feat: add Intern-Decision text training cookbook on Ascend

[Liuchenbing-2026]

PR type

• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Add a text-only Qwen3.5-4B Intern-Decision training cookbook for four Ascend NPUs. Joint questions use a complete JSON skeleton and <decision> markers; answer tokens appear only in labels. Unlike HF causal loss, Twinkle's external cross-entropy does not shift labels, so this adapter shifts supervision exactly once during encoding.

The cookbook contains a decision processor that selects supervised output positions while retaining full-vocabulary cross-entropy, deterministic case-level preparation, frozen-vision/full-language training, validation accounting, and save/reload checks. FSDP2 optimizer state is saved with PyTorch distributed checkpointing; resharding before export preserves optimizer parameter identities. The final model is saved before validation so a later validation error does not discard completed training.

The cookbook is listed in Community Components and uses the existing TransformersModel/InputProcessor APIs rather than adding a duplicate Qwen architecture. Repository-wide dependencies are unchanged: the earlier NumPy lower-bound relaxation has been removed. Historical device validation used NumPy 1.26.4 through an explicit no-deps install, outside the upstream NumPy >=2.0 requirement; a dependency-compliant NPU run remains pending. The guide distinguishes that limitation, and launch paths/device assignment are configurable.

No datasets, weights, private machine details or raw predictions are committed. Public Typed Decisions cases are split independently for training, validation and calibration; the test split is not used for training or selection. This is an Intern-Decision method reproduction, not StartLux training or a reproduction of the unpublished original training corpus. AI assistance was used in implementation, testing and documentation.

Experiment results

Checks on the PR source:

uvx pre-commit run --files README.md $(git diff --name-only upstream/main...HEAD)
PYTHONPATH="$PWD/src" "$PYTHON_BIN" -m pytest tests/processor/test_intern_decision.py -q
cd cookbook/transformers/intern_decision_npu
PYTHONPATH="$TWINKLE_SOURCE/src:$PWD" "$PYTHON_BIN" -m twinkle_adapter.check_contract --output "$NEW_AUDIT_PATH"

• Seven standard-directory CPU pytest cases passed: no target leakage, one causal shift, malformed-target and truncation rejection, supervised-token counting, and actual processor/CE loss-gradient equivalence. No model or dataset download is required.
• Three launch argument checks passed without starting training.
• Applicable pre-commit hooks passed; the repository configuration excludes cookbook Python from flake8/isort/yapf. Python syntax and shell syntax checks passed separately.
• Real tokenizer/processor contract passed for 1,600 cases and 8,000 supervised fields, including answer exclusion from input and single causal shift.
• Selected-position loss and gradients match dense causal cross-entropy at rtol=atol=1e-12.
• Label counting covers both Python rows and singleton tensor rows. The check runs on CPU and explicitly overrides only test device placement; it is not an NPU training test.

Historical four-NPU validation at 6e4a8bc3acec1323f7e7ba5bb0935263800876e0: short training, optimizer-state resume, 120-step training, final model reload and seven-suite evaluation completed. Optimizer moments and step restoration were checked exactly in a separate four-rank test. The trained model remained below the official aggregate quality target; final metrics and raw predictions remain local.

Keep this PR draft pending upstream CI/review and device validation of the final PR head. Full training was not repeated for this submission. Images, sequence parallelism, cross-example packing and 35B training are outside the validated scope.