memset0/vsqa-preview-14b-v2

VSQA preview 14B v2 — DMD 200 steps (critic LR 2e-6, linear NVFP4 global scale 1)

Preview checkpoint, not yet evaluated. Final student of a 200-step DMD run (E0029 Variant V0403). It differs from memset0/vsqa-preview-14b (V0362) in two ways:

• critic (fake score) learning rate 2e-6 instead of 4e-7 (critic update ratio 1 unchanged);
• the linear layers' NVFP4 QAT uses global scales fixed to 1.0 (nvfp4_qat_train_gs1) for activations and weights, instead of the dynamic 448·6/max|x| recomputed every forward.

Training lineage

All at 77 frames, 1280×768, global batch 16, from the original Wan2.1-T2V-14B weights:

1. Dense NVFP4 per-layer reconstruction, 250 steps.
2. Plain ODE initialization (endpoint regression), 500 steps, VSA C256 cube (4,8,8), sparsity 0.9, NVFP4 attention + linear QAT.
3. DMD distillation, 200 steps from step 2: student LR 2e-6, critic LR 2e-6, generator updated every iteration, no EMA; teacher and critic are the original dense Wan2.1-T2V-14B.

Sampling contract

• 3 denoising steps at timesteps [1000.0, 941.1763916015625, 800.0], no additional timestep warp.
• flow_shift = 8.0, guidance scale 1.0 (distilled), 77 frames at 1280×768.
• Training attention backend VSA_QAT_TRAIN_C256 (sparsity 0.9); linear quantization nvfp4_qat_train_gs1. Reproducing the trained numerics needs the FastVideo VSA/QAT kernels plus the global-scale-1 patch; plain dense BF16 inference is a different, untested configuration.

Files

Diffusers layout; transformer/ is the exported student, other components are unchanged from the base model. metadata.json, artifact_manifest.json, resolved.yaml, source.yaml and export-verification.json record provenance and the strict-reload export verification. License follows the base model (Apache-2.0).