[PCG] fix piecewise cuda graph for Qwen3.5

[zminglei]

Motivation

fix piecewise cuda graph for Qwen3.5

Modifications

1. fix piecewise cuda graph for Qwen3.5
2. clean up legacy code gdn_with_output as it's not used anymore.

Accuracy Tests

main:

SGLANG_ENABLE_JIT_DEEPGEMM=0 python -m sglang.launch_server --model-path /shared/public/elr-models/Qwen/Qwen3.5-397B-A17B-FP8 --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --enable-piecewise-cuda-graph

[2026-02-24 04:43:15 TP4] Scheduler hit an exception: Traceback (most recent call last):
  File "/home/jobuser/zminglei/sglang/venv/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py", line 2755, in _dispatch_impl
    r = func(*args, **kwargs)
  File "/home/jobuser/zminglei/sglang/venv/lib/python3.10/site-packages/torch/_ops.py", line 841, in __call__
    return self._op(*args, **kwargs)
NotImplementedError: sgl_kernel::fp8_blockwise_scaled_mm: attempted to run this operator with Meta tensors, but there was no fake impl or Meta kernel registered. You may have run into this message while using an operator with PT2 compilation APIs (torch.compile/torch.export); in order to use this operator with those APIs you'll need to add a fake impl. Please see the following for next steps:  https://pytorch.org/tutorials/advanced/custom_ops_landing_page.html

This PR:

SGLANG_ENABLE_JIT_DEEPGEMM=0 python -m sglang.launch_server --model-path /shared/public/elr-models/Qwen/Qwen3.5-397B-A17B-FP8 --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --enable-piecewise-cuda-graph

python benchmark/gsm8k/bench_sglang.py --data-path /shared/public/data/gsm8k/test.jsonl --port 8000 --num-questions 1319 --parallel 1319
100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [01:56<00:00, 11.35it/s]
Accuracy: 0.948
Invalid: 0.008
Latency: 116.176 s
Output throughput: 1831.956 token/s

Benchmarking 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 Process

1. Ping Merge Oncalls to start the PR flow. 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.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[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!

[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!

[zminglei]

/tag-and-rerun-ci again

[yuan-luo]

why remove this branch?

[zminglei]

This branch was for PCG purpose (most likely copied from previous qwen3_next.py). But now it's not needed anymore since we added split_op inside RadixLinearAttention. It's for the same purpose of this PR #17613

[Oasis-Git]

Maybe split the change for vl model to another pr if possible