[zhouaihui]

Closes #22

Outcome

Add opt-in mtp_conditioning to DSpark and DFlash so a frozen native Qwen3.5 MTP predictor
can condition the Draft through the existing training, recovery and native export workflow.
Token mode supplies the known next-token anchor; hidden mode also adds a normalized,
zero-initialized residual fusion. Inactive features are masked before trainable fusion.

Visible T1 is excluded from supervision. Both algorithms report
train/suffix_simulated_acc_len and eval/suffix_simulated_acc_len, with per-position
suffix_acc_0 / suffix_ploss_0 starting at T2. Checkpoint selection, metadata and reports
retain that suffix scope. These teacher-forced suffix proxies do not measure acceptance
of the native MTP proposal or complete speculative blocks.

Include Qwen3.5-9B Draft configs, a runnable recipe and support documentation. MTP-specific
initialization and binding hooks run only when conditioning is enabled; ordinary
DSpark/DFlash interfaces, supervision and metric names are preserved.

Verification

• Independent nonzero-fusion equations check logits, loss and trainable parameter gradients
  for both algorithms, reusing the retained Draft backbone. Tests also cover activation
  checkpointing, padded inactive NaNs, frozen feature boundaries, native export/reload
  output equality, checkpoint metric selection and exact-resume history.
• Add test_dmtp.py, test_training_resume_rng.py and test_evaluation_contract.py to both
  CPU CI and tools.recipes contracts.
• Isolated Linux CPU/offline validation: CPU CI commands 631 passed and 16 passed;
  recipe contracts 421 passed, 5 skipped; final affected-model/recovery/assets suite
  292 passed, 21 skipped, plus 12 subtests. Lint and local documentation links passed.
• Bounded CUDA checks on one NVIDIA L20X: 14 passed. The explicit-prefix Dense GQA
  fixture used head dimension 16 and the compiled FlexAttention TRITON path: FP32 independent
  logits/loss/gradient reference, plus FP32/BF16 checkpointing and padded-NaN gradient parity.
• All 672 tracked source blobs and file modes matched the Linux-tested tree. Runtime:
  Python 3.12.3, Torch 2.13.0+cu130, Transformers 5.12.1.

Supported conditioned execution is Dense full GQA, CP1, MBS1, single-node REPLICATE,
decay objective, without packing or whole-model compilation. These checks make no new
multi-GPU, throughput or end-to-end decoding acceptance claim.

Compatibility and provenance

• Conditioning is optional. Native assets retain the policy, fusion and frozen predictor
  state; exact resume validates the same-Target binding. Generic SGLang export and the
  ordinary proposal consumer reject conditioned assets pending an MTP-aware consumer.
• The dual-anchor/fusion design is inspired by DMTP.
  This is an independent integration reusing ASpec's retained native predictor and DFlash
  backbone. Source headers and docs/framework/REFERENCES.md record the method and reuse;
  existing upstream notices and license texts are retained.

[Longxmas]

Three suggested follow-ups for 8e7f984:

1. Suffix metric naming: Conditioned runs still report train/eval/simulated_acc_len, although this only measures the suffix given teacher-forced T1. Please use an explicit suffix name and update checkpoint selection, metadata, and per-position metric labels consistently.

2. CI coverage: Please add tests/test_dmtp.py and tests/test_training_resume_rng.py to both the CPU CI suite and tools.recipes contracts; neither explicit test list currently includes them.

3. Fusion validation: Please add an independent forward/backward reference test with nonzero fusion weights, covering logits, loss, and trainable parameter gradients. Also check gradient-checkpointing parity and output equality after native export/reload. For inactive blocks, mask MTP features before fusion and add a padded-NaN regression to verify finite, unchanged gradients on valid data.

[zhouaihui]

Addressed in 6ce63d1:

1. Suffix metric names are consistent across training, evaluation and checkpoint selection.
2. MTP and resume tests are included in CPU CI and recipe contracts.
3. Added nonzero-fusion, checkpointing, export/reload and padded-NaN coverage; inactive features are masked before fusion.

CI, CPU regressions and bounded CUDA checks passed. Unconditioned DSpark/DFlash behavior is preserved.

[Longxmas]

Thanks for the update. Reviewed a234ce3 and the CI results; all three follow-ups are addressed: suffix metric naming, CI coverage, and fusion validation with inactive-feature masking. Looks good to merge.