=True) and not forward_batch.forward_mode.is_draft_extend(include_v2=True)
): ):
assert embeds is not None assert embeds is not None
embeds = torch.cat( last_indices = (
[embeds[:-1], self.embed_tokens(input_ids[-1].unsqueeze(0))] forward_batch.extend_start_loc + forward_batch.extend_seq_lens - 1
) ).long()
embeds[last_indices] = self.embed_tokens(input_ids[last_indices])
if embeds is None: if embeds is None:
embeds = self.embed_tokens(input_ids) embeds = self.embed_tokens(input_ids)
else: else: