[Model][SM70] Support Qwen3.8 Flash Next with PLE offload

[Leonccaa]

Purpose

Add the Qwen3.8-Flash-Next Qwen4Exp model path, PLE CPU offload, and the SM70 compatibility needed to load and execute it on the 1Cat V100 runtime.

This is not a duplicate of an existing 1Cat PR: open and closed 1Cat PRs were searched before submission and no equivalent Qwen4Exp or PLE change was found. It adapts the still-open upstream model PR vllm-project/vllm#53896 and PLE-offload PR vllm-project/vllm#53899 to the 1Cat SM70 runtime.

The scope is intentionally limited to the reusable runtime core:

• Qwen4Exp model, config, registry, attention, cache, and speculative-decode integration
• PLE CPU worker, connector, protocol, and lifecycle integration
• NVIDIA and AMD Qwen4Exp support files carried by the upstream model work
• SM70 compatibility with the existing 1Cat hybrid-cache, FlashQLA, sampler, and worker paths

The separate AWQ g32 metadata adapter and extended performance benchmark remain in the local integration PR Leonccaa#4 and are not part of this upstream PR.

Test Plan

• Build the 1Cat runtime image from this core tree.
• Load the pinned Qwen3.8-Flash-Next NVFP4 checkpoint with TP=4 on four V100 32 GB GPUs.
• Keep PLE on CPU and verify all four GPU workers register with the PLE service.
• Run one real prompt prefill followed by one autoregressive decode step.
• Run repository CI for the imported unit coverage and style checks.

Test Result

Real 4x V100 validation passed on this core tree:

• full checkpoint load completed with PLE CPU offload
• all four GPU workers registered successfully
• prompt prefill and one autoregressive decode completed
• smoke result: SMOKE_OK token_ids=[3709, 95815]

Static validation on the final core tree:

• all-files manual mypy hooks passed for Python 3.10, 3.11, 3.12, and 3.13
• Ruff 0.14.0 check produced no fixes
• the two PR-specific Ruff format differences were applied

CI baseline note: the all-files pre-commit workflow remains red on repository-wide baseline issues. The run after the SM70 QSA commit has the same failing hook set as the immediately preceding core-PR run: Ruff format, typos, clang-format, markdownlint, SPDX normalization, forbidden-import, and legacy torch.cuda checks. Neither QSA file appears in the generated failure diff; file-level pre-commit passes locally. In the same GitHub run, Ruff check and all four Python 3.10-3.13 mypy hooks pass. Unrelated repository-wide rewrites remain intentionally excluded.

SM70 QSA prefill tuning

Commit 533b7a2b9 keeps the existing GB300 and non-SM70 launch profiles unchanged, but uses four warps for exact-SM70 single-split sparse QSA prefill (base_programs > 512). The 72 KiB-per-CTA shape permits only one resident CTA per V100 SM; two warps left that CTA severely under-occupied.

Four-rank CUDA profiling attributed 86.43% of 1K prefill kernel time to _qsa_sparse_paged_gqa_splitk_kernel. At the exact TP4 shape, changing only two to four warps reduced its median from 178.4207 ms to 11.3883 ms (15.67x) with elementwise-exact FP16 output.

End-to-end TP4 V100 results with the same W4A16 Marlin checkpoint path and PLE CPU offload:

Prompt	Baseline prefill	SM70 four-warp prefill	Speedup	Baseline TTFT	New TTFT
1,024	413.64 tok/s	2,263.82 tok/s	5.47x	2.4756 s	0.4523 s
4,096	421.77 tok/s	2,597.56 tok/s	6.16x	9.7114 s	1.5769 s

Both deterministic prompts retained their baseline first token. This is a focused performance/correctness gate rather than full quality acceptance. The pure launch-policy test covers all existing profile boundaries plus the SM70/non-SM70 split; 7/7 tests and all file-level pre-commit hooks pass.

NVFP4 on V100 is an emulation feasibility path, not a production recommendation. The intended production follow-up is activation-aware AWQ W4A16 g32 while keeping PLE on CPU.

This PR is Draft because the upstream vLLM source PRs are still open and because a human maintainer must review the AI-assisted port before it is marked ready.

AI assistance

OpenAI Codex assisted with the port, diagnostics, tests, and PR preparation. Leon must review the code and upstream delta before ready-for-review.

---

Essential Elements of an Effective PR Description Checklist

• Purpose and upstream relationship documented
• Test plan documented
• Real TP4 V100 result documented
• Scope split from AWQ follow-up documented