perf(qwen4): optimize gated residual kernels and PDL

[tuanzhangCS]

Summary

• Move the Qwen4 gated-residual chain behind GPU-only tokenspeed-kernel APIs. Runtime code now owns parameters and checkpoint mapping only; grouped RMSNorm, mix, and combine implementations live in tokenspeed-kernel.
• Fuse the down and inject projections, add a low-M persistent Triton path plus a portable Triton path, and keep the Blackwell CuTeDSL composition available for explicit tuning. Dispatch falls back to the deterministic general path when atomic accumulation is disabled.
• Use stream-private persistent workspaces, preserve reduce-scatter row views, and apply non-power-of-two projection scaling after the projection to avoid checkpoint-weight quantization.
• Make ServerArgs.disable_pdl the authoritative full-chain PDL switch. The effective hardware-gated state is applied to tokenspeed-kernel, TorchInductor, TRT-LLM, spawned workers, and CuTeDSL compilation cache keys.
• Remove redundant runtime CPU/CUDA branches, the registered CPU reference backend, and the public mix-epilogue wrapper. Numerical references remain test-only.

Performance

B200, BF16, CUDA Graph, grouped RMSNorm + persistent mix + combine:

Tokens	PDL off	PDL on	Change
1	18.444 us	16.417 us	-11.0%
8	18.450 us	16.416 us	-11.0%
16	18.446 us	17.073 us	-7.4%

The CuTeDSL composition remains opt-in because current measurements favor the one-launch persistent Triton path through 16 tokens and the general path above it.

Test Plan

• pre-commit run --all-files
• Hyperconnection GPU numerics, dispatch, CUDA Graph, PDL, and concurrent-stream tests: 33 passed
• CuTeDSL low-latency BF16 router tests: 29 passed
• Layernorm tests: 31 passed
• Kernel registry and selection tests: 237 passed, 108 platform skips
• Runtime kernel-boundary and PDL-policy tests: 4 passed
• ServerArgs attention-backend tests: 22 passed

The focused Qwen4 test module is not collectable in the current local environment because its installed tuning dependency does not yet accept the tensor_initializers argument; the changed paths are covered by the focused runtime and kernel suites above.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-31T12:09:49.489042Z	7d29cf4	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.