[docs]: align FastH3 Preview v1 naming
Browse filesUpdate the model card and package self-references after the server-side repository rename.
- README.md +19 -14
- fastvideo_inference.json +1 -1
- modular_model_index.json +8 -8
- provenance.json +1 -1
README.md
CHANGED
|
@@ -16,16 +16,18 @@ tags:
|
|
| 16 |
- minimax-h3
|
| 17 |
- fastvideo
|
| 18 |
- fasth3
|
|
|
|
| 19 |
---
|
| 20 |
|
| 21 |
-
# FastVideo-FastH3-
|
| 22 |
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
|
|
|
| 26 |
|
| 27 |
-
|
| 28 |
-
|
| 29 |
|
| 30 |
## Checkpoint identity
|
| 31 |
|
|
@@ -61,29 +63,32 @@ export FASTVIDEO_DMD_DENOISING_STEPS=999,749,500,250
|
|
| 61 |
export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
|
| 62 |
|
| 63 |
python examples/inference/basic/basic_minimax_h3_t2v.py \
|
| 64 |
-
--model-path FastVideo/FastVideo-FastH3-
|
| 65 |
--prompt '<MiniMax-H3 T2AV prompt>' \
|
| 66 |
--steps 5 \
|
| 67 |
--num-gpus 4
|
| 68 |
```
|
| 69 |
|
| 70 |
The five scheduler grid points execute the four trained jumps. Use guidance
|
| 71 |
-
scale 1.0 and dense FlashAttention (the FA4 route on the pinned
|
| 72 |
-
optimization; first validate output parity
|
|
|
|
| 73 |
|
| 74 |
## Training summary
|
| 75 |
|
| 76 |
-
- Method: data-free DMD2 with carried backward-simulation ODE rollouts.
|
|
|
|
|
|
|
| 77 |
- Student grid: `[999, 749, 500, 250]`.
|
| 78 |
-
- Student attention: dense FlashAttention (the FA4 route on the pinned
|
|
|
|
| 79 |
- Global batch size: 64.
|
| 80 |
- Student and critic learning rate: `2e-06` and
|
| 81 |
`2e-06`.
|
| 82 |
- Precision: fp32 training state and bf16 inference export.
|
| 83 |
-
- Training topology: 32
|
| 84 |
HSDP replicate=1 / shard=32.
|
| 85 |
-
-
|
| 86 |
-
shift `2.4`, warp max `0.999`.
|
| 87 |
|
| 88 |
## Limitations
|
| 89 |
|
|
|
|
| 16 |
- minimax-h3
|
| 17 |
- fastvideo
|
| 18 |
- fasth3
|
| 19 |
+
- preview
|
| 20 |
---
|
| 21 |
|
| 22 |
+
# FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree
|
| 23 |
|
| 24 |
+
FastH3 4-Step Preview v1 checkpoint from FastVideo. This variant is the
|
| 25 |
+
`dmd2_sp4_fsdp32_v12_datafree_mixed_dense_fa4` data-free DMD2 MiniMax-H3
|
| 26 |
+
student at training step 1000. It generates synchronized video and audio with
|
| 27 |
+
four transformer forwards and uses dense FlashAttention.
|
| 28 |
|
| 29 |
+
See the complete [FastH3 4-Step Preview v1
|
| 30 |
+
collection](https://huggingface.co/collections/FastVideo/fastvideo-fasth3).
|
| 31 |
|
| 32 |
## Checkpoint identity
|
| 33 |
|
|
|
|
| 63 |
export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
|
| 64 |
|
| 65 |
python examples/inference/basic/basic_minimax_h3_t2v.py \
|
| 66 |
+
--model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree \
|
| 67 |
--prompt '<MiniMax-H3 T2AV prompt>' \
|
| 68 |
--steps 5 \
|
| 69 |
--num-gpus 4
|
| 70 |
```
|
| 71 |
|
| 72 |
The five scheduler grid points execute the four trained jumps. Use guidance
|
| 73 |
+
scale 1.0 and dense FlashAttention (the FA4 route on the pinned B200 stack).
|
| 74 |
+
Compilation is an optional runtime optimization; first validate output parity
|
| 75 |
+
on the pinned FastVideo commit.
|
| 76 |
|
| 77 |
## Training summary
|
| 78 |
|
| 79 |
+
- Method: data-free DMD2 with carried backward-simulation ODE rollouts.
|
| 80 |
+
Training consumes text conditioning and student-generated latents rather
|
| 81 |
+
than target video latents.
|
| 82 |
- Student grid: `[999, 749, 500, 250]`.
|
| 83 |
+
- Student attention: dense FlashAttention (the FA4 route on the pinned B200
|
| 84 |
+
stack).
|
| 85 |
- Global batch size: 64.
|
| 86 |
- Student and critic learning rate: `2e-06` and
|
| 87 |
`2e-06`.
|
| 88 |
- Precision: fp32 training state and bf16 inference export.
|
| 89 |
+
- Training topology: 32 B200 GPUs, SP=4,
|
| 90 |
HSDP replicate=1 / shard=32.
|
| 91 |
+
- Continuous score-clock shift: `2.4`; warp max: `0.999`.
|
|
|
|
| 92 |
|
| 93 |
## Limitations
|
| 94 |
|
fastvideo_inference.json
CHANGED
|
@@ -11,7 +11,7 @@
|
|
| 11 |
],
|
| 12 |
"fastvideo_commit": "aa6e6c1c1e15f879011a47996c7d6bc8381eaf89",
|
| 13 |
"guidance_scale": 1.0,
|
| 14 |
-
"model_id": "FastVideo/FastVideo-FastH3-
|
| 15 |
"num_gpus": 4,
|
| 16 |
"num_inference_steps": 5,
|
| 17 |
"schema_version": "fasth3-inference-contract-v1",
|
|
|
|
| 11 |
],
|
| 12 |
"fastvideo_commit": "aa6e6c1c1e15f879011a47996c7d6bc8381eaf89",
|
| 13 |
"guidance_scale": 1.0,
|
| 14 |
+
"model_id": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 15 |
"num_gpus": 4,
|
| 16 |
"num_inference_steps": 5,
|
| 17 |
"schema_version": "fasth3-inference-contract-v1",
|
modular_model_index.json
CHANGED
|
@@ -6,7 +6,7 @@
|
|
| 6 |
"diffusers",
|
| 7 |
"MiniMaxH3Scheduler",
|
| 8 |
{
|
| 9 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 10 |
"revision": null,
|
| 11 |
"subfolder": "audio_scheduler",
|
| 12 |
"type_hint": [
|
|
@@ -20,7 +20,7 @@
|
|
| 20 |
"diffusers",
|
| 21 |
"AutoencoderKLMiniMaxH3Audio",
|
| 22 |
{
|
| 23 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 24 |
"revision": null,
|
| 25 |
"subfolder": "audio_vae",
|
| 26 |
"type_hint": [
|
|
@@ -34,7 +34,7 @@
|
|
| 34 |
"transformers",
|
| 35 |
"Qwen3VLProcessor",
|
| 36 |
{
|
| 37 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 38 |
"revision": null,
|
| 39 |
"subfolder": "processor",
|
| 40 |
"type_hint": [
|
|
@@ -48,7 +48,7 @@
|
|
| 48 |
"diffusers",
|
| 49 |
"MiniMaxH3Scheduler",
|
| 50 |
{
|
| 51 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 52 |
"revision": null,
|
| 53 |
"subfolder": "scheduler",
|
| 54 |
"type_hint": [
|
|
@@ -62,7 +62,7 @@
|
|
| 62 |
"transformers",
|
| 63 |
"Qwen3VLForConditionalGeneration",
|
| 64 |
{
|
| 65 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 66 |
"revision": null,
|
| 67 |
"subfolder": "text_encoder",
|
| 68 |
"type_hint": [
|
|
@@ -76,7 +76,7 @@
|
|
| 76 |
"transformers",
|
| 77 |
"Qwen2TokenizerFast",
|
| 78 |
{
|
| 79 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 80 |
"revision": null,
|
| 81 |
"subfolder": "tokenizer",
|
| 82 |
"type_hint": [
|
|
@@ -90,7 +90,7 @@
|
|
| 90 |
"diffusers",
|
| 91 |
"MiniMaxH3Transformer3DModel",
|
| 92 |
{
|
| 93 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 94 |
"revision": null,
|
| 95 |
"subfolder": "transformer",
|
| 96 |
"type_hint": [
|
|
@@ -118,7 +118,7 @@
|
|
| 118 |
"diffusers",
|
| 119 |
"AutoencoderKLMiniMaxH3",
|
| 120 |
{
|
| 121 |
-
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-
|
| 122 |
"revision": null,
|
| 123 |
"subfolder": "vae",
|
| 124 |
"type_hint": [
|
|
|
|
| 6 |
"diffusers",
|
| 7 |
"MiniMaxH3Scheduler",
|
| 8 |
{
|
| 9 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 10 |
"revision": null,
|
| 11 |
"subfolder": "audio_scheduler",
|
| 12 |
"type_hint": [
|
|
|
|
| 20 |
"diffusers",
|
| 21 |
"AutoencoderKLMiniMaxH3Audio",
|
| 22 |
{
|
| 23 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 24 |
"revision": null,
|
| 25 |
"subfolder": "audio_vae",
|
| 26 |
"type_hint": [
|
|
|
|
| 34 |
"transformers",
|
| 35 |
"Qwen3VLProcessor",
|
| 36 |
{
|
| 37 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 38 |
"revision": null,
|
| 39 |
"subfolder": "processor",
|
| 40 |
"type_hint": [
|
|
|
|
| 48 |
"diffusers",
|
| 49 |
"MiniMaxH3Scheduler",
|
| 50 |
{
|
| 51 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 52 |
"revision": null,
|
| 53 |
"subfolder": "scheduler",
|
| 54 |
"type_hint": [
|
|
|
|
| 62 |
"transformers",
|
| 63 |
"Qwen3VLForConditionalGeneration",
|
| 64 |
{
|
| 65 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 66 |
"revision": null,
|
| 67 |
"subfolder": "text_encoder",
|
| 68 |
"type_hint": [
|
|
|
|
| 76 |
"transformers",
|
| 77 |
"Qwen2TokenizerFast",
|
| 78 |
{
|
| 79 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 80 |
"revision": null,
|
| 81 |
"subfolder": "tokenizer",
|
| 82 |
"type_hint": [
|
|
|
|
| 90 |
"diffusers",
|
| 91 |
"MiniMaxH3Transformer3DModel",
|
| 92 |
{
|
| 93 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 94 |
"revision": null,
|
| 95 |
"subfolder": "transformer",
|
| 96 |
"type_hint": [
|
|
|
|
| 118 |
"diffusers",
|
| 119 |
"AutoencoderKLMiniMaxH3",
|
| 120 |
{
|
| 121 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 122 |
"revision": null,
|
| 123 |
"subfolder": "vae",
|
| 124 |
"type_hint": [
|
provenance.json
CHANGED
|
@@ -13,7 +13,7 @@
|
|
| 13 |
],
|
| 14 |
"fastvideo_commit": "aa6e6c1c1e15f879011a47996c7d6bc8381eaf89",
|
| 15 |
"lineage_parent_run_id": "20260825T010407Z-v12-datafree-dense-fa4",
|
| 16 |
-
"model_id": "FastVideo/FastVideo-FastH3-
|
| 17 |
"schema_version": "fasth3-checkpoint-provenance-v1",
|
| 18 |
"source_config_sha256": "50519042643df51fd115dce3294ee38a703a9cf496c28560dd07bb265d60b7dd",
|
| 19 |
"student_dtype": "bfloat16",
|
|
|
|
| 13 |
],
|
| 14 |
"fastvideo_commit": "aa6e6c1c1e15f879011a47996c7d6bc8381eaf89",
|
| 15 |
"lineage_parent_run_id": "20260825T010407Z-v12-datafree-dense-fa4",
|
| 16 |
+
"model_id": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree",
|
| 17 |
"schema_version": "fasth3-checkpoint-provenance-v1",
|
| 18 |
"source_config_sha256": "50519042643df51fd115dce3294ee38a703a9cf496c28560dd07bb265d60b7dd",
|
| 19 |
"student_dtype": "bfloat16",
|