[CLI] Expose --cpu-offload-gb, --tp-size, and --mem-fraction-static on sgl-omni

[edwingao28]

Motivation

sgl-omni serve only exposed high-level options; common SGLang runtime settings like cpu_offload_gb, mem_fraction_static, and tp_size required editing pipeline config YAML by hand. This blocks Ming-flash-omni-2.0 launches, where the generated defaults (cpu_offload_gb=0, mem_fraction_static=0.7) are often too tight, and blocks multi-GPU runs that need tp_size>1.

4.22 put this PR on hold since we are facing major codebase refactor, will apply changes after refactored code pushed to main branch

Modifications

• sglang_omni/cli/serve.py: Add --cpu-offload-gb, --mem-fraction-static, and --tp-size; pass as server_args_overrides via ConfigManager.from_model_path or --config.
• sglang_omni/config/schema.py: Add server_args_overrides and _apply_server_args_overrides to PipelineConfig; route via primary_sglang_stage.
• sglang_omni/config/manager.py: Plumb server_args_overrides through ConfigManager.from_model_path.
• sglang_omni/models/ming_omni/config.py: Set primary_sglang_stage = THINKER_STAGE; auto-inject disable_custom_all_reduce=True when tp_size>1 (Ming custom all-reduce kernel hangs under TP); remove legacy local override logic.

Example

TP=1:

CUDA_VISIBLE_DEVICES=0 \
sgl-omni serve \
  --model-path inclusionAI/Ming-flash-omni-2.0 \
  --port 8000 \
  --model-name ming-omni \
  --cpu-offload-gb 80 \
  --mem-fraction-static 0.92

TP=2 (Ming-Omni):

CUDA_VISIBLE_DEVICES=0,1 \
sgl-omni serve \
  --model-path inclusionAI/Ming-flash-omni-2.0 \
  --port 8000 \
  --model-name ming-omni \
  --tp-size 2 \
  --cpu-offload-gb 0 \
  --mem-fraction-static 0.80

Related Issues

Closes #296

Checklist

• Format your code according with pre-commit.
• Add unit tests.
• Update documentation / docstrings / example tutorials as needed.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.