issue/1565 fix(nvidia): complete runtime support for Qwen MTP

[big-hip]

Problem and changes

Closes #1565. Target: InfiniLM-v0.2.9c.

Running Qwen's built-in greedy MTP on NVIDIA exposes missing E4M3/BOOL ATen conversions, CUDA device changes during TP communicator teardown, and reference-only Prefill dispatch for head size 256.

This six-file patch:

• Maps DataType::F8 and DataType::BOOL to existing ATen scalar types. No new conversion kernel.
• Saves/restores the caller's CUDA device around ncclCommDestroy, so its existing runtime stream remains valid when an engine is destroyed and another is created.
• Instantiates the existing paged Prefill warp kernel at head size 256 and enables that default on NVIDIA only. Existing default dispatch for other vendors and MLA is preserved; this is not a separate attention implementation.
• Adds two focused cast checks and extends existing Prefill coverage with multi-page/long-context cases. Test allocation now scales with the actual shape rather than reserving 8192 pages for every case.

Validation

Fresh build from base 1ab85ef1 with this patch: release, NVIDIA SM86, ATen, CCL, graph, OpenMP and cuDNN. 2 × RTX A6000 48 GiB, driver 580.105.08; native CUDA Toolkit 12.4, Python 3.11.15, PyTorch 2.9.0+cu128. The isolated install was verified with ldd, then tested independently of the development checkout.

• python test/infinicore/ops/paged_attention_prefill.py --nvidia: 88/88 passed, FP16/BF16 and int32/int64 metadata, including head size 256 and existing MLA cases.
• python -m pytest -q test/infinicore/ops/fp8_cast.py: 2 passed, all finite E4M3 byte patterns and BOOL-to-float/int casts against PyTorch exactly.
• Public C API smoke: two TP2 communicator groups, caller device 0 and 1 each preserved after every destroy.
• Project formatting and whitespace checks passed.
• InfiniLM integration on the matching combined runtime: 59 TP2 checks passed, 57 TP1 passed + 1 TP2-only skip; real 27B FP8 packed generation and cancel/re-admit outputs matched ordinary Decode references. This integration additionally uses graph-lifetime PR fix(graph): preserve recurrent graph inputs and allocation lifetimes #1560; the standalone operator tests above do not require it.

Saved logs, commands, exact revisions and measurement provenance

Performance and scope

The associated LM MTP path previously measured 33.91 → 49.57, 33.48 → 53.29, and 30.03 → 48.61 generated tok/s against ordinary Decode graphs on TP2 A6000 (63/127/1023-token prompts, FP8 block/Marlin, BF16 activations, K=2 eager, 40×64-token pages, two repeats). These are whole-stack archived results, not speedups attributed solely to this Core patch. Full definitions/raw measurements are in the evidence link; no standalone kernel speed claim is made here.

No new CUDA Graph runtime, Mamba scan, or vendor-specific backend implementation is included. Other accelerators could not be tested because no allocated usable environment was available; their default Prefill dispatch is unchanged. Please review the shared header mapping for vendor build compatibility.

Review / CI

Per DEV.md, this contribution has a linked issue, issue/1565 branch and commit/title prefix, and is submitted from a fork. Requesting @PanZezhong1725 (paged Prefill) and @spike-zhu (maintainer) review. Fork push starts the existing CPU/Linux/Windows checks; external PR jobs may require maintainer approval.

Merge only after the required checks and reviewer approval. InfiniLM integration: InfiniTensor/InfiniLM#584.

Formal reviewer assignment was rejected by GitHub (RequestReviewsByLogin: insufficient upstream permissions). A maintainer needs to assign the two reviewers above. Fork Ruff passed; the Linux/Windows build and CPU tests are running: https://github.com/big-hip/InfiniCore/actions/runs/35444866036.

Upstream PR workflows currently report action_required and have not executed jobs; maintainer approval is required. Fork checks are reported separately above.