A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 18:30 UTC
[CI Failure] Fix Gemma3 RoPE configuration for sliding attention layers (#29111)
Signed-off-by: Huamin Li <3ericli@gmail.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Huamin Li committed
Nov 20, 2025 at 23:53 UTC
8ac3a4148796648d206a46144aa0dacea8977d55
1 file changed
+4
-2
vllm/model_executor/models/gemma3.py
+4
-2
@@ -166,10 +166,12 @@ class Gemma3Attention(nn.Module):
166
else:
167
# Transformers v4 rope config.
168
# Global attention. Use the values in config.json.
169
- rope_parameters = config.rope_parameters.copy()
169
+ rope_parameters = config.rope_parameters
170
# Local attention. Override the values in config.json.
171
if self.is_sliding:
172
- rope_parameters["rope_theta"] = config.rope_local_base_freq
172
+ rope_parameters = dict(
173
+ rope_type="default", rope_theta=config.rope_local_base_freq
174
+ )
175
176
self.rotary_emb = get_rope(
177
self.head_dim,