[docs]: align FastH3 Preview v1 naming
Browse filesUpdate the model card and package self-references after the server-side repository rename.
- README.md +12 -10
- fastvideo_inference.json +1 -1
- modular_model_index.json +8 -8
- provenance.json +2 -2
README.md
CHANGED
|
@@ -16,16 +16,18 @@ tags:
|
|
| 16 |
- minimax-h3
|
| 17 |
- fastvideo
|
| 18 |
- fasth3
|
|
|
|
| 19 |
---
|
| 20 |
|
| 21 |
-
# FastVideo-FastH3-4-step-v1
|
| 22 |
|
| 23 |
-
|
| 24 |
-
student at training step 1300. It generates
|
| 25 |
-
four transformer forwards
|
|
|
|
| 26 |
|
| 27 |
-
|
| 28 |
-
|
| 29 |
|
| 30 |
## Checkpoint identity
|
| 31 |
|
|
@@ -58,7 +60,7 @@ four-point scheduler grid.
|
|
| 58 |
export FASTVIDEO_DMD_DENOISING_STEPS=999,749,500,250
|
| 59 |
|
| 60 |
python examples/inference/basic/basic_fasth3.py \
|
| 61 |
-
--model-path FastVideo/FastVideo-FastH3-4-step-v1 \
|
| 62 |
--prompt '<MiniMax-H3 T2AV prompt>' \
|
| 63 |
--steps 5 \
|
| 64 |
--num-gpus 4 \
|
|
@@ -69,7 +71,7 @@ python examples/inference/basic/basic_fasth3.py \
|
|
| 69 |
|
| 70 |
The five scheduler grid points execute the four trained jumps. Use guidance
|
| 71 |
scale 1.0, VSA-H3 at 90% sparsity with 64-token tiles, and SP=4/FSDP across one
|
| 72 |
-
four-GPU
|
| 73 |
runs the sparse DiT eagerly; regional compile is disabled for this snapshot.
|
| 74 |
|
| 75 |
## Training summary
|
|
@@ -82,12 +84,12 @@ runs the sparse DiT eagerly; regional compile is disabled for this snapshot.
|
|
| 82 |
- Global batch size: 64.
|
| 83 |
- Student and critic learning rate: `2e-6`.
|
| 84 |
- Precision: fp32 training state and bf16 inference export.
|
| 85 |
-
- Training topology: 32
|
| 86 |
|
| 87 |
## Limitations
|
| 88 |
|
| 89 |
- Intermediate checkpoint at step 1300 of a planned 4000-step run.
|
| 90 |
-
- This historical checkpoint predates the later
|
| 91 |
and FastGen-alignment corrections. Use it to reproduce v10.5 results, not as
|
| 92 |
the reference implementation of the corrected training recipe.
|
| 93 |
- T2AV only; the reference-conditioned transformer was not distilled here.
|
|
|
|
| 16 |
- minimax-h3
|
| 17 |
- fastvideo
|
| 18 |
- fasth3
|
| 19 |
+
- preview
|
| 20 |
---
|
| 21 |
|
| 22 |
+
# FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
|
| 23 |
|
| 24 |
+
FastH3 4-Step Preview v1 checkpoint from FastVideo. This variant is the v10.5
|
| 25 |
+
data-free DMD2 MiniMax-H3 student at training step 1300. It generates
|
| 26 |
+
synchronized video and audio with four transformer forwards and uses VSA-H3 at
|
| 27 |
+
90% sparsity.
|
| 28 |
|
| 29 |
+
See the complete [FastH3 4-Step Preview v1
|
| 30 |
+
collection](https://huggingface.co/collections/FastVideo/fastvideo-fasth3).
|
| 31 |
|
| 32 |
## Checkpoint identity
|
| 33 |
|
|
|
|
| 60 |
export FASTVIDEO_DMD_DENOISING_STEPS=999,749,500,250
|
| 61 |
|
| 62 |
python examples/inference/basic/basic_fasth3.py \
|
| 63 |
+
--model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree \
|
| 64 |
--prompt '<MiniMax-H3 T2AV prompt>' \
|
| 65 |
--steps 5 \
|
| 66 |
--num-gpus 4 \
|
|
|
|
| 71 |
|
| 72 |
The five scheduler grid points execute the four trained jumps. Use guidance
|
| 73 |
scale 1.0, VSA-H3 at 90% sparsity with 64-token tiles, and SP=4/FSDP across one
|
| 74 |
+
four-GPU B200 tray for the validated route. The exact v10.5 execution commit
|
| 75 |
runs the sparse DiT eagerly; regional compile is disabled for this snapshot.
|
| 76 |
|
| 77 |
## Training summary
|
|
|
|
| 84 |
- Global batch size: 64.
|
| 85 |
- Student and critic learning rate: `2e-6`.
|
| 86 |
- Precision: fp32 training state and bf16 inference export.
|
| 87 |
+
- Training topology: 32 B200 GPUs, SP=4, full-world FSDP sharding.
|
| 88 |
|
| 89 |
## Limitations
|
| 90 |
|
| 91 |
- Intermediate checkpoint at step 1300 of a planned 4000-step run.
|
| 92 |
+
- This historical checkpoint predates the later continuous score-clock
|
| 93 |
and FastGen-alignment corrections. Use it to reproduce v10.5 results, not as
|
| 94 |
the reference implementation of the corrected training recipe.
|
| 95 |
- T2AV only; the reference-conditioned transformer was not distilled here.
|
fastvideo_inference.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"schema_version": "fasth3-inference-contract-v1",
|
| 3 |
-
"model_id": "FastVideo/FastVideo-FastH3-4-step-v1",
|
| 4 |
"training_parent_run_id": "20260823T132231Z-v10p5-datafree-sp4",
|
| 5 |
"checkpoint_step": 1300,
|
| 6 |
"checkpoint_content_sha256": "b36987515e4c75fa4c7aaa632a7842c829ea141b235358a54d782b51230497b3",
|
|
|
|
| 1 |
{
|
| 2 |
"schema_version": "fasth3-inference-contract-v1",
|
| 3 |
+
"model_id": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 4 |
"training_parent_run_id": "20260823T132231Z-v10p5-datafree-sp4",
|
| 5 |
"checkpoint_step": 1300,
|
| 6 |
"checkpoint_content_sha256": "b36987515e4c75fa4c7aaa632a7842c829ea141b235358a54d782b51230497b3",
|
modular_model_index.json
CHANGED
|
@@ -10,7 +10,7 @@
|
|
| 10 |
"transformers",
|
| 11 |
"Qwen3VLForConditionalGeneration"
|
| 12 |
],
|
| 13 |
-
|
| 14 |
"subfolder": "text_encoder",
|
| 15 |
"variant": null,
|
| 16 |
"revision": null
|
|
@@ -24,7 +24,7 @@
|
|
| 24 |
"transformers",
|
| 25 |
"Qwen2TokenizerFast"
|
| 26 |
],
|
| 27 |
-
|
| 28 |
"subfolder": "tokenizer",
|
| 29 |
"variant": null,
|
| 30 |
"revision": null
|
|
@@ -38,7 +38,7 @@
|
|
| 38 |
"transformers",
|
| 39 |
"Qwen3VLProcessor"
|
| 40 |
],
|
| 41 |
-
|
| 42 |
"subfolder": "processor",
|
| 43 |
"variant": null,
|
| 44 |
"revision": null
|
|
@@ -52,7 +52,7 @@
|
|
| 52 |
"diffusers",
|
| 53 |
"AutoencoderKLMiniMaxH3"
|
| 54 |
],
|
| 55 |
-
|
| 56 |
"subfolder": "vae",
|
| 57 |
"variant": null,
|
| 58 |
"revision": null
|
|
@@ -66,7 +66,7 @@
|
|
| 66 |
"diffusers",
|
| 67 |
"AutoencoderKLMiniMaxH3Audio"
|
| 68 |
],
|
| 69 |
-
|
| 70 |
"subfolder": "audio_vae",
|
| 71 |
"variant": null,
|
| 72 |
"revision": null
|
|
@@ -80,7 +80,7 @@
|
|
| 80 |
"diffusers",
|
| 81 |
"MiniMaxH3Transformer3DModel"
|
| 82 |
],
|
| 83 |
-
|
| 84 |
"subfolder": "transformer",
|
| 85 |
"variant": null,
|
| 86 |
"revision": null
|
|
@@ -108,7 +108,7 @@
|
|
| 108 |
"diffusers",
|
| 109 |
"MiniMaxH3Scheduler"
|
| 110 |
],
|
| 111 |
-
|
| 112 |
"subfolder": "scheduler",
|
| 113 |
"variant": null,
|
| 114 |
"revision": null
|
|
@@ -122,7 +122,7 @@
|
|
| 122 |
"diffusers",
|
| 123 |
"MiniMaxH3Scheduler"
|
| 124 |
],
|
| 125 |
-
|
| 126 |
"subfolder": "audio_scheduler",
|
| 127 |
"variant": null,
|
| 128 |
"revision": null
|
|
|
|
| 10 |
"transformers",
|
| 11 |
"Qwen3VLForConditionalGeneration"
|
| 12 |
],
|
| 13 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 14 |
"subfolder": "text_encoder",
|
| 15 |
"variant": null,
|
| 16 |
"revision": null
|
|
|
|
| 24 |
"transformers",
|
| 25 |
"Qwen2TokenizerFast"
|
| 26 |
],
|
| 27 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 28 |
"subfolder": "tokenizer",
|
| 29 |
"variant": null,
|
| 30 |
"revision": null
|
|
|
|
| 38 |
"transformers",
|
| 39 |
"Qwen3VLProcessor"
|
| 40 |
],
|
| 41 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 42 |
"subfolder": "processor",
|
| 43 |
"variant": null,
|
| 44 |
"revision": null
|
|
|
|
| 52 |
"diffusers",
|
| 53 |
"AutoencoderKLMiniMaxH3"
|
| 54 |
],
|
| 55 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 56 |
"subfolder": "vae",
|
| 57 |
"variant": null,
|
| 58 |
"revision": null
|
|
|
|
| 66 |
"diffusers",
|
| 67 |
"AutoencoderKLMiniMaxH3Audio"
|
| 68 |
],
|
| 69 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 70 |
"subfolder": "audio_vae",
|
| 71 |
"variant": null,
|
| 72 |
"revision": null
|
|
|
|
| 80 |
"diffusers",
|
| 81 |
"MiniMaxH3Transformer3DModel"
|
| 82 |
],
|
| 83 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 84 |
"subfolder": "transformer",
|
| 85 |
"variant": null,
|
| 86 |
"revision": null
|
|
|
|
| 108 |
"diffusers",
|
| 109 |
"MiniMaxH3Scheduler"
|
| 110 |
],
|
| 111 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 112 |
"subfolder": "scheduler",
|
| 113 |
"variant": null,
|
| 114 |
"revision": null
|
|
|
|
| 122 |
"diffusers",
|
| 123 |
"MiniMaxH3Scheduler"
|
| 124 |
],
|
| 125 |
+
"pretrained_model_name_or_path": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 126 |
"subfolder": "audio_scheduler",
|
| 127 |
"variant": null,
|
| 128 |
"revision": null
|
provenance.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"schema_version": "fasth3-checkpoint-provenance-v1",
|
| 3 |
-
"model_id": "FastVideo/FastVideo-FastH3-4-step-v1",
|
| 4 |
"base_model": "hf://MiniMaxAI/MiniMax-H3@9bfb6693f2cf6de171db46d1aa586f67d773a1da",
|
| 5 |
"training_run_id": "20260823T132231Z-v10p5-datafree-sp4",
|
| 6 |
"wandb_url": "https://wandb.ai/wlsaidhi/h3-dmd2-vsa/runs/a1etree2",
|
|
@@ -15,5 +15,5 @@
|
|
| 15 |
"attention_backend": "VIDEO_SPARSE_ATTN_H3",
|
| 16 |
"vsa_tile_size": 64,
|
| 17 |
"vsa_sparsity": 0.9,
|
| 18 |
-
"known_recipe_deviation": "Predates
|
| 19 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"schema_version": "fasth3-checkpoint-provenance-v1",
|
| 3 |
+
"model_id": "FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree",
|
| 4 |
"base_model": "hf://MiniMaxAI/MiniMax-H3@9bfb6693f2cf6de171db46d1aa586f67d773a1da",
|
| 5 |
"training_run_id": "20260823T132231Z-v10p5-datafree-sp4",
|
| 6 |
"wandb_url": "https://wandb.ai/wlsaidhi/h3-dmd2-vsa/runs/a1etree2",
|
|
|
|
| 15 |
"attention_backend": "VIDEO_SPARSE_ATTN_H3",
|
| 16 |
"vsa_tile_size": 64,
|
| 17 |
"vsa_sparsity": 0.9,
|
| 18 |
+
"known_recipe_deviation": "Predates later continuous score-clock and FastGen-alignment corrections."
|
| 19 |
}
|