Finalize FastH3 OmniRef release model card

README.md

@@ -5,62 +5,42 @@ license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
 base_model: FastVideo/FastVideo-FastH3-Omni-8-Step-V1
 tags:
 - video-generation
-- text-to-audio-video
 - reference-to-video
-- minimax-h3
+- text-to-audio-video
 - fastvideo
+- minimax-h3
+- omniref
 - fp8
 ---
 
 # FastH3 OmniRef FP8
 
-FastH3 OmniRef generates a video with synchronized audio from a prompt and ordered image, video, and audio references. It runs as an eight-forward PDD student.
+FP8 diffusion weights for FastH3 OmniRef, the eight-evaluation reference-conditioned video-and-audio model. This is the FastVideo-format FP8 variant of the [BF16 release](https://huggingface.co/FastVideo/FastVideo-FastH3-Omni-8-Step-V1).
 
-This repository contains an **FP8 W8A8 quantization of the BF16 OmniRef model** [`FastVideo/FastVideo-FastH3-Omni-8-Step-V1`](https://huggingface.co/FastVideo/FastVideo-FastH3-Omni-8-Step-V1). It is intended for GPUs without FP4 tensor cores: RTX 4090 / L40S / RTX 6000 Ada (sm_89), H100 / H200 (sm_90), and Blackwell GPUs when an FP8 deployment is preferred.
+[Release collection](https://huggingface.co/collections/FastVideo/fasth3-omniref) · [FastVideo inference guide](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md)
 
-## Quantization
+## Use with FastVideo
 
-- The attention and MLP projections in all 50 transformer blocks are stored as E4M3 FP8 with per-output-channel weight scales.
-- Activations are quantized dynamically per token at runtime.
-- The remaining transformer tensors stay in BF16: AdaLN and timestep projections, embeddings, norms, VSA compression gates, and the token refiner.
-- The text encoder, video VAE, audio VAE, schedulers, tokenizer, and processor are unchanged from the BF16 release.
-- This is a post-training quantization export. No retraining is applied.
+Follow the [OmniRef device and memory guide](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md#device-profiles). The FP8 transformer is useful on GPUs without FP4 tensor cores, including RTX 4090 and Hopper, and on Blackwell when FP8 is preferred. The RTX 4090 profile uses layerwise transformer offload and the dense video VAE.
 
-## Contents
+The bundled text encoder is NVFP4 and needs Blackwell FP4 kernels. On non-FP4 hardware, compose the transformer with compatible BF16 text-encoder components from MiniMax-H3 rather than selecting that bundled encoder. The guide explains model-directory composition and memory requirements.
 
-| Component | Format |
-|---|---|
-| `transformer/`, `transformer_ref/` | Identical FP8 linears (q, k, v, out, fc_in, fc_out of all 50 blocks) plus BF16 norms, embeddings, gates, AdaLN, and token refiner |
-| `text_encoder/` | NVFP4 Qwen3-VL text encoder |
-| `vae/` | Dense light video VAE for reference encoding and decoding |
-| `audio_vae/`, `scheduler/`, `audio_scheduler/`, `processor/`, `tokenizer/` | Unchanged from the BF16 release |
-| `fastvideo_inference.json` | Eight-forward PDD contract, video/audio shifts 12/3, VSA 0.9 with 128-token tiles |
-| `adaln_table.pt` | Precomputed AdaLN outputs for the fixed timestep ladder; skips loading the large AdaLN projections |
+For UniServe deployments on GB200/B200, RTX PRO 6000 Blackwell or Hopper, start from the [BF16 release](https://huggingface.co/FastVideo/FastVideo-FastH3-Omni-8-Step-V1) and follow the [UniServe guide](https://github.com/hao-ai-lab/uniserve/blob/main/docs/minimax_h3/minimax_h3.md). For Spark and RTX 5090 local inference, use the [NVFP4 QAD variant](https://huggingface.co/FastVideo/FastH3-OmniRef-NVFP4-QAD) with FastVideo's matching profiles.
 
-## Running with FastVideo
-
-Use the OmniRef consumer profile with an FP8 model path. On a 24 GB card, enable layerwise DiT offloading.
-
-```bash
-python examples/inference/basic/basic_fasth3_omniref_pdd.py \
-  --model-path ./FastH3-OmniRef-FP8 \
-  --base-model-path ./FastH3-OmniRef-FP8 \
-  --composed-dir ./FastH3-OmniRef-FP8-composed \
-  --adaln-table ./FastH3-OmniRef-FP8/adaln_table.pt \
-  --image reference.png \
-  --prompt "a person on a sofa, lit by a warm lamp, speaking to camera" \
-  --profile rtx-4090
-```
-
-The `rtx-4090` profile selects FP8 linears, BF16 VSA attention on the 128-token contract, the dense video VAE, and a non-resident encoder. On Hopper or Blackwell data-center GPUs, the same profile applies; alternatively use no profile and disable FP4 attention.
+## Contents
 
-## Validation and limits
+| Component | Purpose |
+| --- | --- |
+| `transformer_ref/` | FP8 reference-conditioned transformer with BF16 sensitive components |
+| `transformer/` | Alternate directory containing the same transformer weights |
+| `text_encoder/` | NVFP4 Qwen3-VL text and vision encoder; Blackwell only |
+| `vae/`, `audio_vae/` | Video and audio VAEs |
+| `scheduler/`, `audio_scheduler/`, `processor/`, `tokenizer/` | Sampling and input processing |
+| `fastvideo_inference.json` | Eight-evaluation OmniRef inference configuration |
+| `adaln_table.pt` | Precomputed modulation table for the supplied schedule |
 
-- The export loads through the FastVideo FP8 loader and reproduces the BF16 model's reference conditioning and eight-forward schedule.
-- FP8 changes the numerical output relative to BF16, as expected for W8A8 execution.
-- On a 24 GB card, the DiT does not fit without layerwise streaming. The example profile enables the validated offload path.
-- The AdaLN table covers the shipped schedule and the reference layouts in the model card. Unsupported layouts fall back to the dense AdaLN projections when the table is not passed.
+FP8 quantization changes numerical output relative to BF16. These weights use the FastVideo FP8 layout; use the [FP8 ComfyUI variant](https://huggingface.co/FastVideo/FastH3-OmniRef-FP8-Comfy) for ComfyUI.
 
 ## License
 
-MiniMax-H3 and this derivative are governed by the MiniMax-H3 Community License Agreement. See the linked license for terms.
+Use of these weights is governed by the [MiniMax-H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE).