Made with @codex

Summary

Enable FlashInfer allreduce fusion by default for Qwen3NextForCausalLM on supported single-node SM90/SM100 TP runs.

Why

Qwen/Qwen3-Coder-Next was running with enable_flashinfer_allreduce_fusion=false on H100, and profiler traces showed prefill time dominated by unfused cross-device reduce kernels.

Change

• add Qwen3NextForCausalLM to the existing FlashInfer allreduce auto-enable whitelist

H100 Evidence

Model: Qwen/Qwen3-Coder-Next

Command:

python -m sglang.launch_server --model-path Qwen/Qwen3-Coder-Next --tp 4 --port 31080

Server args:

• baseline: enable_flashinfer_allreduce_fusion=false
• patch: enable_flashinfer_allreduce_fusion=true

Benchmark (sglang.bench_serving, random 2048/256, 128 prompts, max_concurrency=32):

Full accuracy:

Profiler (TP-0 EXTEND):

• baseline: cross_device_reduce_2stage 136.171 ms (21.89%)
• patch: allreduce_fusion_kernel_oneshot_lamport 57.661 ms (10.41%)

This change activates the fused allreduce path and removes the previous dominant unfused reduce hotspot.

Validation

• H100 before/after throughput benchmark
• H100 before/after sglang.profiler
• H100 before/after full MMLU and full GSM8K server eval

Modifications

Accuracy Tests

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.