Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+852/20250411_21-03-36/epoch1/.adapter_model.safetensors.UkAnlj filter=lfs diff=lfs merge=lfs -text
+epoch2/.adapter_model.safetensors.xd7wPd filter=lfs diff=lfs merge=lfs -text
+global_step26/.mp_rank_00_model_states.pt.xawzDB filter=lfs diff=lfs merge=lfs -text

852/20250411_21-03-36/epoch1/.adapter_model.safetensors.UkAnlj

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3374512ed4a2aab4ecb618826444b294ebe506807a8de17813267270fc315dbd
+size 78643200

852/20250411_21-03-36/epoch1/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k",
+    "ffn.0",
+    "o",
+    "v",
+    "ffn.2",
+    "q"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

852/20250411_21-03-36/events.out.tfevents.1744405901.ip-172-31-72-18.ec2.internal.688981.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae71113042cbe90e21b8e468f8ed264b08656067012910351934fc5e34b53ede
+size 2410

852/20250411_21-03-36/latest

@@ -0,0 +1 @@
+global_step45

852/20250411_21-03-36/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 3
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/640/20250411_15-20-06/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch1/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch1/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cf46892854e655cf605c49400b34350e8111bc5ff3c528b7d383d53bc71a7a2
+size 131309776

epoch1/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch10/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch10/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34c593b1f9b411f88f1225eb8c0f5d5aa92e6fec0ec26041cef30f07cb8cbebb
+size 131309776

epoch10/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch11/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch11/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cabdbef940fed88ca816517b7c8407089ddd4df137827dc2aa88c5490b0b207
+size 131309776

epoch11/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch12/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch12/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1836e1d99ce48ae7ab405d109d81d2e868fa61d507de216ddd7b5d252729639
+size 131309776

epoch12/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch13/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch13/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6207a69da9033073c04f18c78fb55145dba733876014c10cfcdde20b0e1eb7cf
+size 131309776

epoch13/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch14/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch14/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7764dd7fcc86abaf41899c9cb6df56da64484d6903ed8c179250637db7216f8
+size 131309776

epoch14/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch15/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch15/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:451f68156cc76949c7d2a70831b093d926635cbe8253ff1cb2a7382335a3bd7b
+size 131309776

epoch15/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch16/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch16/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08f75e871f448c45369de1963e36c708bbcc80911bd2882fac87fe1555d5b4b1
+size 131309776

epoch16/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch2/.adapter_model.safetensors.xd7wPd

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10c533ac4f4e367c1fe51f90aa2d12549c4081541346e8ccc278b082b95d55b0
+size 12845056

epoch2/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch2/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1579443bf437b9e4bbcf20632d92d8c8c0f04fa1c4c58b65f60d8524e4c7e4dc
+size 131309776

epoch2/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch3/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch3/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9b74d4d1ac0d6be9574a9705cebb1e7ca24f5bda8aacbb578f5a211676c6483
+size 131309776

epoch3/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch4/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch4/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0324be93b40dbc5c6bd7375d132167b631e4cc0caf572d7893d185c3adbef1b
+size 131309776

epoch4/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch5/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch5/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:448b2adefb9592047faa04100b8d87a78898ceaca39eba85346f68bef0dc8a70
+size 131309776

epoch5/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch6/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch6/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:589e70f7de5d53123a61feea4128834ff0f06e5ef855ec98bbbe7f4a3fae5de8
+size 131309776

epoch6/main_example.toml

@@ -0,0 +1,112 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/data/diffusion_pipe_training_runs/852'
+
+# Dataset config file.
+dataset = '/data/shiv/diffusion-pipe/examples/852_dataset.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 16
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 8
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 5
+
+# Block swapping is supported for Wan, HunyuanVideo, Flux, and Chroma. This value controls the number
+# of blocks kept offloaded to RAM. Increasing it lowers VRAM use, but has a performance penalty. The
+# exactly performance penalty depends on the model and the type of training you are doing (e.g. images vs video).
+# Block swapping only works for LoRA training, and requires pipeline_stages=1.
+blocks_to_swap = 24 
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = false
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+# If using block swap, you can disable it for eval. Eval uses less memory, so depending on block swapping amount you can maybe get away with
+# doing this, and then eval is much faster.
+disable_block_swap_for_eval = true
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 1
+# Can checkpoint the training state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 60
+# Always set to true unless you have a huge amount of VRAM.
+# This can also be 'unsloth' to reduce VRAM even more, with a slight performance hit.
+activation_checkpointing = true
+
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# Alternatively you can use 'manual' in combination with partition_split, which specifies the split points for dividing
+# layers between GPUs. For example, with two GPUs, partition_split=[10] puts layers 0-9 on GPU 0, and the rest on GPU 1.
+# With three GPUs, partition_split=[10, 20] puts layers 0-9 on GPU 0, layers 10-19 on GPU 1, and the rest on GPU 2.
+# Length of partition_split must be pipeline_stages-1.
+#partition_split = [N]
+
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# If experiencing CUDA OOM errors with memory fragmentation, set this env var: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 8 
+# How often deepspeed logs to console.
+steps_per_print = 100
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_beginning
+video_clip_mode = 'single_beginning'
+
+# This is how you configure HunyuanVideo. Other models will be different. See docs/supported_models.md for
+# details on the configuration and options for each model.
+[model]
+type = 'wan'
+ckpt_path = '/data/shiv/diffusion-pipe/models/wan1.3'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Wan supports fp8 for the transformer when training LoRA.
+# transformer_dtype = 'float8'
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+# For models that support full fine tuning, simply delete or comment out the [adapter] table to FFT.
+[adapter]
+type = 'lora'
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# LoRA rank - determines capacity to learn (higher = more capacity, more VRAM usage)
+rank = 48 
+
+# Higher rank = more capacity to learn but uses more VRAM and may overfit
+# The alpha value (scaling factor) is automatically set equal to rank in this codebase
+# You can initialize the lora weights from a previously trained lora.
+init_from_existing = '/data/diffusion_pipe_training_runs/852/20250411_21-03-36/epoch3'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 1e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01

epoch7/adapter_config.json

@@ -0,0 +1,38 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 48,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 48,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.0",
+    "v",
+    "q",
+    "o",
+    "k",
+    "ffn.2"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch7/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fa4f972166cc91d677a117f7c01634bddf286878bb5b393cf2730312b31308a
+size 131309776