yitongl/sparse_quant_exp

1# sfp4_v4_sparse09_hpo_on_ours_p_init2050 checkpoint-7002 3This upload contains the consolidated WanTransformer3DModel transformer weights4from:5 6`checkpoints/sfp4_v4_sparse09_hpo_on_ours_p_init2050_1n_interactive/checkpoint-700`7 8Contents:9 10- `transformer/config.json`11- `transformer/diffusion_pytorch_model.safetensors`12- `backend_snapshot/`13- `standalone_inference/`14 15Training run:16 17- run name: `sfp4_v4_sparse09_hpo_on_ours_p_init2050_1n_interactive`18- source init: `sfp4_v4_sparse06_hpo_on_ours_p_1n_interactive_v2 checkpoint-2050`19- attention backend: `SPARSE_FP4_OURS_P_ATTN`20- high precision output for backward: enabled21- VSA sparsity: `0.9`22 23This package does not include the distributed optimizer/training-state24checkpoint. Use the original `distributed_checkpoint/` directory if exact25training resume state is required.26 27`backend_snapshot/` contains the local FastVideo backend code used by this28checkpoint, including `SPARSE_FP4_OURS_P_ATTN`, its Triton forward/backward29kernel, FP4 quant helpers, VSA metadata helper, backend wiring, and the exact30SFT launch scripts.31 32It also includes the inference entrypoint snapshot and an example script:33 34- `backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_700.sh`35- `backend_snapshot/training_attention_settings.json`36- `standalone_inference/`37 38Attention setup for this checkpoint:39 40- self-attention: `SPARSE_FP4_OURS_P_ATTN`, FP4 Q/K/V, sparse 64-token VSA41  tiles, group-local P quant, dropped-tile mean compensation42- cross-attention: dense SDPA fallback, not FP4/sparse43- force-dense paths: dense SDPA44 45`standalone_inference/` is a portable helper for normal inference. It contains46an overlay installer, a runner that downloads/loads the checkpoint-70047transformer weights, and the sparse FP4 backend files required by this48checkpoint.49