nodes/aigc-apps-VideoX-Fun
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge pull request #246 from aigc-apps/comfyui_json

Browse Source 
Update Readme and json
This commit is contained in:
Bubbliiiing
2025-07-04 15:52:53 +08:00
committed by GitHub
parent 21ec5bf326
commit 9702fc9db6
46 changed files with 827 additions and 110 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/cogvideox_fun/README_TRAIN.md
Regular → Executable
+9 -6
View File
@@ -27,8 +27,9 @@ CogVideoX-Fun without deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/cogvideox_fun/train.py \
@@ -70,8 +71,9 @@ CogVideoX-Fun with deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/cogvideox_fun/train.py \
@@ -113,8 +115,9 @@ CogVideoX-Fun with multi machines:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
NUM_PROCESS=$((WORLD_SIZE * 8))
scripts/cogvideox_fun/README_TRAIN_CONTROL.md
Regular → Executable
+9 -6
View File
@@ -46,8 +46,9 @@ CogVideoX-Fun without deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-V1.1-2b-Pose"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/cogvideox_fun/train_control.py \
@@ -87,8 +88,9 @@ CogVideoX-Fun with deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-V1.1-2b-Pose"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/cogvideox_fun/train.py \
@@ -129,8 +131,9 @@ CogVideoX-Fun with multi machines:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-V1.1-2b-Pose"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
NUM_PROCESS=$((WORLD_SIZE * 8))
scripts/cogvideox_fun/README_TRAIN_LORA.md
Regular → Executable
+9 -6
View File
@@ -26,8 +26,9 @@ CogVideoX-Fun without deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/cogvideox_fun/train_lora.py \
@@ -66,8 +67,9 @@ CogVideoX-Fun with deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/cogvideox_fun/train_lora.py \
@@ -107,8 +109,9 @@ CogVideoX-Fun with multi machines:
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
NUM_PROCESS=$((WORLD_SIZE * 8))
scripts/cogvideox_fun/README_TRAIN_REWARD.md
Regular → Executable
View File
scripts/cogvideox_fun/train.sh
Regular → Executable
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/cogvideox_fun/train.py \
scripts/cogvideox_fun/train_control.sh
Regular → Executable
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-V1.1-2b-Pose"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata_control.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/cogvideox_fun/train_control.py \
scripts/cogvideox_fun/train_lora.sh
Regular → Executable
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/CogVideoX-Fun-2b-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/cogvideox_fun/train_lora.py \
scripts/wan2.1/README_TRAIN.md
+58 -6
View File
@@ -29,8 +29,9 @@ Wan without DeepSpeed is more suitable for 1.3B Wan, as using it with 14B Wan ma
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1/train.py \
@@ -77,8 +78,9 @@ Wan with DeepSpeed Zero-2 is suitable for training 1.3B Wan and 14B Wan at low r
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1/train.py \
@@ -130,8 +132,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1/train.py \
@@ -169,4 +172,53 @@ accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag
--use_deepspeed \
--train_mode="normal" \
--trainable_modules "."
```
Wan T2V with FSDP:
Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows:
```sh
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=WanAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/wan2.1/train.py \
--config_path="config/wan2.1/wan_civitai.yaml" \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_data_dir=$DATASET_NAME \
--train_data_meta=$DATASET_META_NAME \
--image_sample_size=1024 \
--video_sample_size=256 \
--token_sample_size=512 \
--video_sample_stride=2 \
--video_sample_n_frames=81 \
--train_batch_size=1 \
--video_repeat=1 \
--gradient_accumulation_steps=1 \
--dataloader_num_workers=8 \
--num_train_epochs=100 \
--checkpointing_steps=50 \
--learning_rate=2e-05 \
--lr_scheduler="constant_with_warmup" \
--lr_warmup_steps=100 \
--seed=42 \
--output_dir="output_dir" \
--gradient_checkpointing \
--mixed_precision="bf16" \
--adam_weight_decay=3e-2 \
--adam_epsilon=1e-10 \
--vae_mini_batch=1 \
--max_grad_norm=0.05 \
--random_hw_adapt \
--training_with_video_token_length \
--enable_bucket \
--uniform_sampling \
--low_vram \
--use_deepspeed \
--train_mode="normal" \
--trainable_modules "."
```
scripts/wan2.1/README_TRAIN_LORA.md
+54 -6
View File
@@ -27,8 +27,9 @@ Wan without DeepSpeed is more suitable for 1.3B Wan, as using it with 14B Wan ma
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1/train_lora.py \
@@ -71,8 +72,9 @@ Wan with DeepSpeed Zero-2 is suitable for training 1.3B Wan and 14B Wan at low r
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1/train_lora.py \
@@ -120,8 +122,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1/train_lora.py \
@@ -155,4 +158,49 @@ accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag
--uniform_sampling \
--use_deepspeed \
--low_vram
```
Wan T2V with FSDP:
Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows:
```sh
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=WanAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/wan2.1/train_lora.py \
--config_path="config/wan2.1/wan_civitai.yaml" \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_data_dir=$DATASET_NAME \
--train_data_meta=$DATASET_META_NAME \
--image_sample_size=1024 \
--video_sample_size=256 \
--token_sample_size=512 \
--video_sample_stride=2 \
--video_sample_n_frames=81 \
--train_batch_size=1 \
--video_repeat=1 \
--gradient_accumulation_steps=1 \
--dataloader_num_workers=8 \
--num_train_epochs=100 \
--checkpointing_steps=50 \
--learning_rate=1e-04 \
--seed=42 \
--output_dir="output_dir" \
--gradient_checkpointing \
--mixed_precision="bf16" \
--adam_weight_decay=3e-2 \
--adam_epsilon=1e-10 \
--vae_mini_batch=1 \
--max_grad_norm=0.05 \
--random_hw_adapt \
--training_with_video_token_length \
--enable_bucket \
--uniform_sampling \
--use_deepspeed \
--low_vram
```
scripts/wan2.1/train.sh
Regular → Executable
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1/train.py \
scripts/wan2.1/train_lora.sh
Regular → Executable
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-T2V-14B"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1/train_lora.py \
scripts/wan2.1_fun/README_TRAIN.md
+58 -6
View File
@@ -27,8 +27,9 @@ Wan without DeepSpeed is more suitable for 1.3B Wan, as using it with 14B Wan ma
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train.py \
@@ -75,8 +76,9 @@ Wan with DeepSpeed Zero-2 is suitable for training 1.3B Wan and 14B Wan at low r
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train.py \
@@ -128,8 +130,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1/train.py \
@@ -167,4 +170,53 @@ accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag
--use_deepspeed \
--train_mode="inpaint" \
--trainable_modules "."
```
Wan T2V with FSDP:
Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows:
```sh
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=WanAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/wan2.1/train.py \
--config_path="config/wan2.1/wan_civitai.yaml" \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_data_dir=$DATASET_NAME \
--train_data_meta=$DATASET_META_NAME \
--image_sample_size=1024 \
--video_sample_size=256 \
--token_sample_size=512 \
--video_sample_stride=2 \
--video_sample_n_frames=81 \
--train_batch_size=1 \
--video_repeat=1 \
--gradient_accumulation_steps=1 \
--dataloader_num_workers=8 \
--num_train_epochs=100 \
--checkpointing_steps=50 \
--learning_rate=2e-05 \
--lr_scheduler="constant_with_warmup" \
--lr_warmup_steps=100 \
--seed=42 \
--output_dir="output_dir" \
--gradient_checkpointing \
--mixed_precision="bf16" \
--adam_weight_decay=3e-2 \
--adam_epsilon=1e-10 \
--vae_mini_batch=1 \
--max_grad_norm=0.05 \
--random_hw_adapt \
--training_with_video_token_length \
--enable_bucket \
--uniform_sampling \
--low_vram \
--use_deepspeed \
--train_mode="inpaint" \
--trainable_modules "."
```
scripts/wan2.1_fun/README_TRAIN_CONTROL.md
+69 -12
View File
@@ -62,8 +62,9 @@ Wan-Fun-Control-V1.1 without deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_control.py \
@@ -109,8 +110,9 @@ Wan-Fun-Control-V1.1 with deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control.py \
@@ -164,8 +166,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control.py \
@@ -207,6 +210,57 @@ accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag
--trainable_modules "."
```
Wan-Fun-Control-V1.1 with FSDP:
Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows:
```sh
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=WanAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/wan2.1_fun/train_control.py \
--config_path="config/wan2.1/wan_civitai.yaml" \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_data_dir=$DATASET_NAME \
--train_data_meta=$DATASET_META_NAME \
--image_sample_size=1024 \
--video_sample_size=256 \
--token_sample_size=512 \
--video_sample_stride=2 \
--video_sample_n_frames=81 \
--train_batch_size=1 \
--video_repeat=1 \
--gradient_accumulation_steps=1 \
--dataloader_num_workers=8 \
--num_train_epochs=100 \
--checkpointing_steps=50 \
--learning_rate=2e-05 \
--lr_scheduler="constant_with_warmup" \
--lr_warmup_steps=100 \
--seed=42 \
--output_dir="output_dir" \
--gradient_checkpointing \
--mixed_precision="bf16" \
--adam_weight_decay=3e-2 \
--adam_epsilon=1e-10 \
--vae_mini_batch=1 \
--max_grad_norm=0.05 \
--random_hw_adapt \
--training_with_video_token_length \
--enable_bucket \
--uniform_sampling \
--low_vram \
--use_deepspeed \
--train_mode="control_ref" \
--control_ref_image="random" \
--add_full_ref_image_in_self_attention \
--trainable_modules "."
```
<details>
<summary>(Obsolete) V1.0:</summary>
@@ -215,8 +269,9 @@ Wan-Fun-Control-V1.0 without deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_control.py \
@@ -261,8 +316,9 @@ Wan-Fun-Control-V1.0 with deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control.py \
@@ -315,8 +371,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control.py \
scripts/wan2.1_fun/README_TRAIN_CONTROL_LORA.md
+67 -12
View File
@@ -62,8 +62,9 @@ Wan-Fun-Control-V1.1 without deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_control_lora.py \
@@ -106,8 +107,9 @@ Wan-Fun-Control-V1.1 with deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control_lora.py \
@@ -158,8 +160,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control_lora.py \
@@ -199,6 +202,55 @@ accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag
--low_vram
```
Wan-Fun-Control-V1.1 with FSDP:
Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows:
```sh
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=WanAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/wan2.1_fun/train_control_lora.py \
--config_path="config/wan2.1/wan_civitai.yaml" \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_data_dir=$DATASET_NAME \
--train_data_meta=$DATASET_META_NAME \
--image_sample_size=1024 \
--video_sample_size=256 \
--token_sample_size=512 \
--video_sample_stride=2 \
--video_sample_n_frames=81 \
--train_batch_size=1 \
--video_repeat=1 \
--gradient_accumulation_steps=1 \
--dataloader_num_workers=8 \
--num_train_epochs=100 \
--checkpointing_steps=50 \
--learning_rate=1e-04 \
--seed=42 \
--output_dir="output_dir" \
--gradient_checkpointing \
--mixed_precision="bf16" \
--adam_weight_decay=3e-2 \
--adam_epsilon=1e-10 \
--vae_mini_batch=1 \
--max_grad_norm=0.05 \
--random_hw_adapt \
--training_with_video_token_length \
--enable_bucket \
--uniform_sampling \
--save_state \
--use_fsdp \
--train_mode="control_ref" \
--control_ref_image="random" \
--add_full_ref_image_in_self_attention \
--low_vram
```
<details>
<summary>(Obsolete) V1.0:</summary>
@@ -207,8 +259,9 @@ Wan-Fun-Control-V1.0 without deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_control_lora.py \
@@ -250,8 +303,9 @@ Wan-Fun-Control-V1.0 with deepspeed:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control_lora.py \
@@ -301,8 +355,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_control_lora.py \
scripts/wan2.1_fun/README_TRAIN_LORA.md
+56 -6
View File
@@ -27,8 +27,9 @@ Wan without DeepSpeed is more suitable for 1.3B Wan, as using it with 14B Wan ma
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_lora.py \
@@ -72,8 +73,9 @@ Wan with DeepSpeed Zero-2 is suitable for training 1.3B Wan and 14B Wan at low r
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_lora.py \
@@ -122,8 +124,9 @@ Training shell command is as follows:
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.1_fun/train_lora.py \
@@ -159,4 +162,51 @@ accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag
--use_deepspeed \
--train_mode="inpaint" \
--low_vram
```
Wan T2V with FSDP:
Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows:
```sh
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=WanAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/wan2.1_fun/train_lora.py \
--config_path="config/wan2.1/wan_civitai.yaml" \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_data_dir=$DATASET_NAME \
--train_data_meta=$DATASET_META_NAME \
--image_sample_size=1024 \
--video_sample_size=256 \
--token_sample_size=512 \
--video_sample_stride=2 \
--video_sample_n_frames=81 \
--train_batch_size=1 \
--video_repeat=1 \
--gradient_accumulation_steps=1 \
--dataloader_num_workers=8 \
--num_train_epochs=100 \
--checkpointing_steps=50 \
--learning_rate=1e-04 \
--seed=42 \
--output_dir="output_dir" \
--gradient_checkpointing \
--mixed_precision="bf16" \
--adam_weight_decay=3e-2 \
--adam_epsilon=1e-10 \
--vae_mini_batch=1 \
--max_grad_norm=0.05 \
--random_hw_adapt \
--training_with_video_token_length \
--enable_bucket \
--uniform_sampling \
--save_state \
--use_deepspeed \
--train_mode="inpaint" \
--low_vram
```
scripts/wan2.1_fun/train.sh
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train.py \
scripts/wan2.1_fun/train_control.sh
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_control.py \
scripts/wan2.1_fun/train_control_lora.sh
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-Control"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_control_lora.py \
scripts/wan2.1_fun/train_lora.sh
+3 -2
View File
@@ -1,8 +1,9 @@
export MODEL_NAME="models/Diffusion_Transformer/Wan2.1-Fun-V1.1-14B-InP"
export DATASET_NAME="datasets/internal_datasets/"
export DATASET_META_NAME="datasets/internal_datasets/metadata.json"
export NCCL_IB_DISABLE=1
export NCCL_P2P_DISABLE=1
# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA.
# export NCCL_IB_DISABLE=1
# export NCCL_P2P_DISABLE=1
NCCL_DEBUG=INFO
accelerate launch --mixed_precision="bf16" scripts/wan2.1_fun/train_lora.py \