#2105 feat(mtp): support native Q4_0 and Q8_0 experts

本文

## Summary

Teach the Flash-Next MTP drafter to load routed experts in native GGUF Q4_0 or Q8_0 row layout, in addition to the existing Q2_0 S2 runtime.

This is opt-in through a runtime generated by `tools/mtp_rt.py`. Existing runtimes without `experts.txt` keep the legacy Q2_0 S2 path and default behavior.

## What changed

- `tools/mtp_rt.py` now:
  - detects the routed experts' GGML type;
  - preserves Q4_0 and Q8_0 rows in native `[gate | up | down]` order;
  - keeps Q2_0 in the existing S2 relayout;
  - writes `experts.txt` with a version, layout, types, and bytes per expert.
- `MtpDrafter` reads and validates that metadata, checks the exact `experts.bin` size, allocates the correct expert blob size, and dispatches native runtimes through the existing grouped native-expert kernels.
- `tools/mtp_fetch.py` accepts opt-in `STRATA_MTP_REPO` and `STRATA_MTP_REVISION` values for a compatible fine-tune's MTP tensors. With neither variable set, the original Qwen URL, pinned revision, and SHA-256 verification path are unchanged.
- Shared slot drafters inherit the selected layout and types from their owner.

The model geometry, server defaults, CLI defaults, legacy Q2_0 runtime, and non-MTP paths are unchanged.

## Validation

Rebased cleanly onto `main` at `7695859a`; the four touched files had no intervening upstream changes since the tested 0.1.42 base (`61b3fb5d`).

Static/tool checks:

- `python -m py_compile tools/mtp_fetch.py tools/mtp_rt.py`
- `python -m unittest tools.test_mtp_fetch`: 8 passed
- `git diff --check`

Conversion check:

- Converted a 1,596,806,016-byte Q4_0 MTP GGUF with this branch.
- The generated `experts.bin` (1,415,577,600 bytes), `experts.txt`, `dense.bin`, and `dense.txt` were byte-identical to the runtime used for the HIP tests.
- `experts.bin` SHA-256: `cb101ab35f946698ef09dddc942c41620ac03de40483a661ea07052b59ff16c5`
- `experts.txt` SHA-256: `c6fed393954caefd784ba689f7cc23f045f82051631406333c3b37c34e4ba44f`
- A Q8_0 runtime was also converted successfully; it was not used for the end-to-end run below.

End-to-end run:

- Windows 11, Ryzen AI Max+ 395 / Radeon 8060S (`gfx1151`), 128 GB unified RAM, Windows HIP Release build from the tested 0.1.42 source.
- The engine reported `native types 2/2`, loaded 1,350 MiB of Q4_0 MTP experts, reached `session is up`, and served repeated text, long-context, tool-use, and vision-integrated requests.
- The deployed Q4_0 runtime has been used with `--spec 6..8`, an 8,192-token MTP window, INT8 KV, and up to 262,144 total context capacity.

## Compatibility and limits

- A missing `experts.txt` means the legacy Q2_0 S2 layout, preserving old runtimes.
- Metadata rejects unknown versions, keys, layouts, types, shapes, blob sizes, and truncated/oversized expert files before inference.
- Q4_0 was built and run on Windows HIP. Q8_0 conversion was checked, but Q8_0 inference was not run.
- CUDA build/runtime validation is still outstanding, so this is a draft under the repository's backend-validation rule.
- SYCL has its separate engine path and was not changed.

Related performance/configuration evidence is kept separate from this code change in #1947.