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

[Bugfix] Fix GLM-ASR audio encoder RoPE dim (#32540)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2026-01-18 19:17:59 +08:00
committed by GitHub
parent c826c72a96
commit 38bf2ffb21
2 changed files with 40 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/models/glmasr.py
View File

@@ -181,6 +181,12 @@ class GlmAsrEncoderAttention(nn.Module):
# Use vLLM's ApplyRotaryEmb CustomOp
# enforce_enable=True ensures the op is always enabled (important for ViT)
rope_params = getattr(config, "rope_parameters", None)
if rope_params:
partial_rotary_factor = rope_params.get("partial_rotary_factor", 0.5)
else:
partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
self.rotary_dim = int(self.head_dim * partial_rotary_factor)
self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)
# Use vLLM's MMEncoderAttention for hardware-optimized attention
@@ -226,8 +232,12 @@ class GlmAsrEncoderAttention(nn.Module):
# Apply rotary position embeddings using vLLM's ApplyRotaryEmb
# ApplyRotaryEmb expects x: [batch, seq, heads, head_dim]
# cos/sin: [seq_len, rotary_dim/2]
q = self.apply_rotary_emb(q, rotary_pos_emb_cos, rotary_pos_emb_sin)
k = self.apply_rotary_emb(k, rotary_pos_emb_cos, rotary_pos_emb_sin)
q[..., : self.rotary_dim] = self.apply_rotary_emb(
q[..., : self.rotary_dim], rotary_pos_emb_cos, rotary_pos_emb_sin
)
k[..., : self.rotary_dim] = self.apply_rotary_emb(
k[..., : self.rotary_dim], rotary_pos_emb_cos, rotary_pos_emb_sin
)
# MMEncoderAttention expects [batch, seq, num_heads, head_dim]
# It handles GQA internally via repeat_interleave