fix hardcode

Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

.buildkite/pipeline.yml

@@ -104,6 +104,18 @@ steps:
                   - TEST_TYPE=distillation_dmd
                 agents:
                   queue: "default"
+            - path:
+                - "fastvideo/training/*self_forcing_distillation_pipeline.py"
+                - "fastvideo/tests/training/self-forcing/**"
+                - "pyproject.toml"
+                - "docker/Dockerfile.python3.12"
+              config:
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                label: "Self-Forcing Tests"
+                env:
+                  - TEST_TYPE=self_forcing
+                agents:
+                  queue: "default"
             - path:
                 - "fastvideo/**"
                 - "pyproject.toml"

.buildkite/scripts/pr_test.sh

@@ -110,6 +110,10 @@ case "$TEST_TYPE" in
         MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_distill_dmd_tests"
         ;;
         # run_inference_tests_vmoba
+    "self_forcing")
+        log "Running self-forcing tests..."
+        MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_self_forcing_tests"
+        ;;
     "inference_vmoba")
         log "Running V-MoBA inference tests..."
         MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_inference_tests_vmoba"

.pre-commit-config.yaml

@@ -12,9 +12,6 @@ exclude: |
     scripts/.*|
     fastvideo/data_preprocess/.*|
     fastvideo/dataset/.*|
-    fastvideo/distill/.*|
-    fastvideo/distill\.py|
-    fastvideo/distill_adv\.py|
     fastvideo/models/.*|
     fastvideo/sample/.*|
     fastvideo/train\.py|
@@ -44,10 +41,10 @@ repos:
   - id: codespell
     additional_dependencies: ['tomli']
     args: ['--toml', 'pyproject.toml']
-- repo: https://github.com/PyCQA/isort
-  rev: 6.0.1
-  hooks:
-  - id: isort
+# - repo: https://github.com/PyCQA/isort
+#   rev: 6.0.1
+#   hooks:
+#   - id: isort
 - repo: https://github.com/jackdewinter/pymarkdown
   rev: v0.9.30
   hooks:

README.md

@@ -7,7 +7,7 @@
 FastVideo features an end-to-end unified pipeline for accelerating diffusion models, starting from data preprocessing to model training, finetuning, distillation, and inference. FastVideo is designed to be modular and extensible, allowing users to easily add new optimizations and techniques. Whether it is training-free optimizations or post-training optimizations, FastVideo has you covered.
 
 <p align="center">
-    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3csdw1isz-Euq8_Q8~baewG8hxjXs2gQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/S7HLCSTh" target="_blank"> <b> WeChat </b> </a> |
+    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3csdw1isz-Euq8_Q8~baewG8hxjXs2gQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/q46BbX6" target="_blank"> <b> WeChat </b> </a> |
 </p>
 
 <div align="center">
@@ -155,8 +155,8 @@ If you find FastVideo useful, please considering citing our work:
 }
 
 @article{zhang2025vsa,
-  title={VSA: Faster Video Diffusion with Trainable Sparse Attention},
-  author={Zhang, Peiyuan and Huang, Haofeng and Chen, Yongqi and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
+  title={Vsa: Faster video diffusion with trainable sparse attention},
+  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
   journal={arXiv preprint arXiv:2505.13389},
   year={2025}
 }

docs/source/inference/optimizations.md

@@ -1,32 +1,40 @@
 (inference-optimizations)=
+
 # Optimizations
 
 This page describes the various options for speeding up generation times in FastVideo.
 
 ## Table of Contents
+
 - Optimized Attention Backends
+
   - [Flash Attention](#optimizations-flash)
   - [Sliding Tile Attention](#optimizations-sta)
   - [Sage Attention](#optimizations-sage)
+  - [Sage Attention 3](#optimizations-sage3)
 
 - Caching Techniques
   - [TeaCache](#optimizations-teacache)
 
 (optimizations-backends)=
+
 ## Attention Backends
 
 ### Available Backends
+
 - Torch SDPA: `FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA`
 - Flash Attention 2 and 3: `FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN`
 - Sliding Tile Attention: `FASTVIDEO_ATTENTION_BACKEND=SLIDING_TILE_ATTN`
 - Video Sparse Attention: `FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN`
 - Sage Attention: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN`
+- Sage Attention 3: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN_THREE`
 
 ### Configuring Backends
 
 There are two ways to configure the attention backend in FastVideo.
 
 #### 1. In Python
+
 In python, set the `FASTVIDEO_ATTENTION_BACKEND` environment variable before instantiating `VideoGenerator` like this:
 
 ```python
@@ -34,6 +42,7 @@ os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLIDING_TILE_ATTN"
 ```
 
 #### 2. In CLI
+
 You can also set the environment variable on the command line:
 
 ```bash
@@ -41,6 +50,7 @@ FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN python example.py
 ```
 
 (optimizations-flash)=
+
 ### Flash Attention
 
 **`FLASH_ATTN`**
@@ -57,7 +67,7 @@ And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://git
 git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention
 
 cd hopper
-pip install ninja 
+pip install ninja
 python setup.py install
 ```
 
@@ -66,7 +76,9 @@ FastVideo will automatically detect and use `FA3` if it is installed when using
 :::
 
 (optimizations-sta)=
+
 ### Sliding Tile Attention
+
 **`SLIDING_TILE_ATTN`**
 
 ```bash
@@ -76,7 +88,9 @@ pip install st_attn==0.0.4
 Please see [this page](#sta-installation) for more installation instructions.
 
 (optimizations-vsa)=
+
 ### Video Sparse Attention
+
 **`VIDEO_SPARSE_ATTN`**
 
 ```bash
@@ -87,19 +101,45 @@ python setup_vsa.py install
 Please see [this page](#vsa-installation) for more installation instructions.
 
 (optimizations-sage)=
+
 ### Sage Attention
+
 **`SAGE_ATTN`**
 
 To use [SageAttention](https://github.com/thu-ml/SageAttention) 2.1.1, please compile from source:
 
 ```bash
 git clone https://github.com/thu-ml/SageAttention.git
-cd sageattention 
+cd sageattention
 python setup.py install  # or pip install -e .
 ```
 
+(optimizations-sage3)=
+
+### Sage Attention 3
+
+**`SAGE_ATTN_THREE`**
+
+[SageAttention 3](https://huggingface.co/jt-zhang/SageAttention3) is an advanced attention mechanism that leverages FP4 quantization and Blackwell GPU Tensor Cores for significant performance improvements.
+
+#### Hardware Requirements
+
+- RTX5090
+
+#### Installation
+
+Note that Sage Attention 3 requires `python>=3.13`, `torch>=2.8.0`, `CUDA >=12.8`. If you are using `uv` and using `torch==2.8.0` make sure that `sentencepiece==0.2.1` in the pyproject.toml file.
+
+To use Sage Attention 3 in FastVideo, first get access to the SageAttention3 code, then move `sageattn/` and `setup.py` to the directory `fastvideo/attention/backends`, then install from using:
+
+```bash
+python setup.py install
+```
+
 (optimizations-teacache)=
+
 ## Teacache
+
 TeaCache is an optimization technique supported in FastVideo that can significantly speed up video generation by skipping redundant calculations across diffusion steps. This guide explains how to enable and configure TeaCache for optimal performance in FastVideo.
 
 ### What is TeaCache?

examples/datasets/crush_smol/download_dataset.sh

@@ -1,3 +0,0 @@
-#!/bin/bash
-
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"

examples/datasets/crush_smol/validation.json

@@ -1,76 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-034.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-027.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
-      "image_path": null,
-      "video_path": "validation_dataset/yYcK4nANZz4-Scene-030.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-016.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-056.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-059.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.",
-      "image_path": null,
-      "video_path": "validation_dataset/EJqsC21GSBY-Scene-059.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": "validation_dataset/GBSfpTcKegk-Scene-003.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}

examples/datasets/crush_smol/validation_single.json

@@ -1,13 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
-      "image_path": null,
-      "video_path": "validation_dataset/1gGQy4nxyUo-Scene-016.mp4",
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}

examples/distill/SFWan2.1-A1.3B/distill_dmd.slurm

@@ -0,0 +1,182 @@
+#!/bin/bash
+#SBATCH --job-name=1.3B_moe_4n_sf_distill_no_2nd_update
+#SBATCH --partition=main
+#SBATCH --nodes=4
+#SBATCH --ntasks=4
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=1.3B_moe_4n_sf_distill_output/moe_sf_distill_no_2nd_update.out
+#SBATCH --error=1.3B_moe_4n_sf_distill_output/moe_sf_distill_no_2nd_update.err
+#SBATCH --exclusive
+set -e -x
+
+
+# Environment Setup
+# source ~/conda/miniconda/bin/activate
+# conda activate wei-fv
+# export HOME="/mnt/weka/home/hao.zhang/wei"
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=32
+
+# Model paths for Self-Forcing DMD distillation:
+# GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+GENERATOR_MODEL_PATH="rand0nmr/SFWan2.1-T2V-A1.3B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
+# REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"  # Teacher model
+# FAKE_SCORE_MODEL_PATH="rand0nmr/SFWan2.1-T2V-A1.3B-Diffusers"  # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
+# DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
+# VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_2.json"
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
+  --output_dir "/mnt/sharefs/users/hao.zhang/wei/1.3B_MoE_SFwan_t2v_finetune"
+  --wandb_run_name "1.3B_moe_sf_distill"
+  # --use_sf_wan
+  # --sf_ode_init_path "checkpoints/ode_init.pt"
+  --max_train_steps 5000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
+  --enable_gradient_checkpointing_type "full"
+  --log_visualization
+  --simulate_generator_forward
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  --enable_gradient_masking
+  --gradient_mask_last_n_frames 21
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+# Model arguments
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --generator_model_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 8
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-6
+  --fake_score_learning_rate 4e-7
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 10.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 5
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 200
+  --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  --init_weights_from_safetensors_2 "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  # --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/FastVideo2/vidprom_8b16k_1e-5_gn1/checkpoint-3500/transformer/diffusion_pytorch_model.safetensors"
+#   --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/sf2/checkpoint_v3_7k/model.safetensors"
+)
+
+# Self-forcing DMD arguments
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 4.0
+  --real_score_guidance_scale_2 3.0
+  --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
+)
+
+# Self-forcing specific arguments
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+  --validate_cache_structure False  # Set to True for debugging KV cache issues
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+--nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"

examples/distill/SFWan2.1-T2V/distill_dmd_t2v_1.3B.sh

@@ -16,30 +16,29 @@ export NCCL_P2P_DISABLE=1
 export TORCH_NCCL_ENABLE_MONITORING=0
 # different cache dir for different processes
 export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29501
+export MASTER_PORT=29503
 export TOKENIZERS_PARALLELISM=false
-export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+export WANDB_API_KEY=your_wandb_api_key
 export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=offline
+export WANDB_MODE=online
 export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
 
 # Configs
-NUM_GPUS=4
+NUM_GPUS=1
 
 # Model paths for Self-Forcing DMD distillation:
 GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
 REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
 FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
 
-DATA_DIR="data/mixkit-64_processed/Node_0_GPU_1_File_1/combined_parquet_dataset"
-VALIDATION_DATASET_FILE="data/mixkit-64_processed/validation.json"
+DATA_DIR=your_data_dir
+VALIDATION_DATASET_FILE=your_validation_data_dir
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
-# Training arguments
 training_args=(
-  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
-  --output_dir "/mnt/sharefs/users/hao.zhang/SFwan_t2v_finetune"
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  
+  --output_dir your_output_dir
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -47,16 +46,15 @@ training_args=(
   --num_latent_t 21
   --num_height 480
   --num_width 832
-  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
   --enable_gradient_checkpointing_type "full"
   --log_visualization
   --simulate_generator_forward
+  --num_frames 81
   --num_frame_per_block 3  # Frame generation block size for self-forcing
   --enable_gradient_masking
   --gradient_mask_last_n_frames 21
 )
 
-# Parallel arguments
 parallel_args=(
   --num_gpus $NUM_GPUS # 64
   --sp_size 1
@@ -65,31 +63,26 @@ parallel_args=(
   --hsdp_shard_dim $NUM_GPUS
 )
 
-# Model arguments
 model_args=(
   --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
   --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
-  --generator_model_path $GENERATOR_MODEL_PATH
   --real_score_model_path $REAL_SCORE_MODEL_PATH
   --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
-# Dataset arguments
 dataset_args=(
   --data_path "$DATA_DIR"
   --dataloader_num_workers 4
 )
 
-# Validation arguments
 validation_args=(
   --log_validation
   --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 100
+  --validation_steps 50
   --validation_sampling_steps "4"
   --validation_guidance_scale "6.0" # not used for dmd inference
 )
 
-# Optimizer arguments
 optimizer_args=(
   --learning_rate 1e-5
   --mixed_precision "bf16"
@@ -100,7 +93,6 @@ optimizer_args=(
   --max_grad_norm 1.0
 )
 
-# Miscellaneous arguments
 miscellaneous_args=(
   --inference_mode False
   --checkpoints_total_limit 3
@@ -111,10 +103,9 @@ miscellaneous_args=(
   --use_ema True
   --ema_decay 0.99
   --ema_start_step 100
-  --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  --init_weights_from_safetensors your_ode_init_weights_path
 )
 
-# Self-forcing DMD arguments
 dmd_args=(
   --dmd_denoising_steps '1000,750,500,250'
   --min_timestep_ratio 0.02
@@ -126,13 +117,11 @@ dmd_args=(
   --warp_denoising_step
 )
 
-# Self-forcing specific arguments
 self_forcing_args=(
   --independent_first_frame False  # Whether to treat first frame independently
   --same_step_across_blocks True  # Whether to use same denoising step across all blocks
   --last_step_only False  # Whether to only use the last denoising step
   --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
-  --validate_cache_structure False  # Set to True for debugging KV cache issues
 )
 
 torchrun \
@@ -148,4 +137,4 @@ torchrun \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
     "${dmd_args[@]}" \
-    "${self_forcing_args[@]}"
+    "${self_forcing_args[@]}"

examples/distill/SFWan2.1-T2V/distill_dmd_t2v_1.3B.slurm

@@ -0,0 +1,180 @@
+#!/bin/bash
+#SBATCH --job-name=sf_distill_profile
+#SBATCH --partition=main
+#SBATCH --nodes=2
+#SBATCH --ntasks=2
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=sf_distill_output/sf_distill_profile.out
+#SBATCH --error=sf_distill_output/sf_distill_profile.err
+#SBATCH --exclusive
+set -e -x
+
+
+# Environment Setup
+source ~/conda/miniconda/bin/activate
+conda activate wei-fv
+export HOME="/mnt/weka/home/hao.zhang/wei"
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29503
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=offline
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+export FASTVIDEO_TORCH_PROFILER_DIR=/mnt/weka/home/hao.zhang/wei/FastVideo/traces
+export FASTVIDEO_TORCH_PROFILE_REGIONS=profiler_region_model_loading
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=16
+
+# Model paths for Self-Forcing DMD distillation:
+GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
+# DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
+# VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_2.json"
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
+  --output_dir "/mnt/sharefs/users/hao.zhang/wei/SFwan_t2v_finetune"
+  --wandb_run_name "sf_distill_2e-6_4e-7_4n"
+  # --use_sf_wan
+  # --sf_ode_init_path "checkpoints/ode_init.pt"
+  --max_train_steps 2
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
+  --enable_gradient_checkpointing_type "full"
+  --log_visualization
+  --simulate_generator_forward
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  --enable_gradient_masking
+  --gradient_mask_last_n_frames 21
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+# Model arguments
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --generator_model_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 8
+)
+
+# Validation arguments
+validation_args=(
+#   --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 100
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-6
+  --fake_score_learning_rate 4e-7
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 10.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 5
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 200
+  --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  # --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/FastVideo2/vidprom_8b16k_1e-5_gn1/checkpoint-3500/transformer/diffusion_pytorch_model.safetensors"
+#   --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/sf2/checkpoint_v3_7k/model.safetensors"
+)
+
+# Self-forcing DMD arguments
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 3.0
+  --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
+)
+
+# Self-forcing specific arguments
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+  --validate_cache_structure False  # Set to True for debugging KV cache issues
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+--nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"

examples/distill/SFWan2.1-T2V/download_dataset.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
 
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"

examples/distill/SFWan2.1-T2V/preprocess_data.sh

@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
     --samples_per_file 8 \
     --flush_frequency 8 \
     --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 

examples/distill/SFWan2.2-A14B/distill_dmd.sh

@@ -0,0 +1,180 @@
+#!/bin/bash
+#SBATCH --job-name=14B_moe_4n_sf_distill_no_2nd_update
+#SBATCH --partition=main
+#SBATCH --nodes=4
+#SBATCH --ntasks=4
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=14B_moe_4n_sf_distill_output/moe_sf_distill_no_2nd_update.out
+#SBATCH --error=14B_moe_4n_sf_distill_output/moe_sf_distill_no_2nd_update.err
+#SBATCH --exclusive
+set -e -x
+
+
+# Environment Setup
+# source ~/conda/miniconda/bin/activate
+# conda activate wei-fv
+# export HOME="/mnt/weka/home/hao.zhang/wei"
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=32
+
+# Model paths for Self-Forcing DMD distillation:
+# GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+GENERATOR_MODEL_PATH="rand0nmr/SFWan2.2-T2V-A14B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
+# REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"  # Teacher model
+# FAKE_SCORE_MODEL_PATH="rand0nmr/SFWan2.1-T2V-A1.3B-Diffusers"  # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
+# DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
+# VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_2.json"
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
+  --output_dir "/mnt/sharefs/users/hao.zhang/wei/14B_MoE_SFwan_t2v_finetune"
+  --wandb_run_name "14B_moe_sf_distill"
+  # --use_sf_wan
+  # --sf_ode_init_path "checkpoints/ode_init.pt"
+  --max_train_steps 5000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
+  --enable_gradient_checkpointing_type "full"
+  --log_visualization
+  --simulate_generator_forward
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  --enable_gradient_masking
+  --gradient_mask_last_n_frames 21
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+# Model arguments
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --generator_model_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 8
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-6
+  --fake_score_learning_rate 4e-7
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 10.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 12
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 200
+  --init_weights_from_safetensors "/mnt/sharefs/users/hao.zhang/wl/release/self_forcing_ode_init_wan22_high_bz128_1e-5/diffusers_2000/"
+  --init_weights_from_safetensors_2 "/mnt/sharefs/users/hao.zhang/wl/release/self_forcing_ode_init_wan22_low_bz128_1e-5/diffusers_2000/"
+)
+
+# Self-forcing DMD arguments
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 4.0
+  --real_score_guidance_scale_2 3.0
+  --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
+)
+
+# Self-forcing specific arguments
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+  --validate_cache_structure False  # Set to True for debugging KV cache issues
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+--nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"

examples/distill/SFWan2.2-A14B/distill_dmd.slurm

@@ -0,0 +1,165 @@
+#!/bin/bash
+#SBATCH --job-name=moe_4n_sf_distill
+#SBATCH --partition=main
+#SBATCH --nodes=4
+#SBATCH --ntasks=4
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=moe_4n_sf_distill_output/moe_sf_distill_%j.out
+#SBATCH --error=moe_4n_sf_distill_output/moe_sf_distill_%j.err
+#SBATCH --exclusive
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export NCCL_DEBUG_SUBSYS=INIT,NET
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export TOKENIZERS_PARALLELISM=false
+# export WANDB_API_KEY="8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde"
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+# export WANDB_API_KEY='your_wandb_api_key_here'
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+source ~/conda/miniconda/bin/activate
+conda activate wei-fv
+export HOME="/mnt/weka/home/hao.zhang/wei"
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=32
+
+# Model paths for Self-Forcing DMD distillation with Wan2.2:
+# GENERATOR_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"
+GENERATOR_MODEL_PATH="rand0nmr/SFWan2.2-T2V-A14B-Diffusers"  # Updated to Wan2.2
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers" # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers" # Critic model
+# GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers" # Updated to Wan2.2
+# REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Teacher model
+# FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
+# DATA_DIR=data/crush-smol_processed_t2v/combined_parquet_dataset
+# DATA_DIR="/mnt/sharefs/users/hao.zhang/Vchitect-2M/Wan-Syn-upload/latents_i2v/train/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
+
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+training_args=(
+  --tracker_project_name SFwan2.2_t2v_distill_self_forcing_dmd  # Updated for Wan2.2
+  --output_dir "/mnt/sharefs/users/hao.zhang/wei/SFwan2.2_distill_self_forcing_dmd"
+  --max_train_steps 4000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 448  # Updated to match Wan2.2 config
+  --num_width 832   # Updated to match Wan2.2 config
+  --enable_gradient_checkpointing_type "full"
+  --simulate_generator_forward
+  --log_visualization
+  --num_frames 81
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  --enable_gradient_masking
+  --gradient_mask_last_n_frames 21
+  --init_weights_from_safetensors /mnt/sharefs/users/hao.zhang/wl/models/sf_ode_init_wan22_checkpoints/high/3k/
+  --init_weights_from_safetensors_2 /mnt/sharefs/users/hao.zhang/wl/models/sf_ode_init_wan22_checkpoints/low/3k/
+  # --init_weights_from_safetensors /mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors
+)
+
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 40
+  --validation_sampling_steps "8"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+optimizer_args=(
+  --learning_rate 2e-6
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 10.0
+)
+
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 12
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 200
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,850,700,550,350,275,200,125'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 4.0
+  --real_score_guidance_scale_2 3.0
+  --fake_score_learning_rate 4e-7
+  --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
+)
+
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"

examples/distill/SFWan2.2-A14B/distill_low_dmd.sh

@@ -0,0 +1,181 @@
+#!/bin/bash
+#SBATCH --job-name=14B_moe_4n_sf_distill_no_2nd_update
+#SBATCH --partition=main
+#SBATCH --nodes=4
+#SBATCH --ntasks=4
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=14B_moe_4n_sf_distill_output/moe_sf_distill_no_2nd_update.out
+#SBATCH --error=14B_moe_4n_sf_distill_output/moe_sf_distill_no_2nd_update.err
+#SBATCH --exclusive
+set -e -x
+
+
+# Environment Setup
+# source ~/conda/miniconda/bin/activate
+# conda activate wei-fv
+# export HOME="/mnt/weka/home/hao.zhang/wei"
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=32
+
+# Model paths for Self-Forcing DMD distillation:
+# GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+GENERATOR_MODEL_PATH="rand0nmr/SFWan2.2-T2V-A14B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
+# REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"  # Teacher model
+# FAKE_SCORE_MODEL_PATH="rand0nmr/SFWan2.1-T2V-A1.3B-Diffusers"  # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+DATA_DIR="/mnt/weka/home/hao.zhang/matthew/FastVideo/data/test-text-preprocessing"
+# DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
+# VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_2.json"
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
+  --output_dir "/mnt/sharefs/users/hao.zhang/wei/14B_MoE_SFwan_t2v_finetune"
+  --wandb_run_name "14B_moe_sf_distill_low_dmd"
+  # --use_sf_wan
+  # --sf_ode_init_path "checkpoints/ode_init.pt"
+  --max_train_steps 5000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
+  --enable_gradient_checkpointing_type "full"
+  --log_visualization
+  --simulate_generator_forward
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
+  --enable_gradient_masking
+  --gradient_mask_last_n_frames 21
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+# Model arguments
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH  # TODO: check if you can remove this in this script
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --generator_model_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 8
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-6
+  --fake_score_learning_rate 4e-7
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 10.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 12
+  --seed 1000
+  --use_ema True
+  --ema_decay 0.99
+  --ema_start_step 200
+  --init_weights_from_safetensors "/mnt/sharefs/users/hao.zhang/wl/release/self_forcing_ode_init_wan22_low_bz128_1e-5/diffusers_2000/"
+  --init_weights_from_safetensors_2 "/mnt/sharefs/users/hao.zhang/wl/release/self_forcing_ode_init_wan22_low_bz128_1e-5/diffusers_2000/"
+)
+
+# Self-forcing DMD arguments
+dmd_args=(
+  --dmd_denoising_steps '350,275,200,125'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --dfake_gen_update_ratio 5
+  --real_score_guidance_scale 4.0
+  --real_score_guidance_scale_2 3.0
+  --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
+)
+
+# Self-forcing specific arguments
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+  --validate_cache_structure False  # Set to True for debugging KV cache issues
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+--nproc_per_node $NUM_GPUS \
+    fastvideo/training/wan_self_forcing_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"
+

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_VSA_t2v_1.3B.slurm

@@ -39,15 +39,18 @@ echo "NODE_RANK: $NODE_RANK"
 # Configs
 NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir"checkpoints/wan_t2v_finetune"
+  --output_dir $OUTPUT_DIR
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -72,6 +75,8 @@ parallel_args=(
 model_args=(
   --model_path $MODEL_PATH
   --pretrained_model_name_or_path $MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
 # Dataset arguments
@@ -91,7 +96,7 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
   --weight_only_checkpointing_steps 500
@@ -134,4 +139,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_VSA_t2v_14B.slurm

@@ -39,15 +39,18 @@ echo "NODE_RANK: $NODE_RANK"
 # Configs
 NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir "checkpoints/wan_t2v_finetune"
+  --output_dir "$OUTPUT_DIR"
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -72,6 +75,8 @@ parallel_args=(
 model_args=(
   --model_path $MODEL_PATH
   --pretrained_model_name_or_path $MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
 # Dataset arguments
@@ -91,7 +96,7 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
   --weight_only_checkpointing_steps 500
@@ -134,4 +139,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B.slurm

@@ -39,15 +39,18 @@ echo "NODE_RANK: $NODE_RANK"
 # Configs
 NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd
-  --output_dir "checkpoints/wan_t2v_finetune"
+  --output_dir "$OUTPUT_DIR"
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -72,6 +75,8 @@ parallel_args=(
 model_args=(
   --model_path $MODEL_PATH
   --pretrained_model_name_or_path $MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
 # Dataset arguments
@@ -91,7 +96,7 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
   --weight_only_checkpointing_steps 500
@@ -133,4 +138,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
 
-python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"

examples/distill/Wan2.2-A14B/distill_dmd.slurm

@@ -0,0 +1,144 @@
+#!/bin/bash
+#SBATCH --job-name=moe_dmd_distill
+#SBATCH --partition=main
+#SBATCH --nodes=2
+#SBATCH --ntasks=2
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=moe_dmd_distill_output/moe_dmd_distill_%j.out
+#SBATCH --error=moe_dmd_distill_output/moe_dmd_distill_%j.err
+#SBATCH --exclusive
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export NCCL_DEBUG_SUBSYS=INIT,NET
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+# export WANDB_API_KEY='your_wandb_api_key_here'
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+source ~/conda/miniconda/bin/activate
+conda activate wei-fv
+export HOME="/mnt/weka/home/hao.zhang/wei"
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=16
+
+# Model paths for Self-Forcing DMD distillation with Wan2.2:
+GENERATOR_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"  # Updated to Wan2.2
+# REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers" # Teacher model
+# FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers" # Critic model
+# GENERATOR_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Updated to Wan2.2
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Teacher model
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers" # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+DATA_DIR="/mnt/weka/home/hao.zhang/wei/FastVideo/data/crush-smol_processed_t2v/combined_parquet_dataset"
+# DATA_DIR=data/crush-smol_processed_t2v/combined_parquet_dataset
+# DATA_DIR="/mnt/sharefs/users/hao.zhang/Vchitect-2M/Wan-Syn-upload/latents_i2v/train/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wl/FastVideo/examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/validation_64.json"
+
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+training_args=(
+  --tracker_project_name SFwan2.2_t2v_distill_dmd  # Updated for Wan2.2
+  --output_dir "/mnt/sharefs/users/hao.zhang/wei/SFwan2.2_distill_dmd"
+  --max_train_steps 4000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 448  # Updated to match Wan2.2 config
+  --num_width 832   # Updated to match Wan2.2 config
+  --enable_gradient_checkpointing_type "full"
+  # --log_visualization
+  --simulate_generator_forward
+  --num_frames 81
+  # --init_weights_from_safetensors /mnt/sharefs/users/hao.zhang/wl/models/sf_ode_init_wan22_checkpoints/high/3k/
+  # --init_weights_from_safetensors_2 /mnt/sharefs/users/hao.zhang/wl/models/sf_ode_init_wan22_checkpoints/low/3k/
+)
+
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
+)
+
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 20
+  --validation_sampling_steps "4"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+optimizer_args=(
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 5
+  --seed 1000
+  --ema_start_step 100
+)
+
+dmd_args=(
+  --dmd_denoising_steps '1000,750,500,250'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --generator_update_interval 5
+  --real_score_guidance_scale 3.0
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/distill_dmd_t2v_5B.sh

@@ -40,15 +40,18 @@ echo "NODE_RANK: $NODE_RANK"
 # Configs
 NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DIR=your_validation_path  #(example:validation_64.json)
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name Wan_distillation
-  --output_dir "your_output_dir"
+  --output_dir "$OUTPUT_DIR"
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -73,6 +76,8 @@ parallel_args=(
 model_args=(
   --model_path $MODEL_PATH
   --pretrained_model_name_or_path $MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
 # Dataset arguments
@@ -92,11 +97,11 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 4e-6
   --lr_scheduler "cosine_with_min_lr"
   --min_lr_ratio 0.5
   --lr_warmup_steps 100
-  --fake_score_learning_rate 1e-5
+  --fake_score_learning_rate 2e-6
   --fake_score_lr_scheduler "cosine_with_min_lr"
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
@@ -141,4 +146,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/distill_dmd_t2v_5B_VSA.sh

@@ -40,6 +40,8 @@ echo "NODE_RANK: $NODE_RANK"
 # Configs
 NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DIR=your_validation_path  #(example:validation_64.json)
 # export CUDA_VISIBLE_DEVICES=4,5
@@ -73,6 +75,8 @@ parallel_args=(
 model_args=(
   --model_path $MODEL_PATH
   --pretrained_model_name_or_path $MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
 # Dataset arguments
@@ -92,11 +96,11 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 4e-6
   --lr_scheduler "cosine_with_min_lr"
   --min_lr_ratio 0.5
   --lr_warmup_steps 100
-  --fake_score_learning_rate 1e-5
+  --fake_score_learning_rate 2e-6
   --fake_score_lr_scheduler "cosine_with_min_lr"
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
@@ -142,4 +146,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/distill_dmd_VSA_t2v_5B.sh

@@ -14,26 +14,29 @@ export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
 # Configs
 NUM_GPUS=1
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers" 
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
 VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir="checkpoints/wan_t2v_finetune"
-  --max_train_steps=4000
-  --train_batch_size=1
+  --output_dir "$OUTPUT_DIR"
+  --max_train_steps 4000
+  --train_batch_size 1
   --train_sp_batch_size 1
-  --gradient_accumulation_steps=1
+  --gradient_accumulation_steps 1
   --num_latent_t 31
   --num_height 704
   --num_width 1280
   --num_frames 121
   --enable_gradient_checkpointing_type "full"
-  --training_state_checkpointing_steps=500
-  --weight_only_checkpointing_steps=500
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
 )
 
 # Parallel arguments
@@ -49,6 +52,8 @@ parallel_args=(
 model_args=(
   --model_path $MODEL_PATH
   --pretrained_model_name_or_path $MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
 # Dataset arguments
@@ -68,8 +73,8 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate=1e-5
-  --mixed_precision="bf16"
+  --learning_rate 2e-6
+  --mixed_precision "bf16"
   --weight_decay 0.01
   --max_grad_norm 1.0
 )
@@ -107,4 +112,4 @@ torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/distill_dmd_VSA_t2v_5B_lora.sh

@@ -14,6 +14,8 @@ export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
 # Configs
 NUM_GPUS=1
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
 VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
 # export CUDA_VISIBLE_DEVICES=4,5
@@ -51,6 +53,8 @@ parallel_args=(
 model_args=(
   --model_path $MODEL_PATH
   --pretrained_model_name_or_path $MODEL_PATH
+  --real_score_model_path $REAL_SCORE_MODEL_PATH
+  --fake_score_model_path $FAKE_SCORE_MODEL_PATH
 )
 
 # Dataset arguments
@@ -109,4 +113,4 @@ torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/download_dataset.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
 
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/preprocess_wan_data_ti2v_5b.sh

@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
     --samples_per_file 8 \
     --flush_frequency 8 \
     --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 

examples/inference/basic/basic.py

@@ -1,6 +1,6 @@
 from fastvideo import VideoGenerator
 
-from fastvideo.configs.sample import SamplingParam
+# from fastvideo.configs.sample import SamplingParam
 
 OUTPUT_PATH = "video_samples"
 def main():
@@ -17,13 +17,12 @@ def main():
         vae_cpu_offload=False,
         text_encoder_cpu_offload=True,
         pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument"
-        ti2v_task=True,
         # image_encoder_cpu_offload=False,
     )
 
-    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
+    # sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
     # sampling_param.num_frames = 45
-    sampling_param.image_path = "test.jpg"
+    # sampling_param.image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
     # Generate videos with the same simple API, regardless of GPU count
     prompt = (
         "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
@@ -45,4 +44,4 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    main()

examples/inference/basic/basic_self_forcing_causal.py

@@ -19,15 +19,13 @@ def main():
     )
 
     sampling_param = SamplingParam.from_pretrained(model_name)
-    sampling_param.num_frames = 81
 
-    prompts = [
-        "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
-        "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
-    ]
-
-    for prompt in prompts:
-        video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
+    prompt = (
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+    )
+    video = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True, sampling_param=sampling_param)
 
 if __name__ == "__main__":
     main()

examples/training/consistency_finetune/causal_ode_init/finetune_ode_init.sh

@@ -62,7 +62,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 6e-6
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/consistency_finetune/causal_ode_init/finetune_ode_init_16k_vidprom.slurm

@@ -1,5 +1,5 @@
 #!/bin/bash
-#SBATCH --job-name=1e5B2_16kFV_warp_ode_vidprom
+#SBATCH --job-name=2e6B8_16kFV_ode_vidprom
 #SBATCH --partition=main
 #SBATCH --nodes=1
 #SBATCH --ntasks=1
@@ -7,14 +7,14 @@
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
 #SBATCH --mem=1440G
-#SBATCH --output=ode_vidprom16k_warp/Dode_vidprom8b16k_1e-5.out
-#SBATCH --error=ode_vidprom16k_warp/Dode_vidprom8b16k_1e-5.err
+#SBATCH --output=ode_vidprom16k/ode_vidprom8b16k_2e-6.out
+#SBATCH --error=ode_vidprom16k/ode_vidprom8b16k_2e-6.err
 #SBATCH --exclusive
 set -e -x
 
 # Environment Setup
 source ~/conda/miniconda/bin/activate
-conda activate will-fv2
+conda activate your-conda-env
 
 export WANDB_MODE="online"
 export NCCL_P2P_DISABLE=1
@@ -28,18 +28,18 @@ export MASTER_ADDR=${nodes[0]}
 export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
 export TOKENIZERS_PARALLELISM=false
 export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+export WANDB_API_KEY=your-wandb-api-key
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
 
 echo "MASTER_ADDR: $MASTER_ADDR"
 echo "NODE_RANK: $NODE_RANK"
 
 
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-16k-t2v-1-3b-81/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
+MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+DATA_DIR="your-data-dir"
+VALIDATION_DATASET_FILE="examples/training/consistency_finetune/causal_ode_init/validation.json"
+OUTPUT_DIR="your-output-dir"
+INIT_WEIGHTS_FROM_SAFETENSORS="your-init-weights-from-safetensors" # bidirectional weights from Wan2.1-T2V-1.3B-Diffusers
 NUM_GPUS=8
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
@@ -47,9 +47,9 @@ NUM_GPUS=8
 # Training arguments
 training_args=(
   --tracker_project_name "wan_ode_init"
-  --output_dir "Dwarp_vidprom_8b16k_test_warp_1e-5"
+  --output_dir $OUTPUT_DIR
   --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "Dwarp_vidprom_8b16k_wan_ode_init_1e-5"
+  --wandb_run_name "vidprom_8b16k_ode_init_2e-6"
   # --resume_from_checkpoint "ode_init_diffusers/"
   --warp_denoising_step
   --log_visualization
@@ -60,7 +60,7 @@ training_args=(
   --num_latent_t 21
   --num_height 480
   --num_width 832
-  --num_frames 77
+  --num_frames 81
   --dmd_denoising_steps "1000,750,500,250"
   --enable_gradient_checkpointing_type "full"
 )
@@ -93,14 +93,15 @@ validation_args=(
   --validation_steps 50
   --validation_sampling_steps "50"
   --validation_guidance_scale "6.0"
-  # --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  --init_weights_from_safetensors $INIT_WEIGHTS_FROM_SAFETENSORS
 )
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
   --mixed_precision "bf16"
-  --checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --training_state_checkpointing_steps 500
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/consistency_finetune/causal_ode_init/validation.json

@@ -1,14 +1,5 @@
 {
   "data": [
-    {
-      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
     {
       "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
       "image_path": null,
@@ -28,7 +19,52 @@
       "num_frames": 77
     },
     {
-      "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
+      "caption": "Elon Musk, dressed in a sleek white spacesuit with a reflective visor, walks confidently across the lunar surface. His posture is upright, and he moves steadily with purpose. The moon's rocky terrain and scattered boulders surround him, casting shadows under the dim sunlight. The background shows vast stretches of the moon's barren landscape with craters and dust clouds kicked up by his boots. The scene captures a wide shot, emphasizing the vastness and desolation of the lunar environment. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "In a dynamic action-packed sequence set in the Marvel multiverse, Spider-Man and Venom engage in an intense battle. Spider-Man, in his classic red and blue suit, swings and dodges venomous attacks from the black symbiote-covered Venom. Both characters display a range of acrobatic moves and powerful strikes. The environment is a chaotic urban landscape with crumbling buildings and neon lights, reflecting the multiversal theme. The camera captures the epic fight from various angles, including wide shots to show the scale of destruction and close-ups to highlight their fierce expressions and physical combat. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A warm, family-oriented scene depicting a father getting ready to leave the house to buy milk. The father, a middle-aged man with a kind face and a casual outfit, picks up a jacket from the coat rack. His posture is upright as he bends down slightly to put on his shoes. In the background, there are glimpses of a cozy living room with a family photograph on the wall. The camera focuses closely on the father, capturing his gentle smile and reassuring nod towards the camera before he opens the front door and steps outside. Static medium close-up shot. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Close-up shot of a man with a prosthetic hand that functions as a rocket launcher. He looks at his new hand with a mix of amazement and concern, his facial expression showing a blend of curiosity and apprehension. The prosthetic hand is sleek and metallic, with intricate details that resemble a high-tech weapon. The background is a dimly lit laboratory with various scientific equipment and monitors displaying data. The man stands in a relaxed posture, his other hand resting on his hip, as he inspects his new limb. The scene is rendered in a realistic sci-fi style, emphasizing the futuristic technology and the man's emotional response to his new appendage. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Realistic CCTV footage style, Kim Taehyung from the band BTS is involved in a drug deal, caught on camera. Kim Taehyung appears nervous and cautious, wearing casual clothing typical of a public space. He exchanges items discreetly with another person, who is partially obscured. Both individuals maintain a watchful demeanor, occasionally glancing around to ensure no one is watching them. The lighting is dim, with flickering fluorescent lights casting shadows on their faces. The background shows a typical urban setting with blurred figures moving in the distance. Static camera angle, medium close-up shot focusing on the interaction between Taehyung and the other individual. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Photorealistic studio setup with professional lighting, showcasing detailed cubic dissections of experimental plastic and felt-like materials on a pristine white background. Each cube reveals intricate layers and textures of the materials, emphasizing their unique properties. The scene has a shallow depth of field initially, then slowly pulls out to reveal the full arrangement of cubes, maintaining a wide depth of field throughout the transition. ",
       "image_path": null,
       "video_path": null,
       "num_inference_steps": 40,

examples/training/consistency_finetune/ode_init/finetune_ode_init.sh

@@ -1,99 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-# DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-extended-t2v-1-3b/"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_70k"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "fixed_wan_ode_init_70k_6e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 6000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --warp_denoising_step
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 50
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 6e-6
-  --mixed_precision "bf16"
-  --checkpointing_steps 1000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/finetune_ode_init_crush.slurm

@@ -1,132 +0,0 @@
-#!/bin/bash
-#SBATCH --job-name=ode_crush
-#SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
-#SBATCH --ntasks-per-node=1
-#SBATCH --gres=gpu:8
-#SBATCH --cpus-per-task=128
-#SBATCH --mem=1440G
-#SBATCH --output=ode_crush_output/ode_crush.out
-#SBATCH --error=ode_crush_output/ode_crush.err
-#SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate will-fv2
-
-export WANDB_MODE="online"
-export NCCL_P2P_DISABLE=1
-export TORCH_NCCL_ENABLE_MONITORING=0
-# different cache dir for different processes
-export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
-export TOKENIZERS_PARALLELISM=false
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
-
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
-NUM_GPUS=2
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_warp_2"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "2warp_fixed_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  # --warp_denoising_step
-  --max_train_steps 6001
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim $NUM_GPUS
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-6
-  --mixed_precision "bf16"
-  --checkpointing_steps 2000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/finetune_ode_init_lmdb.sh

@@ -1,103 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=offline
-export TOKENIZERS_PARALLELISM=false
-export MASTER_PORT=29501
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/mixkit-64_processed/Node_0_GPU_1_File_1/combined_parquet_dataset"
-# DATA_DIR="/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_single_lmdb_sf/"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "Dwarp_vidprom_8b16k_test_warp_1e-5"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "Dwarp_vidprom_8b16k_wan_ode_init_1e-5"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --warp_denoising_step
-  --log_visualization
-  --max_train_steps 10
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --dmd_denoising_steps "1000,750,500,250"
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim $NUM_GPUS
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-#   --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 50
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-  # --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 500
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --seed 1024
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-torchrun \
-  --nnodes 1 \
-  --master_port $MASTER_PORT \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/finetune_ode_init_overfit.sh

@@ -1,98 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/weka/home/hao.zhang/wl/FastVideo2/data/crush-smol_processed_t2v_1_3b_ode_init_single"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_crush_smol"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "overfitwan_ode_init_crush_smol"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 2001
-  # --warp_denoising_step
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 20
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 500
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/finetune_ode_init_single_lmdb.sh

@@ -1,100 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-# DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init_5/combined_parquet_dataset/"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing-16k-t2v-1-3b/"
-VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
-NUM_GPUS=1
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
-# export CUDA_VISIBLE_DEVICES=4,5
-# IP=[MASTER NODE IP]
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "debug_ode_init"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "debug_ode_init"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 1000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 1
-  --num_latent_t 21
-  --num_height 480
-  --num_width 832
-  --num_frames 81
-  --warp_denoising_step
-  # --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS
-  --sp_size 1
-  --tp_size 1
-  --hsdp_replicate_dim 1
-  --hsdp_shard_dim 1
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 10
-  --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 1e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 1000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-)
-
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/ode_causal_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/ode_init/preprocess_wan_data_ode.sh

@@ -1,25 +0,0 @@
-#!/bin/bash
-
-GPU_NUM=1 # 2,4,8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-MODEL_TYPE="wan"
-DATA_MERGE_PATH="data/vidprom_1.txt"
-OUTPUT_DIR="data/ode_vidprom_1_fv/"
-
-torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
-    --model_path $MODEL_PATH \
-    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 1 \
-    --seed 42 \
-    --max_height 480 \
-    --max_width 832 \
-    --num_frames 81 \
-    --flow_shift 5.0 \
-    --dataloader_num_workers 0 \
-    --output_dir=$OUTPUT_DIR \
-    --train_fps 16 \
-    --samples_per_file 1 \
-    --flush_frequency 1 \
-    --video_length_tolerance_range 5 \
-    --preprocess_task "ode_trajectory" 

examples/training/consistency_finetune/ode_init/validation.json

@@ -1,76 +0,0 @@
-{
-  "data": [
-    {
-      "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Elon Musk, dressed in a sleek white spacesuit with a reflective visor, walks confidently across the lunar surface. His posture is upright, and he moves steadily with purpose. The moon's rocky terrain and scattered boulders surround him, casting shadows under the dim sunlight. The background shows vast stretches of the moon's barren landscape with craters and dust clouds kicked up by his boots. The scene captures a wide shot, emphasizing the vastness and desolation of the lunar environment. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "In a dynamic action-packed sequence set in the Marvel multiverse, Spider-Man and Venom engage in an intense battle. Spider-Man, in his classic red and blue suit, swings and dodges venomous attacks from the black symbiote-covered Venom. Both characters display a range of acrobatic moves and powerful strikes. The environment is a chaotic urban landscape with crumbling buildings and neon lights, reflecting the multiversal theme. The camera captures the epic fight from various angles, including wide shots to show the scale of destruction and close-ups to highlight their fierce expressions and physical combat. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "A warm, family-oriented scene depicting a father getting ready to leave the house to buy milk. The father, a middle-aged man with a kind face and a casual outfit, picks up a jacket from the coat rack. His posture is upright as he bends down slightly to put on his shoes. In the background, there are glimpses of a cozy living room with a family photograph on the wall. The camera focuses closely on the father, capturing his gentle smile and reassuring nod towards the camera before he opens the front door and steps outside. Static medium close-up shot. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Close-up shot of a man with a prosthetic hand that functions as a rocket launcher. He looks at his new hand with a mix of amazement and concern, his facial expression showing a blend of curiosity and apprehension. The prosthetic hand is sleek and metallic, with intricate details that resemble a high-tech weapon. The background is a dimly lit laboratory with various scientific equipment and monitors displaying data. The man stands in a relaxed posture, his other hand resting on his hip, as he inspects his new limb. The scene is rendered in a realistic sci-fi style, emphasizing the futuristic technology and the man's emotional response to his new appendage. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Realistic CCTV footage style, Kim Taehyung from the band BTS is involved in a drug deal, caught on camera. Kim Taehyung appears nervous and cautious, wearing casual clothing typical of a public space. He exchanges items discreetly with another person, who is partially obscured. Both individuals maintain a watchful demeanor, occasionally glancing around to ensure no one is watching them. The lighting is dim, with flickering fluorescent lights casting shadows on their faces. The background shows a typical urban setting with blurred figures moving in the distance. Static camera angle, medium close-up shot focusing on the interaction between Taehyung and the other individual. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    },
-    {
-      "caption": "Photorealistic studio setup with professional lighting, showcasing detailed cubic dissections of experimental plastic and felt-like materials on a pristine white background. Each cube reveals intricate layers and textures of the materials, emphasizing their unique properties. The scene has a shallow depth of field initially, then slowly pulls out to reveal the full arrangement of cubes, maintaining a wide depth of field throughout the transition. ",
-      "image_path": null,
-      "video_path": null,
-      "num_inference_steps": 40,
-      "height": 480,
-      "width": 832,
-      "num_frames": 77
-    }
-  ]
-}

examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/finetune_i2v.sh

@@ -61,7 +61,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 2e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 2000
+  --weight_only_checkpointing_steps 2000
+  --training_state_checkpointing_steps 2000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/finetune_i2v.slurm

@@ -95,7 +95,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 1e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/I2V-14B-VSA.slurm

@@ -93,7 +93,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 1e-6
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 0.01
   --max_grad_norm 1.0
 )

examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/T2V-1.3B-VSA.slurm

@@ -1,21 +1,22 @@
 #!/bin/bash
-#SBATCH --job-name=ode_vidprom2k
+#SBATCH --job-name=t2v
 #SBATCH --partition=main
-#SBATCH --nodes=1
-#SBATCH --ntasks=1
+#SBATCH --nodes=8
+#SBATCH --ntasks=8
 #SBATCH --ntasks-per-node=1
 #SBATCH --gres=gpu:8
 #SBATCH --cpus-per-task=128
 #SBATCH --mem=1440G
-#SBATCH --output=ode_vidprom2k_output/ode_vidprom2k.out
-#SBATCH --error=ode_vidprom2k_output/ode_vidprom2k.err
+#SBATCH --output=VSA_t2v_output/t2v_%j.out
+#SBATCH --error=VSA_t2v_output/t2v_%j.err
 #SBATCH --exclusive
 set -e -x
 
 # Environment Setup
 source ~/conda/miniconda/bin/activate
-conda activate will-fv2
+conda activate your_env
 
+# Basic Info
 export WANDB_MODE="online"
 export NCCL_P2P_DISABLE=1
 export TORCH_NCCL_ENABLE_MONITORING=0
@@ -29,45 +30,41 @@ export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
 export TOKENIZERS_PARALLELISM=false
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
-export WANDB_API_KEY='8d9f4b39abd68eb4e29f6fc010b7ee71a2207cde'
+export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
 
 echo "MASTER_ADDR: $MASTER_ADDR"
 echo "NODE_RANK: $NODE_RANK"
 
-
-# MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="/mnt/sharefs/users/hao.zhang/klin/preproc/data/test-ode-preprocessing/"
-VALIDATION_DATASET_FILE="examples/training/consistency_finetune/ode_init/validation.json"
+# Configs
 NUM_GPUS=8
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+DATA_DIR=your_data_dir
+VALIDATION_DATASET_FILE=your_validation_dataset_file
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
-  --tracker_project_name "wan_ode_init"
-  --output_dir "wan_ode_init_vidprom2k"
-  --override_transformer_cls_name "CausalWanTransformer3DModel"
-  --wandb_run_name "vidprom2k_wan_ode_init_5e-6"
-  # --resume_from_checkpoint "ode_init_diffusers/"
-  --max_train_steps 6001
+  --tracker_project_name wan_t2v_VSA
+  --output_dir "checkpoints/wan_t2v_finetune_VSA"
+  --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
   --gradient_accumulation_steps 1
   --num_latent_t 21
   --num_height 480
   --num_width 832
-  --num_frames 77
-  # --enable_gradient_checkpointing_type "full"
+  --num_frames 81
+  # --enable_gradient_checkpointing_type "full" # if OOM enable this
 )
 
 # Parallel arguments
 parallel_args=(
-  --num_gpus $NUM_GPUS
+  --num_gpus 64
   --sp_size 1
   --tp_size 1
-  --hsdp_replicate_dim 8
+  --hsdp_replicate_dim 64
   --hsdp_shard_dim 1
 )
 
@@ -79,25 +76,26 @@ model_args=(
 
 # Dataset arguments
 dataset_args=(
-  --data_path "$DATA_DIR" 
-  --dataloader_num_workers 1
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
 )
 
 # Validation arguments
 validation_args=(
   --log_validation
-  --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 100
+  --validation_dataset_file $VALIDATION_DATASET_FILE
+  --validation_steps 200
   --validation_sampling_steps "50"
-  --validation_guidance_scale "6.0"
+  --validation_guidance_scale "5.0"
 )
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 5e-6
+  --learning_rate 1e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 2000
-  --weight_decay 1e-4
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
+  --weight_decay 0.01
   --max_grad_norm 1.0
 )
 
@@ -106,26 +104,31 @@ miscellaneous_args=(
   --inference_mode False
   --checkpoints_total_limit 3
   --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
   --dit_precision "fp32"
-  --num_euler_timesteps 50
   --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
+  --flow_shift 1
+  --seed 1000
+)
+
+# VSA arguments
+vsa_args=(
+  --VSA_decay_rate 0.03 \
+  --VSA_decay_interval_steps 50 \
+  --VSA_sparsity 0.9 \
 )
 
-# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
 srun torchrun \
-  --nnodes $SLURM_JOB_NUM_NODES \
-  --nproc_per_node $NUM_GPUS \
-  --node_rank $SLURM_PROCID \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
-    fastvideo/training/ode_causal_pipeline.py \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_training_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
     "${dataset_args[@]}" \
     "${training_args[@]}" \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"
+    "${miscellaneous_args[@]}" \
+    "${vsa_args[@]}"  

examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/T2V-14B-VSA.slurm

@@ -93,7 +93,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 1e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 0.01
   --max_grad_norm 1.0
 )

examples/training/finetune/Wan2.2-A14B/finetune.slurm

@@ -0,0 +1,131 @@
+#!/bin/bash
+#SBATCH --job-name=moe_finetune
+#SBATCH --partition=main
+#SBATCH --nodes=2
+#SBATCH --ntasks=2
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=moe_output/moe_%j.out
+#SBATCH --error=moe_output/moe_%j.err
+#SBATCH --exclusive
+
+# Basic Info
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export NCCL_DEBUG_SUBSYS=INIT,NET
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY="50632ebd88ffd970521cec9ab4a1a2d7e85bfc45"
+# export WANDB_API_KEY='your_wandb_api_key_here'
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+
+source ~/conda/miniconda/bin/activate
+conda activate wei-fv
+export HOME="/mnt/weka/home/hao.zhang/wei"
+
+
+# Configs
+NUM_GPUS=8
+NUM_TOTAL_GPUS=16
+
+# Model paths for Self-Forcing DMD distillation with Wan2.2:
+GENERATOR_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers"  # Updated to Wan2.2
+# REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers" # Teacher model
+# FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.2-T2V-A14B-Diffusers" # Critic model
+
+# DATA_DIR="data/test-text-preprocessing/Node_0_GPU_1_File_1/combined_parquet_dataset/"
+DATA_DIR="/mnt/weka/home/hao.zhang/wei/FastVideo/data/crush-smol_processed_t2v/combined_parquet_dataset"
+# DATA_DIR=data/crush-smol_processed_t2v/combined_parquet_dataset
+# DATA_DIR="/mnt/sharefs/users/hao.zhang/Vchitect-2M/Wan-Syn-upload/latents_i2v/train/"
+# VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+VALIDATION_DATASET_FILE="/mnt/weka/home/hao.zhang/wei/FastVideo/data/crush-smol-single_processed_t2v/validation.json"
+
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+training_args=(
+  --tracker_project_name VSA_finetune  # Updated for Wan2.2
+  --output_dir "/mnt/sharefs/users/hao.zhang/wei/Wan2.2-MoE-finetune"
+  --max_train_steps 200
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 448  # Updated to match Wan2.2 config
+  --num_width 832   # Updated to match Wan2.2 config
+  --enable_gradient_checkpointing_type "full"
+  # --log_visualization
+  --num_frames 81
+  # --init_weights_from_safetensors /mnt/sharefs/users/hao.zhang/wl/models/sf_ode_init_wan22_checkpoints/high/3k/
+  # --init_weights_from_safetensors_2 /mnt/sharefs/users/hao.zhang/wl/models/sf_ode_init_wan22_checkpoints/low/3k/
+)
+
+parallel_args=(
+  --num_gpus $NUM_TOTAL_GPUS # 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1 # 64
+  --hsdp_shard_dim $NUM_TOTAL_GPUS
+)
+
+model_args=(
+  --model_path $GENERATOR_MODEL_PATH
+  --pretrained_model_name_or_path $GENERATOR_MODEL_PATH
+)
+
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+validation_args=(
+#   --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "40"
+  --validation_guidance_scale "5.0" # not used for dmd inference
+)
+
+optimizer_args=(
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --dit_cpu_offload True
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
+  --weight_decay 0.01
+  --betas '0.0,0.999'
+  --max_grad_norm 1.0
+)
+
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --flow_shift 5
+  --seed 1000
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_training_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"

examples/training/finetune/wan_i2v_14B_480p/crush_smol/finetune_i2v.sh

@@ -61,7 +61,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 1e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/wan_i2v_14B_480p/crush_smol/finetune_i2v.slurm

@@ -95,7 +95,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 1e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/wan_i2v_14B_480p/crush_smol/finetune_i2v_lora.sh

@@ -61,7 +61,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 1e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.sh

@@ -6,8 +6,8 @@ export TOKENIZERS_PARALLELISM=false
 # export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
 
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_old"
-VALIDATION_DATASET_FILE="examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
+DATA_DIR="data/crush-smol_processed_t2v/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
 NUM_GPUS=4
 # export CUDA_VISIBLE_DEVICES=4,5
 
@@ -52,7 +52,7 @@ dataset_args=(
 validation_args=(
   --log_validation 
   --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 50
+  --validation_steps 200
   --validation_sampling_steps "50" 
   --validation_guidance_scale "6.0"
 )
@@ -61,7 +61,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 5e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 1000
+  --weight_only_checkpointing_steps 1000
+  --training_state_checkpointing_steps 1000
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v.slurm

@@ -92,7 +92,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 5e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 400
+  --weight_only_checkpointing_steps 400
+  --training_state_checkpointing_steps 400
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/wan_t2v_1.3B/crush_smol/finetune_t2v_lora.sh

@@ -61,7 +61,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 5e-5
   --mixed_precision "bf16"
-  --checkpointing_steps 400
+  --weight_only_checkpointing_steps 400
+  --training_state_checkpointing_steps 400
   --weight_decay 1e-4
   --max_grad_norm 1.0
 )

examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v.sh

@@ -4,7 +4,7 @@ GPU_NUM=1 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 MODEL_TYPE="wan"
 DATA_MERGE_PATH="data/crush-smol/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v_old/"
+OUTPUT_DIR="data/crush-smol_processed_t2v/"
 
 torchrun --nproc_per_node=$GPU_NUM \
     fastvideo/pipelines/preprocess/v1_preprocess.py \

examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v_new.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-GPU_NUM=1 # 2,4,8
+GPU_NUM=2 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATASET_PATH="data/crush-smol/"
 OUTPUT_DIR="data/crush-smol_processed_t2v/"
@@ -14,7 +14,7 @@ torchrun --nproc_per_node=$GPU_NUM \
     --preprocess.dataset_type merged \
     --preprocess.dataset_path $DATASET_PATH \
     --preprocess.dataset_output_dir $OUTPUT_DIR \
-    --preprocess.preprocess_video_batch_size 8 \
+    --preprocess.preprocess_video_batch_size 2 \
     --preprocess.dataloader_num_workers 0 \
     --preprocess.max_height 480 \
     --preprocess.max_width 832 \

examples/training/t2v_to_i2v/wan_t2v/finetune_t2v_to_i2v.sh

@@ -1,94 +0,0 @@
-#!/bin/bash
-
-export WANDB_BASE_URL="https://api.wandb.ai"
-export WANDB_MODE=online
-export TOKENIZERS_PARALLELISM=false
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-DATA_DIR="data/crush-smol_processed_t2v_old"
-VALIDATION_DATASET_FILE="examples/datasets/crush_smol/validation.json"
-NUM_GPUS=8
-# export CUDA_VISIBLE_DEVICES=4,5
-
-
-# Training arguments
-training_args=(
-  --tracker_project_name "wan_t2v_i2v_finetune"
-  --output_dir "checkpoints/wan_t2v_i2v_finetune"
-  --max_train_steps 5000
-  --train_batch_size 1
-  --train_sp_batch_size 1
-  --gradient_accumulation_steps 2
-  --num_latent_t 20
-  --num_height 480
-  --num_width 832
-  --num_frames 77
-  --enable_gradient_checkpointing_type "full"
-)
-
-# Parallel arguments
-parallel_args=(
-  --num_gpus $NUM_GPUS 
-  --sp_size 4
-  --tp_size 1
-  --hsdp_replicate_dim 2
-  --hsdp_shard_dim 4
-)
-
-# Model arguments
-model_args=(
-  --model_path $MODEL_PATH
-  --pretrained_model_name_or_path $MODEL_PATH
-)
-
-# Dataset arguments
-dataset_args=(
-  --data_path $DATA_DIR
-  --dataloader_num_workers 1
-)
-
-# Validation arguments
-validation_args=(
-  --log_validation 
-  --validation_dataset_file $VALIDATION_DATASET_FILE
-  --validation_steps 50
-  --validation_sampling_steps "50" 
-  --validation_guidance_scale "6.0"
-)
-
-# Optimizer arguments
-optimizer_args=(
-  --learning_rate 5e-5
-  --mixed_precision "bf16"
-  --checkpointing_steps 1000
-  --weight_decay 1e-4
-  --max_grad_norm 1.0
-)
-
-# Miscellaneous arguments
-miscellaneous_args=(
-  --inference_mode False
-  --checkpoints_total_limit 3
-  --training_cfg_rate 0.1
-  --multi_phased_distill_schedule "4000-1"
-  --not_apply_cfg_solver
-  --dit_precision "fp32"
-  --num_euler_timesteps 50
-  --ema_start_step 0
-  --enable_gradient_checkpointing_type "full"
-  --t2v_as_i2v_task True
-  # --resume_from_checkpoint "checkpoints/wan_t2v_finetune/checkpoint-2500"
-)
-
-torchrun \
-  --nnodes 1 \
-  --nproc_per_node $NUM_GPUS \
-    fastvideo/training/wan_t2v_i2v_training_pipeline.py \
-    "${parallel_args[@]}" \
-    "${model_args[@]}" \
-    "${dataset_args[@]}" \
-    "${training_args[@]}" \
-    "${optimizer_args[@]}" \
-    "${validation_args[@]}" \
-    "${miscellaneous_args[@]}"

examples/training/t2v_to_i2v/wan_t2v/preprocess_wan_data_t2v.sh

@@ -1,24 +0,0 @@
-#!/bin/bash
-
-GPU_NUM=1 # 2,4,8
-MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
-MODEL_TYPE="wan"
-DATA_MERGE_PATH="data/crush-smol/merge.txt"
-OUTPUT_DIR="data/crush-smol_processed_t2v_i2v_1_3b/"
-
-torchrun --nproc_per_node=$GPU_NUM \
-    fastvideo/pipelines/preprocess/v1_preprocess.py \
-    --model_path $MODEL_PATH \
-    --data_merge_path $DATA_MERGE_PATH \
-    --preprocess_video_batch_size 2 \
-    --seed 42 \
-    --max_height 480 \
-    --max_width 832 \
-    --num_frames 77 \
-    --dataloader_num_workers 0 \
-    --output_dir=$OUTPUT_DIR \
-    --train_fps 16 \
-    --samples_per_file 8 \
-    --flush_frequency 8 \
-    --video_length_tolerance_range 5 \
-    --preprocess_task "t2v_ode_trajectory" 

fastvideo/attention/backends/sage_attn3.py

@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+from fastvideo.attention.backends.sageattn.api import sageattn_blackwell
+
+from fastvideo.attention.backends.abstract import (AttentionBackend,
+                                                   AttentionImpl,
+                                                   AttentionMetadata,
+                                                   AttentionMetadataBuilder)
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class SageAttention3Backend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [64, 128, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "SAGE_ATTN_THREE"
+
+    @staticmethod
+    def get_impl_cls() -> type["SageAttention3Impl"]:
+        return SageAttention3Impl
+
+    @staticmethod
+    def get_metadata_cls() -> type["AttentionMetadata"]:
+        raise NotImplementedError
+
+    @staticmethod
+    def get_builder_cls() -> type["AttentionMetadataBuilder"]:
+        raise NotImplementedError
+
+    # @staticmethod
+    # def get_metadata_cls() -> Type["AttentionMetadata"]:
+    #     return FlashAttentionMetadata
+
+
+class SageAttention3Impl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout = extra_impl_args.get("dropout_p", 0.0)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+        output = sageattn_blackwell(query, key, value, is_causal=self.causal)
+        output = output.transpose(1, 2)
+        return output

fastvideo/attention/layer.py

@@ -3,7 +3,6 @@
 import torch
 import torch.nn as nn
 
-import fastvideo.envs as envs
 from fastvideo.attention.selector import backend_name_to_enum, get_attn_backend
 from fastvideo.distributed.communication_op import (
     sequence_model_parallel_all_gather, sequence_model_parallel_all_to_all_4D)
@@ -37,7 +36,7 @@ class DistributedAttention(nn.Module):
         if num_kv_heads is None:
             num_kv_heads = num_heads
 
-        dtype = torch.bfloat16 if envs.FASTVIDEO_FORCE_ATTN_BF16 else get_compute_dtype()
+        dtype = get_compute_dtype()
         attn_backend = get_attn_backend(
             head_size,
             dtype,
@@ -222,7 +221,7 @@ class LocalAttention(nn.Module):
         if num_kv_heads is None:
             num_kv_heads = num_heads
 
-        dtype = torch.bfloat16 if envs.FASTVIDEO_FORCE_ATTN_BF16 else get_compute_dtype()
+        dtype = get_compute_dtype()
         attn_backend = get_attn_backend(
             head_size,
             dtype,

fastvideo/configs/models/dits/base.py

@@ -15,13 +15,10 @@ class DiTArchConfig(ArchConfig):
     reverse_param_names_mapping: dict = field(default_factory=dict)
     lora_param_names_mapping: dict = field(default_factory=dict)
     _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
-        AttentionBackendEnum.SLIDING_TILE_ATTN,
-        AttentionBackendEnum.SAGE_ATTN,
-        AttentionBackendEnum.FLASH_ATTN,
-        AttentionBackendEnum.TORCH_SDPA,
-        AttentionBackendEnum.VIDEO_SPARSE_ATTN,
-        AttentionBackendEnum.VMOBA_ATTN,
-    )
+        AttentionBackendEnum.SLIDING_TILE_ATTN, AttentionBackendEnum.SAGE_ATTN,
+        AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA,
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN, AttentionBackendEnum.VMOBA_ATTN,
+        AttentionBackendEnum.SAGE_ATTN_THREE)
 
     hidden_size: int = 0
     num_attention_heads: int = 0

fastvideo/configs/pipelines/base.py

@@ -45,13 +45,10 @@ class PipelineConfig:
     embedded_cfg_scale: float = 6.0
     flow_shift: float | None = None
     disable_autocast: bool = False
-    ti2v_task: bool = False
-    t2v_as_i2v_task: bool = False
 
     # Model configuration
     dit_config: DiTConfig = field(default_factory=DiTConfig)
     dit_precision: str = "bf16"
-    dit_forward_precision: str = "bf16"
 
     # VAE configuration
     vae_config: VAEConfig = field(default_factory=VAEConfig)
@@ -218,24 +215,6 @@ class PipelineConfig:
             "Comma-separated list of denoising steps (e.g., '1000,757,522')",
         )
 
-        # TI2V task
-        parser.add_argument(
-            f"--{prefix_with_dot}ti2v-task",
-            action=StoreBoolean,
-            dest=f"{prefix_with_dot.replace('-', '_')}ti2v_task",
-            default=PipelineConfig.ti2v_task,
-            help="Enable TI2V",
-        )
-
-        # T2V to I2V task
-        parser.add_argument(
-            f"--{prefix_with_dot}t2v-as-i2v-task",
-            action=StoreBoolean,
-            dest=f"{prefix_with_dot.replace('-', '_')}t2v_as_i2v_task",
-            default=PipelineConfig.t2v_as_i2v_task,
-            help="Enable T2V to I2V task",
-        )
-
         # Add VAE configuration arguments
         from fastvideo.configs.models.vaes.base import VAEConfig
         VAEConfig.add_cli_args(parser, prefix=f"{prefix_with_dot}vae-config")
@@ -267,9 +246,7 @@ class PipelineConfig:
         """
         from fastvideo.configs.pipelines.registry import (
             get_pipeline_config_cls_from_name)
-        logger.info("WTF model_path: %s", model_path)
         pipeline_config_cls = get_pipeline_config_cls_from_name(model_path)
-        logger.info("pipeline_config_cls: %s", pipeline_config_cls)
 
         return cast(PipelineConfig, pipeline_config_cls(model_path=model_path))
 

fastvideo/configs/pipelines/registry.py

@@ -13,7 +13,7 @@ from fastvideo.configs.pipelines.wan import (
     FastWan2_1_T2V_480P_Config, FastWan2_2_TI2V_5B_Config,
     Wan2_2_I2V_A14B_Config, Wan2_2_T2V_A14B_Config, Wan2_2_TI2V_5B_Config,
     WanI2V480PConfig, WanI2V720PConfig, WanT2V480PConfig, WanT2V720PConfig,
-    SelfForcingWanT2V480PConfig)
+    SelfForcingWanT2V480PConfig, SelfForcingMoEWanT2V480PConfig)
 # isort: on
 from fastvideo.logger import init_logger
 from fastvideo.utils import (maybe_download_model_index,
@@ -36,6 +36,8 @@ PIPE_NAME_TO_CONFIG: dict[str, type[PipelineConfig]] = {
     "FastVideo/stepvideo-t2v-diffusers": StepVideoT2VConfig,
     "FastVideo/Wan2.1-VSA-T2V-14B-720P-Diffusers": WanT2V720PConfig,
     "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers": SelfForcingWanT2V480PConfig,
+    "rand0nmr/SFWan2.1-T2V-A1.3B-Diffusers": SelfForcingMoEWanT2V480PConfig,
+    "rand0nmr/SFWan2.2-T2V-A14B-Diffusers": SelfForcingMoEWanT2V480PConfig,
     "Wan-AI/Wan2.2-TI2V-5B-Diffusers": Wan2_2_TI2V_5B_Config,
     "Wan-AI/Wan2.2-T2V-A14B-Diffusers": Wan2_2_T2V_A14B_Config,
     "Wan-AI/Wan2.2-I2V-A14B-Diffusers": Wan2_2_I2V_A14B_Config,
@@ -50,7 +52,6 @@ PIPELINE_DETECTOR: dict[str, Callable[[str], bool]] = {
     "wandmdpipeline": lambda id: "wandmdpipeline" in id.lower(),
     "wancausaldmdpipeline": lambda id: "wancausaldmdpipeline" in id.lower(),
     "stepvideo": lambda id: "stepvideo" in id.lower(),
-    "wancausaldmdpipeline": lambda id: "wancausaldmdpipeline" in id.lower(),
     # Add other pipeline architecture detectors
 }
 
@@ -138,4 +139,4 @@ def get_pipeline_config_cls_from_name(
             f"No match found for pipeline {pipeline_name_or_path}, please check the pipeline name or path."
         )
 
-    return pipeline_config_cls
+    return pipeline_config_cls

fastvideo/configs/pipelines/wan.py

@@ -54,6 +54,9 @@ class WanT2V480PConfig(PipelineConfig):
     text_encoder_precisions: tuple[str, ...] = field(
         default_factory=lambda: ("fp32", ))
 
+    # self-forcing params
+    warp_denoising_step: bool = True
+
     # WanConfig-specific added parameters
 
     def __post_init__(self):
@@ -133,6 +136,11 @@ class Wan2_2_T2V_A14B_Config(WanT2V480PConfig):
     flow_shift: float | None = 12.0
     boundary_ratio: float | None = 0.875
 
+    # self-forcing params
+    dmd_denoising_steps: list[int] | None = field(
+        default_factory=lambda: [1000, 750, 500, 250])
+    warp_denoising_step: bool = True
+
     def __post_init__(self) -> None:
         self.dit_config.boundary_ratio = self.boundary_ratio
 
@@ -157,3 +165,10 @@ class SelfForcingWanT2V480PConfig(WanT2V480PConfig):
     dmd_denoising_steps: list[int] | None = field(
         default_factory=lambda: [1000, 750, 500, 250])
     warp_denoising_step: bool = True
+
+@dataclass
+class SelfForcingMoEWanT2V480PConfig(SelfForcingWanT2V480PConfig):
+    boundary_ratio: float | None = 0.875
+
+    def __post_init__(self) -> None:
+        self.dit_config.boundary_ratio = self.boundary_ratio

fastvideo/configs/sample/registry.py

@@ -55,6 +55,8 @@ SAMPLING_PARAM_REGISTRY: dict[str, Any] = {
 
     # Causal Self-Forcing Wan2.1
     "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers": SelfForcingWanT2V480PConfig,
+    "rand0nmr/SFWan2.1-T2V-A1.3B-Diffusers": SelfForcingWanT2V480PConfig,
+    "rand0nmr/SFWan2.2-T2V-A14B-Diffusers": SelfForcingWanT2V480PConfig,
     # Add other specific weight variants
 }
 

fastvideo/dataset/lmdb_datasets.py

@@ -1,43 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-from fastvideo.dataset.lmdb_utils import get_array_shape_from_lmdb, retrieve_row_from_lmdb
-from torch.utils.data import Dataset
-import numpy as np
-import torch
-import lmdb
-
-# from Self-Forcing: https://github.com/guandeh17/Self-Forcing/blob/main/utils/dataset.py
-class ODERegressionLMDBDataset(Dataset):
-    def __init__(self, data_path: str, max_pair: int = int(1e8)):
-        print(f"data_path: {data_path}")
-        self.env = lmdb.open(data_path, readonly=True,
-                             lock=False, readahead=False, meminit=False)
-
-        self.latents_shape = get_array_shape_from_lmdb(self.env, 'latents')
-        self.max_pair = max_pair
-
-    def __len__(self):
-        return min(self.latents_shape[0], self.max_pair)
-
-    def __getitem__(self, idx):
-        """
-        Outputs:
-            - prompts: List of Strings
-            - latents: Tensor of shape (num_denoising_steps, num_frames, num_channels, height, width). It is ordered from pure noise to clean image.
-        """
-        latents = retrieve_row_from_lmdb(
-            self.env,
-            "latents", np.float16, idx, shape=self.latents_shape[1:]
-        )
-
-        if len(latents.shape) == 4:
-            latents = latents[None, ...]
-
-        prompts = retrieve_row_from_lmdb(
-            self.env,
-            "prompts", str, idx
-        )
-        return {
-            "prompts": prompts,
-            "ode_latent": torch.tensor(latents, dtype=torch.float32)
-        }
-

fastvideo/dataset/lmdb_utils.py

@@ -1,75 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# from Self-Forcing: https://github.com/guandeh17/Self-Forcing/blob/main/utils/lmdb.py
-
-import numpy as np
-
-
-def get_array_shape_from_lmdb(env, array_name):
-    with env.begin() as txn:
-        image_shape = txn.get(f"{array_name}_shape".encode()).decode()
-        image_shape = tuple(map(int, image_shape.split()))
-    return image_shape
-
-
-def store_arrays_to_lmdb(env, arrays_dict, start_index=0):
-    """
-    Store rows of multiple numpy arrays in a single LMDB.
-    Each row is stored separately with a naming convention.
-    """
-    with env.begin(write=True) as txn:
-        for array_name, array in arrays_dict.items():
-            for i, row in enumerate(array):
-                # Convert row to bytes
-                if isinstance(row, str):
-                    row_bytes = row.encode()
-                else:
-                    row_bytes = row.tobytes()
-
-                data_key = f'{array_name}_{start_index + i}_data'.encode()
-
-                txn.put(data_key, row_bytes)
-
-
-def process_data_dict(data_dict, seen_prompts):
-    output_dict = {}
-
-    all_videos = []
-    all_prompts = []
-    for prompt, video in data_dict.items():
-        if prompt in seen_prompts:
-            continue
-        else:
-            seen_prompts.add(prompt)
-
-        video = video.half().numpy()
-        all_videos.append(video)
-        all_prompts.append(prompt)
-
-    if len(all_videos) == 0:
-        return {"latents": np.array([]), "prompts": np.array([])}
-
-    all_videos = np.concatenate(all_videos, axis=0)
-
-    output_dict['latents'] = all_videos
-    output_dict['prompts'] = np.array(all_prompts)
-
-    return output_dict
-
-
-def retrieve_row_from_lmdb(lmdb_env, array_name, dtype, row_index, shape=None):
-    """
-    Retrieve a specific row from a specific array in the LMDB.
-    """
-    data_key = f'{array_name}_{row_index}_data'.encode()
-
-    with lmdb_env.begin() as txn:
-        row_bytes = txn.get(data_key)
-
-    if dtype == str:
-        array = row_bytes.decode()
-    else:
-        array = np.frombuffer(row_bytes, dtype=dtype)
-
-    if shape is not None and len(shape) > 0:
-        array = array.reshape(shape)
-    return array

fastvideo/dataset/utils.py

@@ -3,9 +3,6 @@ from typing import Any, cast
 
 import numpy as np
 import torch
-from fastvideo.logger import init_logger
-
-logger = init_logger(__name__)
 
 
 def pad(t: torch.Tensor, padding_length: int) -> tuple[torch.Tensor, torch.Tensor]:

fastvideo/envs.py

@@ -18,7 +18,6 @@ if TYPE_CHECKING:
     FASTVIDEO_LOGGING_PREFIX: str = ""
     FASTVIDEO_LOGGING_CONFIG_PATH: str | None = None
     FASTVIDEO_TRACE_FUNCTION: int = 0
-    FASTVIDEO_FORCE_ATTN_BF16: bool = False
     FASTVIDEO_ATTENTION_BACKEND: str | None = None
     FASTVIDEO_ATTENTION_CONFIG: str | None = None
     FASTVIDEO_WORKER_MULTIPROC_METHOD: str = "fork"
@@ -169,10 +168,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "FASTVIDEO_TRACE_FUNCTION":
     lambda: int(os.getenv("FASTVIDEO_TRACE_FUNCTION", "0")),
 
-    # if set, fastvideo will force attention to be computed in bfloat16
-    "FASTVIDEO_FORCE_ATTN_BF16":
-    lambda: bool(int(os.getenv("FASTVIDEO_FORCE_ATTN_BF16", "0"))),
-
     # Backend for attention computation
     # Available options:
     # - "TORCH_SDPA": use torch.nn.MultiheadAttention
@@ -180,6 +175,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # - "SLIDING_TILE_ATTN" : use Sliding Tile Attention
     # - "VIDEO_SPARSE_ATTN": use Video Sparse Attention
     # - "SAGE_ATTN": use Sage Attention
+    # - "SAGE_ATTN_THREE": use Sage Attention 3
     "FASTVIDEO_ATTENTION_BACKEND":
     lambda: os.getenv("FASTVIDEO_ATTENTION_BACKEND", None),
 

fastvideo/fastvideo_args.py

@@ -133,6 +133,7 @@ class FastVideoArgs:
 
     # Compilation
     enable_torch_compile: bool = False
+    torch_compile_kwargs: dict[str, Any] = field(default_factory=dict)
 
     disable_autocast: bool = False
 
@@ -164,7 +165,7 @@ class FastVideoArgs:
     # dmd_denoising_steps: List[int] | None = field(default=None)
 
     # MoE parameters used by Wan2.2
-    boundary_ratio: float | None = None
+    boundary_ratio: float | None = 0.875
 
     @property
     def training_mode(self) -> bool:
@@ -330,6 +331,13 @@ class FastVideoArgs:
             help="Use torch.compile to speed up DiT inference." +
             "However, will likely cause precision drifts. See (https://github.com/pytorch/pytorch/issues/145213)",
         )
+        parser.add_argument(
+            "--torch-compile-kwargs",
+            type=str,
+            default=None,
+            help=
+            "JSON string of kwargs to pass to torch.compile. Example: '{\"backend\":\"inductor\",\"mode\":\"reduce-overhead\"}'",
+        )
 
         parser.add_argument(
             "--dit-cpu-offload",
@@ -431,6 +439,21 @@ class FastVideoArgs:
                 mode_value = getattr(args, attr, FastVideoArgs.mode.value)
                 kwargs['mode'] = ExecutionMode.from_string(
                     mode_value) if isinstance(mode_value, str) else mode_value
+            elif attr == 'torch_compile_kwargs':
+                # Parse JSON string for torch.compile kwargs
+                torch_compile_kwargs_str = getattr(args, 'torch_compile_kwargs',
+                                                   None)
+                if torch_compile_kwargs_str:
+                    try:
+                        import json
+                        kwargs['torch_compile_kwargs'] = json.loads(
+                            torch_compile_kwargs_str)
+                    except json.JSONDecodeError as e:
+                        raise ValueError(
+                            f"Invalid JSON for torch_compile_kwargs: {e}"
+                        ) from e
+                else:
+                    kwargs['torch_compile_kwargs'] = {}
             elif attr == 'workload_type':
                 # Convert string to WorkloadType enum
                 workload_type_value = getattr(args, 'workload_type',
@@ -610,10 +633,8 @@ class TrainingArgs(FastVideoArgs):
 
     # text encoder & vae & diffusion model
     pretrained_model_name_or_path: str = ""
-    dit_model_name_or_path: str = ""
 
     # DMD model paths - separate paths for each network
-    generator_model_path: str = ""  # path for generator (student) model
     real_score_model_path: str = ""  # path for real score (teacher) model
     fake_score_model_path: str = ""  # path for fake score (critic) model
 
@@ -637,9 +658,9 @@ class TrainingArgs(FastVideoArgs):
     # output
     output_dir: str = ""
     checkpoints_total_limit: int = 0
-    checkpointing_steps: int = 0
     resume_from_checkpoint: str = ""  # specify the checkpoint folder to resume from
     init_weights_from_safetensors: str = ""  # path to safetensors file for initial weight loading
+    init_weights_from_safetensors_2: str = ""  # path to safetensors file for initial weight loading for transformer_2
 
     # optimizer & scheduler
     num_train_epochs: int = 0
@@ -696,6 +717,7 @@ class TrainingArgs(FastVideoArgs):
     min_timestep_ratio: float = 0.2
     max_timestep_ratio: float = 0.98
     real_score_guidance_scale: float = 3.5
+    real_score_guidance_scale_2: float = 3.5
     fake_score_learning_rate: float = 0.0  # separate learning rate for fake_score_transformer, if 0.0, use learning_rate
     fake_score_lr_scheduler: str = "constant"  # separate lr scheduler for fake_score_transformer, if not set, use lr_scheduler
     fake_score_betas: str = "0.9,0.999"  # betas for fake score optimizer, format: "beta1,beta2"
@@ -705,14 +727,12 @@ class TrainingArgs(FastVideoArgs):
     # simulate generator forward to match inference
     simulate_generator_forward: bool = False
     warp_denoising_step: bool = False
-    intermediate_latents_visualization: bool = False
 
     # Self-forcing specific arguments
     num_frame_per_block: int = 3
     independent_first_frame: bool = False
     enable_gradient_masking: bool = True
     gradient_mask_last_n_frames: int = 21
-    validate_cache_structure: bool = False  # Debug flag for cache validation
     same_step_across_blocks: bool = False  # Use same exit timestep for all blocks
     last_step_only: bool = False  # Only use the last timestep for training
     context_noise: int = 0  # Context noise level for cache updates
@@ -886,9 +906,6 @@ class TrainingArgs(FastVideoArgs):
         parser.add_argument("--checkpoints-total-limit",
                             type=int,
                             help="Maximum number of checkpoints to keep")
-        parser.add_argument("--checkpointing-steps",
-                            type=int,
-                            help="Steps between checkpoints")
         parser.add_argument(
             "--training-state-checkpointing-steps",
             type=int,
@@ -905,6 +922,10 @@ class TrainingArgs(FastVideoArgs):
             "--init-weights-from-safetensors",
             type=str,
             help="Path to safetensors file for initial weight loading")
+        parser.add_argument(
+            "--init-weights-from-safetensors-2",
+            type=str,
+            help="Path to safetensors file for initial weight loading")
         parser.add_argument("--logging-dir",
                             type=str,
                             help="Directory for logging")
@@ -1082,6 +1103,10 @@ class TrainingArgs(FastVideoArgs):
                             type=float,
                             default=TrainingArgs.real_score_guidance_scale,
                             help="Teacher guidance scale")
+        parser.add_argument("--real-score-guidance-scale-2",
+                            type=float,
+                            default=TrainingArgs.real_score_guidance_scale_2,
+                            help="Teacher guidance scale")
         parser.add_argument("--fake-score-learning-rate",
                             type=float,
                             default=TrainingArgs.fake_score_learning_rate,
@@ -1141,11 +1166,10 @@ class TrainingArgs(FastVideoArgs):
             "--last-step-only",
             action=StoreBoolean,
             help="Whether to only use the last timestep for training")
-        parser.add_argument(
-            "--context-noise",
-            type=int,
-            default=TrainingArgs.context_noise,
-            help="Context noise level for cache updates")
+        parser.add_argument("--context-noise",
+                            type=int,
+                            default=TrainingArgs.context_noise,
+                            help="Context noise level for cache updates")
 
         return parser
 
@@ -1153,4 +1177,4 @@ class TrainingArgs(FastVideoArgs):
 def parse_int_list(value: str) -> list[int]:
     if not value:
         return []
-    return [int(x.strip()) for x in value.split(",")]
+    return [int(x.strip()) for x in value.split(",")]

fastvideo/layers/layernorm.py

@@ -212,9 +212,9 @@ class ScaleResidualLayerNormScaleShift(nn.Module):
             frame_seqlen = normalized.shape[1] // num_frames
             modulated = (
                 normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
-                (1 + scale) + shift).flatten(1, 2)
+                (1.0 + scale) + shift).flatten(1, 2)
         else:
-            modulated = normalized * (1 + scale) + shift
+            modulated = normalized * (1.0 + scale) + shift
         return modulated, residual_output
 
 
@@ -267,11 +267,11 @@ class LayerNormScaleShift(nn.Module):
             frame_seqlen = normalized.shape[1] // num_frames
             output = (
                 normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
-                (1 + scale) + shift).flatten(1, 2)
+                (1.0 + scale) + shift).flatten(1, 2)
         else:
             # scale.shape: [batch_size, 1, inner_dim]
             # shift.shape: [batch_size, 1, inner_dim]
-            output = normalized * (1 + scale) + shift
+            output = normalized * (1.0 + scale) + shift
 
         if self.compute_dtype == torch.float32:
             output = output.to(x.dtype)

fastvideo/models/dits/causal_wanvideo.py

@@ -147,8 +147,8 @@ class CausalWanSelfAttention(nn.Module):
                 # Assign new keys/values directly up to current_end
                 local_end_index = kv_cache["local_end_index"].item() + current_end - kv_cache["global_end_index"].item()
                 local_start_index = local_end_index - num_new_tokens
-                # kv_cache["k"] = kv_cache["k"].detach()
-                # kv_cache["v"] = kv_cache["v"].detach()
+                kv_cache["k"] = kv_cache["k"].detach()
+                kv_cache["v"] = kv_cache["v"].detach()
                 # logger.info("kv_cache['k'] is in comp graph: %s", kv_cache["k"].requires_grad or kv_cache["k"].grad_fn is not None)
                 kv_cache["k"][:, local_start_index:local_end_index] = roped_key
                 kv_cache["v"][:, local_start_index:local_end_index] = v
@@ -262,14 +262,9 @@ class CausalWanTransformerBlock(nn.Module):
         # *_msa.shape: [batch_size, num_frames, 1, inner_dim]
         # assert shift_msa.dtype == torch.float32
 
-        # logger.info("temb sum: %s, dtype: %s", temb.float().sum().item(), temb.dtype)
-        # logger.info("scale_msa sum: %s, dtype: %s", scale_msa.float().sum().item(), scale_msa.dtype)
-        # logger.info("shift_msa sum: %s, dtype: %s", shift_msa.float().sum().item(), shift_msa.dtype)
-
         # 1. Self-attention
         norm_hidden_states = (self.norm1(hidden_states).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
                         (1 + scale_msa) + shift_msa).flatten(1, 2)
-        # logger.info("norm_hidden_states sum: %s, shape: %s", norm_hidden_states.float().sum().item(), norm_hidden_states.shape)
         query, _ = self.to_q(norm_hidden_states)
         key, _ = self.to_k(norm_hidden_states)
         value, _ = self.to_v(norm_hidden_states)
@@ -375,7 +370,8 @@ class CausalWanTransformer3DModel(BaseDiT):
 
         # Causal-specific
         self.block_mask = None
-        self.num_frame_per_block = 3
+        self.num_frame_per_block = config.arch_config.num_frames_per_block
+        assert self.num_frame_per_block <= 3
         self.independent_first_frame = False
 
         self.__post_init__()
@@ -556,8 +552,6 @@ class CausalWanTransformer3DModel(BaseDiT):
                 start_frame: int = 0,
                 **kwargs) -> torch.Tensor:
 
-        logger.info("timestep dtype: %s, timestep sum: %s", timestep.dtype, timestep.float().sum().item())
-
         orig_dtype = hidden_states.dtype
         if not isinstance(encoder_hidden_states, torch.Tensor):
             encoder_hidden_states = encoder_hidden_states[0]
@@ -652,11 +646,10 @@ class CausalWanTransformer3DModel(BaseDiT):
         **kwargs
     ):
         if kwargs.get('kv_cache', None) is not None:
-            noise_pred = self._forward_inference(*args, **kwargs)
+            return self._forward_inference(*args, **kwargs)
         else:
-            noise_pred = self._forward_train(*args, **kwargs)
+            return self._forward_train(*args, **kwargs)
 
-        return noise_pred
 
     def unpatchify(self, x, grid_sizes):
         r"""
@@ -682,4 +675,4 @@ class CausalWanTransformer3DModel(BaseDiT):
             # u = torch.einsum('fhwpqrc->cfphqwr', u.contiguous())
             u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
             out.append(u)
-        return out
+        return out

fastvideo/models/dits/wanvideo.py

@@ -1,5 +1,3 @@
-import torch
-import torch.nn as nn
 # SPDX-License-Identifier: Apache-2.0
 
 import math
@@ -39,14 +37,16 @@ class WanImageEmbedding(torch.nn.Module):
     def __init__(self, in_features: int, out_features: int):
         super().__init__()
 
-        self.norm1 = nn.LayerNorm(in_features)
+        self.norm1 = FP32LayerNorm(in_features)
         self.ff = MLP(in_features, in_features, out_features, act_type="gelu")
-        self.norm2 = nn.LayerNorm(out_features)
+        self.norm2 = FP32LayerNorm(out_features)
 
-    def forward(self, encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+    def forward(self,
+                encoder_hidden_states_image: torch.Tensor) -> torch.Tensor:
+        dtype = encoder_hidden_states_image.dtype
         hidden_states = self.norm1(encoder_hidden_states_image)
         hidden_states = self.ff(hidden_states)
-        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.norm2(hidden_states).to(dtype)
         return hidden_states
 
 
@@ -62,7 +62,7 @@ class WanTimeTextImageEmbedding(nn.Module):
         super().__init__()
 
         self.time_embedder = TimestepEmbedder(
-            dim, frequency_embedding_size=time_freq_dim, act_layer="silu", freq_dtype=torch.float64)
+            dim, frequency_embedding_size=time_freq_dim, act_layer="silu")
         self.time_modulation = ModulateProjection(dim,
                                                   factor=6,
                                                   act_layer="silu")
@@ -156,12 +156,12 @@ class WanT2VCrossAttention(WanSelfAttention):
         b, n, d = x.size(0), self.num_heads, self.head_dim
 
         # compute query, key, value
-        q = self.norm_q.forward_native(self.to_q(x)[0]).view(b, -1, n, d)
+        q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d)
 
         if crossattn_cache is not None:
             if not crossattn_cache["is_init"]:
                 crossattn_cache["is_init"] = True
-                k = self.norm_k.forward_native(self.to_k(context)[0]).view(b, -1, n, d)
+                k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
                 v = self.to_v(context)[0].view(b, -1, n, d)
                 crossattn_cache["k"] = k
                 crossattn_cache["v"] = v
@@ -169,16 +169,11 @@ class WanT2VCrossAttention(WanSelfAttention):
                 k = crossattn_cache["k"]
                 v = crossattn_cache["v"]
         else:
-            k = self.norm_k.forward_native(self.to_k(context)[0]).view(b, -1, n, d)
+            k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
             v = self.to_v(context)[0].view(b, -1, n, d)
 
-        if envs.FASTVIDEO_FORCE_ATTN_BF16:
-            out_dtype = v.dtype
-            # compute attention
-            x = self.attn(q.to(torch.bfloat16), k.to(torch.bfloat16), v.to(torch.bfloat16)).to(out_dtype)
-        else:
-            # compute attention
-            x = self.attn(q, k, v)
+        # compute attention
+        x = self.attn(q, k, v)
 
         # output
         x = x.flatten(2)
@@ -218,10 +213,10 @@ class WanI2VCrossAttention(WanSelfAttention):
         b, n, d = x.size(0), self.num_heads, self.head_dim
 
         # compute query, key, value
-        q = self.norm_q.forward_native(self.to_q(x)[0]).view(b, -1, n, d)
-        k = self.norm_k.forward_native(self.to_k(context)[0]).view(b, -1, n, d)
+        q = self.norm_q(self.to_q(x)[0]).view(b, -1, n, d)
+        k = self.norm_k(self.to_k(context)[0]).view(b, -1, n, d)
         v = self.to_v(context)[0].view(b, -1, n, d)
-        k_img = self.norm_added_k.forward_native(self.add_k_proj(context_img)[0]).view(
+        k_img = self.norm_added_k(self.add_k_proj(context_img)[0]).view(
             b, -1, n, d)
         v_img = self.add_v_proj(context_img)[0].view(b, -1, n, d)
         img_x = self.attn(q, k_img, v_img)
@@ -252,7 +247,7 @@ class WanTransformerBlock(nn.Module):
         super().__init__()
 
         # 1. Self-attention
-        self.norm1 = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
         self.to_q = ReplicatedLinear(dim, dim, bias=True)
         self.to_k = ReplicatedLinear(dim, dim, bias=True)
         self.to_v = ReplicatedLinear(dim, dim, bias=True)
@@ -283,29 +278,29 @@ class WanTransformerBlock(nn.Module):
             norm_type="layer",
             eps=eps,
             elementwise_affine=True,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)
 
         # 2. Cross-attention
         if added_kv_proj_dim is not None:
             # I2V
-            self.attn2 = WanI2VCrossAttention(dim, 
-                                              num_heads, 
-                                              qk_norm=qk_norm, 
+            self.attn2 = WanI2VCrossAttention(dim,
+                                              num_heads,
+                                              qk_norm=qk_norm,
                                               eps=eps)
-
         else:
             # T2V
-            self.attn2 = WanT2VCrossAttention(dim, 
-                                              num_heads, 
-                                              qk_norm=qk_norm, 
+            self.attn2 = WanT2VCrossAttention(dim,
+                                              num_heads,
+                                              qk_norm=qk_norm,
                                               eps=eps)
-
         self.cross_attn_residual_norm = ScaleResidualLayerNormScaleShift(
-            dim, 
-            norm_type="layer", 
-            eps=eps, 
-            elementwise_affine=False, 
-            dtype=torch.float32)
+            dim,
+            norm_type="layer",
+            eps=eps,
+            elementwise_affine=False,
+            dtype=torch.float32,
+            compute_dtype=torch.float32)
 
         # 3. Feed-forward
         self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -324,11 +319,12 @@ class WanTransformerBlock(nn.Module):
             hidden_states = hidden_states.squeeze(1)
         bs, seq_length, _ = hidden_states.shape
         orig_dtype = hidden_states.dtype
+        # assert orig_dtype != torch.float32
 
         if temb.dim() == 4:
             # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
             shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
-                self.scale_shift_table.unsqueeze(0) + temb
+                self.scale_shift_table.unsqueeze(0) + temb.float()
             ).chunk(6, dim=2)
             # batch_size, seq_len, 1, inner_dim
             shift_msa = shift_msa.squeeze(2)
@@ -339,20 +335,22 @@ class WanTransformerBlock(nn.Module):
             c_gate_msa = c_gate_msa.squeeze(2)
         else:
             # temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
-            e = self.scale_shift_table + temb
+            e = self.scale_shift_table + temb.float()
             shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
                 6, dim=1)
+        assert shift_msa.dtype == torch.float32
 
         # 1. Self-attention
-        norm_hidden_states = self.norm1(hidden_states) * (1 + scale_msa) + shift_msa
+        norm_hidden_states = (self.norm1(hidden_states.float()) *
+                              (1 + scale_msa) + shift_msa).to(orig_dtype)
         query, _ = self.to_q(norm_hidden_states)
         key, _ = self.to_k(norm_hidden_states)
         value, _ = self.to_v(norm_hidden_states)
 
         if self.norm_q is not None:
-            query = self.norm_q.forward_native(query)
+            query = self.norm_q(query)
         if self.norm_k is not None:
-            key = self.norm_k.forward_native(key)
+            key = self.norm_k(key)
 
         query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
         key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -364,12 +362,7 @@ class WanTransformerBlock(nn.Module):
                                        is_neox_style=False), _apply_rotary_emb(
                                            key, cos, sin, is_neox_style=False)
 
-        if envs.FASTVIDEO_FORCE_ATTN_BF16:
-            out_dtype = value.dtype
-            attn_output, _ = self.attn1(query.to(torch.bfloat16), key.to(torch.bfloat16), value.to(torch.bfloat16))
-            attn_output = attn_output.to(out_dtype)
-        else:
-            attn_output, _ = self.attn1(query, key, value)
+        attn_output, _ = self.attn1(query, key, value)
         attn_output = attn_output.flatten(2)
         attn_output, _ = self.to_out(attn_output)
         attn_output = attn_output.squeeze(1)
@@ -377,20 +370,26 @@ class WanTransformerBlock(nn.Module):
         null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
         norm_hidden_states, hidden_states = self.self_attn_residual_norm(
             hidden_states, attn_output, gate_msa, null_shift, null_scale)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)
 
         # 2. Cross-attention
-        attn_output = self.attn2(norm_hidden_states, 
-                                 context=encoder_hidden_states, 
+        attn_output = self.attn2(norm_hidden_states,
+                                 context=encoder_hidden_states,
                                  context_lens=None)
         norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
             hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)
 
         # 3. Feed-forward
         ff_output = self.ffn(norm_hidden_states)
         hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
+        hidden_states = hidden_states.to(orig_dtype)
 
         return hidden_states
 
+
 class WanTransformerBlock_VSA(nn.Module):
 
     def __init__(self,
@@ -407,7 +406,7 @@ class WanTransformerBlock_VSA(nn.Module):
         super().__init__()
 
         # 1. Self-attention
-        self.norm1 = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
         self.to_q = ReplicatedLinear(dim, dim, bias=True)
         self.to_k = ReplicatedLinear(dim, dim, bias=True)
         self.to_v = ReplicatedLinear(dim, dim, bias=True)
@@ -439,7 +438,8 @@ class WanTransformerBlock_VSA(nn.Module):
             norm_type="layer",
             eps=eps,
             elementwise_affine=True,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)
 
         # 2. Cross-attention
         if added_kv_proj_dim is not None:
@@ -459,7 +459,8 @@ class WanTransformerBlock_VSA(nn.Module):
             norm_type="layer",
             eps=eps,
             elementwise_affine=False,
-            dtype=torch.float32)
+            dtype=torch.float32,
+            compute_dtype=torch.float32)
 
         # 3. Feed-forward
         self.ffn = MLP(dim, ffn_dim, act_type="gelu_pytorch_tanh")
@@ -479,22 +480,23 @@ class WanTransformerBlock_VSA(nn.Module):
         bs, seq_length, _ = hidden_states.shape
         orig_dtype = hidden_states.dtype
         # assert orig_dtype != torch.float32
-        e = self.scale_shift_table + temb
+        e = self.scale_shift_table + temb.float()
         shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
             6, dim=1)
+        assert shift_msa.dtype == torch.float32
 
         # 1. Self-attention
-        norm_hidden_states = (self.norm1(hidden_states) *
-                              (1 + scale_msa) + shift_msa)
+        norm_hidden_states = (self.norm1(hidden_states.float()) *
+                              (1 + scale_msa) + shift_msa).to(orig_dtype)
         query, _ = self.to_q(norm_hidden_states)
         key, _ = self.to_k(norm_hidden_states)
         value, _ = self.to_v(norm_hidden_states)
         gate_compress, _ = self.to_gate_compress(norm_hidden_states)
 
         if self.norm_q is not None:
-            query = self.norm_q.forward_native(query)
+            query = self.norm_q(query)
         if self.norm_k is not None:
-            key = self.norm_k.forward_native(key)
+            key = self.norm_k(key)
 
         query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
         key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
@@ -519,6 +521,8 @@ class WanTransformerBlock_VSA(nn.Module):
         null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
         norm_hidden_states, hidden_states = self.self_attn_residual_norm(
             hidden_states, attn_output, gate_msa, null_shift, null_scale)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)
 
         # 2. Cross-attention
         attn_output = self.attn2(norm_hidden_states,
@@ -526,15 +530,17 @@ class WanTransformerBlock_VSA(nn.Module):
                                  context_lens=None)
         norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
             hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)
 
         # 3. Feed-forward
         ff_output = self.ffn(norm_hidden_states)
         hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
+        hidden_states = hidden_states.to(orig_dtype)
 
         return hidden_states
 
 
-
 class WanTransformer3DModel(CachableDiT):
     _fsdp_shard_conditions = WanVideoConfig()._fsdp_shard_conditions
     _compile_conditions = WanVideoConfig()._compile_conditions
@@ -592,7 +598,8 @@ class WanTransformer3DModel(CachableDiT):
                                             norm_type="layer",
                                             eps=config.eps,
                                             elementwise_affine=False,
-                                            dtype=torch.float32)
+                                            dtype=torch.float32,
+                                            compute_dtype=torch.float32)
         self.proj_out = nn.Linear(
             inner_dim, config.out_channels * math.prod(config.patch_size))
         self.scale_shift_table = nn.Parameter(
@@ -652,12 +659,10 @@ class WanTransformer3DModel(CachableDiT):
             rope_theta=10000)
         freqs_cos = freqs_cos.to(hidden_states.device)
         freqs_sin = freqs_sin.to(hidden_states.device)
-        freqs_cis = (freqs_cos,
-                     freqs_sin) if freqs_cos is not None else None
+        freqs_cis = (freqs_cos.float(),
+                     freqs_sin.float()) if freqs_cos is not None else None
 
         hidden_states = self.patch_embedding(hidden_states)
-        grid_sizes = torch.stack(
-            [torch.tensor(hidden_states[0].shape[1:], dtype=torch.long)])
         hidden_states = hidden_states.flatten(2).transpose(1, 2)
 
         # timestep shape: batch_size, or batch_size, seq_len (wan 2.2 ti2v)
@@ -667,8 +672,6 @@ class WanTransformer3DModel(CachableDiT):
         else:
             ts_seq_len = None
 
-        encoder_hidden_states = torch.cat([encoder_hidden_states, encoder_hidden_states.new_zeros(1, self.text_len - encoder_hidden_states.size(1), encoder_hidden_states.size(2))], dim=1)
-
         temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
             timestep, encoder_hidden_states, encoder_hidden_states_image, timestep_seq_len=ts_seq_len)
         if ts_seq_len is not None:
@@ -725,35 +728,14 @@ class WanTransformer3DModel(CachableDiT):
         hidden_states = self.norm_out(hidden_states, shift, scale)
         hidden_states = self.proj_out(hidden_states)
 
-        output = self.unpatchify(hidden_states, grid_sizes)
+        hidden_states = hidden_states.reshape(batch_size, post_patch_num_frames,
+                                              post_patch_height,
+                                              post_patch_width, p_t, p_h, p_w,
+                                              -1)
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
 
-        return torch.stack(output)
-    
-    def unpatchify(self, x, grid_sizes):
-        r"""
-
-
-        Args:
-            x (List[Tensor]):
-                List of patchified features, each with shape [L, C_out * prod(patch_size)]
-            grid_sizes (Tensor):
-                Original spatial-temporal grid dimensions before patching,
-
-
-        Returns:
-            Tensor:
-                Reconstructed video tensors with shape [B, C_out, F, H / 8, W / 8]
-        """
-
-        c = self.out_channels
-        out = []
-        for u, v in zip(x, grid_sizes.tolist()):
-            u = u[:math.prod(v)].view(*v, *self.patch_size, c)
-            u = u.permute(6, 0, 3, 1, 4, 2, 5)
-            # u = torch.einsum('fhwpqrc->cfphqwr', u.contiguous())
-            u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
-            out.append(u)
-        return out
+        return output
 
     def maybe_cache_states(self, hidden_states: torch.Tensor,
                            original_hidden_states: torch.Tensor) -> None:
@@ -845,4 +827,5 @@ class WanTransformer3DModel(CachableDiT):
         if self.is_even:
             return hidden_states + self.previous_residual_even
         else:
-            return hidden_states + self.previous_residual_odd
+            return hidden_states + self.previous_residual_odd
+        

fastvideo/models/loader/component_loader.py

@@ -238,7 +238,6 @@ class TextEncoderLoader(ComponentLoader):
                 1]
 
         target_device = get_local_torch_device()
-        logger.info("Loading text encoder in %s precision", encoder_precision)
         # TODO(will): add support for other dtypes
         return self.load_model(model_path, encoder_config, target_device,
                                fastvideo_args, encoder_precision)
@@ -416,6 +415,7 @@ class TransformerLoader(ComponentLoader):
             raise ValueError(
                 "Model config does not contain a _class_name attribute. "
                 "Only diffusers format is supported.")
+
         logger.info("transformer cls_name: %s", cls_name)
         if fastvideo_args.override_transformer_cls_name is not None:
             cls_name = fastvideo_args.override_transformer_cls_name
@@ -442,16 +442,29 @@ class TransformerLoader(ComponentLoader):
                             not hasattr(fastvideo_args, '_loading_teacher_critic_model'))
 
         if use_custom_weights:
-            logger.info("Using custom initialization weights from: %s", custom_weights_path)
-            safetensors_list = [custom_weights_path]
+            if 'transformer_2' in model_path:
+                custom_weights_path = getattr(fastvideo_args, 'init_weights_from_safetensors_2', None)
+            assert custom_weights_path is not None, "Custom initialization weights must be provided"
+            
+            # Handle both directory and single file cases
+            if os.path.isdir(custom_weights_path):
+                # Directory: collect all safetensors files
+                safetensors_list = glob.glob(
+                    os.path.join(str(custom_weights_path), "*.safetensors"))
+                if not safetensors_list:
+                    raise ValueError(f"No safetensors files found in directory {custom_weights_path}")
+            elif os.path.isfile(custom_weights_path):
+                # Single file: verify it's a safetensors file
+                assert custom_weights_path.endswith(".safetensors"), "Custom initialization weights must be a safetensors file"
+                safetensors_list = [custom_weights_path]
+            else:
+                raise ValueError(f"Custom weights path does not exist or is not accessible: {custom_weights_path}")
 
-        logger.info("Loading model from %s safetensors files in %s",
-                    len(safetensors_list), model_path)
+        logger.info("Loading model from %s safetensors files: %s",
+                    len(safetensors_list), safetensors_list)
 
         default_dtype = PRECISION_TO_TYPE[
             fastvideo_args.pipeline_config.dit_precision]
-        param_dtype = PRECISION_TO_TYPE[
-            fastvideo_args.pipeline_config.dit_forward_precision]
 
         # Load the model using FSDP loader
         logger.info("Loading model from %s, default_dtype: %s", cls_name,
@@ -472,20 +485,19 @@ class TransformerLoader(ComponentLoader):
             fsdp_inference=fastvideo_args.use_fsdp_inference,
             # TODO(will): make these configurable
             default_dtype=default_dtype,
-            param_dtype=param_dtype,
+            param_dtype=torch.bfloat16,
             reduce_dtype=torch.float32,
             output_dtype=None,
-            training_mode=fastvideo_args.training_mode)
+            training_mode=fastvideo_args.training_mode,
+            enable_torch_compile=fastvideo_args.enable_torch_compile,
+            torch_compile_kwargs=fastvideo_args.torch_compile_kwargs)
 
 
         total_params = sum(p.numel() for p in model.parameters())
         logger.info("Loaded model with %.2fB parameters", total_params / 1e9)
 
-        # Need to convert the model to the default_dtype
-        # Otherwise, the model master weights will be in param_dtype, and the gradients will also be in param_dtype
-        # This means the param update will be in lower precision, causing precision loss
-        logger.info("Converting model to dtype: %s", default_dtype)
-        model = model.to(default_dtype)
+        assert next(model.parameters()).dtype == default_dtype, "Model dtype does not match default dtype"
+
         model = model.eval()
         return model
 

fastvideo/models/loader/fsdp_load.py

@@ -54,7 +54,7 @@ def set_default_dtype(dtype: torch.dtype) -> Generator[None, None, None]:
         torch.set_default_dtype(old_dtype)
 
 
-# TODO(PY): add compile option
+# Supports optional torch.compile for FSDP-wrapped models during training
 def maybe_load_fsdp_model(
     model_cls: type[nn.Module],
     init_params: dict[str, Any],
@@ -70,6 +70,8 @@ def maybe_load_fsdp_model(
     output_dtype: torch.dtype | None = None,
     training_mode: bool = True,
     pin_cpu_memory: bool = True,
+    enable_torch_compile: bool = False,
+    torch_compile_kwargs: dict[str, Any] | None = None,
 ) -> torch.nn.Module:
     """
     Load the model with FSDP if is training, else load the model without FSDP.
@@ -88,6 +90,7 @@ def maybe_load_fsdp_model(
         mp_policy=mp_policy,
     )
 
+    logger.info("Loading model with default_dtype: %s", default_dtype)
     with set_default_dtype(default_dtype), torch.device("meta"):
         model = model_cls(**init_params)
 
@@ -138,6 +141,14 @@ def maybe_load_fsdp_model(
         # Avoid unintended computation graph accumulation during inference
         if isinstance(p, torch.nn.Parameter):
             p.requires_grad = False
+
+    compile_in_loader = enable_torch_compile and training_mode
+    if compile_in_loader:
+        compile_kwargs = torch_compile_kwargs or {}
+        logger.info("Enabling torch.compile for FSDP training module with kwargs=%s",
+                    compile_kwargs)
+        model = torch.compile(model, **compile_kwargs)
+        logger.info("torch.compile enabled for %s", type(model).__name__)
     return model
 
 

fastvideo/models/utils.py

@@ -181,24 +181,49 @@ def pred_noise_to_pred_video(pred_noise: torch.Tensor,
     pred_video = noise_input_latent - sigma_t * pred_noise
     return pred_video.to(dtype)
 
-def pred_video_to_pred_noise(x0_pred: torch.Tensor, xt: torch.Tensor, timestep: torch.Tensor, scheduler: Any) -> torch.Tensor:
+def pred_noise_to_x_bound(pred_noise: torch.Tensor,
+                             noise_input_latent: torch.Tensor,
+                             timestep: torch.Tensor,
+                             boundary_timestep: torch.Tensor,
+                             scheduler: Any) -> torch.Tensor:
     """
-    Convert x0 prediction to flow matching's prediction.
-    x0_pred: the x0 prediction with shape [B, C, H, W]
-    xt: the input noisy data with shape [B, C, H, W]
-    timestep: the timestep with shape [B]
+    Convert predicted noise to clean latent.
 
-    pred = (x_t - x_0) / sigma_t
+    Args:
+    pred_noise: the predicted noise with shape [B, C, H, W]
+        where B is batch_size or batch_size * num_frames
+    noise_input_latent: the noisy latent with shape [B, C, H, W],
+    timestep: the timestep with shape [1] or [bs * num_frames] or [bs, num_frames]
+    boundary_timestep: the boundary timestep with shape [1] or [bs * num_frames] or [bs, num_frames]
+    scheduler: the scheduler
+
+    Returns:
+        the predicted video with shape [B, C, H, W]
     """
-    # use higher precision for calculations
-    original_dtype = x0_pred.dtype
-    x0_pred, xt, sigmas, timesteps = map(
-        lambda x: x.double().to(x0_pred.device), [x0_pred, xt,
-                                                    scheduler.sigmas,
-                                                    scheduler.timesteps]
-    )
+    # If timestep is [bs, num_frames]
+    if timestep.ndim == 2:
+        timestep = timestep.flatten(0, 1)
+        assert timestep.numel() == noise_input_latent.shape[0]
+    elif timestep.ndim == 1:
+        # If timestep is [1]
+        if timestep.shape[0] == 1:
+            timestep = timestep.expand(noise_input_latent.shape[0])
+        else:
+            assert timestep.numel() == noise_input_latent.shape[0]
+    else:
+        raise ValueError(f"[pred_noise_to_pred_video] Invalid timestep shape: {timestep.shape}")
+    # timestep shape should be [B]
+    dtype = pred_noise.dtype
+    device = pred_noise.device
+    pred_noise = pred_noise.double().to(device)
+    noise_input_latent = noise_input_latent.double().to(device)
+    sigmas = scheduler.sigmas.double().to(device)
+    timesteps = scheduler.timesteps.double().to(device)
     timestep_id = torch.argmin(
         (timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
     sigma_t = sigmas[timestep_id].reshape(-1, 1, 1, 1)
-    flow_pred = (xt - x0_pred) / sigma_t
-    return flow_pred.to(original_dtype)
+    boundary_timestep_id = torch.argmin(
+        (timesteps.unsqueeze(0) - boundary_timestep.unsqueeze(1)).abs(), dim=1)
+    sigma_t_boundary = sigmas[boundary_timestep_id].reshape(-1, 1, 1, 1)
+    pred_video = noise_input_latent - (sigma_t - sigma_t_boundary) * pred_noise
+    return pred_video.to(dtype)

fastvideo/pipelines/basic/wan/wan_causal_dmd_pipeline.py

@@ -7,8 +7,6 @@ This module wires the causal DMD denoising stage into the modular pipeline.
 
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import (
-    FlowMatchEulerDiscreteScheduler)
 from fastvideo.pipelines import ComposedPipelineBase, LoRAPipeline
 
 # isort: off
@@ -51,6 +49,7 @@ class WanCausalDMDPipeline(LoRAPipeline, ComposedPipelineBase):
         self.add_stage(stage_name="denoising_stage",
                        stage=CausalDMDDenosingStage(
                            transformer=self.get_module("transformer"),
+                           transformer_2=self.get_module("transformer_2", None),
                            scheduler=self.get_module("scheduler")))
 
         self.add_stage(stage_name="decoding_stage",

fastvideo/pipelines/composed_pipeline_base.py

@@ -71,6 +71,7 @@ class ComposedPipelineBase(ABC):
 
         # Load modules directly in initialization
         logger.info("Loading pipeline modules...")
+        fastvideo_args.dit_cpu_offload = False
         self.modules = self.load_modules(fastvideo_args, loaded_modules)
 
     def set_trainable(self) -> None:
@@ -99,9 +100,30 @@ class ComposedPipelineBase(ABC):
 
         self.initialize_pipeline(self.fastvideo_args)
         if self.fastvideo_args.enable_torch_compile:
-            self.modules["transformer"] = torch.compile(
-                self.modules["transformer"])
-            logger.info("Torch Compile enabled for DiT")
+            transformer_module = self.modules["transformer"]
+            if self.fastvideo_args.training_mode:
+                logger.info(
+                    "Torch Compile enabled via FSDP loader for training; skipping additional pipeline compile"
+                )
+            else:
+                fsdp_module_cls = None
+                try:
+                    from torch.distributed.fsdp import FSDPModule  # type: ignore
+                    fsdp_module_cls = FSDPModule
+                except Exception:  # pragma: no cover - FSDP not always available
+                    fsdp_module_cls = None
+                if fsdp_module_cls is not None and isinstance(
+                        transformer_module, fsdp_module_cls):
+                    logger.info(
+                        "Transformer is already FSDP-wrapped; skipping torch.compile in pipeline"
+                    )
+                else:
+                    compile_kwargs = self.fastvideo_args.torch_compile_kwargs or {}
+                    logger.info("Enabling torch.compile for DiT with kwargs=%s",
+                                compile_kwargs)
+                    self.modules["transformer"] = torch.compile(
+                        transformer_module, **compile_kwargs)
+                    logger.info("Torch Compile enabled for DiT")
 
         if not self.fastvideo_args.training_mode:
             logger.info("Creating pipeline stages...")
@@ -121,25 +143,14 @@ class ComposedPipelineBase(ABC):
                         model_path: str,
                         device: str | None = None,
                         torch_dtype: torch.dtype | None = None,
-                        pipeline_config: PipelineConfig | None = None,
+                        pipeline_config: str | PipelineConfig | None = None,
                         args: argparse.Namespace | None = None,
                         required_config_modules: list[str] | None = None,
                         loaded_modules: dict[str, torch.nn.Module]
                         | None = None,
                         **kwargs) -> "ComposedPipelineBase":
         """
-        Load a pipeline from a pretrained model. 
-        Few different patterns are supported:
-        - Only provide model_path:
-            - This will load the pipeline in inference mode.
-            - The pipeline will be initialized with the default config.
-            - The pipeline will be initialized with the default modules.
-            - The pipeline will be initialized with the default stages.
-            - The pipeline will be initialized with the default stages.
-        - override the default config using pipeline_config or args or kwargs
-        - override the default modules using loaded_modules
-        - override the pipelineconfig
-
+        Load a pipeline from a pretrained model.
         loaded_modules: Optional[Dict[str, torch.nn.Module]] = None,
         If provided, loaded_modules will be used instead of loading from config/pretrained weights.
         """
@@ -147,30 +158,20 @@ class ComposedPipelineBase(ABC):
 
             kwargs['model_path'] = model_path
             fastvideo_args = FastVideoArgs.from_kwargs(**kwargs)
-            if pipeline_config is not None:
-                fastvideo_args.pipeline_config = pipeline_config
-            if fastvideo_args.override_transformer_cls_name is not None:
-                pipeline_config = PipelineConfig.from_pretrained("wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers")
-                fastvideo_args.pipeline_config = pipeline_config
         else:
             assert args is not None, "args must be provided for training mode"
             fastvideo_args = TrainingArgs.from_cli_args(args)
-            if fastvideo_args.override_transformer_cls_name is not None:
-                pipeline_config = PipelineConfig.from_pretrained("wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers")
-                fastvideo_args.pipeline_config = pipeline_config
-                logger.info("in 2 Overriding transformer cls name to %s", fastvideo_args.override_transformer_cls_name)
             # TODO(will): fix this so that its not so ugly
             fastvideo_args.model_path = model_path
             for key, value in kwargs.items():
                 setattr(fastvideo_args, key, value)
 
-            fastvideo_args.dit_cpu_offload = False
+            fastvideo_args.dit_cpu_offload = True
             # we hijack the precision to be the master weight type so that the
             # model is loaded with the correct precision. Subsequently we will
             # use FSDP2's MixedPrecisionPolicy to set the precision for the
             # fwd, bwd, and other operations' precision.
-            fastvideo_args.pipeline_config.dit_precision = 'fp32'
-            # assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'
+            assert fastvideo_args.pipeline_config.dit_precision == 'fp32', 'only fp32 is supported for training'
 
         logger.info("fastvideo_args in from_pretrained: %s", fastvideo_args)
 

fastvideo/pipelines/pipeline_batch_info.py

@@ -151,7 +151,7 @@ class ForwardBatch:
     output: torch.Tensor | None = None
     return_trajectory_latents: bool = False
     return_trajectory_decoded: bool = False
-    trajectory_timesteps: list[int] | None = None
+    trajectory_timesteps: list[torch.Tensor] | None = None
     trajectory_latents: torch.Tensor | None = None
     trajectory_decoded: list[torch.Tensor] | None = None
 
@@ -246,7 +246,7 @@ class TrainingBatch:
     fake_score_loss: float = 0.0
 
     dmd_latent_vis_dict: dict[str, Any] = field(default_factory=dict)
-    latent_vis_dict: dict[str, torch.Tensor] = field(default_factory=dict)
+    latent_vis_dict: dict[str, Any] = field(default_factory=dict)
     fake_score_latent_vis_dict: dict[str, Any] = field(default_factory=dict)
 
 

fastvideo/pipelines/preprocess/preprocess_pipeline_ode_trajectory.py

@@ -12,10 +12,8 @@ import os
 from collections.abc import Iterator
 from typing import Any
 
-import numpy as np
 import pyarrow as pa
 import torch
-from PIL import Image
 from torch.utils.data import DataLoader
 from torchdata.stateful_dataloader import StatefulDataLoader
 from tqdm import tqdm
@@ -42,11 +40,6 @@ from fastvideo.pipelines.stages import (DecodingStage, DenoisingStage,
                                         TimestepPreparationStage)
 from fastvideo.utils import save_decoded_latents_as_video, shallow_asdict
 
-from fastvideo.forward_context import set_forward_context
-from fastvideo.models.utils import pred_noise_to_pred_video, pred_video_to_pred_noise
-from fastvideo.models.loader.component_loader import TransformerLoader
-from fastvideo.distributed import get_local_torch_device
-
 logger = init_logger(__name__)
 
 
@@ -75,15 +68,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
         self.modules["scheduler"].set_timesteps(num_inference_steps=48,
                                                 denoising_strength=1.0)
 
-        fastvideo_args.model_loaded["transformer"] = False
-        loader = TransformerLoader()
-        fastvideo_args.pipeline_config.dit_precision = "fp32" # Overwrite the precision to fp32 for transformer
-        fastvideo_args.pipeline_config.dit_forward_precision = "fp32"
-        self.transformer = loader.load(
-            fastvideo_args.model_paths["transformer"], fastvideo_args)
-        self.add_module("transformer", self.transformer)
-        fastvideo_args.model_loaded["transformer"] = True
-
         self.add_stage(stage_name="input_validation_stage",
                        stage=InputValidationStage())
         self.add_stage(stage_name="prompt_encoding_stage",
@@ -112,7 +96,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
         """Preprocess text-only data and generate trajectory information."""
 
         for batch_idx, data in enumerate(self.pbar):
-            # logger.info("transformer weight sum: %s", sum(p.float().sum().item() for p in self.transformer.parameters()))
             if data is None:
                 continue
 
@@ -143,7 +126,6 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                     ]
 
                 batch_captions = valid_data["text"]
-                self.prompt_encoding_stage.text_encoders[0] = self.prompt_encoding_stage.text_encoders[0].to(dtype=torch.bfloat16).to(dtype=torch.float32)
                 # Encode text using the standalone TextEncodingStage API
                 prompt_embeds_list, prompt_masks_list = self.prompt_encoding_stage.encode_text(
                     batch_captions,
@@ -152,25 +134,22 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                     return_attention_mask=True,
                 )
                 prompt_embeds = prompt_embeds_list[0]
-                logger.info("prompt_embeds sum: %s, prompt_embeds shape: %s, prompt_embeds dtype: %s", prompt_embeds.float().sum(), prompt_embeds.shape, prompt_embeds.dtype)
                 prompt_attention_masks = prompt_masks_list[0]
                 assert prompt_embeds.shape[0] == prompt_attention_masks.shape[0]
 
                 sampling_params = SamplingParam.from_pretrained(args.model_path)
 
-                negative_prompt = '色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走'
-
                 # encode negative prompt for trajectory collection
                 if sampling_params.guidance_scale > 1 and sampling_params.negative_prompt is not None:
                     negative_prompt_embeds_list, negative_prompt_masks_list = self.prompt_encoding_stage.encode_text(
-                        negative_prompt,
+                        sampling_params.negative_prompt,
                         fastvideo_args,
                         encoder_index=[0],
                         return_attention_mask=True,
                     )
-                    negative_prompt_embed = negative_prompt_embeds_list[0]
+                    negative_prompt_embed = negative_prompt_embeds_list[0][0]
                     negative_prompt_attention_mask = negative_prompt_masks_list[
-                        0]
+                        0][0]
                 else:
                     negative_prompt_embed = None
                     negative_prompt_attention_mask = None
@@ -197,129 +176,30 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
                     batch.return_trajectory_latents = True
                     # Enabling this will save the decoded trajectory videos.
                     # Used for debugging.
-                    batch.return_trajectory_decoded = True
+                    batch.return_trajectory_decoded = False
                     batch.height = args.max_height
                     batch.width = args.max_width
                     batch.fps = args.train_fps
-                    batch.guidance_scale = 3.0
+                    batch.guidance_scale = 6.0
                     batch.do_classifier_free_guidance = True
 
                     result_batch = self.input_validation_stage(
                         batch, fastvideo_args)
                     result_batch = self.timestep_preparation_stage(
                         batch, fastvideo_args)
-                    # result_batch = self.latent_preparation_stage(
-                    #     result_batch, fastvideo_args)
-                    # result_batch = self.denoising_stage(result_batch,
-                    #                                     fastvideo_args)
-                    noisy_input = []
-                    # latents = result_batch.latents.permute(0, 2, 1, 3, 4)
-                    latents = torch.randn(
-                        [1, 21, 16, 60, 104], dtype=torch.float32, device=get_local_torch_device()
-                    )
-                    # logger.info("transformer weight sum: %s", sum(p.float().sum().item() for p in self.transformer.parameters()))
-                    logger.info("latents sum: %s, latents shape: %s, latents dtype: %s", latents.float().sum(), latents.shape, latents.dtype)
-
-                    logger.info("scheduler timesteps: %s", self.get_module("scheduler").timesteps)
-                    for progress_id, t in enumerate(tqdm(self.get_module("scheduler").timesteps)):
-                        timestep = t * \
-                            torch.ones([1, 21], device=latents.device, dtype=torch.float32)
-
-                        noisy_input.append(latents)
-
-                        with set_forward_context(
-                            current_timestep=0,
-                            attn_metadata=None,
-                            forward_batch=None,
-                        ):
-                            # logger.info("prompt_embed sum: %s, prompt_embed shape: %s, prompt_embed dtype: %s", prompt_embed.float().sum(), prompt_embed.shape, prompt_embed.dtype)
-                            # logger.info("timestep: %s", timestep[:, 0])
-                            # Run transformer
-                            cond_pred_noise_btchw = self.transformer(
-                                hidden_states=latents.permute(0, 2, 1, 3, 4),
-                                encoder_hidden_states=prompt_embed,
-                                timestep=timestep[:, 0]
-                            ).permute(0, 2, 1, 3, 4)
-
-                        # logger.info("cond_pred_noise_btchw sum: %s, cond_pred_noise_btchw shape: %s, cond_pred_noise_btchw dtype: %s", cond_pred_noise_btchw.float().sum(), cond_pred_noise_btchw.shape, cond_pred_noise_btchw.dtype)
-
-                        cond_pred_video_btchw = pred_noise_to_pred_video(
-                            pred_noise=cond_pred_noise_btchw.flatten(0, 1),
-                            noise_input_latent=latents.flatten(0, 1),
-                            timestep=timestep.flatten(0, 1),
-                            scheduler=self.get_module("scheduler")).unflatten(
-                                0, cond_pred_noise_btchw.shape[:2])
-
-                        # logger.info("cond_pred_video_btchw sum: %s, cond_pred_video_btchw shape: %s, cond_pred_video_btchw dtype: %s", cond_pred_video_btchw.float().sum(), cond_pred_video_btchw.shape, cond_pred_video_btchw.dtype)
-
-                        with set_forward_context(
-                            current_timestep=t,
-                            attn_metadata=None,
-                            forward_batch=result_batch,
-                        ):
-                            # logger.info("latents sum: %s, latents shape: %s, latents dtype: %s", latents.float().sum(), latents.shape, latents.dtype)
-                            # Run transformer
-                            uncond_pred_noise_btchw = self.transformer(
-                                latents.permute(0, 2, 1, 3, 4),
-                                negative_prompt_embed,
-                                timestep[:, 0]
-                            ).permute(0, 2, 1, 3, 4)
-
-                        # logger.info("uncond_pred_noise_btchw sum: %s, uncond_pred_noise_btchw shape: %s, uncond_pred_noise_btchw dtype: %s", uncond_pred_noise_btchw.float().sum(), uncond_pred_noise_btchw.shape, uncond_pred_noise_btchw.dtype)
-
-                        uncond_pred_video_btchw = pred_noise_to_pred_video(
-                            pred_noise=uncond_pred_noise_btchw.flatten(0, 1),
-                            noise_input_latent=latents.flatten(0, 1),
-                            timestep=timestep.flatten(0, 1),
-                            scheduler=self.get_module("scheduler")).unflatten(
-                                0, uncond_pred_noise_btchw.shape[:2])
-
-                        pred_video_btchw = uncond_pred_video_btchw + batch.guidance_scale * (
-                            cond_pred_video_btchw - uncond_pred_video_btchw
-                        )
-
-                        # logger.info("pred_video_btchw sum: %s, pred_video_btchw shape: %s, pred_video_btchw dtype: %s", pred_video_btchw.float().sum(), pred_video_btchw.shape, pred_video_btchw.dtype)
-
-                        pred_noise_btchw = pred_video_to_pred_noise(
-                            x0_pred=pred_video_btchw.flatten(0, 1),
-                            xt=latents.flatten(0, 1),
-                            timestep=timestep.flatten(0, 1),
-                            scheduler=self.get_module("scheduler")).unflatten(
-                                0, pred_video_btchw.shape[:2])
-
-                        # logger.info("pred_noise_btchw sum: %s, pred_noise_btchw shape: %s, pred_noise_btchw dtype: %s", pred_noise_btchw.float().sum(), pred_noise_btchw.shape, pred_noise_btchw.dtype)
-
-                        latents = self.get_module("scheduler").step(
-                            pred_noise_btchw.flatten(0, 1),
-                            self.get_module("scheduler").timesteps[progress_id] * torch.ones(
-                                [1, 21], device=latents.device, dtype=torch.long).flatten(0, 1),
-                            latents.flatten(0, 1)
-                        )[0].unflatten(dim=0, sizes=pred_noise_btchw.shape[:2])
-
-                        # logger.info("latents sum: %s, latents shape: %s, latents dtype: %s", latents.float().sum(), latents.shape, latents.dtype)
-
-                    noisy_input.append(latents)
-
-                    noisy_inputs = torch.stack(noisy_input, dim=1)
-
-                    noisy_inputs = noisy_inputs[:, [0, 12, 24, 36, -1]].half()
-
-                    logger.info("noisy inputs sum: %s, noisy inputs shape: %s, noisy inputs dtype: %s", noisy_inputs.float().sum(), noisy_inputs.shape, noisy_inputs.dtype)
-
-                    result_batch.trajectory_latents = noisy_inputs.permute(0, 1, 3, 2, 4, 5)
-                    result_batch.trajectory_timesteps = torch.tensor([self.get_module("scheduler").timesteps[i] for i in [0, 12, 24, 36, -1]])
-                    result_batch.latents = latents.permute(0, 2, 1, 3, 4)
+                    result_batch = self.latent_preparation_stage(
+                        result_batch, fastvideo_args)
+                    result_batch = self.denoising_stage(result_batch,
+                                                        fastvideo_args)
                     result_batch = self.decoding_stage(result_batch,
                                                        fastvideo_args)
+
                     trajectory_latents.append(
                         result_batch.trajectory_latents.cpu())
                     trajectory_timesteps.append(
                         result_batch.trajectory_timesteps.cpu())
                     trajectory_decoded.append(result_batch.trajectory_decoded)
 
-                trajectory_latents = torch.stack(trajectory_latents, dim=0).squeeze(0)
-                # trajecotry_latents = trajectory_latents[:, [0, 12, 24, 36, -1]]
-
                 # Prepare extra features for text-only processing
                 extra_features = {
                     "trajectory_latents": trajectory_latents,

fastvideo/pipelines/stages/causal_denoising.py

@@ -1,4 +1,5 @@
 import torch  # type: ignore
+import gc
 
 from fastvideo.distributed import get_local_torch_device
 from fastvideo.fastvideo_args import FastVideoArgs
@@ -34,13 +35,15 @@ class CausalDMDDenosingStage(DenoisingStage):
     Denoising stage for causal diffusion.
     """
 
-    def __init__(self, transformer, scheduler) -> None:
-        super().__init__(transformer, scheduler)
+    def __init__(self, transformer, scheduler, transformer_2=None) -> None:
+        super().__init__(transformer, scheduler, transformer_2)
         # KV and cross-attention cache state (initialized on first forward)
-        self.kv_cache1: list | None = None
-        self.crossattn_cache: list | None = None
+        self.transformer = transformer
+        self.transformer_2 = transformer_2
+        # self.kv_cache1: list | None = None
+        # self.crossattn_cache: list | None = None
         # Model-dependent constants (aligned with causal_inference.py assumptions)
-        self.num_transformer_blocks = self.transformer.config.arch_config.num_layers
+        self.num_transformer_blocks = len(self.transformer.blocks)
         self.num_frames_per_block = self.transformer.config.arch_config.num_frames_per_block
         self.sliding_window_num_frames = self.transformer.config.arch_config.sliding_window_num_frames
 
@@ -65,23 +68,18 @@ class CausalDMDDenosingStage(DenoisingStage):
             -1] * self.transformer.config.arch_config.patch_size[-2]
         self.frame_seq_length = latent_seq_length // patch_ratio
         # TODO(will): make this a parameter once we add i2v support
-        independent_first_frame = self.transformer.independent_first_frame
-
+        independent_first_frame = self.transformer.independent_first_frame if hasattr(
+            self.transformer, 'independent_first_frame') else False
         # Timesteps for DMD
         timesteps = torch.tensor(
             fastvideo_args.pipeline_config.dmd_denoising_steps,
             dtype=torch.long).cpu()
-
         if fastvideo_args.pipeline_config.warp_denoising_step:
-            logger.info("Warping timesteps...")
             scheduler_timesteps = torch.cat((self.scheduler.timesteps.cpu(),
                                              torch.tensor([0],
                                                           dtype=torch.float32)))
             timesteps = scheduler_timesteps[1000 - timesteps]
-        else:
-            assert False, "warp_denoising_step must be true"
         timesteps = timesteps.to(get_local_torch_device())
-        logger.info("Using timesteps: %s", timesteps)
 
         # Image kwargs (kept empty unless caller provides compatible args)
         image_kwargs: dict = {}
@@ -106,34 +104,43 @@ class CausalDMDDenosingStage(DenoisingStage):
         assert torch.isnan(prompt_embeds[0]).sum() == 0
 
         # Initialize or reset caches
-        if self.kv_cache1 is None:
-            self._initialize_kv_cache(batch_size=latents.shape[0],
-                                      dtype=target_dtype,
-                                      device=latents.device)
-            self._initialize_crossattn_cache(
-                batch_size=latents.shape[0],
-                max_text_len=fastvideo_args.pipeline_config.
-                text_encoder_configs[0].arch_config.text_len,
-                dtype=target_dtype,
-                device=latents.device)
-        else:
-            assert self.crossattn_cache is not None
-            # reset cross-attention cache
-            for block_index in range(self.num_transformer_blocks):
-                self.crossattn_cache[block_index][
-                    "is_init"] = False  # type: ignore
-            # reset kv cache pointers
-            for block_index in range(len(self.kv_cache1)):
-                self.kv_cache1[block_index][
-                    "global_end_index"] = torch.tensor(  # type: ignore
-                        [0],
-                        dtype=torch.long,
-                        device=latents.device)
-                self.kv_cache1[block_index][
-                    "local_end_index"] = torch.tensor(  # type: ignore
-                        [0],
-                        dtype=torch.long,
-                        device=latents.device)
+        kv_cache1 = self._initialize_kv_cache(batch_size=latents.shape[0],
+                                              dtype=target_dtype,
+                                              device=latents.device)
+        crossattn_cache = self._initialize_crossattn_cache(
+            batch_size=latents.shape[0],
+            max_text_len=fastvideo_args.pipeline_config.text_encoder_configs[0].
+            arch_config.text_len,
+            dtype=target_dtype,
+            device=latents.device)
+        # if self.kv_cache1 is None:
+        #     self._initialize_kv_cache(batch_size=latents.shape[0],
+        #                               dtype=target_dtype,
+        #                               device=latents.device)
+        #     self._initialize_crossattn_cache(
+        #         batch_size=latents.shape[0],
+        #         max_text_len=fastvideo_args.pipeline_config.
+        #         text_encoder_configs[0].arch_config.text_len,
+        #         dtype=target_dtype,
+        #         device=latents.device)
+        # else:
+        #     assert self.crossattn_cache is not None
+        #     # reset cross-attention cache
+        #     for block_index in range(self.num_transformer_blocks):
+        #         self.crossattn_cache[block_index][
+        #             "is_init"] = False  # type: ignore
+        #     # reset kv cache pointers
+        #     for block_index in range(len(self.kv_cache1)):
+        #         self.kv_cache1[block_index][
+        #             "global_end_index"] = torch.tensor(  # type: ignore
+        #                 [0],
+        #                 dtype=torch.long,
+        #                 device=latents.device)
+        #         self.kv_cache1[block_index][
+        #             "local_end_index"] = torch.tensor(  # type: ignore
+        #                 [0],
+        #                 dtype=torch.long,
+        #                 device=latents.device)
 
         # Optional: cache context features from provided image latents prior to generation
         current_start_frame = 0
@@ -156,8 +163,8 @@ class CausalDMDDenosingStage(DenoisingStage):
                         image_first_btchw,
                         prompt_embeds,
                         t_zero,
-                        kv_cache=self.kv_cache1,
-                        crossattn_cache=self.crossattn_cache,
+                        kv_cache=kv_cache1,
+                        crossattn_cache=crossattn_cache,
                         current_start=current_start_frame *
                         self.frame_seq_length,
                         **image_kwargs,
@@ -182,8 +189,8 @@ class CausalDMDDenosingStage(DenoisingStage):
                         ref_btchw,
                         prompt_embeds,
                         t_zero,
-                        kv_cache=self.kv_cache1,
-                        crossattn_cache=self.crossattn_cache,
+                        kv_cache=kv_cache1,
+                        crossattn_cache=crossattn_cache,
                         current_start=current_start_frame *
                         self.frame_seq_length,
                         **image_kwargs,
@@ -225,6 +232,10 @@ class CausalDMDDenosingStage(DenoisingStage):
                 video_raw_latent_shape = noise_latents_btchw.shape
 
                 for i, t_cur in enumerate(timesteps):
+                    if self.transformer_2 is not None and fastvideo_args.pipeline_config.dit_config.boundary_ratio is not None and t_cur < fastvideo_args.pipeline_config.dit_config.boundary_ratio * self.scheduler.num_train_timesteps:
+                        current_model = self.transformer_2
+                    else:
+                        current_model = self.transformer
                     # Copy for pred conversion
                     noise_latents = noise_latents_btchw.clone()
                     latent_model_input = current_latents.to(target_dtype)
@@ -275,12 +286,12 @@ class CausalDMDDenosingStage(DenoisingStage):
                             (latent_model_input.shape[0], 1),
                             device=latent_model_input.device,
                             dtype=torch.long)
-                        pred_noise_btchw = self.transformer(
+                        pred_noise_btchw = current_model(
                             latent_model_input,
                             prompt_embeds,
                             t_expanded_noise,
-                            kv_cache=self.kv_cache1,
-                            crossattn_cache=self.crossattn_cache,
+                            kv_cache=kv_cache1,
+                            crossattn_cache=crossattn_cache,
                             current_start=(pos_start_base + start_index) *
                             self.frame_seq_length,
                             start_frame=start_index,
@@ -340,12 +351,12 @@ class CausalDMDDenosingStage(DenoisingStage):
                                         attn_metadata=attn_metadata,
                                         forward_batch=batch):
                     t_expanded_context = t_context.unsqueeze(1)
-                    _ = self.transformer(
+                    _ = current_model(
                         context_bcthw,
                         prompt_embeds,
                         t_expanded_context,
-                        kv_cache=self.kv_cache1,
-                        crossattn_cache=self.crossattn_cache,
+                        kv_cache=kv_cache1,
+                        crossattn_cache=crossattn_cache,
                         current_start=(pos_start_base + start_index) *
                         self.frame_seq_length,
                         start_frame=start_index,
@@ -354,6 +365,11 @@ class CausalDMDDenosingStage(DenoisingStage):
                     )
                 start_index += current_num_frames
 
+        del kv_cache1
+        del crossattn_cache
+        gc.collect()
+        torch.cuda.empty_cache()
+
         batch.latents = latents
         return batch
 
@@ -391,7 +407,8 @@ class CausalDMDDenosingStage(DenoisingStage):
                 torch.tensor([0], dtype=torch.long, device=device),
             })
 
-        self.kv_cache1 = kv_cache1
+        # self.kv_cache1 = kv_cache1
+        return kv_cache1
 
     def _initialize_crossattn_cache(self, batch_size, max_text_len, dtype,
                                     device) -> None:
@@ -420,7 +437,8 @@ class CausalDMDDenosingStage(DenoisingStage):
                 "is_init":
                 False,
             })
-        self.crossattn_cache = crossattn_cache
+        # self.crossattn_cache = crossattn_cache
+        return crossattn_cache
 
     def verify_input(self, batch: ForwardBatch,
                      fastvideo_args: FastVideoArgs) -> VerificationResult:
@@ -444,4 +462,4 @@ class CausalDMDDenosingStage(DenoisingStage):
         result.add_check(
             "negative_prompt_embeds", batch.negative_prompt_embeds, lambda x:
             not batch.do_classifier_free_guidance or V.list_not_empty(x))
-        return result
+        return result

fastvideo/pipelines/stages/denoising.py

@@ -85,8 +85,9 @@ class DenoisingStage(PipelineStage):
                 AttentionBackendEnum.SLIDING_TILE_ATTN,
                 AttentionBackendEnum.VIDEO_SPARSE_ATTN,
                 AttentionBackendEnum.VMOBA_ATTN,
-                AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA
-            )  # hack
+                AttentionBackendEnum.FLASH_ATTN,
+                AttentionBackendEnum.TORCH_SDPA,
+                AttentionBackendEnum.SAGE_ATTN_THREE)  # hack
         )
 
     def forward(
@@ -140,12 +141,11 @@ class DenoisingStage(PipelineStage):
             latents = latents[:, :, rank_in_sp_group, :, :, :]
             batch.latents = latents
             if batch.image_latent is not None:
-                if not fastvideo_args.pipeline_config.ti2v_task and not fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    image_latent = rearrange(batch.image_latent,
-                                             "b c (n t) h w -> b c n t h w",
-                                             n=sp_world_size).contiguous()
-                    image_latent = image_latent[:, :, rank_in_sp_group, :, :, :]
-                    batch.image_latent = image_latent
+                image_latent = rearrange(batch.image_latent,
+                                         "b c (n t) h w -> b c n t h w",
+                                         n=sp_world_size).contiguous()
+                image_latent = image_latent[:, :, rank_in_sp_group, :, :, :]
+                batch.image_latent = image_latent
         # Get timesteps and calculate warmup steps
         timesteps = batch.timesteps
         # TODO(will): remove this once we add input/output validation for stages
@@ -254,7 +254,8 @@ class DenoisingStage(PipelineStage):
                                                              patch_size[2])
             seq_len = int(math.ceil(seq_len / sp_world_size)) * sp_world_size
 
-        trajectory_timesteps: list[int] = []
+        # Initialize lists for ODE trajectory
+        trajectory_timesteps: list[torch.Tensor] = []
         trajectory_latents: list[torch.Tensor] = []
 
         # Run denoising loop
@@ -284,27 +285,14 @@ class DenoisingStage(PipelineStage):
 
                 # Expand latents for I2V
                 latent_model_input = latents.to(target_dtype)
-                if batch.image_latent is not None and not fastvideo_args.pipeline_config.t2v_as_i2v_task:
+                if batch.image_latent is not None:
                     assert not fastvideo_args.pipeline_config.ti2v_task, "image latents should not be provided for TI2V task"
                     latent_model_input = torch.cat(
                         [latent_model_input, batch.image_latent],
                         dim=1).to(target_dtype)
-                elif batch.image_latent is not None and fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    assert batch.image_latent is not None, "image latents should be provided for T2V to I2V task"
-                    if rank_in_sp_group == 0:
-                        logger.info("latent_model_input.shape: %s",
-                                    latent_model_input.shape)
-                        latent_model_input = torch.cat([
-                            batch.image_latent,
-                            latent_model_input[:, :, 1:, :, :],
-                        ],
-                                                       dim=2).to(target_dtype)
-                        logger.info("latent_model_input.shape: %s",
-                                    latent_model_input.shape)
 
                 assert not torch.isnan(
                     latent_model_input).any(), "latent_model_input contains nan"
-
                 if fastvideo_args.pipeline_config.ti2v_task and batch.pil_image is not None:
                     timestep = torch.stack([t]).to(get_local_torch_device())
                     temp_ts = (mask2[0][0][:, ::2, ::2] * timestep).flatten()
@@ -319,13 +307,6 @@ class DenoisingStage(PipelineStage):
 
                 latent_model_input = self.scheduler.scale_model_input(
                     latent_model_input, t)
-                if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-                    if rank_in_sp_group == 0:
-                        latent_model_input = torch.cat([
-                            batch.image_latent,
-                            latent_model_input[:, :, 1:, :, :],
-                        ],
-                                                       dim=2).to(target_dtype)
 
                 # Prepare inputs for transformer
                 guidance_expand = (
@@ -488,17 +469,10 @@ class DenoisingStage(PipelineStage):
                 trajectory_tensor = sequence_model_parallel_all_gather(
                     trajectory_tensor, dim=3)
 
-        if trajectory_tensor is not None:
-            batch.trajectory_timesteps = torch.tensor(trajectory_timesteps).cpu()
+        if trajectory_tensor is not None and trajectory_timesteps_tensor is not None:
+            batch.trajectory_timesteps = trajectory_timesteps_tensor.cpu()
             batch.trajectory_latents = trajectory_tensor.cpu()
 
-        if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            latents = torch.cat([
-                batch.image_latent,
-                latents[:, :, 1:, :, :],
-            ],
-                                dim=2)
-
         # Update batch with final latents
         batch.latents = latents
 

fastvideo/pipelines/stages/image_encoding.py

@@ -105,81 +105,6 @@ class ImageVAEEncodingStage(PipelineStage):
     def __init__(self, vae: ParallelTiledVAE) -> None:
         self.vae: ParallelTiledVAE = vae
 
-    def encode_image(self,
-                     image: PIL.Image.Image,
-                     height: int,
-                     width: int,
-                     fastvideo_args: FastVideoArgs,
-                     generator: torch.Generator | None = None) -> torch.Tensor:
-        """
-        Encode image into latent space.
-        """
-        image = self.preprocess(
-            image,
-            vae_scale_factor=self.vae.spatial_compression_ratio,
-            height=height,
-            width=width).to(get_local_torch_device(), dtype=torch.float32)
-
-        # (B, C, H, W) -> (B, C, 1, H, W)
-        print(f"image.shape: {image.shape}")
-        image = image.unsqueeze(2)
-        print(f"after unsqueeze image.shape: {image.shape}")
-        return self.encode_tensor(image, fastvideo_args, generator)
-
-    def encode_tensor(self,
-                      video_condition: torch.Tensor,
-                      fastvideo_args: FastVideoArgs,
-                      generator: torch.Generator | None = None) -> torch.Tensor:
-        """
-        Encode frames into latent space.
-        """
-        self.vae = self.vae.to(get_local_torch_device())
-        video_condition = video_condition.to(device=get_local_torch_device(),
-                                             dtype=torch.float32)
-
-        # Setup VAE precision
-        vae_dtype = PRECISION_TO_TYPE[
-            fastvideo_args.pipeline_config.vae_precision]
-        vae_autocast_enabled = (
-            vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
-
-        # Encode Image
-        with torch.autocast(device_type="cuda",
-                            dtype=vae_dtype,
-                            enabled=vae_autocast_enabled):
-            if fastvideo_args.pipeline_config.vae_tiling:
-                self.vae.enable_tiling()
-            # if fastvideo_args.vae_sp:
-            #     self.vae.enable_parallel()
-            if not vae_autocast_enabled:
-                video_condition = video_condition.to(vae_dtype)
-            encoder_output = self.vae.encode(video_condition)
-
-        if fastvideo_args.mode == ExecutionMode.PREPROCESS:
-            latent_condition = encoder_output.mean
-        else:
-            generator = generator
-            if generator is None:
-                raise ValueError("Generator must be provided")
-            latent_condition = self.retrieve_latents(encoder_output, generator)
-
-        # Apply shifting if needed
-        if (hasattr(self.vae, "shift_factor")
-                and self.vae.shift_factor is not None):
-            if isinstance(self.vae.shift_factor, torch.Tensor):
-                latent_condition -= self.vae.shift_factor.to(
-                    latent_condition.device, latent_condition.dtype)
-            else:
-                latent_condition -= self.vae.shift_factor
-
-        if isinstance(self.vae.scaling_factor, torch.Tensor):
-            latent_condition = latent_condition * self.vae.scaling_factor.to(
-                latent_condition.device, latent_condition.dtype)
-        else:
-            latent_condition = latent_condition * self.vae.scaling_factor
-
-        return latent_condition
-
     def forward(
         self,
         batch: ForwardBatch,
@@ -232,28 +157,57 @@ class ImageVAEEncodingStage(PipelineStage):
         # (B, C, H, W) -> (B, C, 1, H, W)
         image = image.unsqueeze(2)
 
-        if fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            # repeat the image self.vae.temporal_compression_ratio times
-            video_condition = image.repeat(1, 1,
-                                           self.vae.temporal_compression_ratio,
-                                           1, 1)
-            # video_condition = image
-            logger.info("video_condition.shape: %s", video_condition.shape)
-        else:
-            video_condition = torch.cat([
-                image,
-                image.new_zeros(image.shape[0], image.shape[1], num_frames - 1,
-                                image.shape[3], image.shape[4])
-            ],
-                                        dim=2)
+        video_condition = torch.cat([
+            image,
+            image.new_zeros(image.shape[0], image.shape[1], num_frames - 1,
+                            image.shape[3], image.shape[4])
+        ],
+                                    dim=2)
+        video_condition = video_condition.to(device=get_local_torch_device(),
+                                             dtype=torch.float32)
 
-        latent_condition = self.encode_tensor(video_condition, fastvideo_args,
-                                              batch.generator)
+        # Setup VAE precision
+        vae_dtype = PRECISION_TO_TYPE[
+            fastvideo_args.pipeline_config.vae_precision]
+        vae_autocast_enabled = (
+            vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
+
+        # Encode Image
+        with torch.autocast(device_type="cuda",
+                            dtype=vae_dtype,
+                            enabled=vae_autocast_enabled):
+            if fastvideo_args.pipeline_config.vae_tiling:
+                self.vae.enable_tiling()
+            # if fastvideo_args.vae_sp:
+            #     self.vae.enable_parallel()
+            if not vae_autocast_enabled:
+                video_condition = video_condition.to(vae_dtype)
+            encoder_output = self.vae.encode(video_condition)
+
+        if fastvideo_args.mode == ExecutionMode.PREPROCESS:
+            latent_condition = encoder_output.mean
+        else:
+            generator = batch.generator
+            if generator is None:
+                raise ValueError("Generator must be provided")
+            latent_condition = self.retrieve_latents(encoder_output, generator)
+
+        # Apply shifting if needed
+        if (hasattr(self.vae, "shift_factor")
+                and self.vae.shift_factor is not None):
+            if isinstance(self.vae.shift_factor, torch.Tensor):
+                latent_condition -= self.vae.shift_factor.to(
+                    latent_condition.device, latent_condition.dtype)
+            else:
+                latent_condition -= self.vae.shift_factor
+
+        if isinstance(self.vae.scaling_factor, torch.Tensor):
+            latent_condition = latent_condition * self.vae.scaling_factor.to(
+                latent_condition.device, latent_condition.dtype)
+        else:
+            latent_condition = latent_condition * self.vae.scaling_factor
 
         if fastvideo_args.mode == ExecutionMode.PREPROCESS:
-            batch.image_latent = latent_condition
-        elif fastvideo_args.pipeline_config.t2v_as_i2v_task:
-            logger.info("latent_condition.shape: %s", latent_condition.shape)
             batch.image_latent = latent_condition
         else:
             mask_lat_size = torch.ones(1, 1, num_frames, latent_height,

fastvideo/pipelines/stages/input_validation.py

@@ -35,15 +35,9 @@ class InputValidationStage(PipelineStage):
         """Generate seeds for the inference"""
         seed = batch.seed
         num_videos_per_prompt = batch.num_videos_per_prompt
-        if isinstance(batch.prompt, list):
-            num_prompts = len(batch.prompt)
-        else:
-            num_prompts = 1
-
-        total_num_videos = num_prompts * num_videos_per_prompt
 
         assert seed is not None
-        seeds = [seed + i for i in range(total_num_videos)]
+        seeds = [seed + i for i in range(num_videos_per_prompt)]
         batch.seeds = seeds
         # Peiyuan: using GPU seed will cause A100 and H100 to generate different results...
         batch.generator = [

fastvideo/pipelines/stages/latent_preparation.py

@@ -3,7 +3,6 @@
 Latent preparation stage for diffusion pipelines.
 """
 
-import torch
 from diffusers.utils.torch_utils import randn_tensor
 
 from fastvideo.distributed import get_local_torch_device
@@ -94,7 +93,7 @@ class LatentPreparationStage(PipelineStage):
         # Generate or use provided latents
         if latents is None:
             latents = randn_tensor(shape,
-                                   generator=torch.Generator(device="cuda").manual_seed(1024),
+                                   generator=generator,
                                    device=device,
                                    dtype=dtype)
         else:

fastvideo/platforms/__init__.py

@@ -82,8 +82,8 @@ def rocm_platform_plugin() -> str | None:
                 logger.info("ROCm platform is available")
         finally:
             amdsmi.amdsmi_shut_down()
-    except Exception:
-        pass
+    except Exception as e:
+        logger.info("ROCm platform is unavailable: %s", e)
 
     return "fastvideo.platforms.rocm.RocmPlatform" if is_rocm else None
 

fastvideo/platforms/cuda.py

@@ -115,6 +115,7 @@ class CudaPlatformBase(Platform):
 
         logger.info("Trying FASTVIDEO_ATTENTION_BACKEND=%s",
                     envs.FASTVIDEO_ATTENTION_BACKEND)
+        logger.info("Selected backend: %s", selected_backend)
         if selected_backend == AttentionBackendEnum.SLIDING_TILE_ATTN:
             try:
                 from st_attn import sliding_tile_attention  # noqa: F401
@@ -144,6 +145,20 @@ class CudaPlatformBase(Platform):
                 logger.info(
                     "Sage Attention backend is not installed. Fall back to Flash Attention."
                 )
+        elif selected_backend == AttentionBackendEnum.SAGE_ATTN_THREE:
+            try:
+                from fastvideo.attention.backends.sageattn.api import sageattn_blackwell  # noqa: F401
+
+                from fastvideo.attention.backends.sage_attn3 import (  # noqa: F401
+                    SageAttention3Backend)
+                logger.info("Using Sage Attention 3 backend.")
+
+                return "fastvideo.attention.backends.sage_attn3.SageAttention3Backend"
+            except ImportError as e:
+                logger.info(e)
+                logger.info(
+                    "Sage Attention 3 backend is not installed. Fall back to Flash Attention."
+                )
         elif selected_backend == AttentionBackendEnum.VIDEO_SPARSE_ATTN:
             try:
                 from vsa import block_sparse_attn  # noqa: F401

fastvideo/platforms/interface.py

@@ -18,6 +18,7 @@ class AttentionBackendEnum(enum.Enum):
     SLIDING_TILE_ATTN = enum.auto()
     TORCH_SDPA = enum.auto()
     SAGE_ATTN = enum.auto()
+    SAGE_ATTN_THREE = enum.auto()
     VIDEO_SPARSE_ATTN = enum.auto()
     VMOBA_ATTN = enum.auto()
     NO_ATTENTION = enum.auto()

fastvideo/tests/encoders/test_t5_encoder.py

@@ -8,9 +8,6 @@ from torch.distributed.tensor import DTensor
 from torch.testing import assert_close
 from transformers import AutoConfig, AutoTokenizer, UMT5EncoderModel
 
-from fastvideo.wan.modules.tokenizers import HuggingfaceTokenizer
-from fastvideo.wan.modules.t5 import umt5_xxl
-
 from fastvideo.configs.pipelines import PipelineConfig
 from fastvideo.forward_context import set_forward_context
 from fastvideo.logger import init_logger
@@ -41,22 +38,9 @@ def test_t5_encoder():
     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
     precision_str = "fp32"
     precision = PRECISION_TO_TYPE[precision_str]
-    # model1 = UMT5EncoderModel.from_pretrained(TEXT_ENCODER_PATH).to(
-    #     precision).to(device).eval()
-    # tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_PATH)
-    model1 = umt5_xxl(
-        encoder_only=True,
-        return_tokenizer=False,
-        dtype=torch.float32,
-        device=device,
-    ).eval().requires_grad_(False)
-    model1.load_state_dict(
-        torch.load("/mnt/weka/home/hao.zhang/wei/Self-Forcing-clean/wan_models/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth",
-                    map_location='cpu', weights_only=False)
-    )
-
-    tokenizer1 = HuggingfaceTokenizer(
-        name="/mnt/weka/home/hao.zhang/wei/Self-Forcing-clean/wan_models/Wan2.1-T2V-1.3B/google/umt5-xxl/", seq_len=512, clean='whitespace')
+    model1 = UMT5EncoderModel.from_pretrained(TEXT_ENCODER_PATH).to(
+        precision).to(device).eval()
+    tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_PATH)
 
 
     args = FastVideoArgs(model_path=TEXT_ENCODER_PATH,
@@ -65,9 +49,8 @@ def test_t5_encoder():
                         pin_cpu_memory=False)
     loader = TextEncoderLoader()
     model2 = loader.load(TEXT_ENCODER_PATH, args)
-    model2 = model2.to(dtype=torch.bfloat16).to(precision)
+    model2 = model2.to(precision)
     model2.eval()
-    tokenizer2 = AutoTokenizer.from_pretrained(TOKENIZER_PATH)
 
     # Sanity check weights between the two models
     logger.info("Comparing model weights for sanity check...")
@@ -78,13 +61,8 @@ def test_t5_encoder():
     logger.info("Model1 has %s parameters", len(params1))
     logger.info("Model2 has %s parameters", len(params2))
 
-    model1_weight_sum = sum(p.float().sum().item() for p in model1.parameters())
-    model2_weight_sum = sum(p.float().sum().item() for p in model2.parameters())
-    logger.info("Model1 weight sum: %s", model1_weight_sum)
-    logger.info("Model2 weight sum: %s", model2_weight_sum)
-
-    # weight_diffs = []
-    # # check if embed_tokens are the same
+    weight_diffs = []
+    # check if embed_tokens are the same
     weights = ["encoder.block.{}.layer.0.layer_norm.weight", "encoder.block.{}.layer.0.SelfAttention.relative_attention_bias.weight", \
                "encoder.block.{}.layer.0.SelfAttention.o.weight", "encoder.block.{}.layer.1.DenseReluDense.wi_0.weight", "encoder.block.{}.layer.1.DenseReluDense.wi_1.weight",\
                 "encoder.block.{}.layer.1.DenseReluDense.wo.weight", \
@@ -92,21 +70,18 @@ def test_t5_encoder():
     
     for idx in range(hf_config.num_hidden_layers):
         for w in weights:
-            # name1 = w.format(idx)
+            name1 = w.format(idx)
             name2 = w.format(idx)
-            # p1 = params1[name1]
+            p1 = params1[name1]
             p2 = params2[name2]
-            p2 = (p2.to_local() if isinstance(p2, DTensor) else p2).to(device)
-            # assert_close(p1, p2, atol=1e-4, rtol=1e-4)
+            p2 = (p2.to_local() if isinstance(p2, DTensor) else p2).to(p1)
+            assert_close(p1, p2, atol=1e-4, rtol=1e-4)
     
 
     # Test with some sample prompts
-    # prompts = [
-    #     "Once upon a time", "The quick brown fox jumps over",
-    #     "In a galaxy far, far away"
-    # ]
     prompts = [
-        "A vibrant scene of Kenyan golfers at a lush green golf course on a sunny day. The golfers, dressed in casual yet stylish attire, are teeing off with animated expressions, showcasing their enthusiasm for the game. Rolling hills and pristine greens stretch out behind them, creating a picturesque backdrop. In the foreground, a golf buggy and a caddy stand ready, adding to the serene atmosphere. The camera captures the action from a mid-shot angle, focusing on the golfers' dynamic motions as they swing their clubs."
+        "Once upon a time", "The quick brown fox jumps over",
+        "In a galaxy far, far away"
     ]
 
     logger.info("Testing T5 encoder with sample prompts")
@@ -116,8 +91,7 @@ def test_t5_encoder():
             logger.info("Testing prompt: %s", prompt)
 
             # Tokenize the prompt
-            tokens1, mask = tokenizer1(prompt, return_mask=True, add_special_tokens=True)
-            tokens2 = tokenizer2(prompt,
+            tokens = tokenizer(prompt,
                                padding="max_length",
                                max_length=512,
                                truncation=True,
@@ -127,23 +101,22 @@ def test_t5_encoder():
             # filter out padding input_ids
             # tokens.input_ids = tokens.input_ids[tokens.attention_mask==1]
             # tokens.attention_mask = tokens.attention_mask[tokens.attention_mask==1]
-            outputs1 = model1(tokens1.to(device),
-                              mask.to(device))
+            outputs1 = model1(input_ids=tokens.input_ids,
+                              attention_mask=tokens.attention_mask,
+                              output_hidden_states=True).last_hidden_state
             print("--------------------------------")
             logger.info("Testing model2")
 
             # Get outputs from our implementation
             with set_forward_context(current_timestep=0, attn_metadata=None):
                 outputs2 = model2(
-                    input_ids=tokens2.input_ids,
-                    attention_mask=tokens2.attention_mask,
+                    input_ids=tokens.input_ids,
+                    attention_mask=tokens.attention_mask,
                 ).last_hidden_state
 
             # Compare last hidden states
-            last_hidden_state1 = outputs1[mask == 1]
-            last_hidden_state2 = outputs2[tokens2.attention_mask == 1]
-            logger.info("last_hidden_state1 sum: %s", last_hidden_state1.float().sum())
-            logger.info("last_hidden_state2 sum: %s", last_hidden_state2.float().sum())
+            last_hidden_state1 = outputs1[tokens.attention_mask == 1]
+            last_hidden_state2 = outputs2[tokens.attention_mask == 1]
 
             assert last_hidden_state1.shape == last_hidden_state2.shape, \
                 f"Hidden state shapes don't match: {last_hidden_state1.shape} vs {last_hidden_state2.shape}"

fastvideo/tests/modal/pr_test.py

@@ -62,7 +62,7 @@ def run_test(pytest_command: str):
     
     sys.exit(result.returncode)
 
-@app.function(gpu="L40S:1", image=image, timeout=900)
+@app.function(gpu="H100:1", image=image, timeout=900)
 def run_encoder_tests():
     run_test("pytest ./fastvideo/tests/encoders -vs")
 
@@ -118,6 +118,10 @@ def run_inference_lora_tests():
 def run_distill_dmd_tests():
     run_test("pytest ./fastvideo/tests/training/distill/test_distill_dmd.py -vs")
 
+@app.function(gpu="L40S:2", image=image, timeout=900, secrets=[modal.Secret.from_dict({"WANDB_API_KEY": os.environ.get("WANDB_API_KEY", "")})])
+def run_self_forcing_tests():
+    run_test("wandb login $WANDB_API_KEY && pytest ./fastvideo/tests/training/self-forcing/test_self_forcing.py -vs")
+
 @app.function(gpu="L40S:1", image=image, timeout=900)
 def run_unit_test():
     run_test("pytest ./fastvideo/tests/dataset/ ./fastvideo/tests/workflow/ -vs")

fastvideo/tests/nightly/test_e2e_dmd_t2v_crush_smol.py

@@ -0,0 +1,184 @@
+import os
+from pathlib import Path
+from huggingface_hub import snapshot_download
+import subprocess
+import sys
+from fastvideo.tests.ssim.test_inference_similarity import compute_video_ssim_torchvision
+import shutil
+
+# Import the training pipeline
+sys.path.append(str(Path(__file__).parent.parent.parent.parent.parent))
+
+NUM_NODES = "1"
+MODEL_PATH = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+
+# preprocessing
+DATA_DIR = "data"
+LOCAL_RAW_DATA_DIR = Path(os.path.join(DATA_DIR, "crush-smol"))
+NUM_GPUS_PER_NODE_PREPROCESSING = "1"
+PREPROCESSING_ENTRY_FILE_PATH = "fastvideo/pipelines/preprocess/v1_preprocess.py"
+
+LOCAL_PREPROCESSED_DATA_DIR = Path(os.path.join(DATA_DIR, "crush-smol_processed_t2v"))
+
+
+# training
+NUM_GPUS_PER_NODE_TRAINING = "4"
+TRAINING_ENTRY_FILE_PATH = "fastvideo/training/wan_distillation_pipeline.py"
+LOCAL_TRAINING_DATA_DIR = os.path.join(LOCAL_PREPROCESSED_DATA_DIR, "combined_parquet_dataset")
+LOCAL_VALIDATION_DATASET_FILE = "examples/training/finetune/Wan2.1-Fun-1.3B-InP/crush_smol/validation.json"
+LOCAL_OUTPUT_DIR = Path(os.path.join(DATA_DIR, "outputs"))
+
+def download_data():
+    # create the data dir if it doesn't exist
+    data_dir = Path(DATA_DIR)
+
+    print(f"Creating data directory at {data_dir}")
+    os.makedirs(data_dir, exist_ok=True)
+
+    print(f"Downloading raw dataset to {LOCAL_RAW_DATA_DIR}...")
+    try:
+        result = snapshot_download(
+            repo_id="wlsaidhi/crush-smol-merged",
+            local_dir=str(LOCAL_RAW_DATA_DIR),
+            repo_type="dataset",
+            resume_download=True,
+            token=os.environ.get("HF_TOKEN"),  # In case authentication is needed
+        )
+        print(f"Download completed successfully. Files downloaded to: {result}")
+        
+        # Verify the download
+        if not LOCAL_RAW_DATA_DIR.exists():
+            raise RuntimeError(f"Download appeared to succeed but {LOCAL_RAW_DATA_DIR} does not exist")
+            
+        # List downloaded files
+        print("Downloaded files:")
+        for file in LOCAL_RAW_DATA_DIR.rglob("*"):
+            if file.is_file():
+                print(f"  - {file.relative_to(LOCAL_RAW_DATA_DIR)}")
+                
+    except Exception as e:
+        print(f"Error during download: {str(e)}")
+        raise
+
+
+def run_preprocessing():
+    # remove the local_preprocessed_data_dir if it exists
+    if LOCAL_PREPROCESSED_DATA_DIR.exists():
+        print(f"Removing local_preprocessed_data_dir: {LOCAL_PREPROCESSED_DATA_DIR}")
+        shutil.rmtree(LOCAL_PREPROCESSED_DATA_DIR)
+
+    # Run torchrun command
+    cmd = [
+        "torchrun",
+        "--nnodes", NUM_NODES,
+        "--nproc_per_node", NUM_GPUS_PER_NODE_PREPROCESSING,
+        PREPROCESSING_ENTRY_FILE_PATH,
+        "--model_path", MODEL_PATH,
+        "--seed", "42",
+        "--data_merge_path", os.path.join(LOCAL_RAW_DATA_DIR, "merge.txt"),
+        "--preprocess_video_batch_size", "1",
+        "--max_height", "480",
+        "--max_width", "832",
+        "--num_frames", "81",
+        "--dataloader_num_workers", "0",
+        "--output_dir", LOCAL_PREPROCESSED_DATA_DIR,
+        "--train_fps", "16",
+        "--samples_per_file", "1",
+        "--flush_frequency", "1",
+        "--video_length_tolerance_range", "5",
+        "--preprocess_task", "t2v",
+    ]
+
+    process = subprocess.run(cmd, check=True)
+
+
+def run_training():
+    cmd = [
+        "torchrun",
+        "--nnodes", NUM_NODES,
+        "--nproc_per_node", NUM_GPUS_PER_NODE_TRAINING,
+        TRAINING_ENTRY_FILE_PATH,
+        "--model_path", MODEL_PATH,
+        "--inference_mode", "False",
+        "--pretrained_model_name_or_path", MODEL_PATH,
+        "--data_path", LOCAL_TRAINING_DATA_DIR,
+        "--validation_dataset_file", LOCAL_VALIDATION_DATASET_FILE,
+        "--train_batch_size", "1",
+        "--num_latent_t", "8",
+        "--num_gpus", NUM_GPUS_PER_NODE_TRAINING,
+        "--sp_size", "1",
+        "--tp_size", "1",
+        "--hsdp_replicate_dim", "1",
+        "--hsdp_shard_dim", NUM_GPUS_PER_NODE_TRAINING,
+        "--train_sp_batch_size", "1",
+        "--dataloader_num_workers", "10",
+        "--gradient_accumulation_steps", "1",
+        "--max_train_steps", "501",
+        "--learning_rate", "2e-6",
+        "--fake_score_learning_rate", "2e-6",
+        "--mixed_precision", "bf16",
+        "--training_state_checkpointing_steps", "1000",
+        "--weight_only_checkpointing_steps", "1000",
+        "--validation_steps", "50",
+        "--validation_sampling_steps", "3",
+        "--log_validation",
+        "--checkpoints_total_limit", "3",
+        "--ema_start_step", "0",
+        "--training_cfg_rate", "0.0",
+        "--output_dir", LOCAL_OUTPUT_DIR,
+        "--tracker_project_name", "ci_wan_t2v_dmd_overfit",
+        "--num_height", "480",
+        "--num_width", "832",
+        "--num_frames", "81",
+        "--flow_shift", "8",
+        "--validation_guidance_scale", "6.0",
+        "--weight_decay", "0.01",
+        "--generator_update_interval", "5",
+        "--dmd_denoising_steps", "1000,757,522",
+        "--min_timestep_ratio", "0.02",
+        "--max_timestep_ratio", "0.98",
+        "--seed", "1000",
+        "--real_score_guidance_scale", "3.5",
+        "--dit_precision", "fp32",
+        "--max_grad_norm", "1.0",
+        "--enable_gradient_checkpointing_type", "full",
+    ]
+
+    print(f"Running training with command: {cmd}")
+    process = subprocess.run(cmd, check=True)
+
+
+def test_e2e_overfit_single_sample():
+    os.environ["WANDB_MODE"] = "online"
+
+    download_data()
+    run_preprocessing()
+    run_training()
+
+    reference_video_file = os.path.join(os.path.dirname(__file__), "reference_video_1_sample_v0.mp4")
+    print(f"reference_video_file: {reference_video_file}")
+    final_validation_video_file = os.path.join(LOCAL_OUTPUT_DIR, "validation_step_900_inference_steps_50_video_0.mp4")
+    print(f"final_validation_video_file: {final_validation_video_file}")
+
+
+    # Ensure both files exist
+    assert os.path.exists(reference_video_file), f"Reference video not found at {reference_video_file}"
+    assert os.path.exists(final_validation_video_file), f"Validation video not found at {final_validation_video_file}"
+
+    # Compute SSIM
+    mean_ssim, min_ssim, max_ssim = compute_video_ssim_torchvision(
+        reference_video_file,
+        final_validation_video_file,
+        use_ms_ssim=True  # Using MS-SSIM for better quality assessment
+    )
+
+    print("\n===== SSIM Results for Step 900 Validation =====")
+    print(f"Mean MS-SSIM: {mean_ssim:.4f}")
+    print(f"Min MS-SSIM: {min_ssim:.4f}")
+    print(f"Max MS-SSIM: {max_ssim:.4f}")
+
+    assert max_ssim > 0.5, f"Max SSIM is below 0.5: {max_ssim}"
+
+
+if __name__ == "__main__":
+    test_e2e_overfit_single_sample()

fastvideo/tests/nightly/test_e2e_i2v_overfit_single_sample.py

@@ -111,7 +111,8 @@ def run_training():
         "--max_train_steps", "901",
         "--learning_rate", "1e-5",
         "--mixed_precision", "bf16",
-        "--checkpointing_steps", "6000",
+        "--weight_only_checkpointing_steps", "6000",
+        "--training_state_checkpointing_steps", "6000",
         "--validation_steps", "100",
         "--validation_sampling_steps", "50",
         "--log_validation",

fastvideo/tests/nightly/test_e2e_overfit_single_sample.py

@@ -62,6 +62,11 @@ def download_data():
 
 
 def run_preprocessing():
+    # remove the local_preprocessed_data_dir if it exists
+    if LOCAL_PREPROCESSED_DATA_DIR.exists():
+        print(f"Removing local_preprocessed_data_dir: {LOCAL_PREPROCESSED_DATA_DIR}")
+        shutil.rmtree(LOCAL_PREPROCESSED_DATA_DIR)
+
     # Run torchrun command
     cmd = [
         "torchrun",
@@ -111,7 +116,8 @@ def run_training():
         "--max_train_steps", "901",
         "--learning_rate", "1e-5",
         "--mixed_precision", "bf16",
-        "--checkpointing_steps", "6000",
+        "--weight_only_checkpointing_steps", "6000",
+        "--training_state_checkpointing_steps", "6000",
         "--validation_steps", "100",
         "--validation_sampling_steps", "50",
         "--log_validation",

fastvideo/tests/training/VSA/reference_wandb_summary_VSA.json

@@ -1 +1 @@
-{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":0.50390625,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.19960195198655128,"_runtime":107.325113071}
+{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":1.260593056678772,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.2620866410434246,"_runtime":107.325113071}