claude_refactor_v3: De-duplicate gin-configs, + "in/config_files" with common and cpecific experiment-presets gins

in/config_files/hardware.gin

@@ -1 +0,0 @@
-

in/config_files/models.gin

@@ -1 +0,0 @@
-

in/config_files/pipeline.gin

@@ -1 +0,0 @@
-

in/config_files/tracking.gin

@@ -1 +1,16 @@
+# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
+TrackingConfig.use_wandb = False
+TrackingConfig.use_tb = True
+TrackingConfig.wandb_project = 'caption-test-gtauav'
+TrackingConfig.wandb_run_name = None
+TrackingConfig.wandb_entity = None
+
+TrackingConfig.log_grad_norms = True
+TrackingConfig.use_gradcam = False
+TrackingConfig.gradcam_every = 5
+TrackingConfig.gradcam_samples = 8
+
+TrackingConfig.use_profiler = False
+TrackingConfig.profiler_warmup = 3
+TrackingConfig.profiler_active = 5
 

in/config_files/training.gin

@@ -1 +1,31 @@
+# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
+TrainingConfig.loss_type = 'symmetric'
+TrainingConfig.tau_init = 0.07
+TrainingConfig.tau_min = 0.01
+TrainingConfig.tau_max = 0.1
+TrainingConfig.learnable_temperature = True
+TrainingConfig.label_smoothing = 0.1
 
+TrainingConfig.tau_final = 0.01
+TrainingConfig.weight_q2g = 0.6
+TrainingConfig.weight_g2q = 0.4
+TrainingConfig.hard_mining_k = 0
+TrainingConfig.neg_bank_size = 0
+
+# WeightedInfoNCE-only (unused when loss_type='symmetric').
+TrainingConfig.weighted_loss_k = 5.0
+
+TrainingConfig.learning_rate = 1e-4
+TrainingConfig.text_lr_factor = 0.1
+TrainingConfig.weight_decay = 1e-4
+TrainingConfig.grad_clip = 1.0
+
+TrainingConfig.sampler_type = 'mutex'
+TrainingConfig.dss_warmup_epochs = 1
+TrainingConfig.dss_reembed_every = 1
+TrainingConfig.dss_knn_device = 'cuda'
+TrainingConfig.dss_use_lsh = False
+TrainingConfig.dss_lsh_num_tables = 8
+TrainingConfig.dss_lsh_num_bits = 14
+TrainingConfig.dss_cache_dir = None
+TrainingConfig.use_mutex_sampler = True

presets/gtauav_balanced/tracking.gin

@@ -1,15 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5

presets/gtauav_balanced/training.gin

@@ -1,31 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True

presets/gtauav_balanced_asym/hardware.gin

@@ -1,10 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_balanced_asym/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_balanced_asym/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_balanced_sofia/hardware.gin

@@ -1,9 +0,0 @@
-# SOFIA v7.1 from-scratch — keep activations live (no gradient checkpointing).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = False
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0

presets/gtauav_balanced_sofia/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_balanced_sofia/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_balanced_sofia_v1/hardware.gin

@@ -1,8 +0,0 @@
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = False
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0

presets/gtauav_balanced_sofia_v1/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_balanced_sofia_v1/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_balanced_stripnet/hardware.gin

@@ -1,10 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_balanced_stripnet/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_balanced_stripnet/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_balanced_stripnet_unfrozen/hardware.gin

@@ -1,10 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_balanced_stripnet_unfrozen/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_balanced_stripnet_unfrozen/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_baseline/hardware.gin

@@ -1,10 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_baseline/tracking.gin

@@ -1,15 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5

presets/gtauav_baseline/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_baseline_asym/hardware.gin

@@ -1,9 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0

presets/gtauav_baseline_asym/tracking.gin

@@ -1,15 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5

presets/gtauav_baseline_asym/training.gin

@@ -1,31 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True

presets/gtauav_baseline_sofia/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_baseline_sofia/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_baseline_sofia_v1/hardware.gin

@@ -1,9 +0,0 @@
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = False
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_baseline_sofia_v1/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_baseline_sofia_v1/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_baseline_stripnet/hardware.gin

@@ -1,10 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_baseline_stripnet/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_baseline_stripnet/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_baseline_stripnet_unfrozen/hardware.gin

@@ -1,10 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_baseline_stripnet_unfrozen/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_baseline_stripnet_unfrozen/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_image_heavy/hardware.gin

@@ -1,10 +0,0 @@
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0
-

presets/gtauav_image_heavy/tracking.gin

@@ -1,16 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5
-

presets/gtauav_image_heavy/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

presets/gtauav_text_heavy/hardware.gin

@@ -1,10 +0,0 @@
-
-# RTX 4090 profile, shared encoder (DINOv3).
-HardwareConfig.device = 'cuda'
-HardwareConfig.batch_size = 8
-HardwareConfig.grad_accum_steps = 8
-HardwareConfig.num_workers = 4
-HardwareConfig.use_amp = True
-HardwareConfig.gradient_checkpointing = True
-HardwareConfig.reserve_gb = 2.0
-HardwareConfig.total_vram_gb = 24.0

presets/gtauav_text_heavy/tracking.gin

@@ -1,15 +0,0 @@
-# Default tracking: TensorBoard on, W&B off, no Grad-CAM, no profiler.
-TrackingConfig.use_wandb = False
-TrackingConfig.use_tb = True
-TrackingConfig.wandb_project = 'caption-test-gtauav'
-TrackingConfig.wandb_run_name = None
-TrackingConfig.wandb_entity = None
-
-TrackingConfig.log_grad_norms = True
-TrackingConfig.use_gradcam = False
-TrackingConfig.gradcam_every = 5
-TrackingConfig.gradcam_samples = 8
-
-TrackingConfig.use_profiler = False
-TrackingConfig.profiler_warmup = 3
-TrackingConfig.profiler_active = 5

presets/gtauav_text_heavy/training.gin

@@ -1,32 +0,0 @@
-# Loss + optimizer + sampler — symmetric InfoNCE, AdamW, mutex sampler.
-TrainingConfig.loss_type = 'symmetric'
-TrainingConfig.tau_init = 0.07
-TrainingConfig.tau_min = 0.01
-TrainingConfig.tau_max = 0.1
-TrainingConfig.learnable_temperature = True
-TrainingConfig.label_smoothing = 0.1
-
-TrainingConfig.tau_final = 0.01
-TrainingConfig.weight_q2g = 0.6
-TrainingConfig.weight_g2q = 0.4
-TrainingConfig.hard_mining_k = 0
-TrainingConfig.neg_bank_size = 0
-
-# WeightedInfoNCE-only (unused when loss_type='symmetric').
-TrainingConfig.weighted_loss_k = 5.0
-
-TrainingConfig.learning_rate = 1e-4
-TrainingConfig.text_lr_factor = 0.1
-TrainingConfig.weight_decay = 1e-4
-TrainingConfig.grad_clip = 1.0
-
-TrainingConfig.sampler_type = 'mutex'
-TrainingConfig.dss_warmup_epochs = 1
-TrainingConfig.dss_reembed_every = 1
-TrainingConfig.dss_knn_device = 'cuda'
-TrainingConfig.dss_use_lsh = False
-TrainingConfig.dss_lsh_num_tables = 8
-TrainingConfig.dss_lsh_num_bits = 14
-TrainingConfig.dss_cache_dir = None
-TrainingConfig.use_mutex_sampler = True
-

src/conf/config_loader.py

@@ -1,8 +1,24 @@
 """Single entry point for loading all configs in a training run.
 
-Reads every .gin file in path2cfg (one preset directory) and instantiates
-the 4 universal configs + ModelsCommonConfig + the family-specific Models
-config selected by ModelsCommonConfig.backbone.
+Layout (per REQUIREMENTS_GIN_STYLE.md §1, extended for multi-experiment):
+
+    in/config_files/
+        ├── training.gin                    # common: shared by all training presets
+        ├── hardware_default.gin            # common: for DINOv3 / StripNet
+        ├── hardware_no_gradckpt.gin        # common: for SOFIA backbones
+        ├── tracking.gin                    # common: shared by all
+        └── <preset_name>/
+            ├── pipeline.gin                # local: output_dir, paths, schedule
+            └── models.gin                  # local: backbone-specific bindings
+
+`load_all_configs` parses common files at the path2cfg root + preset locals
+in `path2cfg/<preset_name>/`. Two-pass loading:
+  Pass 1: read <preset_name>/models.gin to learn the backbone.
+  Pass 2: parse common files + preset locals in one batch.
+
+Override semantics: if the preset directory contains its own training.gin /
+hardware.gin / tracking.gin, those files are appended AFTER the common
+versions, so their bindings win (gin: last write wins).
 """
 
 from __future__ import annotations
@@ -33,49 +49,114 @@ _BACKBONE_TO_MODELS_CLS = {
     "sofia_v71": SOFIAv71ModelsConfig,
 }
 
+# Sofia backbones disable gradient checkpointing.
+_NO_GRADCKPT_BACKBONES = {"sofia_v1", "sofia_v71"}
 
-def load_all_configs(path2cfg: str) -> dict[str, Any]:
-    """Parse every .gin in path2cfg and instantiate all configs.
+# Common filenames at the path2cfg root.
+_COMMON_TRAINING = "training.gin"
+_COMMON_TRACKING = "tracking.gin"
+_COMMON_HARDWARE_DEFAULT = "hardware_default.gin"
+_COMMON_HARDWARE_NO_GRADCKPT = "hardware_no_gradckpt.gin"
+
+# Files a preset must always have locally (in <path2cfg>/<preset_name>/).
+_LOCAL_REQUIRED = ("pipeline.gin", "models.gin")
+
+# Files that can optionally be overridden locally; if present in the preset
+# directory, they win over the common version.
+_LOCAL_OVERRIDABLE = ("training.gin", "hardware.gin", "tracking.gin")
+
+
+def load_all_configs(path2cfg: str, preset_name: str) -> dict[str, Any]:
+    """Parse common gin files + preset gin files and return all configs.
 
     Args:
-        path2cfg: Path to a preset directory ending with '/', e.g.
-                  '/home/user/caption-test/in/config_files/'.
+        path2cfg: Path to in/config_files/ (per REQUIREMENTS_GIN_STYLE.md §5).
+                  Must contain training.gin, hardware_default.gin,
+                  hardware_no_gradckpt.gin, tracking.gin at its root, plus
+                  per-preset subdirectories.
+        preset_name: Name of the preset subdirectory under path2cfg, e.g.
+                     'gtauav_balanced'.
 
     Returns:
-        Dict with keys:
-            'pipeline'      → PipelineConfig
-            'hardware'      → HardwareConfig
-            'training'      → TrainingConfig
-            'tracking'      → TrackingConfig
-            'models_common' → ModelsCommonConfig
-            'models'        → DINOv3 | StripNet | SOFIAv1 | SOFIAv71 ModelsConfig
-                              (selected by models_common.backbone)
+        Dict with keys: 'pipeline', 'hardware', 'training', 'tracking',
+        'models_common', 'models'.
 
     Raises:
-        FileNotFoundError: If path2cfg contains no .gin files.
-        ValueError: If models_common.backbone is not one of the known values.
+        FileNotFoundError: If path2cfg, preset directory, or a required file is missing.
+        ValueError: If models_common.backbone is not a known value.
     """
-    cfg_dir = Path(path2cfg)
-    gin_files = sorted(cfg_dir.glob("*.gin"))
-    if not gin_files:
-        raise FileNotFoundError(f"No .gin files found in {cfg_dir}")
+    common_dir = Path(path2cfg)
+    preset_dir = common_dir / preset_name
 
-    # MANDATORY: reset gin global state — without clear_config(), bindings
-    # from previous parses accumulate.
+    # Sanity checks.
+    if not common_dir.is_dir():
+        raise FileNotFoundError(
+            f"Config root not found: {common_dir}. "
+            f"Per REQUIREMENTS_GIN_STYLE.md §5, this should be "
+            f"<proj_dir>/in/config_files/.",
+        )
+    if not preset_dir.is_dir():
+        raise FileNotFoundError(
+            f"Preset directory not found: {preset_dir}. "
+            f"Available presets in {common_dir}: "
+            f"{sorted(d.name for d in common_dir.iterdir() if d.is_dir())}",
+        )
+    for required in _LOCAL_REQUIRED:
+        if not (preset_dir / required).is_file():
+            raise FileNotFoundError(
+                f"Preset {preset_name} is missing required file '{required}' "
+                f"(looked in {preset_dir})",
+            )
+
+    # ===== Pass 1: peek at models.gin to learn the backbone. =====
+    gin.clear_config()
+    gin.parse_config_file(str(preset_dir / "models.gin"))
+    backbone = ModelsCommonConfig().backbone
+    if backbone not in _BACKBONE_TO_MODELS_CLS:
+        raise ValueError(
+            f"Unknown backbone={backbone!r} in {preset_dir / 'models.gin'}; "
+            f"expected one of {sorted(_BACKBONE_TO_MODELS_CLS)}",
+        )
+
+    # ===== Pass 2: build the full file list and parse in one batch. =====
+    # Order: common first, preset locals last (so locals win on overrides).
+    gin_files: list[Path] = [
+        common_dir / _COMMON_TRAINING,
+        common_dir / (
+            _COMMON_HARDWARE_NO_GRADCKPT if backbone in _NO_GRADCKPT_BACKBONES
+            else _COMMON_HARDWARE_DEFAULT
+        ),
+        common_dir / _COMMON_TRACKING,
+        # Always-required preset locals.
+        preset_dir / "pipeline.gin",
+        preset_dir / "models.gin",
+    ]
+
+    # Optional preset overrides (rare).
+    for overridable in _LOCAL_OVERRIDABLE:
+        local = preset_dir / overridable
+        if local.is_file():
+            gin_files.append(local)
+            logger.info("Preset %s overrides %s locally", preset_name, overridable)
+
+    # Sanity: all chosen files must exist.
+    for f in gin_files:
+        if not f.is_file():
+            raise FileNotFoundError(f"Required gin file not found: {f}")
+
+    # MANDATORY: clear gin global state before parsing.
     gin.clear_config()
     gin.parse_config_files_and_bindings(
         config_files=[str(f) for f in gin_files],
         bindings=[],
     )
-    logger.info("Loaded %d gin files from %s", len(gin_files), cfg_dir)
+    logger.info(
+        "Loaded preset %s with %d gin files (backbone=%s)",
+        preset_name, len(gin_files), backbone,
+    )
 
-    # Build common first to learn which backbone to use.
+    # Build all configs from gin global state.
     models_common = ModelsCommonConfig()
-    if models_common.backbone not in _BACKBONE_TO_MODELS_CLS:
-        raise ValueError(
-            f"Unknown backbone={models_common.backbone!r}; "
-            f"expected one of {sorted(_BACKBONE_TO_MODELS_CLS)}",
-        )
     models_specific = _BACKBONE_TO_MODELS_CLS[models_common.backbone]()
 
     return {

tests/test_config_loader.py

@@ -0,0 +1,125 @@
+"""Tests for config_loader.load_all_configs with _common/ + preset structure."""
+
+from __future__ import annotations
+from pathlib import Path
+import pytest
+from src.conf.config_loader import load_all_configs
+
+
+def _make_minimal_preset_layout(root: Path) -> None:
+    """Create _common/ + base files matching the production layout."""
+    common = root / "_common"
+    common.mkdir()
+    (common / "training.gin").write_text("TrainingConfig.tau_init = 0.07\n")
+    (common / "hardware_default.gin").write_text("HardwareConfig.batch_size = 8\n")
+    (common / "hardware_no_gradckpt.gin").write_text(
+        "HardwareConfig.batch_size = 8\n"
+        "HardwareConfig.gradient_checkpointing = False\n",
+    )
+    (common / "tracking.gin").write_text("TrackingConfig.use_tb = True\n")
+
+
+def _make_preset(root: Path, name: str, models_body: str, output_dir: str) -> None:
+    """Create a preset directory with required pipeline.gin + models.gin."""
+    preset = root / name
+    preset.mkdir()
+    (preset / "pipeline.gin").write_text(
+        f"PipelineConfig.output_dir = '{output_dir}'\n",
+    )
+    (preset / "models.gin").write_text(models_body)
+
+
+def test_load_returns_six_keys(tmp_path: Path) -> None:
+    """load_all_configs returns the 6 expected keys."""
+    _make_minimal_preset_layout(tmp_path)
+    _make_preset(
+        tmp_path, "test_dinov3",
+        models_body="ModelsCommonConfig.backbone = 'dinov3'\n",
+        output_dir="out/test_dinov3",
+    )
+
+    cfgs = load_all_configs("test_dinov3", str(tmp_path))
+    assert set(cfgs.keys()) == {
+        "pipeline", "hardware", "training", "tracking", "models_common", "models",
+    }
+    assert cfgs["pipeline"].output_dir == "out/test_dinov3"
+    assert cfgs["models_common"].backbone == "dinov3"
+
+
+def test_sofia_uses_no_gradckpt_hardware(tmp_path: Path) -> None:
+    """Sofia backbones get hardware_no_gradckpt.gin (gradient_checkpointing=False)."""
+    _make_minimal_preset_layout(tmp_path)
+    _make_preset(
+        tmp_path, "test_sofia",
+        models_body="ModelsCommonConfig.backbone = 'sofia_v71'\n",
+        output_dir="out/test_sofia",
+    )
+
+    cfgs = load_all_configs("test_sofia", str(tmp_path))
+    assert cfgs["hardware"].gradient_checkpointing is False
+
+
+def test_dinov3_uses_default_hardware(tmp_path: Path) -> None:
+    """DINOv3/StripNet backbones get hardware_default.gin (gradient_checkpointing=True default)."""
+    _make_minimal_preset_layout(tmp_path)
+    # Override gradient_checkpointing=True in default profile to test routing.
+    (tmp_path / "_common" / "hardware_default.gin").write_text(
+        "HardwareConfig.batch_size = 8\n"
+        "HardwareConfig.gradient_checkpointing = True\n",
+    )
+    _make_preset(
+        tmp_path, "test_dinov3_hw",
+        models_body="ModelsCommonConfig.backbone = 'dinov3'\n",
+        output_dir="out/test_dinov3_hw",
+    )
+    cfgs = load_all_configs("test_dinov3_hw", str(tmp_path))
+    assert cfgs["hardware"].gradient_checkpointing is True
+
+
+def test_local_override_wins(tmp_path: Path) -> None:
+    """Local training.gin in preset dir overrides _common/training.gin."""
+    _make_minimal_preset_layout(tmp_path)
+    _make_preset(
+        tmp_path, "test_override",
+        models_body="ModelsCommonConfig.backbone = 'dinov3'\n",
+        output_dir="out/test_override",
+    )
+    # Drop a local training.gin that overrides tau_init.
+    (tmp_path / "test_override" / "training.gin").write_text(
+        "TrainingConfig.tau_init = 0.123\n",
+    )
+
+    cfgs = load_all_configs("test_override", str(tmp_path))
+    assert cfgs["training"].tau_init == 0.123
+
+
+def test_unknown_backbone_raises(tmp_path: Path) -> None:
+    """Invalid backbone string in models.gin → ValueError."""
+    _make_minimal_preset_layout(tmp_path)
+    _make_preset(
+        tmp_path, "test_bad",
+        models_body="ModelsCommonConfig.backbone = 'mistral_42b'\n",
+        output_dir="out/bad",
+    )
+
+    with pytest.raises(ValueError, match="Unknown backbone"):
+        load_all_configs("test_bad", str(tmp_path))
+
+
+def test_missing_preset_raises(tmp_path: Path) -> None:
+    """Non-existent preset directory → FileNotFoundError."""
+    _make_minimal_preset_layout(tmp_path)
+    with pytest.raises(FileNotFoundError, match="Preset directory not found"):
+        load_all_configs("nonexistent", str(tmp_path))
+
+
+def test_missing_required_file_raises(tmp_path: Path) -> None:
+    """Preset missing pipeline.gin → FileNotFoundError."""
+    _make_minimal_preset_layout(tmp_path)
+    bad_preset = tmp_path / "test_bad_preset"
+    bad_preset.mkdir()
+    (bad_preset / "models.gin").write_text("ModelsCommonConfig.backbone = 'dinov3'\n")
+
+    with pytest.raises(FileNotFoundError, match="missing required file 'pipeline.gin'"):
+        load_all_configs("test_bad_preset", str(tmp_path))
+