refactor: minor refactors regarding multimodal processing (#6187)

2025-05-18 13:53:20 +08:00
parent b3f3d610fd
commit 01dd39bac1
15 changed files with 140 additions and 98 deletions
--- a/python/sglang/srt/models/qwen2_5_vl.py
+++ b/python/sglang/srt/models/qwen2_5_vl.py
@@ -146,6 +146,8 @@ class Qwen2_5_VisionBlock(nn.Module):
            num_heads=num_heads,
            projection_size=dim,
            use_qkv_parallel=True,
+            rotary_embed="normal",
+            proj_bias=True,
            qkv_backend=qkv_backend,
            softmax_in_single_precision=softmax_in_single_precision,
            flatten_batch=flatten_batch,