JustinTong commited on
Commit
70b221c
·
verified ·
1 Parent(s): ee0f7be

Trim SGLang launch command to the verified minimal recipe

Browse files

Trim the SGLang launch command to the verified minimal recipe.

All removed flags were verified as unnecessary on the current `lmsysorg/sglang:dev-Ling-3.0-flash-VL` image: serving smoke plus full MMMU-Pro / GSM8K / bench_serving were measured without them. The MXFP4 MoE backend and the ling3 parsers are auto-selected; `--mem-fraction-static` uses the engine default with no measured perf/accuracy impact.

See the cookbook for the per-hardware verified matrix: https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-flash-VL

Files changed (1) hide show
  1. README.md +0 -1
README.md CHANGED
@@ -70,7 +70,6 @@ docker run --rm --gpus all --ipc=host --shm-size 32g \
70
  --tp 4 \
71
  --context-length 262144 \
72
  --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":2.0,"rope_theta":6000000,"partial_rotary_factor":0.5,"original_max_position_embeddings":131072}}' \
73
- --mem-fraction-static 0.85 \
74
  --trust-remote-code \
75
  --reasoning-parser auto \
76
  --tool-call-parser auto \
 
70
  --tp 4 \
71
  --context-length 262144 \
72
  --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":2.0,"rope_theta":6000000,"partial_rotary_factor":0.5,"original_max_position_embeddings":131072}}' \
 
73
  --trust-remote-code \
74
  --reasoning-parser auto \
75
  --tool-call-parser auto \