Upload folder using huggingface_hub

epoch10/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch10/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b26d913f569d3c63995bfac613be1a6ba1e4702522b574a6a0948a84a5cb6db
+size 322519480

epoch10/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch12/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": 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:ef99d203f220b971927de7f35a92faef61f2bb2fcc39348439fdc50173310ace
+size 322519480

epoch12/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch14/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch14/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e50816eb1aeedf56409603fae068041b9de60dc646a8df257a1500ceb6644d9
+size 322519480

epoch14/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch16/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch16/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98485752533efdf979c98f7ab5ac0bcfcc6602d99371f1bc44ec4b2373a88844
+size 322519480

epoch16/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch18/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch18/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4414c1f6b4c2624ad69b74ce150f0185db84324d1ac5e7f60a5a7b0a4ed25dc8
+size 322519480

epoch18/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch2/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch2/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f48bddaa694882e0c7021e79b2bec040890a53ffd35ddb4a1479767b949f14b
+size 322519480

epoch2/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch20/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch20/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b787305efddd090a4f9877768fd04f60e08dd1f31a23eef4f2749fcc4b1b8b4
+size 322519480

epoch20/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch22/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch22/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a9d7962c914d71343aa491fcaf0a9e970bd47d97da5498c127945d8f8624b9f
+size 322519480

epoch22/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch4/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch4/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f0d5df7fc57c3bd8b4b8b73a1f738d6d9e86ff82918152e88aa5fde5995049e
+size 322519480

epoch4/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch6/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch6/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8cc5e1ed33955907134f267adb243a4bba935378bd15c764350f20f11790c76b
+size 322519480

epoch6/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch8/adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "txt_attn_proj",
+    "img_mod.linear",
+    "img_attn_qkv",
+    "img_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear",
+    "linear2",
+    "txt_mlp.fc1",
+    "txt_mlp.fc2",
+    "linear1"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch8/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7601a1323a60c43c627d3b7b2eab33e938bfb62a1c892313a1f89a004633764
+size 322519480

epoch8/hunyuan_video.toml

@@ -0,0 +1,109 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/diffusion-pipe/outputs'
+
+# Dataset config file.
+dataset = 'examples/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 = 22
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+# 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 = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# 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
+
+# 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 = 2
+# 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 = 120
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = true
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# 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 = 16
+# How often deepspeed logs to console.
+steps_per_print = 1
+# 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_middle
+video_clip_mode = 'single_middle'
+
+[model]
+# flux, ltx-video, or hunyuan-video
+type = 'hunyuan-video'
+transformer_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+vae_path = '/diffusion-pipe/models/hunyuan/hunyuan_video_vae_bf16.safetensors'
+llm_path = '/diffusion-pipe/models/llm'
+clip_path = '/diffusion-pipe/models/clip'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+# Hunyuan Video 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'
+
+# flux example
+# [model]
+# type = 'flux'
+# # Path to Huggingface Diffusers directory for Flux
+# diffusers_path = '/data2/imagegen_models/FLUX.1-dev'
+# # You can override the transformer from a BFL format checkpoint.
+# transformer_path = '/data2/imagegen_models/flux-dev-single-files/consolidated_s6700-schnell.safetensors'
+# dtype = 'bfloat16'
+# flux_shift = true
+
+# LTV-Video example
+# [model]
+# type = 'ltx-video'
+# diffusers_path = '/data2/imagegen_models/LTX-Video'
+# dtype = 'bfloat16'
+# timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 32
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[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 = 5e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

events.out.tfevents.1737020282.modal.3348.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7f1f4d0844c2e911fca3e4662a9f074af00289e8efb57687938282c7e314b13
+size 76609

global_step1027/layer_00-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c682c9cc8e731c37498845f3a980635b04094468e55e77553c7a12ccff998f5
+size 920

global_step1027/layer_01-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80399ca739b97c0eca95fbb87e44c5fcc8c49505076fbbfcf6498ab0436de61b
+size 9051114

global_step1027/layer_02-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f347631ea6f6ce794e6f22286f2af971bd27ebe60ce2da38050cd8171b2e32a0
+size 9051114

global_step1027/layer_03-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58b895a100fa34c755fbf8596fd19f47483620ea4df21e959e994bc694f33071
+size 9051114

global_step1027/layer_04-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:967e02c29c288a25bb9984a5aee9098791fa788750bba40e073adbd899966951
+size 9051114

global_step1027/layer_05-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b847886f074e4acfe50cbb328579b48d8e443982eaa277074461a35e32485ba
+size 9051114

global_step1027/layer_06-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6afb834c929f7ef67d945dfd77cdd931e13b793d9802320ebe33fa57f87d10c7
+size 9051114

global_step1027/layer_07-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:032712b5389ace4ca7dc99fc8f966775bd5bb873faf31521eae1f1c0d5652d55
+size 9051114

global_step1027/layer_08-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3411ef035dc92a30c3a8d1ea9ec3cfe8a87a0cea3d8866b24f97f9952c487661
+size 9051114

global_step1027/layer_09-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a5f76554415ac46ec06ab2211a32c459b842fe7d5360e8be1e54e63cf508e15
+size 9051114

global_step1027/layer_10-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a841549789ec903285d6290842d02ff3818faa37f281742a5dd56661179c3eb5
+size 9051114

global_step1027/layer_11-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fae9c030f68f2515b9eb0f362dc11abb135365c758d59a19629272ee4f83c86
+size 9051114

global_step1027/layer_12-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6490b678de916e4bf5880da964701e139d5930b43074901e6a340d6924752667
+size 9051114

global_step1027/layer_13-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09bfeba167db18ce306108b15ea8676cb8b50a847a1c286d83c9b77a85c882fc
+size 9051114

global_step1027/layer_14-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297a60e0d65f0d02051821a208c0663aa48445d5f32d50fbf10ed4c408c375bb
+size 9051114

global_step1027/layer_15-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1829c786458eb632d700b3f4d41ec532473ab74114efcdacac776b9402e9d6e8
+size 9051114