Document self-describing loading and generation validation
Browse files
README.md
CHANGED
|
@@ -23,21 +23,20 @@ This repository is a packed UniServe checkpoint derived from [FastVideo/FastVide
|
|
| 23 |
|
| 24 |
## Run with UniServe
|
| 25 |
|
| 26 |
-
UniServe detects `modelopt_manifest.json` automatically
|
| 27 |
|
| 28 |
```bash
|
| 29 |
uniserve serve skx618/FastVideo-FastH3-8-Step-V2-NVFP4 \
|
| 30 |
--model-description minimax-h3 \
|
| 31 |
--served-model-name MiniMax-H3 \
|
| 32 |
-
--dtype bfloat16
|
| 33 |
-
--quantization-config '{"mode":"calibrated"}'
|
| 34 |
```
|
| 35 |
|
| 36 |
The checkpoint targets the source model's four-B200 execution contract. The GPU count must divide H3's 56 attention heads.
|
| 37 |
|
| 38 |
## Validation status
|
| 39 |
|
| 40 |
-
The denoiser and Video VAE components were exported after their complete 1,000-record Phase A and Phase B calibration jobs.
|
| 41 |
|
| 42 |
## License
|
| 43 |
|
|
|
|
| 23 |
|
| 24 |
## Run with UniServe
|
| 25 |
|
| 26 |
+
UniServe detects and validates `modelopt_manifest.json` automatically. The packed checkpoint does not require NVIDIA Model Optimizer at deployment time and rejects runtime precision presets or component overrides that would conflict with its calibrated weights and scales.
|
| 27 |
|
| 28 |
```bash
|
| 29 |
uniserve serve skx618/FastVideo-FastH3-8-Step-V2-NVFP4 \
|
| 30 |
--model-description minimax-h3 \
|
| 31 |
--served-model-name MiniMax-H3 \
|
| 32 |
+
--dtype bfloat16
|
|
|
|
| 33 |
```
|
| 34 |
|
| 35 |
The checkpoint targets the source model's four-B200 execution contract. The GPU count must divide H3's 56 attention heads.
|
| 36 |
|
| 37 |
## Validation status
|
| 38 |
|
| 39 |
+
The denoiser and Video VAE components were exported after their complete 1,000-record Phase A and Phase B calibration jobs. An executable-checkpoint review generated five 5-second, 1344×768, 24-fps synchronized audio/video samples from this packed checkpoint and five same-prompt, same-seed quality-reference samples successfully while preserving the source model's VSA sparsity 0.8 contract. Across those five pairs, mean RGB PSNR was 11.92 dB, SSIM 0.3647, AlexNet LPIPS 0.5184, and audio log-mel cosine 0.9346. These prompts came from the calibration-source dataset, so the measurements describe trajectory drift and generation viability, not independent held-out quality acceptance; no perceptual-equivalence claim is made.
|
| 40 |
|
| 41 |
## License
|
| 42 |
|