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

Wire YARN rope_parameters through LFM2 and LFM2-MoE attention (#26187)

Browse Source
This commit is contained in:
Piotr Mazurek
2026-05-26 23:00:22 +00:00
committed by GitHub
parent 6989fede3c
commit 468c565168
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/lfm2.py
+2 -2
View File
@@ -125,13 +125,13 @@ class Lfm2Attention(nn.Module):
if rope_parameters is not None and "rope_theta" in rope_parameters: if rope_parameters is not None and "rope_theta" in rope_parameters:
rope_theta = rope_parameters["rope_theta"] rope_theta = rope_parameters["rope_theta"]
else: else:
rope_theta = config.rope_parameters["rope_theta"] rope_theta = getattr(config, "rope_theta", 1000000.0)
self.rotary_emb = get_rope( self.rotary_emb = get_rope(
head_size=self.head_dim, head_size=self.head_dim,
rotary_dim=self.head_dim, rotary_dim=self.head_dim,
max_position=getattr(config, "max_position_embeddings", 8192), max_position=getattr(config, "max_position_embeddings", 8192),
rope_scaling=config.rope_parameters, rope_scaling=rope_parameters or getattr(config, "rope_scaling", None),
base=rope_theta, base=rope_theta,
is_neox_style=True, is_neox_style=True,
dtype=torch.get_default_dtype(), dtype=torch.get_default_dtype(),
python/sglang/srt/models/lfm2_moe.py
+1 -1
View File
@@ -196,7 +196,7 @@ class Lfm2MoeAttention(nn.Module):
head_size=self.head_dim, head_size=self.head_dim,
rotary_dim=self.head_dim, rotary_dim=self.head_dim,
max_position=getattr(config, "max_position_embeddings", 128000), max_position=getattr(config, "max_position_embeddings", 128000),
rope_scaling=getattr(config, "rope_scaling", None), rope_scaling=rope_parameters or getattr(config, "rope_scaling", None),
base=rope_theta, base=rope_theta,
is_neox_style=True, is_neox_style=True,
dtype=torch.get_default_dtype(), dtype=torch.get_default_dtype(),