Add training

fastvideo/training/training_pipeline.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+import gc
 import dataclasses
 import math
 import os
@@ -286,7 +287,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
     def _sample_timesteps(self, batch_size, device):
         # Determine which model to train based on the boundary timestep
         if (self.transformer_2 is not None and self.boundary_timestep is not None and
-                torch.rand(1, generator=self.noise_random_generator).item() > self.training_args.boundary_ratio):
+                torch.rand(1, generator=self.noise_random_generator).item() <= self.training_args.boundary_ratio):
             self.train_transformer_2 = True
         else:
             self.train_transformer_2 = False
@@ -308,9 +309,9 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
         boundary_ratio = self.training_args.boundary_ratio
         if self.train_transformer_2:
-            u = boundary_ratio + u * (1.0 - boundary_ratio)
+            u = (1 - boundary_ratio) + u * boundary_ratio # min: 1 - boundary_ratio, max: 1
         else:
-            u = u * boundary_ratio
+            u = u * (1 - boundary_ratio) # min: 0, max: 1 - boundary_ratio
 
         indices = (u * self.noise_scheduler.config.num_train_timesteps).long()
         return self.noise_scheduler.timesteps[indices].to(device=device)
@@ -490,6 +491,11 @@ class TrainingPipeline(LoRAPipeline, ABC):
         logger.info("Starting training with %s B trainable parameters",
                     round(num_trainable_params / 1e9, 3))
 
+        if getattr(self, "transformer_2", None) is not None:
+            num_trainable_params = _get_trainable_params(self.transformer_2)
+            logger.info("Transformer 2: Starting training with %s B trainable parameters",
+                        round(num_trainable_params / 1e9, 3))
+
         # Set random seeds for deterministic training
         self.noise_random_generator = torch.Generator(device="cpu").manual_seed(
             self.seed)
@@ -510,7 +516,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
         self._log_training_info()
 
-        self._log_validation(self.transformer, self.training_args,
+        self._log_validation(self.training_args,
                              self.init_steps)
 
         # Train!
@@ -572,7 +578,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 self.transformer.train()
                 self.sp_group.barrier()
             if self.training_args.log_validation and step % self.training_args.validation_steps == 0:
-                self._log_validation(self.transformer, self.training_args, step)
+                self._log_validation(self.training_args, step)
                 gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
                 trainable_params = round(
                     _get_trainable_params(self.transformer) / 1e9, 3)
@@ -653,12 +659,12 @@ class TrainingPipeline(LoRAPipeline, ABC):
         return batch
 
     @torch.no_grad()
-    def _log_validation(self, transformer, training_args, global_step) -> None:
+    def _log_validation(self, training_args, global_step) -> None:
         """
         Generate a validation video and log it to wandb to check the quality during training.
         """
         training_args.inference_mode = True
-        training_args.dit_cpu_offload = True
+        training_args.dit_cpu_offload = False
         if not training_args.log_validation:
             return
         if self.validation_pipeline is None:
@@ -680,7 +686,9 @@ class TrainingPipeline(LoRAPipeline, ABC):
                                            batch_size=None,
                                            num_workers=0)
 
-        transformer.eval()
+        self.transformer.eval()
+        if getattr(self, "transformer_2", None) is not None:
+            self.transformer_2.eval()
 
         validation_steps = training_args.validation_sampling_steps.split(",")
         validation_steps = [int(step) for step in validation_steps]
@@ -772,4 +780,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
         # Re-enable gradients for training
         training_args.inference_mode = False
-        transformer.train()
+        self.transformer.train()
+        if getattr(self, "transformer_2", None) is not None:
+            self.transformer_2.train()

fastvideo/training/wan_training_pipeline.py

@@ -42,6 +42,7 @@ class WanTrainingPipeline(TrainingPipeline):
             inference_mode=True,
             loaded_modules={
                 "transformer": self.get_module("transformer"),
+                "transformer_2": self.get_module("transformer_2")
             },
             tp_size=training_args.tp_size,
             sp_size=training_args.sp_size,

wan2.2_train.sh

@@ -0,0 +1,56 @@
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=offline
+export WANDB_API_KEY='50632ebd88ffd970521cec9ab4a1a2d7e85bfc45'
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+export TRITON_CACHE_DIR=/tmp/triton_cache
+# DATA_DIR=/mnt/sharefs/users/hao.zhang/Vchitect-2M/Wan-Syn-upload/latents_i2v/train/
+DATA_DIR=/mnt/weka/home/hao.zhang/wei/FastVideo/data/crush-smol_processed_t2v/combined_parquet_dataset
+# VALIDATION_DIR=/mnt/sharefs/users/hao.zhang/Vchitect-2M/mixkit/validation_8.json
+VALIDATION_DIR=/mnt/weka/home/hao.zhang/wei/FastVideo/data/crush-smol-single_processed_t2v/validation.json
+NUM_GPUS=8
+export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+# export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+CHECKPOINT_PATH="outputs_train_test/wan_finetune/checkpoint-10"
+# 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 $NUM_GPUS \
+    fastvideo/training/wan_training_pipeline.py \
+    --model_path Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+    --inference_mode False\
+    --pretrained_model_name_or_path Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+    --cache_dir "/home/ray/.cache" \
+    --data_path "$DATA_DIR" \
+    --validation_dataset_file  "$VALIDATION_DIR" \
+    --train_batch_size 1 \
+    --num_latent_t 16\
+    --sp_size 4 \
+    --tp_size 1 \
+    --num_gpus $NUM_GPUS \
+    --hsdp_replicate_dim 1 \
+    --hsdp-shard-dim 8 \
+    --train_sp_batch_size 1 \
+    --dataloader_num_workers 4 \
+    --gradient_accumulation_steps 1 \
+    --max_train_steps 30000 \
+    --learning_rate 1e-5 \
+    --mixed_precision "bf16" \
+    --checkpointing_steps 1000 \
+    --validation_steps 30 \
+    --validation_sampling_steps "40" \
+    --checkpoints_total_limit 3 \
+    --ema_start_step 0 \
+    --training_cfg_rate 0.1 \
+    --seed 1024 \
+    --output_dir "outputs_train_test/wan_finetune_v1" \
+    --tracker_project_name VSA_finetune \
+    --num_height 448 \
+    --num_width 832 \
+    --num_frames 61 \
+    --flow_shift 5 \
+    --validation_guidance_scale "5.0" \
+    --num_euler_timesteps 50 \
+    --master_weight_type "fp32" \
+    --dit_precision "fp32" \
+    --weight_decay 0.01 \
+    --max_grad_norm 1.0