[RL] [FlashInfer] Integrate FlashInfer trtllm_fp4_block_scale_routed_moe

[zianglih]

Motivation

@HumansAnd

This PR largely mirrors existing routed MoE integration:

• [FlashInfer v0.6.6][RL] Support fp8-last-n-bf16 RL for flashinfer_trtllm_routed moe backend #20214
• [FlashInfer v0.6.4] [RL] Integrate FlashInfer mxfp8 gemm, MoE, and routed MoE #19537

This PR also depends on #22204 for FlashInfer trtllm moe refactoring.

Modifications

• Rename and expand test_update_weights_from_disk_blackwell.py, now it covers both mxfp8 and nvfp4
• Expand test_flashinfer_trtllm_gen_moe_backend.py for nvfp4 coverage
• Add integration for trtllm_fp4_block_scale_routed_moe

Accuracy Tests

gsm8k

python3 -m sglang.launch_server --kv-cache-dtype bf16 --model nvidia/Qwen3-30B-A3B-NVFP4
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1209 --parallel 1209 --platinum
Accuracy: 0.945
Invalid: 0.001
Latency: 8.517 s
Output throughput: 17180.418 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: 0.942
Invalid: 0.001
Latency: 7.789 s
Output throughput: 18788.124 token/s

python3 -m pytest -s -q test/registered/backends/test_flashinfer_trtllm_gen_moe_backend.py -k NVFP4
============================================================================ warnings summary ============================================================================
../../usr/local/lib/python3.12/dist-packages/_pytest/config/__init__.py:1428
  /usr/local/lib/python3.12/dist-packages/_pytest/config/__init__.py:1428: PytestConfigWarning: Unknown config option: asyncio_mode
  
    self._warn_or_fail_if_strict(f"Unknown config option: {key}\n")

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
2 passed, 6 deselected, 3 warnings in 137.38s (0:02:17)
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute

python3 -m pytest -s -q test/registered/rl/test_update_weights_from_disk_blackwell.py -k NVFP4
============================================================================ warnings summary ============================================================================
../../usr/local/lib/python3.12/dist-packages/_pytest/config/__init__.py:1428
  /usr/local/lib/python3.12/dist-packages/_pytest/config/__init__.py:1428: PytestConfigWarning: Unknown config option: asyncio_mode
  
    self._warn_or_fail_if_strict(f"Unknown config option: {key}\n")

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
1 passed, 1 deselected, 3 warnings, 3 subtests passed in 92.21s (0:01:32)
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute

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]

Code Review

This pull request introduces support for FP4 MoE using the FlashInfer/TRT-LLM backend, including a new routed MoE wrapper and integration with the model optimization quantization path. The implementation refactors weight handling to use standard parameter names and adds comprehensive tests for NVFP4 backends and weight updates. Feedback was provided to refactor the new FP4 MoE wrapper using a keyword argument dictionary to ensure consistency with existing wrappers in the codebase.

[zianglih]

Set to 0.89 according to #22136

[nvpohanh]

cc @trevor-m

[trevor-m]

We also have #21240

[zianglih]

@trevor-m do you have plan on merging the PR? I can close this one since the implementation looks identical.

[zianglih]

I will strip this PR to weight update and test changes and hold untill #21240 merges.

[zianglih]

Closing this PR since flashinfer trtllm nvfp4 routed moe implementation is duplicated with #21240

Moving weight update refactoring and test file changes to:

• [RL] Refactor NVFP4 shuffling/swizzling to in-place replacement #22204