biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

PR #12718 (#12718)

Browse Source
This commit is contained in:
Amit Garg
2025-02-07 06:22:37 -08:00
committed by GitHub
parent ce26b16268
commit 538fab93cd
2 changed files with 15 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/models/llama.py
View File

@@ -128,6 +128,9 @@ class LlamaAttention(nn.Module):
# MistralConfig has an optional head_dim introduced by Mistral-Nemo
self.head_dim = getattr(config, "head_dim",
self.hidden_size // self.total_num_heads)
# Phi models introduced a partial_rotary_factor parameter in the config
partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
self.rotary_dim = int(partial_rotary_factor * self.head_dim)
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
@@ -159,7 +162,7 @@ class LlamaAttention(nn.Module):
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
rotary_dim=self.rotary_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,