Upload folder using huggingface_hub

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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "linear1",
+    "txt_mlp.fc2",
+    "txt_mlp.fc1",
+    "img_mod.linear",
+    "img_mlp.fc1",
+    "img_mlp.fc2",
+    "linear2",
+    "txt_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear"
+  ],
+  "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 sha256:9f79ce4a0b604a811410df7a6d0515d46f04a045d914191967aed9c7be1618d6
+size 322519480

epoch12/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "linear1",
+    "txt_mlp.fc2",
+    "txt_mlp.fc1",
+    "img_mod.linear",
+    "img_mlp.fc1",
+    "img_mlp.fc2",
+    "linear2",
+    "txt_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear"
+  ],
+  "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 sha256:674bacb3a2536f3de017d3429ab96a618519e89712e21258c132494f96593163
+size 322519480

epoch16/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "linear1",
+    "txt_mlp.fc2",
+    "txt_mlp.fc1",
+    "img_mod.linear",
+    "img_mlp.fc1",
+    "img_mlp.fc2",
+    "linear2",
+    "txt_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear"
+  ],
+  "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:ed53e4252c9df9559103c7d7dc5f5711edd09a68933b8496e0be70c135e6a857
+size 322519480

epoch20/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch24/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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "linear1",
+    "txt_mlp.fc2",
+    "txt_mlp.fc1",
+    "img_mod.linear",
+    "img_mlp.fc1",
+    "img_mlp.fc2",
+    "linear2",
+    "txt_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch24/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:396b9badd672e338ff1a9823f9ebdc8ad98375327828a3cce2199a5e228a91d4
+size 322519480

epoch24/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch28/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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "linear1",
+    "txt_mlp.fc2",
+    "txt_mlp.fc1",
+    "img_mod.linear",
+    "img_mlp.fc1",
+    "img_mlp.fc2",
+    "linear2",
+    "txt_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch28/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77f02941bad70469d5894e6533cb47e5a25234dbb4113e81160aca0b91623972
+size 322519480

epoch28/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch32/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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "modulation.linear",
+    "txt_mlp.fc1",
+    "linear2",
+    "img_mod.linear",
+    "txt_attn_proj",
+    "txt_mlp.fc2",
+    "linear1",
+    "txt_attn_qkv"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch32/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f000c862806221f764eb17d0bad9a88d64fc2a97a7f2e36bb5b38c63da701338
+size 322519480

epoch32/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch36/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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "modulation.linear",
+    "txt_mlp.fc1",
+    "linear2",
+    "img_mod.linear",
+    "txt_attn_proj",
+    "txt_mlp.fc2",
+    "linear1",
+    "txt_attn_qkv"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch36/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a10363795547eeed342390b1a52cafc5794e054c950214eb55afb13f6f08ae7b
+size 322519480

epoch36/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "linear1",
+    "txt_mlp.fc2",
+    "txt_mlp.fc1",
+    "img_mod.linear",
+    "img_mlp.fc1",
+    "img_mlp.fc2",
+    "linear2",
+    "txt_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear"
+  ],
+  "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 sha256:9683cb98d7dba81cb6a78fd62945afab2c52053c195725e6f6cc20d8031a30f5
+size 322519480

epoch4/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

epoch40/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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "img_mlp.fc2",
+    "img_mlp.fc1",
+    "modulation.linear",
+    "txt_mlp.fc1",
+    "linear2",
+    "img_mod.linear",
+    "txt_attn_proj",
+    "txt_mlp.fc2",
+    "linear1",
+    "txt_attn_qkv"
+  ],
+  "task_type": null,
+  "use_dora": false,
+  "use_rslora": false
+}

epoch40/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53ef2171c2995fa4ed9c3e5d6637456aa3dd1f2944e1a986466abd722771e641
+size 322519480

epoch40/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-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": [
+    "img_attn_proj",
+    "img_attn_qkv",
+    "txt_mod.linear",
+    "linear1",
+    "txt_mlp.fc2",
+    "txt_mlp.fc1",
+    "img_mod.linear",
+    "img_mlp.fc1",
+    "img_mlp.fc2",
+    "linear2",
+    "txt_attn_proj",
+    "txt_attn_qkv",
+    "modulation.linear"
+  ],
+  "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:ea055fd36cbe8673ebb25e8b6652b32521257636d4f658b51af0afe6228d8918
+size 322519480

epoch8/hunyuan_config.toml

@@ -0,0 +1,94 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/notebooks/diffusion-pipe/output'
+
+# Dataset config file.
+dataset = '/notebooks/diffusion-pipe/dataset_files/dataset_config.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 = 1
+# 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 = 4
+# 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 = 4
+# 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 = 30
+# 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 = 1
+# 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]
+type = 'hunyuan-video'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+# Or you can load it by pointing to all the ComfyUI files.
+transformer_path = '/notebooks/diffusion-pipe/hunyuan_model_files/hunyuan_dit.safetensors'
+vae_path = '/notebooks/diffusion-pipe/hunyuan_model_files/vae.safetensors'
+llm_path = '/notebooks/diffusion-pipe/hunyuan_model_files/llava-llama-3-8b-text-encoder-tokenizer/'
+clip_path = '/notebooks/diffusion-pipe/hunyuan_model_files/clip-vit-large-patch14/'
+# 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'
+
+[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 = 2e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8

events.out.tfevents.1738141044.nqlftbgczs.2674.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07f381f8faff6d339fcc3bb1d1f4aa2d4f4707da8aa8cae54de8449036022cab
+size 77781

events.out.tfevents.1738162540.nhuesj33qo.1708.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eba4a2fa089fe4ffe2fee7a0232d912d7c2bb7427c33474014cf541dc9f43347
+size 30675

global_step1051/layer_00-model_states.pt

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

global_step1051/layer_01-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c695daf41332c4486e3b25d8df2d04727ec81fe7bd05d4a8d57ffe97c697493
+size 9051114

global_step1051/layer_02-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65c62cf333480de91c555bc0a2a50209437435630c1431c108b3c006c0cc61d6
+size 9051114

global_step1051/layer_03-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b48cd8f9af53872b4fa338df7f44a4c09570263f2b7cd898a5996cb92c2cd9f
+size 9051114

global_step1051/layer_04-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2af99c4bfa9956c8a1f6af627e54c79d2e9878e3b7dec46303ed92387641fd98
+size 9051114

global_step1051/layer_05-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f06b6097c6d2855127af9d79ed49a1fed50831385810047351911d4662505c22
+size 9051114

global_step1051/layer_06-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a5772dfff67ac1f10b20de1ce44603d76460110d8a59dead0611795cd06950e
+size 9051114

global_step1051/layer_07-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c60dfcafbc9d7ba00874f81c8afe89f5961814e6de41d081d689d3297d46755
+size 9051114

global_step1051/layer_08-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77a48a2e81fdd333a5bfd486a104803e27a1b9d37b2cba03615bce13ee858b2c
+size 9051114

global_step1051/layer_09-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e22c3dd6d9db94110b5a94ead58710c93251897ce9932750b82e734fe1ee14a
+size 9051114

global_step1051/layer_10-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b1baf2ef85dbbb81e7ea08ba1292a94ca7da5658d466e25657699f014277a7c
+size 9051114

global_step1051/layer_11-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:602f66ce5a11694ebba492b6e56ebac247e1c615aa84eed9be0bdbdd812b609d
+size 9051114

global_step1051/layer_12-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:afcaf2fc8db137cbd8f3c0aad0fcd286f8b2f23711fa64bf4be814f796c97c09
+size 9051114

global_step1051/layer_13-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68356599409c944e45c74f517dd3af879f749af9a0f4f13f478a5739df198cf7
+size 9051114

global_step1051/layer_14-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7b554d994141e0582aff674b4f3b1ed09070f45bddc675304775b1bdf11f86a
+size 9051114

global_step1051/layer_15-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1209215dc0875ca5854ddd4bcfd0b55957169a3199632872fb8b025bd0770bcd
+size 9051114

global_step1051/layer_16-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2333e65160d7b12b823d78f69351f27e45a14a6490e34ba00536ffe22e0ca1b8
+size 9051114

global_step1051/layer_17-model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90d5f6c4fbb15e0a40e5b2df372c03a00b9b800dd6eb5c6f11245dd679de35f4
+size 9051114