[zhouaihui]

Outcome

Use native cuDNN THD attention with CP2 for the Qwen3.8-27B long-context recipe. The capacity audit checks that the observed backend matches the saved configuration while retaining the full 261,888-token requirements. Allow cold TP engine compilation during initialization.

Verification

• 103 relevant tests passed; lint, documentation links and wheel/sdist audits passed in an isolated Linux environment.
• Eight-GPU SM90 acceptance passed: 261,888-token synthetic inputs, training 0→4, exact resume 4→8, evaluation, native export and strict loading. Both phases completed healthy with cuDNN CP2 and no fallback.
• Acceptance uses accumulation 1; the training recipe defaults to 2. This establishes bounded capacity and workflow behavior, not quality or comparative performance.

Compatibility and provenance

The launcher ID remains qwen27b-long. Direct YAML users should use configs/recipes/qwen3_8_27b_256k_cudnn_cp2_nvlink.yaml. Explicit FA4 configurations remain supported by the audit. Checkpoint and export formats are unchanged; existing AngelSpec attribution is retained.