Motivation

Modifications

• Defer o_proj and share_expert all-reduce, combine with MoE output into a single all-reduce per layer (was 3 separate all-reduces)
• Enable allreduce fusion and reduce-scatter for Step3p5
• Add Step3p5ForCausalLM to flashinfer allreduce fusion whitelist

Performance

Good Perfermance Launch Command:
H200x8
63,213 TPS

  python3 -m sglang.launch_server \
      --model-path stepfun-ai/Step-3.5-Flash-FP8 \
      --tp 8 \
      --ep 4 \
      --trust-remote-code \
      --mem-fraction-static 0.75 \
      --chunked-prefill-size 16384 \
      --port 30000 \
      --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 64}'

Prefill throughput (TP=8, EP=4, input_len=8192, output_len=1, 200 prompts):

• Before: ~46k tok/s
• After: ~56k tok/s (+21%)

Accuracy Tests

GSM8K Full Test (1319 questions)

Server command:

python3 -m sglang.launch_server --model-path stepfun-ai/Step-3.5-Flash-FP8 --tp 8 --ep 4 --trust-remote-code --port 30000

Benchmark command:

python3 -m sglang.test.few_shot_gsm8k --num-q 1319 --port 30000

Difference is 0.4% (~5 questions), within normal sampling variance. No accuracy regression.

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.