Add files using upload-large-folder tool

.gitattributes

@@ -58,3 +58,18 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_b32_1b/models/checkpoint/__5_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__7_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__1_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/.metadata filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__3_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__4_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__2_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__6_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__5_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__7_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__1_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/.metadata filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__3_0.distcp filter=lfs diff=lfs merge=lfs -text
+LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__0_0.distcp filter=lfs diff=lfs merge=lfs -text

LongVid/experiments/lm_aifm_anyflow_replay_smoke_1b/train_anyflow_replay_smoke_1b.yml

@@ -0,0 +1,80 @@
+# GENERATE TIME: Sun Aug  9 12:36:41 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_anyflow_replay_smoke_1b.yml
+
+name: lm_aifm_anyflow_replay_smoke_1b
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VTarDataset
+    meta_path: assets/data/meta/vidprom_dummy/raw.json
+    num_frames: 81
+    data_dtype: .mp4
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      shuffle_chunk_size: 159
+      drop_last: true
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 0.5
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ["attn1.to_q", "attn1.to_k", "attn1.to_v", "attn1.to_out.0", "ffn.net.0.proj", "ffn.net.2", "delta_embedder", "time_embedder"]
+    param_names_to_optimize: ["lora_A", "lora_B"]
+    bidirectional_modeling_prob: 0.0
+  text_encoder_cfg:
+    drop_text_ratio: 0.0
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+  vae_cfg:
+    load: true
+  scheduler_cfg:
+    shift: 5.0
+    weight_type: beta08
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path:
+  pretrain_network: null
+
+train:
+  train_pipeline: FAR_Wan_AnyFlow_Pretrain_Trainer
+  optim_g:
+    type: AdamW
+    lr: 5.0e-5
+    weight_decay: 0.0
+    betas: [0.9, 0.95]
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 0
+  total_iter: 0
+  max_grad_norm: 1.0
+
+val:
+  val_pipeline: FARWanAnyFlowPipeline
+  skip_first_eval: true
+  sample_freq: 1000000
+  eval_freq: 1000000
+
+logger:
+  print_freq: 1
+  save_checkpoint_freq: 1000000
+  save_eval_checkpoint_freq: 1000000
+  clean_cache_freq: 1
+  use_wandb: false

LongVid/experiments/lm_aifm_core_pilot_1b_identity_smoke/train_lm_aifm_core_pilot_1b.yml

@@ -0,0 +1,165 @@
+# GENERATE TIME: Sun Aug  9 12:57:02 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_core_pilot_1b.yml name=lm_aifm_core_pilot_1b_identity_smoke
+
+# LM-AIFM public-asset core integration pilot.
+#
+# This is intentionally a one-update smoke configuration.  It validates an
+# exact full-grid current-window cDMD loop with a same-content base-Wan score.
+# It is NOT a production random-edge or cross-window long-memory experiment.
+# For the bounded 20-update continuation override:
+#   train.total_iter=19 models.rollout_cfg.num_inference_steps_list='[2,4]'
+
+name: lm_aifm_core_pilot_1b_smoke
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VTarDataset
+    meta_path: assets/data/meta/vidprom_dummy/raw.json
+    num_frames: 81
+    data_dtype: .mp4
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      shuffle_chunk_size: 1000
+      drop_last: true
+
+models:
+  # Released, already trained AnyFlow-FAR student.  No unpublished
+  # experiments/.../step_6000.pt is referenced anywhere in this config.
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+    bidirectional_modeling_prob: 0.0
+
+  # Raw public base Wan, frozen.  It sees the same detached history content
+  # concatenated with the current noisy block in bidirectional score mode.
+  real_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+
+  # Independent base-Wan copy for the fake score.  LoRA and the schedule
+  # encoder are the only discriminator-side trainable parameters.
+  discriminator_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2']
+    param_names_to_optimize: ['lora_A', 'lora_B', 'schedule_encoder', 'schedule_projection']
+
+  schedule_encoder_cfg:
+    embedding_dim: 256
+    hidden_dim: 512
+    num_tokens: 1
+    num_heads: 4
+    num_fourier_bands: 8
+    dropout: 0.0
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.0
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+
+  vae_cfg:
+    load: true
+
+  scheduler_cfg:
+    shift: 5.0
+    weight_type: beta08
+
+  dmd_scheduler_cfg:
+    shift: 5.0
+
+  rollout_cfg:
+    rollout_pipeline: FARWanAnyFlowPipeline
+    # Smoke: N=2.  The 20-update continuation should use [2, 4].
+    num_inference_steps_list: [2]
+    current_num_chunks: 3
+    use_kv_cache: true
+
+  dmd_cfg:
+    dmd_weight: 0.1
+    dmd_batch_size: 1
+    min_timestep: 20
+    max_timestep: 980
+    normalizer_epsilon: !!float 1e-5
+    vector_clip: 10.0
+
+  interval_cfg:
+    # Disabled for the first smoke.  Set to 0.1 only after the base-teacher
+    # solver check passes; this remains an off-policy interval proxy.
+    weight: 0.0
+    teacher_substeps: 2
+    current_num_chunks: 3
+    delta_min: 0.05
+    replay_every: 0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path:
+  pretrain_network: null
+
+train:
+  train_pipeline: LM_AIFM_Core_Pilot_Trainer
+
+  optim_g:
+    type: AdamW
+    lr: !!float 2e-6
+    weight_decay: 0.0
+    betas: [0.0, 0.999]
+
+  optim_d:
+    type: AdamW
+    lr: !!float 2e-6
+    weight_decay: 0.0
+    betas: [0.0, 0.999]
+
+  discriminator_update_ratio: 1
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 0
+
+  # BaseTrainer uses `while global_step <= total_iter`; zero means one update.
+  total_iter: 0
+
+val:
+  val_pipeline: FARWanAnyFlowPipeline
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+  sample_cfg:
+    height: 480
+    width: 832
+    num_frames: 81
+    guidance_scale: 1.0
+    num_inference_steps: [2]
+    negative_prompt: 'overexposed, static, blurry, subtitles, low quality, deformed'
+
+logger:
+  print_freq: 1
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 1
+  use_wandb: false

LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/.metadata

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e56e41a95ed9e1950a110953ab75e41f8d6951c700934f0c5dc00301d430670b
+size 7989801

LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__0_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:524a8587bc185882b2f3ffd953bf84d5ecc9d279b485fff75c84d8f21670eb71
+size 1480078748

LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__1_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0da36ace448de27098c4151a438c977ab93af268a61c307b184ef3a1b9feb5ed
+size 1477914257

LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__3_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72c47b24564f16b2c2eaf3fadfb0c8b0e9b895d87df8beab5dbc475d46710b6b
+size 1477927497

LongVid/experiments/lm_aifm_core_pilot_1b_main/models/checkpoint_step20/__7_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:393743de58521b5af75603ece4c7706def1e08ceb1e0031b9ec5dc69f3cdc75e
+size 1477929703

LongVid/experiments/lm_aifm_core_pilot_1b_main/train_lm_aifm_core_pilot_1b.yml

@@ -0,0 +1,165 @@
+# GENERATE TIME: Sun Aug  9 13:14:59 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_core_pilot_1b.yml name=lm_aifm_core_pilot_1b_main train.total_iter=199 logger.save_checkpoint_freq=20 models.rollout_cfg.num_inference_steps_list=[2,4]
+
+# LM-AIFM public-asset core integration pilot.
+#
+# This is intentionally a one-update smoke configuration.  It validates an
+# exact full-grid current-window cDMD loop with a same-content base-Wan score.
+# It is NOT a production random-edge or cross-window long-memory experiment.
+# For the bounded 20-update continuation override:
+#   train.total_iter=19 models.rollout_cfg.num_inference_steps_list='[2,4]'
+
+name: lm_aifm_core_pilot_1b_smoke
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VTarDataset
+    meta_path: assets/data/meta/vidprom_dummy/raw.json
+    num_frames: 81
+    data_dtype: .mp4
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      shuffle_chunk_size: 1000
+      drop_last: true
+
+models:
+  # Released, already trained AnyFlow-FAR student.  No unpublished
+  # experiments/.../step_6000.pt is referenced anywhere in this config.
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+    bidirectional_modeling_prob: 0.0
+
+  # Raw public base Wan, frozen.  It sees the same detached history content
+  # concatenated with the current noisy block in bidirectional score mode.
+  real_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+
+  # Independent base-Wan copy for the fake score.  LoRA and the schedule
+  # encoder are the only discriminator-side trainable parameters.
+  discriminator_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2']
+    param_names_to_optimize: ['lora_A', 'lora_B', 'schedule_encoder', 'schedule_projection']
+
+  schedule_encoder_cfg:
+    embedding_dim: 256
+    hidden_dim: 512
+    num_tokens: 1
+    num_heads: 4
+    num_fourier_bands: 8
+    dropout: 0.0
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.0
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+
+  vae_cfg:
+    load: true
+
+  scheduler_cfg:
+    shift: 5.0
+    weight_type: beta08
+
+  dmd_scheduler_cfg:
+    shift: 5.0
+
+  rollout_cfg:
+    rollout_pipeline: FARWanAnyFlowPipeline
+    # Smoke: N=2.  The 20-update continuation should use [2, 4].
+    num_inference_steps_list: [2]
+    current_num_chunks: 3
+    use_kv_cache: true
+
+  dmd_cfg:
+    dmd_weight: 0.1
+    dmd_batch_size: 1
+    min_timestep: 20
+    max_timestep: 980
+    normalizer_epsilon: !!float 1e-5
+    vector_clip: 10.0
+
+  interval_cfg:
+    # Disabled for the first smoke.  Set to 0.1 only after the base-teacher
+    # solver check passes; this remains an off-policy interval proxy.
+    weight: 0.0
+    teacher_substeps: 2
+    current_num_chunks: 3
+    delta_min: 0.05
+    replay_every: 0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path:
+  pretrain_network: null
+
+train:
+  train_pipeline: LM_AIFM_Core_Pilot_Trainer
+
+  optim_g:
+    type: AdamW
+    lr: !!float 2e-6
+    weight_decay: 0.0
+    betas: [0.0, 0.999]
+
+  optim_d:
+    type: AdamW
+    lr: !!float 2e-6
+    weight_decay: 0.0
+    betas: [0.0, 0.999]
+
+  discriminator_update_ratio: 1
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 0
+
+  # BaseTrainer uses `while global_step <= total_iter`; zero means one update.
+  total_iter: 0
+
+val:
+  val_pipeline: FARWanAnyFlowPipeline
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+  sample_cfg:
+    height: 480
+    width: 832
+    num_frames: 81
+    guidance_scale: 1.0
+    num_inference_steps: [2]
+    negative_prompt: 'overexposed, static, blurry, subtitles, low quality, deformed'
+
+logger:
+  print_freq: 1
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 1
+  use_wandb: false

LongVid/experiments/lm_aifm_core_pilot_1b_smoke/train_lm_aifm_core_pilot_1b.yml

@@ -0,0 +1,165 @@
+# GENERATE TIME: Sun Aug  9 12:46:27 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_core_pilot_1b.yml
+
+# LM-AIFM public-asset core integration pilot.
+#
+# This is intentionally a one-update smoke configuration.  It validates an
+# exact full-grid current-window cDMD loop with a same-content base-Wan score.
+# It is NOT a production random-edge or cross-window long-memory experiment.
+# For the bounded 20-update continuation override:
+#   train.total_iter=19 models.rollout_cfg.num_inference_steps_list='[2,4]'
+
+name: lm_aifm_core_pilot_1b_smoke
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VTarDataset
+    meta_path: assets/data/meta/vidprom_dummy/raw.json
+    num_frames: 81
+    data_dtype: .mp4
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      shuffle_chunk_size: 1000
+      drop_last: true
+
+models:
+  # Released, already trained AnyFlow-FAR student.  No unpublished
+  # experiments/.../step_6000.pt is referenced anywhere in this config.
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+    bidirectional_modeling_prob: 0.0
+
+  # Raw public base Wan, frozen.  It sees the same detached history content
+  # concatenated with the current noisy block in bidirectional score mode.
+  real_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+
+  # Independent base-Wan copy for the fake score.  LoRA and the schedule
+  # encoder are the only discriminator-side trainable parameters.
+  discriminator_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2']
+    param_names_to_optimize: ['lora_A', 'lora_B', 'schedule_encoder', 'schedule_projection']
+
+  schedule_encoder_cfg:
+    embedding_dim: 256
+    hidden_dim: 512
+    num_tokens: 1
+    num_heads: 4
+    num_fourier_bands: 8
+    dropout: 0.0
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.0
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+
+  vae_cfg:
+    load: true
+
+  scheduler_cfg:
+    shift: 5.0
+    weight_type: beta08
+
+  dmd_scheduler_cfg:
+    shift: 5.0
+
+  rollout_cfg:
+    rollout_pipeline: FARWanAnyFlowPipeline
+    # Smoke: N=2.  The 20-update continuation should use [2, 4].
+    num_inference_steps_list: [2]
+    current_num_chunks: 3
+    use_kv_cache: true
+
+  dmd_cfg:
+    dmd_weight: 0.1
+    dmd_batch_size: 1
+    min_timestep: 20
+    max_timestep: 980
+    normalizer_epsilon: !!float 1e-5
+    vector_clip: 10.0
+
+  interval_cfg:
+    # Disabled for the first smoke.  Set to 0.1 only after the base-teacher
+    # solver check passes; this remains an off-policy interval proxy.
+    weight: 0.0
+    teacher_substeps: 2
+    current_num_chunks: 3
+    delta_min: 0.05
+    replay_every: 0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path:
+  pretrain_network: null
+
+train:
+  train_pipeline: LM_AIFM_Core_Pilot_Trainer
+
+  optim_g:
+    type: AdamW
+    lr: !!float 2e-6
+    weight_decay: 0.0
+    betas: [0.0, 0.999]
+
+  optim_d:
+    type: AdamW
+    lr: !!float 2e-6
+    weight_decay: 0.0
+    betas: [0.0, 0.999]
+
+  discriminator_update_ratio: 1
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 0
+
+  # BaseTrainer uses `while global_step <= total_iter`; zero means one update.
+  total_iter: 0
+
+val:
+  val_pipeline: FARWanAnyFlowPipeline
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+  sample_cfg:
+    height: 480
+    width: 832
+    num_frames: 81
+    guidance_scale: 1.0
+    num_inference_steps: [2]
+    negative_prompt: 'overexposed, static, blurry, subtitles, low quality, deformed'
+
+logger:
+  print_freq: 1
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 1
+  use_wandb: false

LongVid/experiments/lm_aifm_formal_gate_a_smoke_1b/train_lm_aifm_formal_gate_a_smoke_1b.yml

@@ -0,0 +1,164 @@
+# GENERATE TIME: Sun Aug  9 16:59:05 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_formal_gate_a_smoke_1b.yml
+
+# Production-path technical smoke only.
+#
+# This exercises cross-window MemSpec commits and detached-LoRA random-edge
+# suffix VJPs on the real 1.3B/FSDP stack. It deliberately disables the
+# interval loss and allows an unaudited public base score, so it cannot be used
+# as a scientific LM-AIFM result. Numerical all-edge/full-BPTT equality is
+# checked separately in FP64; this job covers the real BF16/FSDP/PEFT stack.
+
+name: lm_aifm_formal_gate_a_smoke_1b
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VTarDataset
+    meta_path: assets/data/meta/vidprom_dummy/raw.json
+    num_frames: 81
+    data_dtype: .mp4
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      shuffle_chunk_size: 1000
+      drop_last: true
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    # Hard requirement of detached-LoRA suffix mode. See random_edge_autograd.py.
+    enable_gradient_checkpoint: false
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+    bidirectional_modeling_prob: 0.0
+
+  # Technical Gate-A proxy only; the formal-F config below requires an audited
+  # same-observation long teacher and will fail closed without its gate JSON.
+  real_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+
+  discriminator_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/Wan2.1-T2V-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    # The detached-LoRA suffix restriction applies only to the student above.
+    # Fake-score checkpointing is safe and required for 21-frame bidirectional
+    # training to fit on an H200; without it the discriminator step reaches
+    # roughly 139.7 GiB/rank and OOMs before the first update completes.
+    enable_gradient_checkpoint: true
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2']
+    param_names_to_optimize: ['lora_A', 'lora_B', 'schedule_encoder', 'schedule_projection']
+
+  schedule_encoder_cfg:
+    embedding_dim: 256
+    hidden_dim: 512
+    num_tokens: 1
+    num_heads: 4
+    num_fourier_bands: 8
+    dropout: 0.0
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.0
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+  vae_cfg: {load: true}
+  scheduler_cfg: {shift: 5.0, weight_type: beta08}
+  dmd_scheduler_cfg: {shift: 5.0}
+
+  rollout_cfg:
+    rollout_pipeline: FARWanAnyFlowPipeline
+    num_inference_steps_list: [2]
+    current_num_chunks: 3
+    use_kv_cache: false
+
+  memory_cfg:
+    capacity: 5
+    sink_slots: 1
+    recent_slots: 2
+    context_slots: 5
+    target_chunks: 3
+    history_depths: [1]
+
+  schedule_cfg:
+    train_manifest: assets/lm_aifm/schedules_train_v1.json
+    heldout_manifest: assets/lm_aifm/schedules_test_v1.json
+    active_schedule_ids: [train_u_n2]
+
+  dmd_cfg:
+    dmd_weight: 0.1
+    dmd_batch_size: 1
+    min_timestep: 20
+    max_timestep: 980
+    normalizer_epsilon: !!float 1e-5
+    vector_clip: 10.0
+
+  interval_cfg:
+    enabled: false
+    weight: 0.0
+    gate_manifest: null
+    delta_min: 0.05
+    max_teacher_substeps: 8
+    mixture: {next_knot: 0.25, endpoint: 0.25, interior: 0.35, near_diagonal: 0.15}
+
+  replay_cfg: {every: 0, weight: 0.1}
+  formal_cfg:
+    require_teacher_gate: false
+    suffix_trainable_parameter_fragments: ['lora_A', 'lora_B']
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path: {pretrain_network: null}
+
+train:
+  train_pipeline: LM_AIFM_Formal_Trainer
+  optim_g: {type: AdamW, lr: !!float 2e-6, weight_decay: 0.0, betas: [0.0, 0.999]}
+  optim_d: {type: AdamW, lr: !!float 2e-6, weight_decay: 0.0, betas: [0.0, 0.999]}
+  discriminator_update_ratio: 1
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 0
+  # BaseTrainer uses <=. Two updates are intentional: update 1 makes the fake
+  # score non-identical to the public real score; with seed 0, update 2 then
+  # selects edge 0 and exercises a nonzero gradient through the suffix VJP.
+  total_iter: 1
+
+val:
+  val_pipeline: FARWanAnyFlowPipeline
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+  sample_cfg:
+    height: 480
+    width: 832
+    num_frames: 81
+    guidance_scale: 1.0
+    num_inference_steps: [2]
+    negative_prompt: 'overexposed, static, blurry, subtitles, low quality, deformed'
+
+logger:
+  print_freq: 1
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 1
+  use_wandb: false

LongVid/experiments/lm_aifm_same_observation_teacher_b32_smoke_1b/train_lm_aifm_same_observation_teacher_phase1_1b.yml

@@ -0,0 +1,104 @@
+# GENERATE TIME: Mon Aug 10 07:18:51 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_same_observation_teacher_phase1_1b.yml name=lm_aifm_same_observation_teacher_b32_smoke_1b datasets.train.dataloader_cfg.batch_size_per_gpu=4 train.total_iter=0 train.warmup_iter=0 logger.print_freq=1 logger.save_checkpoint_freq=1
+
+# Main Phase-1 LoRA training on real long-video prefixes. The resulting
+# checkpoint is only a teacher candidate. It must still pass the separate
+# real/student-prefix competence audit and interval-solver gate before F runs.
+
+name: lm_aifm_same_observation_teacher_phase1_1b
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VMemSpecWindowsTarDataset
+    meta_path: /mnt/workspace/fly/lm_aifm_20260809/datasets/formal_sekai_60s_v1/wids-meta.json
+    data_dtype: .mp4
+    temporal_sampling: memspec_sink_slow_recent_29f
+    window_num_frames: 29
+    history_windows: 6
+    target_fps: 16.0
+    temporal_seed: 0
+    decode_threads: 2
+    height: 480
+    width: 832
+    return_temporal_metadata: true
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      seed: 0
+      shuffle_chunk_size: 1000
+      drop_last: true
+      persistent_workers: false
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.1
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+  vae_cfg: {load: true}
+  scheduler_cfg: {shift: 5.0, weight_type: uniform}
+
+  memory_cfg:
+    capacity: 5
+    sink_slots: 1
+    recent_slots: 2
+    context_slots: 5
+    target_chunks: 3
+
+  teacher_cfg:
+    phase: real_prefix_phase1
+    # The dataset preselects sink/slow/recent roles around a fixed current.
+    # Only all-six has an unambiguous Phase-1 observation contract today.
+    history_depths: [6]
+    timestep_weighting: logit_normal
+    logit_mean: 0.0
+    logit_std: 1.0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path: {pretrain_network: null}
+
+train:
+  train_pipeline: LM_AIFM_SameObservation_Teacher_Trainer
+  optim_g: {type: AdamW, lr: !!float 1e-5, weight_decay: 0.0, betas: [0.9, 0.999]}
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 500
+  total_iter: 9999
+  ema_decay: 0.999
+  ema_warmup_step: 500
+
+val:
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+
+logger:
+  print_freq: 10
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 10
+  use_wandb: false

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/.metadata

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a15ff71b41187c432a3221b16065a282eb25747ff139f948ca42d04a1509479
+size 4477919

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__0_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:611e9da8dc7e6ca13b8297ee88afe32760452536b71a401951f656b7ff520251
+size 747963312

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__1_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19d97eef69ceb51601e1940b1fd74ba7080f7ac7a694509ce547ce1d02faa0a0
+size 746883265

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__2_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff5c2279b9f75d3c5b16297b32936cc5027ede8764cd3fe4e823c3a9f77e5f43
+size 746912940

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__3_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd13e91680eacc506c8481605f87e6903865deef00cc408241e79691a7ea2b26
+size 746887427

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__4_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f78e0f9f991f472e83f5a2397cdbfc534ab651179366131cb6dbc58cb7e54d41
+size 746887491

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__5_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fc145378708dca0badabdb500c9cadb42854c7dc60c3c232b2ff4a54e04dcd7
+size 746887427

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__6_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d33234b8a73707f9f98f2372cf8b1c13e2438c2e3a1eec31263d8b0998c856d7
+size 746890894

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/checkpoint/__7_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3af84ae7a5fc326fb747aad3bdcc49092d1e078eb69eb0d49ab4596eb00fa7fe
+size 746890894

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/models/step_1000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:687df0fdd4c2ef4b95bb1301beec706ad5a2bbcd0fea745600138412216a84eb
+size 11507032957

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_1b/train_lm_aifm_same_observation_teacher_phase1_1b.yml

@@ -0,0 +1,104 @@
+# GENERATE TIME: Sun Aug  9 17:12:28 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_same_observation_teacher_phase1_1b.yml models.teacher_cfg.history_depths=[6] logger.print_freq=1 logger.save_checkpoint_freq=100
+
+# Main Phase-1 LoRA training on real long-video prefixes. The resulting
+# checkpoint is only a teacher candidate. It must still pass the separate
+# real/student-prefix competence audit and interval-solver gate before F runs.
+
+name: lm_aifm_same_observation_teacher_phase1_1b
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VMemSpecWindowsTarDataset
+    meta_path: /mnt/workspace/fly/lm_aifm_20260809/datasets/formal_sekai_60s_v1/wids-meta.json
+    data_dtype: .mp4
+    temporal_sampling: memspec_sink_slow_recent_29f
+    window_num_frames: 29
+    history_windows: 6
+    target_fps: 16.0
+    temporal_seed: 0
+    decode_threads: 2
+    height: 480
+    width: 832
+    return_temporal_metadata: true
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      seed: 0
+      shuffle_chunk_size: 1000
+      drop_last: true
+      persistent_workers: false
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.1
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+  vae_cfg: {load: true}
+  scheduler_cfg: {shift: 5.0, weight_type: uniform}
+
+  memory_cfg:
+    capacity: 5
+    sink_slots: 1
+    recent_slots: 2
+    context_slots: 5
+    target_chunks: 3
+
+  teacher_cfg:
+    phase: real_prefix_phase1
+    # The dataset preselects sink/slow/recent roles around a fixed current.
+    # Only all-six has an unambiguous Phase-1 observation contract today.
+    history_depths: [6]
+    timestep_weighting: logit_normal
+    logit_mean: 0.0
+    logit_std: 1.0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path: {pretrain_network: null}
+
+train:
+  train_pipeline: LM_AIFM_SameObservation_Teacher_Trainer
+  optim_g: {type: AdamW, lr: !!float 1e-5, weight_decay: 0.0, betas: [0.9, 0.999]}
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 500
+  total_iter: 9999
+  ema_decay: 0.999
+  ema_warmup_step: 500
+
+val:
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+
+logger:
+  print_freq: 10
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 10
+  use_wandb: false

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_b32_1b/models/checkpoint/__5_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a71805aacae636e0e8597b66b5afb25ffc5086b163f6ecf8fc086a48b6c428d
+size 746887427

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_b32_1b/train_lm_aifm_same_observation_teacher_phase1_b32_1b.yml

@@ -0,0 +1,106 @@
+# GENERATE TIME: Mon Aug 10 07:25:37 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_same_observation_teacher_phase1_b32_1b.yml
+
+# Phase-1 LoRA training on real long-video prefixes with the paper-comparable
+# global batch used by the main experiment: 4 samples/rank x 8 ranks = 32.
+# The resulting checkpoint is a teacher candidate and must still pass the
+# real/student-prefix competence audit and interval-solver gate before F runs.
+
+name: lm_aifm_same_observation_teacher_phase1_b32_1b
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VMemSpecWindowsTarDataset
+    meta_path: /mnt/workspace/fly/lm_aifm_20260809/datasets/formal_sekai_60s_v1/wids-meta.json
+    data_dtype: .mp4
+    temporal_sampling: memspec_sink_slow_recent_29f
+    window_num_frames: 29
+    history_windows: 6
+    target_fps: 16.0
+    temporal_seed: 0
+    decode_threads: 2
+    height: 480
+    width: 832
+    return_temporal_metadata: true
+    dataloader_cfg:
+      batch_size_per_gpu: 4
+      num_workers: 0
+      shuffle: true
+      seed: 0
+      shuffle_chunk_size: 1000
+      drop_last: true
+      persistent_workers: false
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.1
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+  vae_cfg: {load: true}
+  scheduler_cfg: {shift: 5.0, weight_type: uniform}
+
+  memory_cfg:
+    capacity: 5
+    sink_slots: 1
+    recent_slots: 2
+    context_slots: 5
+    target_chunks: 3
+
+  teacher_cfg:
+    phase: real_prefix_phase1
+    history_depths: [6]
+    timestep_weighting: logit_normal
+    logit_mean: 0.0
+    logit_std: 1.0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path: {pretrain_network: null}
+
+train:
+  train_pipeline: LM_AIFM_SameObservation_Teacher_Trainer
+  optim_g: {type: AdamW, lr: !!float 1e-5, weight_decay: 0.0, betas: [0.9, 0.999]}
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  # Preserve the old batch-8 warmup exposure: 500 * 8 = 125 * 32 samples.
+  warmup_iter: 125
+  # BaseTrainer currently uses an inclusive loop; 2499 means 2500 updates.
+  total_iter: 2499
+  ema_decay: 0.999
+  ema_warmup_step: 125
+
+val:
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+
+logger:
+  print_freq: 1
+  # Rolling resumable state roughly hourly; immutable model milestones every 500.
+  save_checkpoint_freq: 50
+  save_eval_checkpoint_freq: 500
+  clean_cache_freq: 10
+  use_wandb: false

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_maincfg_smoke_1b/train_lm_aifm_same_observation_teacher_phase1_1b.yml

@@ -0,0 +1,104 @@
+# GENERATE TIME: Sun Aug  9 17:07:42 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_same_observation_teacher_phase1_1b.yml name=lm_aifm_same_observation_teacher_phase1_maincfg_smoke_1b train.total_iter=0 logger.print_freq=1 logger.save_checkpoint_freq=1
+
+# Main Phase-1 LoRA training on real long-video prefixes. The resulting
+# checkpoint is only a teacher candidate. It must still pass the separate
+# real/student-prefix competence audit and interval-solver gate before F runs.
+
+name: lm_aifm_same_observation_teacher_phase1_1b
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VMemSpecWindowsTarDataset
+    meta_path: /mnt/workspace/fly/lm_aifm_20260809/datasets/formal_sekai_60s_v1/wids-meta.json
+    data_dtype: .mp4
+    temporal_sampling: memspec_sink_slow_recent_29f
+    window_num_frames: 29
+    history_windows: 6
+    target_fps: 16.0
+    temporal_seed: 0
+    decode_threads: 2
+    height: 480
+    width: 832
+    return_temporal_metadata: true
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: true
+      seed: 0
+      shuffle_chunk_size: 1000
+      drop_last: true
+      persistent_workers: false
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.1
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+  vae_cfg: {load: true}
+  scheduler_cfg: {shift: 5.0, weight_type: uniform}
+
+  memory_cfg:
+    capacity: 5
+    sink_slots: 1
+    recent_slots: 2
+    context_slots: 5
+    target_chunks: 3
+
+  teacher_cfg:
+    phase: real_prefix_phase1
+    # The dataset preselects sink/slow/recent roles around a fixed current.
+    # Only all-six has an unambiguous Phase-1 observation contract today.
+    history_depths: [6]
+    timestep_weighting: logit_normal
+    logit_mean: 0.0
+    logit_std: 1.0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path: {pretrain_network: null}
+
+train:
+  train_pipeline: LM_AIFM_SameObservation_Teacher_Trainer
+  optim_g: {type: AdamW, lr: !!float 1e-5, weight_decay: 0.0, betas: [0.9, 0.999]}
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 500
+  total_iter: 9999
+  ema_decay: 0.999
+  ema_warmup_step: 500
+
+val:
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+
+logger:
+  print_freq: 10
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 10
+  use_wandb: false

LongVid/experiments/lm_aifm_same_observation_teacher_phase1_smoke_1b/train_lm_aifm_same_observation_teacher_smoke_1b.yml

@@ -0,0 +1,103 @@
+# GENERATE TIME: Sun Aug  9 17:05:09 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/train/lm_aifm/train_lm_aifm_same_observation_teacher_smoke_1b.yml
+
+# One-update Phase-1 systems smoke for the FAR same-observation long teacher.
+# This checks data -> per-window VAE -> MemSpec -> 13c+8t bidirectional score.
+# It is not a teacher competence result and does not unlock formal-F training.
+# Use the full 799-sample WIDS even for one update: a four-sample shard with
+# drop_last=true yields zero samples/rank under an eight-rank launch.
+
+name: lm_aifm_same_observation_teacher_phase1_smoke_1b
+mode: train
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  train:
+    type: T2VMemSpecWindowsTarDataset
+    meta_path: /mnt/workspace/fly/lm_aifm_20260809/datasets/formal_sekai_60s_v1/wids-meta.json
+    data_dtype: .mp4
+    temporal_sampling: memspec_sink_slow_recent_29f
+    window_num_frames: 29
+    history_windows: 6
+    target_fps: 16.0
+    temporal_seed: 0
+    decode_threads: 2
+    height: 480
+    width: 832
+    return_temporal_metadata: true
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: false
+      seed: 0
+      shuffle_chunk_size: null
+      drop_last: true
+      persistent_workers: false
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: true
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 1.0
+    lora_config:
+      lora_rank: 16
+      lora_alpha: 16
+      target_module_name: ['attn1.to_q', 'attn1.to_k', 'attn1.to_v', 'attn1.to_out.0', 'ffn.net.0.proj', 'ffn.net.2', 'delta_embedder', 'time_embedder']
+    param_names_to_optimize: ['lora_A', 'lora_B']
+
+  text_encoder_cfg:
+    drop_text_ratio: 0.0
+    negative_embedding_path: assets/negative_embedding/wan_negemb_cn.pth
+  vae_cfg: {load: true}
+  scheduler_cfg: {shift: 5.0, weight_type: uniform}
+
+  memory_cfg:
+    capacity: 5
+    sink_slots: 1
+    recent_slots: 2
+    context_slots: 5
+    target_chunks: 3
+
+  teacher_cfg:
+    phase: real_prefix_phase1
+    history_depths: [6]
+    timestep_weighting: logit_normal
+    logit_mean: 0.0
+    logit_std: 1.0
+
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path: {pretrain_network: null}
+
+train:
+  train_pipeline: LM_AIFM_SameObservation_Teacher_Trainer
+  optim_g: {type: AdamW, lr: !!float 1e-5, weight_decay: 0.0, betas: [0.9, 0.999]}
+  max_grad_norm: 1.0
+  lr_scheduler: constant_with_warmup
+  warmup_iter: 0
+  # BaseTrainer uses <=, so this is exactly one update.
+  total_iter: 0
+
+val:
+  skip_first_eval: true
+  sample_freq: 100000
+  eval_freq: 100000
+
+logger:
+  print_freq: 1
+  save_checkpoint_freq: 1000
+  save_eval_checkpoint_freq: 1000
+  clean_cache_freq: 1
+  use_wandb: false

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/test_anyflow_far_1b_smoke8.yml

@@ -0,0 +1,66 @@
+# GENERATE TIME: Sun Aug  9 12:32:01 2026
+# CMD:
+# /mnt/workspace/fly/lm_aifm_20260809/src/AnyFlow/far/main.py config_path=options/test/lm_aifm/test_anyflow_far_1b_smoke8.yml
+
+name: lm_aifm_anyflow_far_1b_smoke8
+mode: eval
+manual_seed: 0
+mixed_precision: true
+
+datasets:
+  sample:
+    type: PromptDataset
+    prompt_path: assets/evaluation/eval_caption_t2v_smoke8.json
+    preprocess_cfg:
+      height: 256
+      width: 448
+    dataloader_cfg:
+      batch_size_per_gpu: 1
+      num_workers: 0
+      shuffle: false
+      drop_last: false
+
+models:
+  transformer_cfg:
+    model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+    model_type: FAR_Wan_Transformer3DModel
+    enable_gradient_checkpoint: false
+    far_config:
+      full_chunk_limit: 3
+      chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+      compressed_patch_size: [1, 4, 4]
+      long_context_training_ratio: 0.5
+  text_encoder_cfg:
+    drop_text_ratio: 0.0
+    fuse_guidance_scale: 3.0
+  vae_cfg:
+    load: true
+  scheduler_cfg:
+    shift: 5.0
+    weight_type: beta08
+  flowmap_cfg:
+    gate_value: 0.25
+    deltatime_type: r
+    diffusion_ratio: 0.5
+    consistency_ratio: 0.25
+    epsilon: 5
+
+path:
+  pretrain_network: null
+
+train:
+  train_pipeline: FAR_Wan_AnyFlow_Pretrain_Trainer
+
+val:
+  val_pipeline: FARWanAnyFlowPipeline
+  sample_cfg:
+    height: 256
+    width: 448
+    seed: [0]
+    guidance_scale: 1.0
+    num_inference_steps: [2]
+    num_frames: 81
+    negative_prompt: "overexposed, static, blurry details, subtitles, low quality, deformed"
+
+logger:
+  use_wandb: false

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/test_lm_aifm_anyflow_far_1b_smoke8_20260809_122740.log

@@ -0,0 +1,78 @@
+2026-08-09 12:27:43,453 INFO: 
+  name: lm_aifm_anyflow_far_1b_smoke8
+  mode: eval
+  manual_seed: 0
+  mixed_precision: True
+  datasets:[
+    sample:[
+      type: PromptDataset
+      prompt_path: assets/evaluation/eval_caption_t2v_smoke8.json
+      preprocess_cfg:[
+        height: 256
+        width: 448
+      ]
+      dataloader_cfg:[
+        batch_size_per_gpu: 1
+        num_workers: 0
+        shuffle: False
+        drop_last: False
+      ]
+    ]
+  ]
+  models:[
+    transformer_cfg:[
+      model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+      model_type: FAR_Wan_Transformer3DModel
+      enable_gradient_checkpoint: False
+      far_config:[
+        full_chunk_limit: 3
+        chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+        compressed_patch_size: [1, 4, 4]
+        long_context_training_ratio: 0.5
+      ]
+    ]
+    text_encoder_cfg:[
+      drop_text_ratio: 0.0
+      fuse_guidance_scale: 3.0
+    ]
+    vae_cfg:[
+      load: True
+    ]
+    scheduler_cfg:[
+      shift: 5.0
+      weight_type: beta08
+    ]
+    flowmap_cfg:[
+      gate_value: 0.25
+      deltatime_type: r
+      diffusion_ratio: 0.5
+      consistency_ratio: 0.25
+      epsilon: 5
+    ]
+  ]
+  path:[
+    pretrain_network: None
+    results_root: results/lm_aifm_anyflow_far_1b_smoke8
+    log: results/lm_aifm_anyflow_far_1b_smoke8
+    visualization: results/lm_aifm_anyflow_far_1b_smoke8/visualization
+  ]
+  train:[
+    train_pipeline: FAR_Wan_AnyFlow_Pretrain_Trainer
+  ]
+  val:[
+    val_pipeline: FARWanAnyFlowPipeline
+    sample_cfg:[
+      height: 256
+      width: 448
+      seed: [0]
+      guidance_scale: 1.0
+      num_inference_steps: [2]
+      num_frames: 81
+      negative_prompt: overexposed, static, blurry details, subtitles, low quality, deformed
+    ]
+  ]
+  logger:[
+    use_wandb: False
+  ]
+  config_path: options/test/lm_aifm/test_anyflow_far_1b_smoke8.yml
+

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/test_lm_aifm_anyflow_far_1b_smoke8_20260809_123201.log

@@ -0,0 +1,78 @@
+2026-08-09 12:32:02,112 INFO: 
+  name: lm_aifm_anyflow_far_1b_smoke8
+  mode: eval
+  manual_seed: 0
+  mixed_precision: True
+  datasets:[
+    sample:[
+      type: PromptDataset
+      prompt_path: assets/evaluation/eval_caption_t2v_smoke8.json
+      preprocess_cfg:[
+        height: 256
+        width: 448
+      ]
+      dataloader_cfg:[
+        batch_size_per_gpu: 1
+        num_workers: 0
+        shuffle: False
+        drop_last: False
+      ]
+    ]
+  ]
+  models:[
+    transformer_cfg:[
+      model_name: /mnt/workspace/fly/lm_aifm_20260809/checkpoints/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+      model_type: FAR_Wan_Transformer3DModel
+      enable_gradient_checkpoint: False
+      far_config:[
+        full_chunk_limit: 3
+        chunk_partition: [1, 3, 3, 3, 3, 3, 3, 2]
+        compressed_patch_size: [1, 4, 4]
+        long_context_training_ratio: 0.5
+      ]
+    ]
+    text_encoder_cfg:[
+      drop_text_ratio: 0.0
+      fuse_guidance_scale: 3.0
+    ]
+    vae_cfg:[
+      load: True
+    ]
+    scheduler_cfg:[
+      shift: 5.0
+      weight_type: beta08
+    ]
+    flowmap_cfg:[
+      gate_value: 0.25
+      deltatime_type: r
+      diffusion_ratio: 0.5
+      consistency_ratio: 0.25
+      epsilon: 5
+    ]
+  ]
+  path:[
+    pretrain_network: None
+    results_root: results/lm_aifm_anyflow_far_1b_smoke8
+    log: results/lm_aifm_anyflow_far_1b_smoke8
+    visualization: results/lm_aifm_anyflow_far_1b_smoke8/visualization
+  ]
+  train:[
+    train_pipeline: FAR_Wan_AnyFlow_Pretrain_Trainer
+  ]
+  val:[
+    val_pipeline: FARWanAnyFlowPipeline
+    sample_cfg:[
+      height: 256
+      width: 448
+      seed: [0]
+      guidance_scale: 1.0
+      num_inference_steps: [2]
+      num_frames: 81
+      negative_prompt: overexposed, static, blurry details, subtitles, low quality, deformed
+    ]
+  ]
+  logger:[
+    use_wandb: False
+  ]
+  config_path: options/test/lm_aifm/test_anyflow_far_1b_smoke8.yml
+

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/visualization/sample/iter_final/task_t2v/sample_step_2/seed_0/00_A_golden_retriever_runs_through_a_sunny_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:347a6c700e0a91215a06c193086ea40f2905b66250d164d60038a7d7688aa64e
+size 303586

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/visualization/sample/iter_final/task_t2v/sample_step_2/seed_0/01_A_red_vintage_car_drives_along_a_winding.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9b3fb157769240d7ba0a00ab3e99c79bb9b8fa512fe704e9e1d9450cfc02e46
+size 213071

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/visualization/sample/iter_final/task_t2v/sample_step_2/seed_0/02_A_chef_mixes_ingredients_in_a_bright_mod.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f28d212a4e12402f4fa5c71070398f3821f6fd81f3f55102416ab87eaa88f38f
+size 135648

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/visualization/sample/iter_final/task_t2v/sample_step_2/seed_0/03_A_panda_cub_climbs_bamboo_on_a_misty_for.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6b11812775af4840b533c71e8fd2821ef5478d0b1a6258e046a64b0789619bc
+size 238897

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/visualization/sample/iter_final/task_t2v/sample_step_2/seed_0/04_An_astronaut_walks_slowly_across_a_rocky.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:449562bf0186cd315ce7e1902ab79a2bf10fb627e7b600ddfc9b47049a934dbd
+size 282367

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/visualization/sample/iter_final/task_t2v/sample_step_2/seed_0/05_A_woman_in_a_red_dress_walks_through_a_n.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2694e8a48ac50d73c996586307e67f0d6c179cc01e8c18c7b0dfa978b6b8958
+size 168657

LongVid/results/lm_aifm_anyflow_far_1b_smoke8/visualization/sample/iter_final/task_t2v/sample_step_2/seed_0/06_A_small_sailboat_crosses_a_calm_mountain.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddea7796ef790502c8cb6913f790ec9146317ca725b703efbc322fb6d1df2333
+size 161150