[RL] Support FlashInfer per-token NVFP4 MoE

[zianglih]

Motivation

@HumansAnd

FlashInfer PR:

• [feat] Trtllm-gen Per-token Nvfp4 MoE flashinfer-ai/flashinfer#3027

Modifications

• When SGLANG_FLASHINFER_PER_TOKEN_NVFP4_MOE is true, ignore fp32 input activation scale in checkpoint, and use online per-token path instead
• Expand test coverage

Accuracy Tests

Baseline

python -m sglang.launch_server \
  --model-path nvidia/Qwen3-30B-A3B-NVFP4 \
  --disable-flashinfer-autotune \
  --kv-cache-dtype bf16 \
  --moe-runner-backend flashinfer_trtllm_routed
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1209 --parallel 1209 --platinum
Accuracy: 0.923
Invalid: 0.000
Latency: 10.510 s
Output throughput: 14359.210 token/s
Accuracy: 0.927
Invalid: 0.000
Latency: 9.423 s
Output throughput: 15836.511 token/s
curl -sS http://localhost:30000/update_weights_from_disk \
  -H 'Content-Type: application/json' \
  -d '{
    "model_path": "nvidia/Qwen3-30B-A3B-NVFP4",
    "flush_cache": true,
    "abort_all_requests": false
  }'
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1209 --parallel 1209 --platinum
Accuracy: 0.926
Invalid: 0.000
Latency: 9.313 s
Output throughput: 16046.278 token/s
Accuracy: 0.935
Invalid: 0.000
Latency: 9.254 s
Output throughput: 16089.310 token/s

With per-token NVFP4

SGLANG_FLASHINFER_PER_TOKEN_NVFP4_MOE=1 \
python -m sglang.launch_server \
  --kv-cache-dtype bf16 \
  --model-path nvidia/Qwen3-30B-A3B-NVFP4 \
  --moe-runner-backend flashinfer_trtllm_routed
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1209 --parallel 1209 --platinum
Accuracy: 0.939
Invalid: 0.000
Latency: 9.873 s
Output throughput: 14751.988 token/s
Accuracy: 0.940
Invalid: 0.000
Latency: 9.291 s
Output throughput: 15633.533 token/s
curl -sS http://localhost:30000/update_weights_from_disk \
  -H 'Content-Type: application/json' \
  -d '{
    "model_path": "nvidia/Qwen3-30B-A3B-NVFP4",
    "flush_cache": true,
    "abort_all_requests": false
  }'
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1209 --parallel 1209 --platinum
Accuracy: 0.940
Invalid: 0.000
Latency: 9.305 s
Output throughput: 15661.988 token/s
Accuracy: 0.939
Invalid: 0.000
Latency: 9.198 s
Output throughput: 15764.530 token/s

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.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!