Add step8064

step8064/adapter_config.json

@@ -0,0 +1,45 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "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": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "ffn.2",
+    "q",
+    "k",
+    "v",
+    "o",
+    "ffn.0"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

step8064/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e0ba54ed8ab8f44648de986737de9518f5328abb3ccb680e0ef3c9b5ba2cf1c
+size 153453160

step8064/train_low.toml

@@ -0,0 +1,54 @@
+# Wan 2.2 I2V 14B - LOW NOISE LoRA training
+# Low noise model handles details, anatomy, fine textures
+
+output_dir = '/workspace/work/output_low'
+dataset = '/workspace/work/diffusion-pipe/configs/dataset_blink.toml'
+
+# training settings
+epochs = 1000
+max_steps = 1875
+micro_batch_size_per_gpu = 2
+pipeline_stages = 1
+gradient_accumulation_steps = 1
+gradient_clipping = 1.0
+warmup_steps = 100
+
+# eval settings
+eval_every_n_epochs = 1
+eval_before_first_step = true
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+
+# misc settings
+save_every_n_steps = 63
+checkpoint_every_n_minutes = 60
+activation_checkpointing = true
+compile = true
+partition_method = 'parameters'
+save_dtype = 'bfloat16'
+caching_batch_size = 1
+steps_per_print = 1
+video_clip_mode = 'single_beginning'
+
+[model]
+type = 'wan'
+ckpt_path = '/workspace/work/models/Wan2.2-I2V-A14B'
+transformer_path = '/workspace/work/models/comfyui/split_files/diffusion_models/wan2.2_i2v_low_noise_14B_fp16.safetensors'
+llm_path = '/workspace/work/models/comfyui/split_files/text_encoders/umt5_xxl_fp16.safetensors'
+dtype = 'bfloat16'
+transformer_dtype = 'float8'
+timestep_sample_method = 'logit_normal'
+# Full timestep range for low noise — helps with high->low switching at inference
+# (advice from PENISLORA author: train low on 0-1 instead of restricted range)
+
+[adapter]
+type = 'lora'
+rank = 16
+dtype = 'bfloat16'
+
+[optimizer]
+type = 'adamw_optimi'
+lr = 2e-5
+betas = [0.9, 0.99]
+weight_decay = 0.01
+eps = 1e-8