minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix Grok-2 nightly: derive image-understanding capability from is_multimodal (#33730)

Browse Source
This commit is contained in:
YC Yen-Ching Tseng
2026-08-19 23:56:51 -07:00
committed by GitHub
parent d287880a7a
commit 7ba3430365
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/configs/model_config.py
+7 -3
View File
@@ -460,13 +460,17 @@ class ModelConfig:
self.is_audio_model = enable_multimodal and is_audio_model( self.is_audio_model = enable_multimodal and is_audio_model(
self.hf_config.architectures self.hf_config.architectures
) )
# TODO: requires further polishing # Gated on `is_multimodal` because this flag is advertised via /model_info
# and drives the VLM warmup request, while the OpenAI serving layer rejects
# media input for models that are not `is_multimodal`. A text-only model
# with an auto-populated `vision_config` (see above) would otherwise warm up
# with an image request that its own serving layer answers with 400.
# Key on the tower, not the attribute: several config classes default # Key on the tower, not the attribute: several config classes default
# vision_config to None, which presence alone would read as image-capable # vision_config to None, which presence alone would read as image-capable
# (MuseGlimmerConfig's text-only layouts are one such case). # (MuseGlimmerConfig's text-only layouts are one such case).
# TODO: requires further polishing
self.is_image_understandable_model = ( self.is_image_understandable_model = (
enable_multimodal self.is_multimodal
and not self.is_lm_only
and getattr(self.hf_config, "vision_config", None) is not None and getattr(self.hf_config, "vision_config", None) is not None
) )