Add fast-8L bf16 decoder weights

fast-8L/config.json

@@ -0,0 +1,25 @@
+{
+  "name": "FastH3 NVFP4 VAE decoder (fast-8L)",
+  "format": "fastvideo_h3_decoder_nvfp4_deploy_v1",
+  "architecture": "MiniMaxH3VideoViTDecoder3d",
+  "decoder_blocks": 8,
+  "hidden_size": 2048,
+  "nvfp4": {
+    "scheme": "W4A4 NVFP4 (E2M1 values, E4M3 per-16 block scales, FP32 per-tensor scale)",
+    "block_size": 16,
+    "activation_global_scale": "unit",
+    "rotation_group": null,
+    "quantized_linears": 48,
+    "quantized_layers": "transformer_blocks.*.{attn.to_q,attn.to_k,attn.to_v,attn.to_out.0,ff.net.0.proj,ff.net.2}",
+    "high_precision": "proj_in, proj_out, norms, register tokens, post_quant_conv"
+  },
+  "stored_dtype_note": "NVFP4 linear weights are stored as bf16 master weights and packed to NVFP4 at load; packing is deterministic, so decoded pixels match training.",
+  "compute_dtype": "bfloat16 autocast",
+  "host_vae": "any MiniMax-H3 video VAE with >= 8 decoder blocks (e.g. the 26-block light VAE of the FastH3 V2/Trim stacks); the loader keeps 8",
+  "tensor_count_by_dtype": {
+    "float32": 89,
+    "bfloat16": 96
+  },
+  "nvfp4_linear_parameters": 537083904,
+  "source_step": 5000
+}

fast-8L/decoder.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:154391a86fbb46ce299f4dc9ed705c924a992790a206dc500527998bc04f8375
+size 1099886160