xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-18 11:16:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Minor][Models] Pass partial_rotary_factor parameter to rope (#17266)

Browse Source 
Signed-off-by: evian <eviantai@u.nus.edu>
Co-authored-by: evian <eviantai@u.nus.edu>
This commit is contained in:
Wanrui Dai 2025-04-28 12:28:59 +08:00 committed by GitHub
parent 8262a3e23b
commit 7fcc4223dc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 10 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/models/llama.py
View File

@ -130,8 +130,8 @@ class LlamaAttention(nn.Module):
self.head_dim = getattr(config, "head_dim", self.head_dim = getattr(config, "head_dim",
self.hidden_size // self.total_num_heads) self.hidden_size // self.total_num_heads)
# Phi models introduced a partial_rotary_factor parameter in the config # Phi models introduced a partial_rotary_factor parameter in the config
partial_rotary_factor = getattr(config, "partial_rotary_factor", 1) self.partial_rotary_factor = getattr(config, "partial_rotary_factor",
self.rotary_dim = int(partial_rotary_factor * self.head_dim) 1)
self.q_size = self.num_heads * self.head_dim self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5 self.scaling = self.head_dim**-0.5
@ -163,11 +163,12 @@ class LlamaAttention(nn.Module):
self.rotary_emb = get_rope( self.rotary_emb = get_rope(
self.head_dim, self.head_dim,
rotary_dim=self.rotary_dim, rotary_dim=self.head_dim,
max_position=max_position_embeddings, max_position=max_position_embeddings,
base=rope_theta, base=rope_theta,
rope_scaling=rope_scaling, rope_scaling=rope_scaling,
is_neox_style=is_neox_style, is_neox_style=is_neox_style,
partial_rotary_factor=self.partial_rotary_factor,
) )
if hasattr(config, "interleaved_sliding_window"): if hasattr(config, "interleaved_sliding_window"):

3
vllm/model_executor/models/persimmon.py
View File

@ -115,9 +115,10 @@ class PersimmonAttention(nn.Module):
self.rotary_emb = get_rope( self.rotary_emb = get_rope(
self.head_dim, self.head_dim,
rotary_dim=int(self.partial_rotary_factor * self.head_dim), rotary_dim=self.head_dim,
max_position=self.max_position_embeddings, max_position=self.max_position_embeddings,
base=self.rope_theta, base=self.rope_theta,
partial_rotary_factor=self.partial_rotary_factor,
) )
self.scaling = self.head_dim**-0.5 self.scaling = self.head_dim**-0.5
self.attn = Attention(self.num_heads, self.attn = Attention(self.num_heads,

8
vllm/model_executor/models/stablelm.py
View File

@ -104,9 +104,8 @@ class StablelmAttention(nn.Module):
1, self.total_num_key_value_heads // tp_size) 1, self.total_num_key_value_heads // tp_size)
self.head_dim = self.hidden_size // self.total_num_heads self.head_dim = self.hidden_size // self.total_num_heads
self.max_position_embeddings = config.max_position_embeddings self.max_position_embeddings = config.max_position_embeddings
rope_pct = getattr(config, "rope_pct", self.partial_rotary_factor = getattr(
getattr(config, "partial_rotary_factor", 1)) config, "rope_pct", getattr(config, "partial_rotary_factor", 1))
self.rotary_ndims = int(self.head_dim * rope_pct)
self.scaling = self.head_dim**-0.5 self.scaling = self.head_dim**-0.5
self.q_size = self.num_heads * self.head_dim self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_key_value_heads * self.head_dim self.kv_size = self.num_key_value_heads * self.head_dim
@ -130,9 +129,10 @@ class StablelmAttention(nn.Module):
prefix=f"{prefix}.o_proj") prefix=f"{prefix}.o_proj")
self.rotary_emb = get_rope( self.rotary_emb = get_rope(
self.head_dim, self.head_dim,
rotary_dim=self.rotary_ndims, rotary_dim=self.head_dim,
max_position=self.config.max_position_embeddings, max_position=self.config.max_position_embeddings,
base=self.config.rope_theta, base=self.config.rope_theta,
partial_rotary_factor=self.partial_rotary_factor,
) )
self.attn = Attention(self.num_heads, self.attn = Attention(self.num_heads,
self.head_dim, self.head_dim,