memset0/vsqa-test-1.3b

VSQA E0029 V0247 — step 1000

Student model exported from E0029 V0247, checkpoint 1000. Includes the trained transformer and the base model's frozen pipeline components. Optimizer and critic states are not included.

• Base model: Wan-AI/Wan2.1-T2V-1.3B-Diffusers.
• Attention: VSA_QAT_TRAIN_C128, logical cube (4, 4, 8), sparsity parameter 0.9, trained compression gates.
• Quantization: NVFP4 Q/K/P/V attention QAT and NVFP4 W4A4 linear QAT. Safetensors contain floating-point master weights, not a packed FP4 deployment model.
• Distillation: DMD2, four-step ladder [1000, 750, 500, 250], CFG 1, flow shift 8.
• Training geometry: 61 frames, 448 × 896; effective batch 16.

Use a compatible FastVideo VSQA implementation supporting this attention backend, compression gates, and linear QAT contract. This export has been strictly reloaded with the frozen training implementation; stock Diffusers inference compatibility is not asserted.

metadata.json and artifact_manifest.json describe the native export. Original training metadata is retained in provenance/training_checkpoint/; provenance/export_relationship.json binds both metadata files by SHA-256. source.yaml and resolved.yaml are audit artifacts and are not portable runtime recipes. The VSQA management-sidecar cutover is not implemented.