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

Don't fuse wk+weight_proj for nextn (#12863)

Browse Source
This commit is contained in:
Trevor Morris
2025-11-11 20:02:52 -08:00
committed by GitHub
parent 2f9952cdbf
commit 151e13687a
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/deepseek_v2.py
+4 -2
View File
@@ -3593,8 +3593,10 @@ class DeepseekV2ForCausalLM(nn.Module):
self.config.q_lora_rank is not None self.config.q_lora_rank is not None
) )
cached_a_proj = {} if fuse_qkv_a_proj else None cached_a_proj = {} if fuse_qkv_a_proj else None
fuse_wk_and_weights_proj = is_nsa_indexer_wk_and_weights_proj_fused( # Fuse wk and weights_proj when NSA Indexer is enabled and quant_config is FP4. For nextn, fp4 is disabled so we cannot fuse.
self.config, self.quant_config fuse_wk_and_weights_proj = (
is_nsa_indexer_wk_and_weights_proj_fused(self.config, self.quant_config)
and not is_nextn
) )
cached_wk_and_weights_proj = {} if fuse_wk_and_weights_proj else None cached_wk_and_weights_proj = {} if fuse_wk_and_weights_proj else None