update lightx2v_train

lightx2v_train/configs/train/consistency/qwen_image_cm_cd_lora.yaml

@@ -0,0 +1,109 @@
+model:
+    name: qwen_image
+    pretrained_model_name_or_path: /root/yongyang/models/Qwen/Qwen-Image-2512
+    # Optional overrides let CD distill a different Qwen-Image checkpoint.
+    # teacher:
+    #     pretrained_model_name_or_path: /root/yongyang/models/Qwen/Qwen-Image-2512
+    max_sequence_length: 1024
+    running_dtype: bf16
+
+data:
+    train:
+        name: image_dataset
+        num_workers: 8
+        prompt_dropout_rate: 0.0
+        target_area: 1048576 # 1024 * 1024
+        shuffle: true
+        data_path:
+            - /root/yongyang/datasets/LightX2V_train_data_examples/dataset_v1/train.jsonl
+    val:
+        name: image_dataset
+        num_workers: 8
+        shuffle: false
+        data_path:
+            - /root/yongyang/datasets/LightX2V_train_data_examples/dataset_v1/val.jsonl
+
+scheduler:
+    num_train_timesteps: 1000
+    timestep_distribution: logitnormal
+    logitnormal_mean: 0.0
+    logitnormal_std: 1.0
+    min_t: 0.001
+    max_t: 1.0
+    time_shift_settings:
+        do_time_shift: true
+        shift_type: exponential
+        time_shift_power: 1.0
+        dynamic_shift: true
+        patch_size: [2, 2]
+        shift_x1: 256
+        shift_x2: 4096
+        shift_y1: 0.5
+        shift_y2: 1.15
+
+training:
+    method: consistency
+    train_type: lora
+    max_train_iters: 3000
+    gradient_accumulation_iters: 1
+    gradient_checkpointing: true
+    max_grad_norm: 1.0
+    lr_scheduler: constant
+    lr_warmup_iters: 10
+    save_every_iters: 100
+    save_total_limit: 10
+    consistency:
+        algorithm: cm
+        mode: cd
+        time_pair:
+            mapping: ect
+            q: 2.0
+            ratio_limit: 0.999
+            kimg_per_stage: 1.0
+            min_r: 0.0
+            safety_epsilon: 0.000001
+        loss:
+            distance: pseudo_huber
+            huber_constant: 0.00000001
+            weighting: inverse_delta
+            normalize_by_numel: false
+            computation_dtype: float32
+        teacher:
+            guidance_scale: 4.0
+            negative_prompt: " "
+            cfg_norm: none
+    lora:
+        rank: 16
+        alpha: 16
+        target_modules:
+            - to_k
+            - to_q
+            - to_v
+            - to_out.0
+    optimizer:
+        learning_rate: 0.0001
+        adam_beta1: 0.9
+        adam_beta2: 0.999
+        weight_decay: 0.01
+        adam_epsilon: 0.00000001
+    output_dir: ./output_train/qwen_image_cm_cd_lora
+
+inference:
+    method: image_infer
+    negative_prompt: " "
+    default_width: 1024
+    default_height: 1024
+    num_inference_steps: 4
+    enable_cfg: true
+    cfg_guidance_scale: 4.0
+    seed: 42
+    output_dir: ./output_infer/qwen_image_cm_cd_lora
+    infer_every_iters: ${training.save_every_iters}
+
+logging:
+    rank_zero_only: true
+    train_log_every_iters: 10
+    infer_log_every_steps: 10
+
+resume:
+    auto_resume: true

lightx2v_train/configs/train/consistency/qwen_image_cm_ct_lora.yaml

@@ -0,0 +1,104 @@
+model:
+    name: qwen_image
+    pretrained_model_name_or_path: /root/yongyang/models/Qwen/Qwen-Image-2512
+    max_sequence_length: 1024
+    running_dtype: bf16
+
+data:
+    train:
+        name: image_dataset
+        num_workers: 8
+        prompt_dropout_rate: 0.1
+        target_area: 1048576 # 1024 * 1024
+        shuffle: true
+        data_path:
+            - /root/yongyang/datasets/LightX2V_train_data_examples/dataset_v1/train.jsonl
+    val:
+        name: image_dataset
+        num_workers: 8
+        shuffle: false
+        data_path:
+            - /root/yongyang/datasets/LightX2V_train_data_examples/dataset_v1/val.jsonl
+
+scheduler:
+    num_train_timesteps: 1000
+    timestep_distribution: logitnormal
+    logitnormal_mean: 0.0
+    logitnormal_std: 1.0
+    min_t: 0.001
+    max_t: 1.0
+    time_shift_settings:
+        do_time_shift: true
+        shift_type: exponential
+        time_shift_power: 1.0
+        dynamic_shift: true
+        patch_size: [2, 2]
+        shift_x1: 256
+        shift_x2: 4096
+        shift_y1: 0.5
+        shift_y2: 1.15
+
+training:
+    method: consistency
+    train_type: lora
+    max_train_iters: 3000
+    gradient_accumulation_iters: 1
+    gradient_checkpointing: true
+    max_grad_norm: 1.0
+    lr_scheduler: constant
+    lr_warmup_iters: 10
+    save_every_iters: 100
+    save_total_limit: 10
+    consistency:
+        algorithm: cm
+        mode: ct
+        time_pair:
+            # FastGen/ECT sigmoid mapping. With global batch 1, one stage is
+            # 1000 optimizer steps; scale this with the actual global batch.
+            mapping: ect
+            q: 2.0
+            ratio_limit: 0.999
+            kimg_per_stage: 1.0
+            min_r: 0.0
+            safety_epsilon: 0.000001
+        loss:
+            distance: pseudo_huber
+            huber_constant: 0.00000001
+            weighting: inverse_delta
+            normalize_by_numel: false
+            computation_dtype: float32
+    lora:
+        rank: 16
+        alpha: 16
+        target_modules:
+            - to_k
+            - to_q
+            - to_v
+            - to_out.0
+    optimizer:
+        learning_rate: 0.0001
+        adam_beta1: 0.9
+        adam_beta2: 0.999
+        weight_decay: 0.01
+        adam_epsilon: 0.00000001
+    output_dir: ./output_train/qwen_image_cm_ct_lora
+
+inference:
+    method: image_infer
+    negative_prompt: " "
+    default_width: 1024
+    default_height: 1024
+    num_inference_steps: 4
+    enable_cfg: true
+    cfg_guidance_scale: 4.0
+    seed: 42
+    output_dir: ./output_infer/qwen_image_cm_ct_lora
+    infer_every_iters: ${training.save_every_iters}
+
+logging:
+    rank_zero_only: true
+    train_log_every_iters: 10
+    infer_log_every_steps: 10
+
+resume:
+    auto_resume: true

lightx2v_train/configs/train/consistency/qwen_image_mean_flow_ct_lora.yaml

@@ -0,0 +1,83 @@
+model:
+    name: qwen_image
+    pretrained_model_name_or_path: /root/yongyang/models/Qwen/Qwen-Image-2512
+    max_sequence_length: 1024
+    running_dtype: bf16
+
+data:
+    train:
+        name: image_dataset
+        num_workers: 8
+        prompt_dropout_rate: 0.1
+        target_area: 1048576
+        shuffle: true
+        data_path:
+            - /root/yongyang/datasets/LightX2V_train_data_examples/dataset_v1/train.jsonl
+
+scheduler:
+    num_train_timesteps: 1000
+    timestep_distribution: logitnormal
+    logitnormal_mean: -1.1
+    logitnormal_std: 2.0
+    min_t: 0.001
+    max_t: 1.0
+    time_shift_settings:
+        do_time_shift: true
+        shift_type: exponential
+        time_shift_power: 1.0
+        dynamic_shift: true
+        patch_size: [2, 2]
+        shift_x1: 256
+        shift_x2: 4096
+        shift_y1: 0.5
+        shift_y2: 1.15
+
+training:
+    method: consistency
+    train_type: lora
+    max_train_iters: 10000
+    gradient_accumulation_iters: 1
+    gradient_checkpointing: true
+    max_grad_norm: 1.0
+    lr_scheduler: constant
+    lr_warmup_iters: 10
+    save_every_iters: 100
+    save_total_limit: 10
+    consistency:
+        algorithm: mean_flow
+        mode: ct
+        sampling:
+            # The remaining samples use r=t and retain a flow-matching anchor.
+            random_endpoint_probability: 0.5
+        jvp:
+            method: finite_difference
+            epsilon: 0.0001
+        loss:
+            type: opt_grad
+            norm_method: poly_1.0
+            norm_constant: 0.1
+            tangent_warmup_steps: 0
+            spatially_normalized_tangent: false
+    lora:
+        rank: 16
+        alpha: 16
+        target_modules: [to_k, to_q, to_v, to_out.0]
+    optimizer:
+        learning_rate: 0.0001
+        adam_beta1: 0.9
+        adam_beta2: 0.999
+        weight_decay: 0.0
+        adam_epsilon: 0.00000001
+    output_dir: ./output_train/qwen_image_mean_flow_ct_lora
+
+inference:
+    # MeanFlow ODE sampling must pass r=t_next on every step; the standard
+    # Qwen pipeline cannot express that endpoint yet.
+    method: none
+
+logging:
+    rank_zero_only: true
+    train_log_every_iters: 10
+
+resume:
+    auto_resume: true