Skip to content

Commit bc92d67

Browse files
committed
gemma-4-31B-it: --disable-fast-image-processor to stop engine crash on bad media
sglang gemma4 crashes the whole engine (unhandled RuntimeError in tokenizer_manager._tokenize_one_request -> gemma4 mm processor -> base_processor.fast_load_mm_data) when a client sends malformed multimodal input (base64 video data-URL or broken image URL as an image). Engine crash-loops under live traffic. Trying the non-fast image-load path to make these fail per-request instead of crashing.
1 parent 950bc96 commit bc92d67

1 file changed

Lines changed: 1 addition & 0 deletions

File tree

small-models.yaml

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -519,6 +519,7 @@ services:
519519
--chunked-prefill-size 8192
520520
--num-continuous-decode-steps 5
521521
--enable-mixed-chunk
522+
--disable-fast-image-processor
522523
--model-loader-extra-config '{"enable_multithread_load": "true", "num_threads": 64}'
523524
--port 8000
524525
--host 0.0.0.0

0 commit comments

Comments
 (0)