[docs]: make the V2.1 model card self-contained

Describe V2.1 and its QAD recipe directly; remove the V2 comparison table and the earlier-run narrative.

README.md

@@ -5,36 +5,32 @@ tags: [video-generation, text-to-video, audio, fastvideo, nvfp4, qad]
 ---
 # FastH3 V2.1, 8-step, NVFP4 for one GPU (QAD)
 
-FastH3 V2.1 is FastH3 V2 (8 distilled steps, video with synchronized audio) with NVFP4 weights recovered by
-**quantization-aware distillation (QAD)**. It is a drop-in replacement for
-[FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer):
-same layout, same speed, same single Blackwell GPU (RTX 5090, RTX PRO 6000) or DGX Spark target. Only
-`transformer/nvfp4_weights.safetensors` and this card differ.
-
-- **Transformer:** FastH3 V2 (50 blocks), NVFP4 W4A4 on all attention q/k/v/out and MLP linears, plus the sparse-attention
-  (VSA) gates. The q/k/v/out and MLP weights were trained with QAD. The VSA gates and all norms are V2's, byte-identical to the
-  Consumer release. Attention itself runs in bf16 sparse attention (VSA, tile 64, sparsity 0.8), as in V2.
+FastH3 V2.1 is an 8-step video-and-audio model with NVFP4 weights, trained with **quantization-aware distillation
+(QAD)**. It runs on one Blackwell GPU (RTX 5090, RTX PRO 6000) or a DGX Spark.
+
+- **Transformer:** 50 blocks, NVFP4 W4A4 on all attention q/k/v/out and MLP linears, plus the sparse-attention
+  (VSA) gates. The q/k/v/out and MLP weights were trained with QAD. The VSA gates and all norms stay in BF16.
+- **Attention:** block-sparse VSA with 64-token tiles and sparsity 0.8. The checkpoint was trained with BF16 VSA.
+  FastVideo also exposes an FP4 VSA inference lane on Blackwell.
 - **Text encoder:** NVFP4 Qwen3-VL trimmed to the 50 layers H3 reads.
 - **VAE:** LynnReal lightweight video VAE with Kijai's INT8 weights; H3 audio VAE.
 - **Sampling:** `fastvideo_inference.json` holds the 8-step DMD schedule, which FastVideo reads automatically.
 
-## What changed from V2 NVFP4
-
-QAD trains the NVFP4 student to match the bf16 V2 teacher, step by step along the teacher's 8-step trajectory (data-free:
-text prompts only, 300 updates, 8 prompts per update, lr 5e-6, 25% of rows on the student's own trajectory). Only the
-quantized linears were trained. The VSA gates and norms were frozen: training the gates improved the metric but caused
-ghosting on fast motion.
+## Quantization-aware distillation
 
-Per-forward error against the bf16 teacher on 30 held-out prompts (relative L2 of the x0 prediction, lower is better):
+QAD trains the NVFP4 student against a frozen BF16 teacher along the teacher's 8-step trajectory. The student runs
+the deployed 4-bit weights and activations, so it learns weights that work with the quantized forward pass. Training
+uses text prompts only: 300 steps, learning rate 5e-6, with 25% of rows on the student's own trajectory. Only the
+quantized linears train. The VSA gates and norms are frozen because changing them changes which video tiles the
+model attends to, not just the arithmetic of attention.
 
-| | V2 NVFP4 | **V2.1 (QAD)** |
-|---|---|---|
-| Mean (video + audio) | 0.431 | **0.351** |
-| Video | 0.471 | 0.382 |
-| Audio | 0.391 | 0.319 |
+Error against the BF16 teacher on 30 held-out prompts (relative L2 of the x0 prediction, lower is better):
 
-Side-by-side renders on a DGX Spark (11 prompts: faces, fast motion, dance, crowds) show no new artifacts versus V2 NVFP4, with
-equal or slightly better faces and fine detail. End-to-end time is unchanged.
+| Metric | V2.1 (QAD) |
+|---|---:|
+| Mean (video + audio) | 0.351 |
+| Video | 0.382 |
+| Audio | 0.319 |
 
 ## Machines with less than 64 GB of system RAM