Skip to content
FastVideo
tfsft
Initializing search
hao-ai-lab/FastVideo
Home
Getting Started
Inference
Training
Distillation
Attention
Utilities
Design
Developer Guide
API Reference
FastVideo
hao-ai-lab/FastVideo
Home
Getting Started
Getting Started
Installation
Quick Start
V1 API
Inference
Inference
Quick Start
Configuration
Offloading
Optimizations
ComfyUI
Support Matrix
CLI
Add Pipeline
Examples
Examples
Basic
Optimizations
STA Mask Search (Archived)
Training
Training
Overview
Training Infrastructure
Experiment Tracking
Data Preprocessing
Fine-tuning
Attn-QAT Training
Examples
Examples
Training Index
Wan T2V 1.3B
Wan I2V 14B 480p
Wan2.1 Fun 1.3B InP
Wan2.1 VSA
Fine-tuning
Distillation
Distillation
Data Preprocessing
DMD
AnyFlow
Attention
Attention
Overview
Video Sparse Attention
Sliding Tile Attention (Archived)
Backend Development
Utilities
Utilities
LoRA
Debugging
Design
Design
Overview
Training Architecture
Server Contracts
Server Contracts
Overview
OpenAI HTTP
Streaming WebSocket
Dynamo Integration
Developer Guide
Developer Guide
Overview
Developer Environment
Developer Environment
Index
Docker
RunPod
Pull Requests
CI/CD Architecture
Testing
Performance Benchmarks
Profiling
Activation Trace
Coding Agents
Dreamverse Development
Backend Development
API Reference
API Reference
Overview
fastvideo
fastvideo
fastvideo
api
api
api
compat
errors
flux
matrixgame2
matrixgame3
overrides
parser
presets
request_metadata
results
sampling_param
schema
attention
attention
attention
backends
backends
backends
abstract
attn_qat_infer
attn_qat_train
bsa_attn
flash_attn
nabla
sage_attn
sage_attn3
sdpa
sla
video_sparse_attn
vmoba
layer
selector
configs
configs
configs
configs
models
models
models
audio
audio
audio
ltx2_audio_vae
base
dits
dits
dits
base
cosmos
cosmos2_5
dreamx_world
flux
flux_2
gen3c
glm_image
hunyuangamecraft
hunyuanvideo
hunyuanvideo15
hyworld
kandinsky5
lingbot_video
lingbotworld
lingbotworld2
longcat
ltx2
magi_human
matrixgame2
matrixgame3
sd3
stable_audio
wanvideo
zimage
encoders
encoders
encoders
base
clip
gemma
lingbot_video
lingbotworld2_t5
llama
mistral3
qwen2_5
qwen3
reason1
siglip
stable_audio_conditioner
t5
t5gemma
upsamplers
upsamplers
upsamplers
base
hunyuan15
vaes
vaes
vaes
autoencoder_kl
base
cosmos2_5vae
cosmosvae
flux2vae
gamecraftvae
gen3cvae
glm_image
hunyuan15vae
hunyuanvae
ltx2vae
oobleck
wanvae
pipelines
pipelines
pipelines
base
cosmos
cosmos2_5
dreamx_world
flux
flux_2
gen3c
glm_image
hunyuan
hunyuan15
hunyuangamecraft
hyworld
kandinsky5
lingbot_video
lingbotworld
lingbotworld2
longcat
matrixgame2
matrixgame3
sd35
stable_audio
turbodiffusion
wan
zimage
utils
dataset
dataset
dataset
latent_datasets
ltx2_precomputed_dataset
parquet_dataset_iterable_style
parquet_dataset_map_style
preprocessing_datasets
transform
upload_to_hf
utils
validation_dataset
distributed
distributed
distributed
communication_op
device_communicators
device_communicators
device_communicators
base_device_communicator
cpu_communicator
cuda_communicator
npu_communicator
pyhccl
pyhccl_wrapper
pynccl
pynccl_wrapper
parallel_state
utils
entrypoints
entrypoints
entrypoints
cli
cli
cli
bench
bench_serving
cli_types
eval
generate
inference_config
main
router_serve
serve
utils
openai
openai
openai
api_server
common_api
image_api
protocol
state
stores
utils
video_api
streaming
streaming
streaming
gpu_pool
health
mock_server
prompt
prompt
prompt
enhancer
providers
providers
providers
base
cerebras
groq
rewrite
safety
protocol
router
router
router
config
main
registry
server
session
session_init_image
session_logger
session_store
stream
worker
streaming_generator
video_generator
envs
eval
eval
eval
api
datasets
datasets
datasets
base
physics_iq
registry
vbench
evaluator
io
io
io
audio
inputs
paths
video
memory
metrics
metrics
metrics
audio
audio
audio
audiobox_aesthetics
audiobox_aesthetics
audiobox_aesthetics
metric
clap_score
clap_score
clap_score
metric
desync
desync
desync
metric
frechet_distance
frechet_distance
frechet_distance
metric
imagebind_score
imagebind_score
imagebind_score
metric
kl_divergence
kl_divergence
kl_divergence
metric
wer
wer
wer
metric
base
common
common
common
fvd
fvd
fvd
extractors
metric
lpips
lpips
lpips
metric
psnr
psnr
psnr
metric
ssim
ssim
ssim
metric
judge
judge
judge
third_person_separation
third_person_separation
third_person_separation
metric
optical_flow
optical_flow
optical_flow
gt_optical_flow
gt_optical_flow
gt_optical_flow
metric
synthetic_optical_flow
synthetic_optical_flow
synthetic_optical_flow
metric
physics_iq
physics_iq
physics_iq
metric
mse
mse
mse
metric
spatial_iou
spatial_iou
spatial_iou
metric
spatiotemporal_iou
spatiotemporal_iou
spatiotemporal_iou
metric
utils
weighted_spatial_iou
weighted_spatial_iou
weighted_spatial_iou
metric
vbench
vbench
vbench
aesthetic_quality
aesthetic_quality
aesthetic_quality
metric
appearance_style
appearance_style
appearance_style
metric
background_consistency
background_consistency
background_consistency
metric
color
color
color
metric
dynamic_degree
dynamic_degree
dynamic_degree
metric
human_action
human_action
human_action
metric
imaging_quality
imaging_quality
imaging_quality
metric
motion_smoothness
motion_smoothness
motion_smoothness
metric
multiple_objects
multiple_objects
multiple_objects
metric
object_class
object_class
object_class
metric
overall_consistency
overall_consistency
overall_consistency
metric
scene
scene
scene
metric
spatial_relationship
spatial_relationship
spatial_relationship
metric
subject_consistency
subject_consistency
subject_consistency
metric
temporal_flickering
temporal_flickering
temporal_flickering
metric
temporal_style
temporal_style
temporal_style
metric
videoscore2
videoscore2
videoscore2
metric
models
pool
registry
types
worker
fastvideo_args
forward_context
hooks
hooks
hooks
activation_trace
hooks
layerwise_offload
image_processor
layers
layers
layers
activation
custom_op
fp4linear
fp8linear
layernorm
linear
mlp
quantization
quantization
quantization
absmax_fp8
base_config
fp8_config
fp8_qat_train_config
nvfp4_config
nvfp4_qat_config
nvfp4_qat_train_config
rotary_embedding
rotary_embedding_3d
utils
visual_embedding
vocab_parallel_embedding
logger
logging_utils
logging_utils
logging_utils
formatter
models
models
models
audio
audio
audio
ltx2_audio_processing
ltx2_audio_vae
camera
camera
camera
trajectory
dits
dits
dits
base
causal_wanvideo
cosmos
cosmos2_5
dreamx_world
dreamx_world_ar
flux
flux_2
gen3c
glm_image
hunyuangamecraft
hunyuanvideo
hunyuanvideo15
hyworld
hyworld
hyworld
camera_rope
data_utils
hyworld
pose
resolution_utils
retrieval_context
trajectory
kandinsky5
lingbot_video
lingbotworld
lingbotworld
lingbotworld
cam_utils
model
lingbotworld2
lingbotworld2
lingbotworld2
cam_utils
causal_fast
longcat
ltx2
magi_human
matrixgame2
matrixgame2
matrixgame2
action_module
causal_model
model
utils
matrixgame3
matrixgame3
matrixgame3
action_module
model
utils
sd3
stable_audio
wanvideo
zimage
encoders
encoders
encoders
base
bert
clip
gemma
glm_image_ar_loader
lingbot_video
lingbotworld2_t5
llama
mistral3
qwen2_5
qwen2_5_vl_custom
qwen3
reason1
siglip
stable_audio_conditioner
t5
t5_hf
t5gemma
vision
hf_transformer_utils
loader
loader
loader
benchmarks
benchmarks
benchmarks
benchmark_weight_loading
benchmark_weight_loading_comparison
component_loader
fsdp_load
utils
weight_utils
mask_utils
parameter
registry
schedulers
schedulers
schedulers
base
scheduling_flow_map_euler_discrete
scheduling_flow_match_euler_discrete
scheduling_flow_unipc_multistep
scheduling_rcm
scheduling_self_forcing_flow_match
scheduling_unipc_multistep
upsamplers
upsamplers
upsamplers
hunyuan15
ltx2_upsampler
utils
vaes
vaes
vaes
autoencoder_kl
common
cosmos25wanvae
flux2_components
flux2vae
gamecraftvae
gamecraftvae_blocks
gen3c_tokenizer_vae
hunyuan15vae
hunyuanvae
hyworldvae
lingbotworld2_wanvae
ltx2vae
oobleck
sa_audio
wanvae
vision_utils
performance
performance
performance
hf_store
metric_policy
performance_dashboard
performance_dashboard
performance_dashboard
api
metrics
service
pipelines
pipelines
pipelines
basic
basic
basic
cosmos
cosmos
cosmos
cosmos2_5_pipeline
cosmos_pipeline
presets
dreamx_world
dreamx_world
dreamx_world
ar_denoising
camera_conditioning
config
dreamx_world_ar_pipeline
dreamx_world_pipeline
presets
stages
flux
flux
flux
flux_pipeline
flux_2
flux_2
flux_2
flux_2_klein_pipeline
flux_2_latent_preparation
flux_2_pipeline
flux_2_text_encoding
flux_2_timestep_preparation
presets
gamecraft
gamecraft
gamecraft
gamecraft_pipeline
presets
gen3c
gen3c
gen3c
cache_3d
camera_utils
depth_estimation
gen3c_pipeline
presets
glm_image
glm_image
glm_image
glm_image_pipeline
stages
stages
stages
before_denoising
condition_encoding
decoding
denoising
hunyuan
hunyuan
hunyuan
hunyuan_pipeline
presets
hunyuan15
hunyuan15
hunyuan15
hunyuan15_2sr_pipeline
hunyuan15_i2v_pipeline
hunyuan15_pipeline
hunyuan15_sr_pipeline
presets
hyworld
hyworld
hyworld
hyworld_pipeline
presets
kandinsky5
kandinsky5
kandinsky5
kandinsky5_i2v_pipeline
kandinsky5_pipeline
presets
lingbot_video
lingbot_video
lingbot_video
lingbot_video_pipeline
presets
stages
lingbotworld
lingbotworld
lingbotworld
lingbotworld_pipeline
presets
lingbotworld2
lingbotworld2
lingbotworld2
causal_fast_pipeline
presets
longcat
longcat
longcat
longcat_i2v_pipeline
longcat_pipeline
longcat_vc_pipeline
presets
ltx2
ltx2
ltx2
continuation
ltx2_pipeline
pipeline_configs
presets
stage_overrides
stages
stages
stages
ltx2_audio_decoding
ltx2_denoising
ltx2_image_conditioning
ltx2_latent_preparation
ltx2_refine
ltx2_text_encoding
magi_human
magi_human
magi_human
magi_human_pipeline
pipeline_configs
presets
stages
stages
stages
audio_decoding
denoising
latent_preparation
reference_image
sr_denoising
sr_latent_preparation
matrixgame2
matrixgame2
matrixgame2
matrixgame2_causal_dmd_pipeline
matrixgame2_i2v_pipeline
presets
matrixgame3
matrixgame3
matrixgame3
matrixgame3_i2v_pipeline
presets
sd35
sd35
sd35
presets
sd35_pipeline
stable_audio
stable_audio
stable_audio
presets
stable_audio_pipeline
stages
stages
stages
conditioning
decoding
denoising
latent_preparation
turbodiffusion
turbodiffusion
turbodiffusion
presets
turbodiffusion_i2v_pipeline
turbodiffusion_pipeline
wan
wan
wan
lucy_edit_pipeline
presets
wan_causal_dmd_pipeline
wan_causal_pipeline
wan_dmd_pipeline
wan_i2v_dmd_pipeline
wan_i2v_pipeline
wan_pipeline
wan_v2v_pipeline
zimage
zimage
zimage
presets
stages
zimage_pipeline
composed_pipeline_base
lora_pipeline
pipeline_batch_info
pipeline_registry
preprocess
preprocess
preprocess
ltx2
ltx2
ltx2
ltx2_preprocess_pipelines
matrixgame2
matrixgame2
matrixgame2
matrixgame2_preprocess_pipeline
matrixgame2_preprocess_pipeline_ode_trajectory
preprocess_cosmos25_overfit
preprocess_cosmos_overfit
preprocess_hunyuan_overfit
preprocess_ltx2_overfit
preprocess_pipeline_base
preprocess_pipeline_i2v
preprocess_pipeline_ode_trajectory
preprocess_pipeline_t2v
preprocess_pipeline_text
preprocess_stages
v1_preprocess
v1_preprocessing_new
wan
wan
wan
wan_preprocess_pipelines
stages
stages
stages
base
causal_denoising
conditioning
decoding
denoising
encoding
flux_stages
gamecraft_denoising
gamecraft_image_encoding
gen3c_stages
hyworld_denoising
image_encoding
input_validation
kandinsky5
latent_preparation
longcat_denoising
longcat_i2v_denoising
longcat_i2v_latent_preparation
longcat_image_vae_encoding
longcat_kv_cache_init
longcat_refine_init
longcat_refine_timestep
longcat_vc_denoising
longcat_video_vae_encoding
matrixgame2_denoising
matrixgame3_denoising
sd35_conditioning
sr_denoising
text_encoding
timestep_preparation
utils
validators
training
platforms
platforms
platforms
cpu
cuda
interface
mps
npu
rocm
profiler
registry
train
train
train
callbacks
callbacks
callbacks
callback
ema
grad_clip
validation
entrypoint
entrypoint
entrypoint
dcp_to_diffusers
misc
misc
misc
wan_ode_init_conversion
train
methods
methods
methods
base
consistency_model
consistency_model
consistency_model
causal_cd
distribution_matching
distribution_matching
distribution_matching
anyflow
anyflow_pretrain
dmd2
self_forcing
streaming_long_tuning
fine_tuning
fine_tuning
fine_tuning
dfsft
finetune
tfsft
tfsft
Table of contents
Classes
knowledge_distillation
knowledge_distillation
knowledge_distillation
kd
rl
rl
rl
common
common
common
prompt_sampling
sampling
validation
diffusion_nft
rewards
rewards
rewards
frame_rewards
media
models
models
models
base
cosmos
cosmos
cosmos
cosmos
hunyuan
hunyuan
hunyuan
hunyuan
longcat
longcat
longcat
longcat
ltx2
ltx2
ltx2
ltx2
wan
wan
wan
wan
wan_causal
trainer
utils
utils
utils
builder
checkpoint
config
dataloader
instantiate
lora
module_state
moduleloader
negative_prompt
optimizer
tracking
training_config
training
training
training
activation_checkpoint
checkpointing_utils
cosmos2_5_training_pipeline
distillation_pipeline
ltx2_training_pipeline
matrixgame2_ar_diffusion_pipeline
matrixgame2_ode_causal_pipeline
matrixgame2_self_forcing_distillation_pipeline
matrixgame2_training_pipeline
ode_causal_pipeline
self_forcing_distillation_pipeline
trackers
training_pipeline
training_utils
wan_distillation_pipeline
wan_i2v_distillation_pipeline
wan_i2v_training_pipeline
wan_self_forcing_distillation_pipeline
wan_training_pipeline
utils
version
worker
worker
worker
executor
gpu_worker
multiproc_executor
ray_distributed_executor
ray_env
ray_utils
worker_base
workflow
workflow
workflow
preprocess
preprocess
preprocess
components
preprocess_workflow
preprocess_workflow_i2v
preprocess_workflow_ltx2_t2v
preprocess_workflow_t2v
workflow_base
Table of contents
Classes
Home
API Reference
fastvideo
train
methods
fine_tuning
tfsft
¶
Teacher-forcing SFT method (TFSFT; algorithm layer).
Classes
¶
Back to top