Training API

Training loops, schedulers, curricula, distributed-training helpers, and performance utilities for discrete and continuous latent models.

Trainers

class medlatents.training.DiscreteLatentTrainer(model_type, model, train_loader, val_loader, args, accelerator=None, use_ema=True, ema_decay=0.9999, log_gradients=True, generate_samples_every=None, enable_profiling=False, profile_log_interval=100, check_gradients_finite=False, use_compile=False, compile_mode='reduce-overhead')

Bases: object

Enhanced unified trainer for discrete latent generative models.

Supports: - autoreg: Autoregressive transformer - maskgit: Bidirectional transformer with masking (MaskGIT) - flow: Discrete flow matching with DiscreteDiT - d3pm: D3PM discrete diffusion (DiscreteDiT backbone) - bayesian_flow: Bayesian Flow Networks (DiscreteDiT backbone)

Features: - EMA (Exponential Moving Average) for stable training - Gradient monitoring and logging - Curriculum learning - Periodic sample generation - Multi-GPU support with FSDP

Parameters:

• model_type (Literal['autoreg', 'maskgit', 'flow', 'd3pm', 'bayesian_flow'])

• model (Module)

• train_loader (DataLoader)

• val_loader (DataLoader)

• accelerator (Accelerator | None, default: None)

• use_ema (bool, default: True)

• ema_decay (float, default: 0.9999)

• log_gradients (bool, default: True)

• generate_samples_every (int | None, default: None)

• enable_profiling (bool, default: False)

• profile_log_interval (int, default: 100)

• check_gradients_finite (bool, default: False)

• use_compile (bool, default: False)

• compile_mode (str, default: 'reduce-overhead')

__init__(model_type, model, train_loader, val_loader, args, accelerator=None, use_ema=True, ema_decay=0.9999, log_gradients=True, generate_samples_every=None, enable_profiling=False, profile_log_interval=100, check_gradients_finite=False, use_compile=False, compile_mode='reduce-overhead')

Parameters:

• model_type (Literal['autoreg', 'maskgit', 'flow', 'd3pm', 'bayesian_flow'])

• model (Module)

• train_loader (DataLoader)

• val_loader (DataLoader)

• accelerator (Accelerator | None, default: None)

• use_ema (bool, default: True)

• ema_decay (float, default: 0.9999)

• log_gradients (bool, default: True)

• generate_samples_every (int | None, default: None)

• enable_profiling (bool, default: False)

• profile_log_interval (int, default: 100)

• check_gradients_finite (bool, default: False)

• use_compile (bool, default: False)

• compile_mode (str, default: 'reduce-overhead')

compute_loss(batch, step=None)

Compute loss based on model type.

Validates token contracts before computing loss. All discrete models expect tokens in [0, vocab_size) with dtype torch.long.

train_epoch(epoch)

Train for one epoch.

validate()

Run validation.

train(resume=False, resume_best=False, reset_wandb=False)

Main training loop.

class medlatents.training.ContinuousLatentTrainer(model_type, model, train_loader, val_loader, args, accelerator=None, use_ema=True, ema_decay=0.999, log_gradients=False)

Bases: object

Unified trainer for continuous diffusion and flow-matching models.

Parameters:

• model_type (Literal['diffusion', 'flow'])

• model (Module)

• train_loader (DataLoader)

• val_loader (DataLoader | None)

• accelerator (Accelerator | None, default: None)

• use_ema (bool, default: True)

• ema_decay (float, default: 0.999)

• log_gradients (bool, default: False)

__init__(model_type, model, train_loader, val_loader, args, accelerator=None, use_ema=True, ema_decay=0.999, log_gradients=False)

Parameters:

• model_type (Literal['diffusion', 'flow'])

• model (Module)

• train_loader (DataLoader)

• val_loader (DataLoader | None)

• accelerator (Accelerator | None, default: None)

• use_ema (bool, default: True)

• ema_decay (float, default: 0.999)

• log_gradients (bool, default: False)

compute_loss(batch, condition=None)

Compute loss based on model type.

Validates continuous latent contracts before computing loss. Expects floating-point tensors in [B, C, H, W] or [B, C, D, H, W] format.

train_epoch(epoch)

Parameters:

epoch (int)

Return type:

None

validate(step=None)

Parameters:

step (int | None, default: None)

Return type:

float | None

save_checkpoint(output_dir, epoch, metric=0.0, wandb_id='')

Save training checkpoint to output_dir.

Parameters:

• output_dir (str)

• epoch (int)

• metric (float, default: 0.0)

• wandb_id (str, default: '')

Return type:

None

load_checkpoint(checkpoint_path)

Load model and optimizer state from a checkpoint file.

Parameters:

checkpoint_path (str)

Return type:

None

resume_from_checkpoint(checkpoint_dir, name='')

Resume training from the latest checkpoint in checkpoint_dir.

Parameters:

• checkpoint_dir (str)

• name (str, default: '')

Return type:

int

Schedules

class medlatents.training.MaskingSchedule(start_mask_ratio=0.5, end_mask_ratio=0.15, num_steps=100000, schedule='cosine', min_ratio=0.1)

Bases: object

Dynamic masking schedule for MaskGIT-style training.

Adaptively adjust mask ratio during training.

Parameters:

• start_mask_ratio (float, default: 0.5)

• end_mask_ratio (float, default: 0.15)

• num_steps (int, default: 100000)

• schedule (str, default: 'cosine')

• min_ratio (float, default: 0.1)

__init__(start_mask_ratio=0.5, end_mask_ratio=0.15, num_steps=100000, schedule='cosine', min_ratio=0.1)

Parameters:

• start_mask_ratio (float, default: 0.5) – initial masking ratio

• end_mask_ratio (float, default: 0.15) – final masking ratio

• num_steps (int, default: 100000) – steps to reach end_mask_ratio

• schedule (str, default: 'cosine') – ‘linear’, ‘cosine’, ‘constant’

• min_ratio (float, default: 0.1) – minimum mask ratio (safety)

get_mask_ratio()

Get current mask ratio.

Return type:

float

step()

Increment step.

class medlatents.training.NoiseSchedule(noise_type='uniform', start_noise=0.1, end_noise=0.01, num_steps=100000, schedule='linear')

Bases: object

Noise scheduling for training with noisy inputs.

Gradually reduce noise during training for better convergence. For discrete tokens, applies random replacement or dropout.

Parameters:

• noise_type (str, default: 'uniform')

• start_noise (float, default: 0.1)

• end_noise (float, default: 0.01)

• num_steps (int, default: 100000)

• schedule (str, default: 'linear')

__init__(noise_type='uniform', start_noise=0.1, end_noise=0.01, num_steps=100000, schedule='linear')

Parameters:

• noise_type (str, default: 'uniform') – ‘uniform’ (random replacement) or ‘dropout’ (set to 0)

• start_noise (float, default: 0.1) – initial noise level (probability of noising each token)

• end_noise (float, default: 0.01) – final noise level

• num_steps (int, default: 100000) – steps to reach end_noise

• schedule (str, default: 'linear') – ‘linear’, ‘exp’, ‘cosine’

get_noise_level()

Get current noise level.

Return type:

float

add_noise(tokens, vocab_size)

Add noise to tokens.

Parameters:

• tokens (Tensor) – [batch, seq_len]

• vocab_size (int) – vocabulary size

Returns:

[batch, seq_len]

Return type:

noisy_tokens

medlatents.training.get_cosine_schedule_with_warmup(step, warmup_steps, total_steps, max_lr, min_lr=0.0)

Cosine schedule with linear warmup for learning rates.

Standard learning rate schedule used in transformers and diffusion models.

Parameters:

• step (int) – Current step

• warmup_steps (int) – Number of warmup steps

• total_steps (int) – Total training steps

• max_lr (float) – Maximum learning rate

• min_lr (float, default: 0.0) – Minimum learning rate

Return type:

float

Returns:

Learning rate at current step

medlatents.training.get_mask_ratio(progress, schedule='cosine')

Get mask ratio based on training progress (0 to 1) and schedule type.

MaskGIT training uses a decreasing mask ratio schedule, starting with high masking (e.g., 90% masked) and ending with low masking (e.g., 10%).

Parameters:

• progress (float) – Training progress from 0.0 (start) to 1.0 (end)

• schedule (Literal['cosine', 'linear', 'square'], default: 'cosine') – Schedule type (‘cosine’, ‘linear’, ‘square’)

Return type:

float

Returns:

Mask ratio at this progress (higher = more tokens masked)

Examples

>>> get_mask_ratio(0.0, "cosine")  # Start: max masking
1.0
>>> get_mask_ratio(1.0, "cosine")  # End: min masking
0.0
>>> get_mask_ratio(0.5, "cosine")  # Mid: half masking
0.5

Curriculum Learning

class medlatents.training.CurriculumScheduler(config)

Bases: abc.ABC

Base class for curriculum schedulers.

Parameters:

config (CurriculumConfig)

__init__(config)

Parameters:

config (CurriculumConfig)

step()

Advance curriculum by one step.

Return type:

None

set_step(step)

Set curriculum to specific step.

Parameters:

step (int)

Return type:

None

abstractmethod get_value()

Get current curriculum value.

Return type:

float

class medlatents.training.MaskingRatioCurriculum(start_ratio=0.9, end_ratio=0.1, total_steps=100000, schedule=CurriculumSchedule.COSINE)

Bases: medlatents.training.curriculum.CurriculumScheduler

Curriculum that decreases masking ratio over training.

For MaskGIT-style training, start with high masking (easier task) and decrease to lower masking (harder, more context needed).

Parameters:

• start_ratio (float, default: 0.9) – Initial masking ratio (e.g., 0.9)

• end_ratio (float, default: 0.1) – Final masking ratio (e.g., 0.1)

• total_steps (int, default: 100000) – Total training steps

• schedule (CurriculumSchedule, default: <CurriculumSchedule.COSINE: 'cosine'>)

__init__(start_ratio=0.9, end_ratio=0.1, total_steps=100000, schedule=CurriculumSchedule.COSINE)

Parameters:

• start_ratio (float, default: 0.9)

• end_ratio (float, default: 0.1)

• total_steps (int, default: 100000)

• schedule (CurriculumSchedule, default: <CurriculumSchedule.COSINE: 'cosine'>)

get_value()

Get current masking ratio.

Return type:

float

get_mask_ratio()

Get current masking ratio.

Return type:

float

create_mask(shape, device)

Create mask with current curriculum ratio.

Parameters:

• shape (tuple[int, int]) – (batch_size, seq_len)

• device (device) – Device for tensor

Return type:

Tensor

Returns:

Boolean mask [batch, seq_len] where True = masked

set_step(step)

Set curriculum to specific step.

Parameters:

step (int)

Return type:

None

step()

Advance curriculum by one step.

Return type:

None

class medlatents.training.NoiseLevelCurriculum(start_t_max=0.5, end_t_max=1.0, total_steps=100000, schedule=CurriculumSchedule.LINEAR)

Bases: medlatents.training.curriculum.CurriculumScheduler

Curriculum that adjusts noise level for diffusion training.

For diffusion models, can start with higher noise levels (easier denoising) and progress to full noise schedule.

Parameters:

• start_t_max (float, default: 0.5) – Initial maximum timestep (fraction of full schedule)

• end_t_max (float, default: 1.0) – Final maximum timestep (1.0 = full schedule)

• total_steps (int, default: 100000) – Total training steps

• schedule (CurriculumSchedule, default: <CurriculumSchedule.LINEAR: 'linear'>)

__init__(start_t_max=0.5, end_t_max=1.0, total_steps=100000, schedule=CurriculumSchedule.LINEAR)

Parameters:

• start_t_max (float, default: 0.5)

• end_t_max (float, default: 1.0)

• total_steps (int, default: 100000)

• schedule (CurriculumSchedule, default: <CurriculumSchedule.LINEAR: 'linear'>)

get_value()

Get current maximum timestep fraction.

Return type:

float

sample_timesteps(batch_size, num_timesteps, device)

Sample timesteps with curriculum-limited maximum.

Parameters:

• batch_size (int) – Number of samples

• num_timesteps (int) – Total timesteps in diffusion schedule

• device (device) – Device for tensor

Return type:

Tensor

Returns:

Timesteps [batch_size] in range [0, max_t]

set_step(step)

Set curriculum to specific step.

Parameters:

step (int)

Return type:

None

step()

Advance curriculum by one step.

Return type:

None

class medlatents.training.SequenceLengthCurriculum(min_length, max_length, total_steps, schedule=CurriculumSchedule.LINEAR, warmup_steps=0)

Bases: medlatents.training.curriculum.CurriculumScheduler

Curriculum that progressively increases sequence length.

Start training with short sequences for faster iteration, then increase to full length for learning long-range dependencies.

Parameters:

• min_length (int) – Starting sequence length

• max_length (int) – Final sequence length

• total_steps (int) – Total training steps

• schedule (CurriculumSchedule, default: <CurriculumSchedule.LINEAR: 'linear'>) – Progression schedule

• warmup_steps (int, default: 0)

__init__(min_length, max_length, total_steps, schedule=CurriculumSchedule.LINEAR, warmup_steps=0)

Parameters:

• min_length (int)

• max_length (int)

• total_steps (int)

• schedule (CurriculumSchedule, default: <CurriculumSchedule.LINEAR: 'linear'>)

• warmup_steps (int, default: 0)

get_value()

Get current sequence length.

Return type:

float

get_length()

Get current sequence length as integer.

Return type:

int

truncate_batch(tokens, random_start=True)

Truncate batch to current curriculum length.

Parameters:

• tokens (Tensor) – [batch, seq_len] token tensor

• random_start (bool, default: True) – If True, sample random start position

Return type:

Tensor

Returns:

Truncated tokens [batch, curr_length]

set_step(step)

Set curriculum to specific step.

Parameters:

step (int)

Return type:

None

step()

Advance curriculum by one step.

Return type:

None

medlatents.training.create_curriculum_from_config(config)

Create curriculum from configuration dict.

Parameters:

config (dict[str, Any]) – Dict with curriculum type and parameters

Return type:

CurriculumScheduler

Returns:

Configured curriculum scheduler

Conditional Training

class medlatents.training.ConditionalTrainingConfig(conditioning_config=None, class_dropout_prob=0.1, text_dropout_prob=0.1, use_repa=False, repa_weight=0.5, repa_timestep_threshold=0.5, repa_encoder_name='dinov2_vitb14', use_vecor=False, vecor_weight=0.1, vecor_temperature=0.1, use_haste=False, haste_termination_step=None, haste_termination_type='linear')

Bases: object

Configuration for conditional training.

Variables:

• conditioning_config – Model conditioning capabilities

• class_dropout_prob – CFG dropout probability for class labels

• text_dropout_prob – CFG dropout probability for text

• use_repa – Whether to use REPA alignment loss

• repa_weight – Weight for REPA loss

• repa_timestep_threshold – Only apply REPA at t > threshold

• use_vecor – Whether to use VeCoR contrastive loss

• vecor_weight – Weight for VeCoR loss

• use_haste – Whether to use HASTE termination scheduling

• haste_termination_step – Step to terminate alignment (for HASTE)

Parameters:

• conditioning_config (ConditioningConfig | None, default: None)

• class_dropout_prob (float, default: 0.1)

• text_dropout_prob (float, default: 0.1)

• use_repa (bool, default: False)

• repa_weight (float, default: 0.5)

• repa_timestep_threshold (float, default: 0.5)

• repa_encoder_name (str, default: 'dinov2_vitb14')

• use_vecor (bool, default: False)

• vecor_weight (float, default: 0.1)

• vecor_temperature (float, default: 0.1)

• use_haste (bool, default: False)

• haste_termination_step (int | None, default: None)

• haste_termination_type (Literal['hard', 'linear', 'cosine'], default: 'linear')

conditioning_config: medlatents.conditioning.bundle.ConditioningConfig | None = None

class_dropout_prob: float = 0.1

text_dropout_prob: float = 0.1

use_repa: bool = False

repa_weight: float = 0.5

repa_timestep_threshold: float = 0.5

repa_encoder_name: str = 'dinov2_vitb14'

use_vecor: bool = False

vecor_weight: float = 0.1

vecor_temperature: float = 0.1

use_haste: bool = False

haste_termination_step: int | None = None

haste_termination_type: Literal['hard', 'linear', 'cosine'] = 'linear'

__init__(conditioning_config=None, class_dropout_prob=0.1, text_dropout_prob=0.1, use_repa=False, repa_weight=0.5, repa_timestep_threshold=0.5, repa_encoder_name='dinov2_vitb14', use_vecor=False, vecor_weight=0.1, vecor_temperature=0.1, use_haste=False, haste_termination_step=None, haste_termination_type='linear')

Parameters:

• conditioning_config (ConditioningConfig | None, default: None)

• class_dropout_prob (float, default: 0.1)

• text_dropout_prob (float, default: 0.1)

• use_repa (bool, default: False)

• repa_weight (float, default: 0.5)

• repa_timestep_threshold (float, default: 0.5)

• repa_encoder_name (str, default: 'dinov2_vitb14')

• use_vecor (bool, default: False)

• vecor_weight (float, default: 0.1)

• vecor_temperature (float, default: 0.1)

• use_haste (bool, default: False)

• haste_termination_step (int | None, default: None)

• haste_termination_type (Literal['hard', 'linear', 'cosine'], default: 'linear')

class medlatents.training.ConditionalLossWrapper(base_loss_fn, config, repa_projection=None, frozen_encoder=None)

Bases: torch.nn.modules.module.Module

Wrapper that adds REPA and VeCoR losses to a base loss.

This wrapper can be added around any diffusion/flow loss to incorporate representation alignment (REPA) and velocity contrastive regularization (VeCoR) losses.

Example

>>> base_loss = MSELoss()
>>> wrapper = ConditionalLossWrapper(
...     base_loss_fn=base_loss,
...     config=ConditionalTrainingConfig(use_repa=True),
...     repa_projection=REPAProjection(1024, 768),
...     frozen_encoder=encoder,
... )
>>> loss = wrapper(pred, target, hidden_states=hidden, original=x, timesteps=t)

Parameters:

• base_loss_fn (Module)

• config (ConditionalTrainingConfig)

• repa_projection (Module | None, default: None)

• frozen_encoder (Module | None, default: None)

__init__(base_loss_fn, config, repa_projection=None, frozen_encoder=None)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

• base_loss_fn (Module)

• config (ConditionalTrainingConfig)

• repa_projection (Module | None, default: None)

• frozen_encoder (Module | None, default: None)

forward(pred, target, hidden_states=None, original_data=None, timesteps=None, **base_loss_kwargs)

Compute combined loss with optional REPA and VeCoR.

Parameters:

• pred (Tensor) – Model prediction (velocity, noise, or x0)

• target (Tensor) – Target for denoising loss

• hidden_states (Tensor | None, default: None) – DiT hidden states for REPA (optional)

• original_data (Tensor | None, default: None) – Clean data for REPA encoder (optional)

• timesteps (Tensor | None, default: None) – Current timesteps for REPA threshold (optional)

• **base_loss_kwargs – Additional args for base loss

Return type:

dict[str, Tensor]

Returns:

Dict with ‘total’, ‘base’, and optional ‘repa’, ‘vecor’ losses

step()

Increment internal step counter (call after each optimizer step).

Return type:

None

T_destination = ~T_destination

add_module(name, module)

Add a child module to the current module.

The module can be accessed as an attribute using the given name.

Parameters:

• name (str) – name of the child module. The child module can be accessed from this module using the given name

• module (Module) – child module to be added to the module.

Return type:

None

apply(fn)

Apply fn recursively to every submodule (as returned by .children()) as well as self.

Typical use includes initializing the parameters of a model (see also torch.nn.init).

Parameters:

fn (Module -> None) – function to be applied to each submodule

Returns:

self

Return type:

Module

Example:

>>> @torch.no_grad()
>>> def init_weights(m):
>>>     print(m)
>>>     if type(m) is nn.Linear:
>>>         m.weight.fill_(1.0)
>>>         print(m.weight)
>>> net = nn.Sequential(nn.Linear(2, 2), nn.Linear(2, 2))
>>> net.apply(init_weights)
Linear(in_features=2, out_features=2, bias=True)
Parameter containing:
tensor([[1., 1.],
        [1., 1.]], requires_grad=True)
Linear(in_features=2, out_features=2, bias=True)
Parameter containing:
tensor([[1., 1.],
        [1., 1.]], requires_grad=True)
Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
)

bfloat16()

Casts all floating point parameters and buffers to bfloat16 datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

buffers(recurse=True)

Return an iterator over module buffers.

Parameters:

recurse (bool) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module.

Yields:

torch.Tensor – module buffer

Return type:

Iterator[Tensor]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for buf in model.buffers():
>>>     print(type(buf), buf.size())
<class 'torch.Tensor'> (20L,)
<class 'torch.Tensor'> (20L, 1L, 5L, 5L)

call_super_init: bool = False

children()

Return an iterator over immediate children modules.

Yields:

Module – a child module

Return type:

Iterator[Module]

compile(*args, **kwargs)

Compile this Module’s forward using torch.compile().

This Module’s __call__ method is compiled and all arguments are passed as-is to torch.compile().

See torch.compile() for details on the arguments for this function.

Return type:

None

cpu()

Move all model parameters and buffers to the CPU.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

cuda(device=None)

Move all model parameters and buffers to the GPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on GPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

double()

Casts all floating point parameters and buffers to double datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

dump_patches: bool = False

eval()

Set the module in evaluation mode.

This has an effect only on certain modules. See the documentation of particular modules for details of their behaviors in training/evaluation mode, i.e. whether they are affected, e.g. Dropout, BatchNorm, etc.

This is equivalent with self.train(False).

See Locally disabling gradient computation for a comparison between .eval() and several similar mechanisms that may be confused with it.

Returns:

self

Return type:

Module

extra_repr()

Return the extra representation of the module.

To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.

Return type:

str

float()

Casts all floating point parameters and buffers to float datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

get_buffer(target)

Return the buffer given by target if it exists, otherwise throw an error.

See the docstring for get_submodule for a more detailed explanation of this method’s functionality as well as how to correctly specify target.

Parameters:

target (str) – The fully-qualified string name of the buffer to look for. (See get_submodule for how to specify a fully-qualified string.)

Returns:

The buffer referenced by target

Return type:

torch.Tensor

Raises:

AttributeError – If the target string references an invalid path or resolves to something that is not a buffer

get_extra_state()

Return any extra state to include in the module’s state_dict.

Implement this and a corresponding set_extra_state() for your module if you need to store extra state. This function is called when building the module’s state_dict().

Note that extra state should be picklable to ensure working serialization of the state_dict. We only provide backwards compatibility guarantees for serializing Tensors; other objects may break backwards compatibility if their serialized pickled form changes.

Returns:

Any extra state to store in the module’s state_dict

Return type:

object

get_parameter(target)

Return the parameter given by target if it exists, otherwise throw an error.

See the docstring for get_submodule for a more detailed explanation of this method’s functionality as well as how to correctly specify target.

Parameters:

target (str) – The fully-qualified string name of the Parameter to look for. (See get_submodule for how to specify a fully-qualified string.)

Returns:

The Parameter referenced by target

Return type:

torch.nn.Parameter

Raises:

AttributeError – If the target string references an invalid path or resolves to something that is not an nn.Parameter

get_submodule(target)

Return the submodule given by target if it exists, otherwise throw an error.

For example, let’s say you have an nn.Module A that looks like this:

A(
    (net_b): Module(
        (net_c): Module(
            (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2))
        )
        (linear): Linear(in_features=100, out_features=200, bias=True)
    )
)

(The diagram shows an nn.Module A. A which has a nested submodule net_b, which itself has two submodules net_c and linear. net_c then has a submodule conv.)

To check whether or not we have the linear submodule, we would call get_submodule("net_b.linear"). To check whether we have the conv submodule, we would call get_submodule("net_b.net_c.conv").

The runtime of get_submodule is bounded by the degree of module nesting in target. A query against named_modules achieves the same result, but it is O(N) in the number of transitive modules. So, for a simple check to see if some submodule exists, get_submodule should always be used.

Parameters:

target (str) – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)

Returns:

The submodule referenced by target

Return type:

torch.nn.Module

Raises:

AttributeError – If at any point along the path resulting from the target string the (sub)path resolves to a non-existent attribute name or an object that is not an instance of nn.Module.

half()

Casts all floating point parameters and buffers to half datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

ipu(device=None)

Move all model parameters and buffers to the IPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on IPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

load_state_dict(state_dict, strict=True, assign=False)

Copy parameters and buffers from state_dict into this module and its descendants.

If strict is True, then the keys of state_dict must exactly match the keys returned by this module’s state_dict() function.

Warning

If assign is True the optimizer must be created after the call to load_state_dict unless get_swap_module_params_on_conversion() is True.

Parameters:

• state_dict (dict) – a dict containing parameters and persistent buffers.

• strict (bool, optional) – whether to strictly enforce that the keys in state_dict match the keys returned by this module’s state_dict() function. Default: True

• assign (bool, optional) – When set to False, the properties of the tensors in the current module are preserved whereas setting it to True preserves properties of the Tensors in the state dict. The only exception is the requires_grad field of Parameter for which the value from the module is preserved. Default: False

Returns:

• missing_keys is a list of str containing any keys that are expected

  by this module but missing from the provided state_dict.

• unexpected_keys is a list of str containing the keys that are not

  expected by this module but present in the provided state_dict.

Return type:

NamedTuple with missing_keys and unexpected_keys fields

Note

If a parameter or buffer is registered as None and its corresponding key exists in state_dict, load_state_dict() will raise a RuntimeError.

modules(remove_duplicate=True)

Return an iterator over all modules in the network.

Parameters:

remove_duplicate (bool, default: True) – whether to remove the duplicated module instances in the result or not.

Yields:

Module – a module in the network

Return type:

Iterator[Module]

Note

Duplicate modules are returned only once by default. In the following example, l will be returned only once.

Example:

>>> l = nn.Linear(2, 2)
>>> net = nn.Sequential(l, l)
>>> for idx, m in enumerate(net.modules()):
...     print(idx, '->', m)

0 -> Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
)
1 -> Linear(in_features=2, out_features=2, bias=True)

mtia(device=None)

Move all model parameters and buffers to the MTIA.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on MTIA while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

named_buffers(prefix='', recurse=True, remove_duplicate=True)

Return an iterator over module buffers, yielding both the name of the buffer as well as the buffer itself.

Parameters:

• prefix (str) – prefix to prepend to all buffer names.

• recurse (bool, optional) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module. Defaults to True.

• remove_duplicate (bool, optional) – whether to remove the duplicated buffers in the result. Defaults to True.

Yields:

(str, torch.Tensor) – Tuple containing the name and buffer

Return type:

Iterator[tuple[str, Tensor]]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, buf in self.named_buffers():
>>>     if name in ['running_var']:
>>>         print(buf.size())

named_children()

Return an iterator over immediate children modules, yielding both the name of the module as well as the module itself.

Yields:

(str, Module) – Tuple containing a name and child module

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, module in model.named_children():
>>>     if name in ['conv4', 'conv5']:
>>>         print(module)

Return type:

Iterator[tuple[str, Module]]

named_modules(memo=None, prefix='', remove_duplicate=True)

Return an iterator over all modules in the network, yielding both the name of the module as well as the module itself.

Parameters:

• memo (set[Module] | None, default: None) – a memo to store the set of modules already added to the result

• prefix (str, default: '') – a prefix that will be added to the name of the module

• remove_duplicate (bool, default: True) – whether to remove the duplicated module instances in the result or not

Yields:

(str, Module) – Tuple of name and module

Note

Duplicate modules are returned only once. In the following example, l will be returned only once.

Example:

>>> l = nn.Linear(2, 2)
>>> net = nn.Sequential(l, l)
>>> for idx, m in enumerate(net.named_modules()):
...     print(idx, '->', m)

0 -> ('', Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
))
1 -> ('0', Linear(in_features=2, out_features=2, bias=True))

named_parameters(prefix='', recurse=True, remove_duplicate=True)

Return an iterator over module parameters, yielding both the name of the parameter as well as the parameter itself.

Parameters:

• prefix (str) – prefix to prepend to all parameter names.

• recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.

• remove_duplicate (bool, optional) – whether to remove the duplicated parameters in the result. Defaults to True.

Yields:

(str, Parameter) – Tuple containing the name and parameter

Return type:

Iterator[tuple[str, Parameter]]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, param in self.named_parameters():
>>>     if name in ['bias']:
>>>         print(param.size())

parameters(recurse=True)

Return an iterator over module parameters.

This is typically passed to an optimizer.

Parameters:

recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.

Yields:

Parameter – module parameter

Return type:

Iterator[Parameter]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for param in model.parameters():
>>>     print(type(param), param.size())
<class 'torch.Tensor'> (20L,)
<class 'torch.Tensor'> (20L, 1L, 5L, 5L)

register_backward_hook(hook)

Register a backward hook on the module.

This function is deprecated in favor of register_full_backward_hook() and the behavior of this function will change in future versions.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

Parameters:

hook (Callable[[Module, tuple[Tensor, ...] | Tensor, tuple[Tensor, ...] | Tensor], tuple[Tensor, ...] | Tensor | None])

register_buffer(name, tensor, persistent=True)

Add a buffer to the module.

This is typically used to register a buffer that should not be considered a model parameter. For example, BatchNorm’s running_mean is not a parameter, but is part of the module’s state. Buffers, by default, are persistent and will be saved alongside parameters. This behavior can be changed by setting persistent to False. The only difference between a persistent buffer and a non-persistent buffer is that the latter will not be a part of this module’s state_dict.

Buffers can be accessed as attributes using given names.

Parameters:

• name (str) – name of the buffer. The buffer can be accessed from this module using the given name

• tensor (Tensor or None) – buffer to be registered. If None, then operations that run on buffers, such as cuda, are ignored. If None, the buffer is not included in the module’s state_dict.

• persistent (bool) – whether the buffer is part of this module’s state_dict.

Return type:

None

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> self.register_buffer('running_mean', torch.zeros(num_features))

register_forward_hook(hook, *, prepend=False, with_kwargs=False, always_call=False)

Register a forward hook on the module.

The hook will be called every time after forward() has computed an output.

If with_kwargs is False or not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to the forward. The hook can modify the output. It can modify the input inplace but it will not have effect on forward since this is called after forward() is called. The hook should have the following signature:

hook(module, args, output) -> None or modified output

If with_kwargs is True, the forward hook will be passed the kwargs given to the forward function and be expected to return the output possibly modified. The hook should have the following signature:

hook(module, args, kwargs, output) -> None or modified output

Parameters:

• hook (Callable) – The user defined hook to be registered.

• prepend (bool) – If True, the provided hook will be fired before all existing forward hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing forward hooks on this torch.nn.Module. Note that global forward hooks registered with register_module_forward_hook() will fire before all hooks registered by this method. Default: False

• with_kwargs (bool) – If True, the hook will be passed the kwargs given to the forward function. Default: False

• always_call (bool) – If True the hook will be run regardless of whether an exception is raised while calling the Module. Default: False

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_forward_pre_hook(hook, *, prepend=False, with_kwargs=False)

Register a forward pre-hook on the module.

The hook will be called every time before forward() is invoked.

If with_kwargs is false or not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to the forward. The hook can modify the input. User can either return a tuple or a single modified value in the hook. We will wrap the value into a tuple if a single value is returned (unless that value is already a tuple). The hook should have the following signature:

hook(module, args) -> None or modified input

If with_kwargs is true, the forward pre-hook will be passed the kwargs given to the forward function. And if the hook modifies the input, both the args and kwargs should be returned. The hook should have the following signature:

hook(module, args, kwargs) -> None or a tuple of modified input and kwargs

Parameters:

• hook (Callable) – The user defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing forward_pre hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing forward_pre hooks on this torch.nn.Module. Note that global forward_pre hooks registered with register_module_forward_pre_hook() will fire before all hooks registered by this method. Default: False

• with_kwargs (bool) – If true, the hook will be passed the kwargs given to the forward function. Default: False

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_full_backward_hook(hook, prepend=False)

Register a backward hook on the module.

The hook will be called every time the gradients with respect to a module are computed, and its firing rules are as follows:

1. Ordinarily, the hook fires when the gradients are computed with respect to the module inputs.

2. If none of the module inputs require gradients, the hook will fire when the gradients are computed with respect to module outputs.

3. If none of the module outputs require gradients, then the hooks will not fire.

The hook should have the following signature:

hook(module, grad_input, grad_output) -> tuple(Tensor) or None

The grad_input and grad_output are tuples that contain the gradients with respect to the inputs and outputs respectively. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the input that will be used in place of grad_input in subsequent computations. grad_input will only correspond to the inputs given as positional arguments and all kwarg arguments are ignored. Entries in grad_input and grad_output will be None for all non-Tensor arguments.

For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.

Warning

Modifying inputs or outputs inplace is not allowed when using backward hooks and will raise an error.

Parameters:

• hook (Callable) – The user-defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing backward hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing backward hooks on this torch.nn.Module. Note that global backward hooks registered with register_module_full_backward_hook() will fire before all hooks registered by this method.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_full_backward_pre_hook(hook, prepend=False)

Register a backward pre-hook on the module.

The hook will be called every time the gradients for the module are computed. The hook should have the following signature:

hook(module, grad_output) -> tuple[Tensor, ...], Tensor or None

The grad_output is a tuple. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the output that will be used in place of grad_output in subsequent computations. Entries in grad_output will be None for all non-Tensor arguments.

For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.

Warning

Modifying inputs inplace is not allowed when using backward hooks and will raise an error.

Parameters:

• hook (Callable) – The user-defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing backward_pre hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing backward_pre hooks on this torch.nn.Module. Note that global backward_pre hooks registered with register_module_full_backward_pre_hook() will fire before all hooks registered by this method.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_load_state_dict_post_hook(hook)

Register a post-hook to be run after module’s load_state_dict() is called.

It should have the following signature::

hook(module, incompatible_keys) -> None

The module argument is the current module that this hook is registered on, and the incompatible_keys argument is a NamedTuple consisting of attributes missing_keys and unexpected_keys. missing_keys is a list of str containing the missing keys and unexpected_keys is a list of str containing the unexpected keys.

The given incompatible_keys can be modified inplace if needed.

Note that the checks performed when calling load_state_dict() with strict=True are affected by modifications the hook makes to missing_keys or unexpected_keys, as expected. Additions to either set of keys will result in an error being thrown when strict=True, and clearing out both missing and unexpected keys will avoid an error.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_load_state_dict_pre_hook(hook)

Register a pre-hook to be run before module’s load_state_dict() is called.

It should have the following signature::

hook(module, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) -> None # noqa: B950

Parameters:

hook (Callable) – Callable hook that will be invoked before loading the state dict.

register_module(name, module)

Alias for add_module().

Parameters:

• name (str)

• module (Optional[Module])

Return type:

None

register_parameter(name, param)

Add a parameter to the module.

The parameter can be accessed as an attribute using given name.

Parameters:

• name (str) – name of the parameter. The parameter can be accessed from this module using the given name

• param (Parameter or None) – parameter to be added to the module. If None, then operations that run on parameters, such as cuda, are ignored. If None, the parameter is not included in the module’s state_dict.

Return type:

None

register_state_dict_post_hook(hook)

Register a post-hook for the state_dict() method.

It should have the following signature::

hook(module, state_dict, prefix, local_metadata) -> None

The registered hooks can modify the state_dict inplace.

register_state_dict_pre_hook(hook)

Register a pre-hook for the state_dict() method.

It should have the following signature::

hook(module, prefix, keep_vars) -> None

The registered hooks can be used to perform pre-processing before the state_dict call is made.

requires_grad_(requires_grad=True)

Change if autograd should record operations on parameters in this module.

This method sets the parameters’ requires_grad attributes in-place.

This method is helpful for freezing part of the module for finetuning or training parts of a model individually (e.g., GAN training).

See Locally disabling gradient computation for a comparison between .requires_grad_() and several similar mechanisms that may be confused with it.

Parameters:

requires_grad (bool) – whether autograd should record operations on parameters in this module. Default: True.

Returns:

self

Return type:

Module

set_extra_state(state)

Set extra state contained in the loaded state_dict.

This function is called from load_state_dict() to handle any extra state found within the state_dict. Implement this function and a corresponding get_extra_state() for your module if you need to store extra state within its state_dict.

Parameters:

state (dict) – Extra state from the state_dict

Return type:

None

set_submodule(target, module, strict=False)

Set the submodule given by target if it exists, otherwise throw an error.

Note

If strict is set to False (default), the method will replace an existing submodule or create a new submodule if the parent module exists. If strict is set to True, the method will only attempt to replace an existing submodule and throw an error if the submodule does not exist.

For example, let’s say you have an nn.Module A that looks like this:

A(
    (net_b): Module(
        (net_c): Module(
            (conv): Conv2d(3, 3, 3)
        )
        (linear): Linear(3, 3)
    )
)

(The diagram shows an nn.Module A. A has a nested submodule net_b, which itself has two submodules net_c and linear. net_c then has a submodule conv.)

To override the Conv2d with a new submodule Linear, you could call set_submodule("net_b.net_c.conv", nn.Linear(1, 1)) where strict could be True or False

To add a new submodule Conv2d to the existing net_b module, you would call set_submodule("net_b.conv", nn.Conv2d(1, 1, 1)).

In the above if you set strict=True and call set_submodule("net_b.conv", nn.Conv2d(1, 1, 1), strict=True), an AttributeError will be raised because net_b does not have a submodule named conv.

Parameters:

• target (str) – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)

• module (Module) – The module to set the submodule to.

• strict (bool, default: False) – If False, the method will replace an existing submodule or create a new submodule if the parent module exists. If True, the method will only attempt to replace an existing submodule and throw an error if the submodule doesn’t already exist.

Raises:

• ValueError – If the target string is empty or if module is not an instance of nn.Module.

• AttributeError – If at any point along the path resulting from the target string the (sub)path resolves to a non-existent attribute name or an object that is not an instance of nn.Module.

Return type:

None

share_memory()

See torch.Tensor.share_memory_().

Return type:

Self

state_dict(*args, destination=None, prefix='', keep_vars=False)

Return a dictionary containing references to the whole state of the module.

Both parameters and persistent buffers (e.g. running averages) are included. Keys are corresponding parameter and buffer names. Parameters and buffers set to None are not included.

Note

The returned object is a shallow copy. It contains references to the module’s parameters and buffers.

Warning

Currently state_dict() also accepts positional arguments for destination, prefix and keep_vars in order. However, this is being deprecated and keyword arguments will be enforced in future releases.

Warning

Please avoid the use of argument destination as it is not designed for end-users.

Parameters:

• destination (dict, optional) – If provided, the state of module will be updated into the dict and the same object is returned. Otherwise, an OrderedDict will be created and returned. Default: None.

• prefix (str, optional) – a prefix added to parameter and buffer names to compose the keys in state_dict. Default: ''.

• keep_vars (bool, optional) – by default the Tensor s returned in the state dict are detached from autograd. If it’s set to True, detaching will not be performed. Default: False.

Returns:

a dictionary containing a whole state of the module

Return type:

dict

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> module.state_dict().keys()
['bias', 'weight']

to(*args, **kwargs)

Move and/or cast the parameters and buffers.

This can be called as

to(device=None, dtype=None, non_blocking=False)

to(dtype, non_blocking=False)

to(tensor, non_blocking=False)

to(memory_format=torch.channels_last)

Its signature is similar to torch.Tensor.to(), but only accepts floating point or complex dtypes. In addition, this method will only cast the floating point or complex parameters and buffers to dtype (if given). The integral parameters and buffers will be moved device, if that is given, but with dtypes unchanged. When non_blocking is set, it tries to convert/move asynchronously with respect to the host if possible, e.g., moving CPU Tensors with pinned memory to CUDA devices.

See below for examples.

Note

This method modifies the module in-place.

Parameters:

• device (torch.device) – the desired device of the parameters and buffers in this module

• dtype (torch.dtype) – the desired floating point or complex dtype of the parameters and buffers in this module

• tensor (torch.Tensor) – Tensor whose dtype and device are the desired dtype and device for all parameters and buffers in this module

• memory_format (torch.memory_format) – the desired memory format for 4D parameters and buffers in this module (keyword only argument)

Returns:

self

Return type:

Module

Examples:

>>> # xdoctest: +IGNORE_WANT("non-deterministic")
>>> linear = nn.Linear(2, 2)
>>> linear.weight
Parameter containing:
tensor([[ 0.1913, -0.3420],
        [-0.5113, -0.2325]])
>>> linear.to(torch.double)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1913, -0.3420],
        [-0.5113, -0.2325]], dtype=torch.float64)
>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CUDA1)
>>> gpu1 = torch.device("cuda:1")
>>> linear.to(gpu1, dtype=torch.half, non_blocking=True)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1914, -0.3420],
        [-0.5112, -0.2324]], dtype=torch.float16, device='cuda:1')
>>> cpu = torch.device("cpu")
>>> linear.to(cpu)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1914, -0.3420],
        [-0.5112, -0.2324]], dtype=torch.float16)

>>> linear = nn.Linear(2, 2, bias=None).to(torch.cdouble)
>>> linear.weight
Parameter containing:
tensor([[ 0.3741+0.j,  0.2382+0.j],
        [ 0.5593+0.j, -0.4443+0.j]], dtype=torch.complex128)
>>> linear(torch.ones(3, 2, dtype=torch.cdouble))
tensor([[0.6122+0.j, 0.1150+0.j],
        [0.6122+0.j, 0.1150+0.j],
        [0.6122+0.j, 0.1150+0.j]], dtype=torch.complex128)

to_empty(*, device, recurse=True)

Move the parameters and buffers to the specified device without copying storage.

Parameters:

• device (torch.device) – The desired device of the parameters and buffers in this module.

• recurse (bool) – Whether parameters and buffers of submodules should be recursively moved to the specified device.

Returns:

self

Return type:

Module

train(mode=True)

Set the module in training mode.

This has an effect only on certain modules. See the documentation of particular modules for details of their behaviors in training/evaluation mode, i.e., whether they are affected, e.g. Dropout, BatchNorm, etc.

Parameters:

mode (bool) – whether to set training mode (True) or evaluation mode (False). Default: True.

Returns:

self

Return type:

Module

type(dst_type)

Casts all parameters and buffers to dst_type.

Note

This method modifies the module in-place.

Parameters:

dst_type (type or string) – the desired type

Returns:

self

Return type:

Module

xpu(device=None)

Move all model parameters and buffers to the XPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing optimizer if the module will live on XPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

zero_grad(set_to_none=True)

Reset gradients of all model parameters.

See similar function under torch.optim.Optimizer for more context.

Parameters:

set_to_none (bool) – instead of setting to zero, set the grads to None. See torch.optim.Optimizer.zero_grad() for details.

Return type:

None

training: bool

medlatents.training.prepare_batch_with_conditioning(batch, timesteps, config, training=True)

Prepare a batch for conditional training.

Extracts data and conditioning from the batch, creates a ConditioningBundle, and optionally applies CFG dropout during training.

Parameters:

• batch (dict[str, Tensor] | tuple | Tensor) – Input batch (dict, tuple, or tensor)

• timesteps (Tensor) – Sampled timesteps for diffusion/flow

• config (ConditionalTrainingConfig) – Training configuration

• training (bool, default: True) – Whether in training mode (applies dropout)

Return type:

tuple[Tensor, ConditioningBundle]

Returns:

Tuple of (data_tensor, conditioning_bundle)

medlatents.training.compute_conditional_loss(model, data, bundle, noise_process, loss_fn, return_hidden=False)

Compute loss for conditional diffusion/flow training.

This is a high-level utility that handles: 1. Sampling noisy data from the noise process 2. Forward pass with conditioning 3. Loss computation (optionally with REPA/VeCoR)

Parameters:

• model (Module) – Denoising model

• data (Tensor) – Clean data tensor

• bundle (ConditioningBundle) – Conditioning bundle

• noise_process (Module) – Diffusion or flow matching process

• loss_fn (Module | ConditionalLossWrapper) – Loss function (optionally ConditionalLossWrapper)

• return_hidden (bool, default: False) – Whether to return hidden states (for REPA)

Return type:

Tensor | tuple[Tensor, dict]

Returns:

Loss tensor, or (loss, info_dict) if return_hidden=True

Representation Alignment (REPA / REPA-E)

class medlatents.training.REPALoss(hidden_dim, target_dim=768, proj_dim=256, timestep_threshold=0.5, weight=0.5, normalize=True, loss_type='cosine', num_proj_layers=2)

Bases: torch.nn.modules.module.Module

REPA: Representation Alignment Loss for accelerated diffusion training.

Aligns DiT hidden states with frozen pretrained encoder features using cosine similarity. Applied only at high-noise timesteps where semantic alignment provides the most benefit.

The key insight from REPA is that early in denoising (high t), the model benefits from semantic guidance from a pretrained encoder. As denoising progresses (low t), the model needs to focus on fine details.

Reference:

Yu et al., “Representation Alignment for Generation” (ICLR’25 Oral) https://arxiv.org/abs/2410.06940

Parameters:

• hidden_dim (int) – DiT hidden dimension

• target_dim (int, default: 768) – Target encoder dimension

• proj_dim (int, default: 256) – Projection dimension for alignment

• timestep_threshold (float, default: 0.5) – Only apply at t > threshold (default 0.5)

• weight (float, default: 0.5) – Loss weight relative to denoising loss

• normalize (bool, default: True) – Whether to L2-normalize before similarity

• loss_type (Literal['cosine', 'mse'], default: 'cosine') – ‘cosine’ or ‘mse’

Usage:

repa = REPALoss(hidden_dim=1024, target_dim=768)

# In training loop: hidden_states = extract_hidden_from_dit(model, x_t, t) target_features = frozen_encoder(clean_images)

repa_loss = repa(hidden_states, target_features, t) total_loss = denoising_loss + repa_loss

Parameters:

num_proj_layers (int, default: 2)

__init__(hidden_dim, target_dim=768, proj_dim=256, timestep_threshold=0.5, weight=0.5, normalize=True, loss_type='cosine', num_proj_layers=2)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

• hidden_dim (int)

• target_dim (int, default: 768)

• proj_dim (int, default: 256)

• timestep_threshold (float, default: 0.5)

• weight (float, default: 0.5)

• normalize (bool, default: True)

• loss_type (Literal['cosine', 'mse'], default: 'cosine')

• num_proj_layers (int, default: 2)

forward(hidden_states, target_features, timesteps)

Compute REPA alignment loss.

Parameters:

• hidden_states (Tensor) – DiT hidden states [batch, seq, hidden_dim]

• target_features (Tensor) – Encoder features [batch, seq, target_dim]

• timesteps (Tensor) – Current timesteps [batch], in [0, 1]

Return type:

Tensor

Returns:

Weighted alignment loss (scalar)

T_destination = ~T_destination

add_module(name, module)

Add a child module to the current module.

The module can be accessed as an attribute using the given name.

Parameters:

• name (str) – name of the child module. The child module can be accessed from this module using the given name

• module (Module) – child module to be added to the module.

Return type:

None

apply(fn)

Apply fn recursively to every submodule (as returned by .children()) as well as self.

Typical use includes initializing the parameters of a model (see also torch.nn.init).

Parameters:

fn (Module -> None) – function to be applied to each submodule

Returns:

self

Return type:

Module

Example:

>>> @torch.no_grad()
>>> def init_weights(m):
>>>     print(m)
>>>     if type(m) is nn.Linear:
>>>         m.weight.fill_(1.0)
>>>         print(m.weight)
>>> net = nn.Sequential(nn.Linear(2, 2), nn.Linear(2, 2))
>>> net.apply(init_weights)
Linear(in_features=2, out_features=2, bias=True)
Parameter containing:
tensor([[1., 1.],
        [1., 1.]], requires_grad=True)
Linear(in_features=2, out_features=2, bias=True)
Parameter containing:
tensor([[1., 1.],
        [1., 1.]], requires_grad=True)
Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
)

bfloat16()

Casts all floating point parameters and buffers to bfloat16 datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

buffers(recurse=True)

Return an iterator over module buffers.

Parameters:

recurse (bool) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module.

Yields:

torch.Tensor – module buffer

Return type:

Iterator[Tensor]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for buf in model.buffers():
>>>     print(type(buf), buf.size())
<class 'torch.Tensor'> (20L,)
<class 'torch.Tensor'> (20L, 1L, 5L, 5L)

call_super_init: bool = False

children()

Return an iterator over immediate children modules.

Yields:

Module – a child module

Return type:

Iterator[Module]

compile(*args, **kwargs)

Compile this Module’s forward using torch.compile().

This Module’s __call__ method is compiled and all arguments are passed as-is to torch.compile().

See torch.compile() for details on the arguments for this function.

Return type:

None

cpu()

Move all model parameters and buffers to the CPU.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

cuda(device=None)

Move all model parameters and buffers to the GPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on GPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

double()

Casts all floating point parameters and buffers to double datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

dump_patches: bool = False

eval()

Set the module in evaluation mode.

This has an effect only on certain modules. See the documentation of particular modules for details of their behaviors in training/evaluation mode, i.e. whether they are affected, e.g. Dropout, BatchNorm, etc.

This is equivalent with self.train(False).

See Locally disabling gradient computation for a comparison between .eval() and several similar mechanisms that may be confused with it.

Returns:

self

Return type:

Module

extra_repr()

Return the extra representation of the module.

To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.

Return type:

str

float()

Casts all floating point parameters and buffers to float datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

get_buffer(target)

Return the buffer given by target if it exists, otherwise throw an error.

See the docstring for get_submodule for a more detailed explanation of this method’s functionality as well as how to correctly specify target.

Parameters:

target (str) – The fully-qualified string name of the buffer to look for. (See get_submodule for how to specify a fully-qualified string.)

Returns:

The buffer referenced by target

Return type:

torch.Tensor

Raises:

AttributeError – If the target string references an invalid path or resolves to something that is not a buffer

get_extra_state()

Return any extra state to include in the module’s state_dict.

Implement this and a corresponding set_extra_state() for your module if you need to store extra state. This function is called when building the module’s state_dict().

Note that extra state should be picklable to ensure working serialization of the state_dict. We only provide backwards compatibility guarantees for serializing Tensors; other objects may break backwards compatibility if their serialized pickled form changes.

Returns:

Any extra state to store in the module’s state_dict

Return type:

object

get_parameter(target)

Return the parameter given by target if it exists, otherwise throw an error.

See the docstring for get_submodule for a more detailed explanation of this method’s functionality as well as how to correctly specify target.

Parameters:

target (str) – The fully-qualified string name of the Parameter to look for. (See get_submodule for how to specify a fully-qualified string.)

Returns:

The Parameter referenced by target

Return type:

torch.nn.Parameter

Raises:

AttributeError – If the target string references an invalid path or resolves to something that is not an nn.Parameter

get_submodule(target)

Return the submodule given by target if it exists, otherwise throw an error.

For example, let’s say you have an nn.Module A that looks like this:

A(
    (net_b): Module(
        (net_c): Module(
            (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2))
        )
        (linear): Linear(in_features=100, out_features=200, bias=True)
    )
)

(The diagram shows an nn.Module A. A which has a nested submodule net_b, which itself has two submodules net_c and linear. net_c then has a submodule conv.)

To check whether or not we have the linear submodule, we would call get_submodule("net_b.linear"). To check whether we have the conv submodule, we would call get_submodule("net_b.net_c.conv").

The runtime of get_submodule is bounded by the degree of module nesting in target. A query against named_modules achieves the same result, but it is O(N) in the number of transitive modules. So, for a simple check to see if some submodule exists, get_submodule should always be used.

Parameters:

target (str) – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)

Returns:

The submodule referenced by target

Return type:

torch.nn.Module

Raises:

AttributeError – If at any point along the path resulting from the target string the (sub)path resolves to a non-existent attribute name or an object that is not an instance of nn.Module.

half()

Casts all floating point parameters and buffers to half datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

ipu(device=None)

Move all model parameters and buffers to the IPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on IPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

load_state_dict(state_dict, strict=True, assign=False)

Copy parameters and buffers from state_dict into this module and its descendants.

If strict is True, then the keys of state_dict must exactly match the keys returned by this module’s state_dict() function.

Warning

If assign is True the optimizer must be created after the call to load_state_dict unless get_swap_module_params_on_conversion() is True.

Parameters:

• state_dict (dict) – a dict containing parameters and persistent buffers.

• strict (bool, optional) – whether to strictly enforce that the keys in state_dict match the keys returned by this module’s state_dict() function. Default: True

• assign (bool, optional) – When set to False, the properties of the tensors in the current module are preserved whereas setting it to True preserves properties of the Tensors in the state dict. The only exception is the requires_grad field of Parameter for which the value from the module is preserved. Default: False

Returns:

• missing_keys is a list of str containing any keys that are expected

  by this module but missing from the provided state_dict.

• unexpected_keys is a list of str containing the keys that are not

  expected by this module but present in the provided state_dict.

Return type:

NamedTuple with missing_keys and unexpected_keys fields

Note

If a parameter or buffer is registered as None and its corresponding key exists in state_dict, load_state_dict() will raise a RuntimeError.

modules(remove_duplicate=True)

Return an iterator over all modules in the network.

Parameters:

remove_duplicate (bool, default: True) – whether to remove the duplicated module instances in the result or not.

Yields:

Module – a module in the network

Return type:

Iterator[Module]

Note

Duplicate modules are returned only once by default. In the following example, l will be returned only once.

Example:

>>> l = nn.Linear(2, 2)
>>> net = nn.Sequential(l, l)
>>> for idx, m in enumerate(net.modules()):
...     print(idx, '->', m)

0 -> Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
)
1 -> Linear(in_features=2, out_features=2, bias=True)

mtia(device=None)

Move all model parameters and buffers to the MTIA.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on MTIA while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

named_buffers(prefix='', recurse=True, remove_duplicate=True)

Return an iterator over module buffers, yielding both the name of the buffer as well as the buffer itself.

Parameters:

• prefix (str) – prefix to prepend to all buffer names.

• recurse (bool, optional) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module. Defaults to True.

• remove_duplicate (bool, optional) – whether to remove the duplicated buffers in the result. Defaults to True.

Yields:

(str, torch.Tensor) – Tuple containing the name and buffer

Return type:

Iterator[tuple[str, Tensor]]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, buf in self.named_buffers():
>>>     if name in ['running_var']:
>>>         print(buf.size())

named_children()

Return an iterator over immediate children modules, yielding both the name of the module as well as the module itself.

Yields:

(str, Module) – Tuple containing a name and child module

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, module in model.named_children():
>>>     if name in ['conv4', 'conv5']:
>>>         print(module)

Return type:

Iterator[tuple[str, Module]]

named_modules(memo=None, prefix='', remove_duplicate=True)

Return an iterator over all modules in the network, yielding both the name of the module as well as the module itself.

Parameters:

• memo (set[Module] | None, default: None) – a memo to store the set of modules already added to the result

• prefix (str, default: '') – a prefix that will be added to the name of the module

• remove_duplicate (bool, default: True) – whether to remove the duplicated module instances in the result or not

Yields:

(str, Module) – Tuple of name and module

Note

Duplicate modules are returned only once. In the following example, l will be returned only once.

Example:

>>> l = nn.Linear(2, 2)
>>> net = nn.Sequential(l, l)
>>> for idx, m in enumerate(net.named_modules()):
...     print(idx, '->', m)

0 -> ('', Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
))
1 -> ('0', Linear(in_features=2, out_features=2, bias=True))

named_parameters(prefix='', recurse=True, remove_duplicate=True)

Return an iterator over module parameters, yielding both the name of the parameter as well as the parameter itself.

Parameters:

• prefix (str) – prefix to prepend to all parameter names.

• recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.

• remove_duplicate (bool, optional) – whether to remove the duplicated parameters in the result. Defaults to True.

Yields:

(str, Parameter) – Tuple containing the name and parameter

Return type:

Iterator[tuple[str, Parameter]]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, param in self.named_parameters():
>>>     if name in ['bias']:
>>>         print(param.size())

parameters(recurse=True)

Return an iterator over module parameters.

This is typically passed to an optimizer.

Parameters:

recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.

Yields:

Parameter – module parameter

Return type:

Iterator[Parameter]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for param in model.parameters():
>>>     print(type(param), param.size())
<class 'torch.Tensor'> (20L,)
<class 'torch.Tensor'> (20L, 1L, 5L, 5L)

register_backward_hook(hook)

Register a backward hook on the module.

This function is deprecated in favor of register_full_backward_hook() and the behavior of this function will change in future versions.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

Parameters:

hook (Callable[[Module, tuple[Tensor, ...] | Tensor, tuple[Tensor, ...] | Tensor], tuple[Tensor, ...] | Tensor | None])

register_buffer(name, tensor, persistent=True)

Add a buffer to the module.

This is typically used to register a buffer that should not be considered a model parameter. For example, BatchNorm’s running_mean is not a parameter, but is part of the module’s state. Buffers, by default, are persistent and will be saved alongside parameters. This behavior can be changed by setting persistent to False. The only difference between a persistent buffer and a non-persistent buffer is that the latter will not be a part of this module’s state_dict.

Buffers can be accessed as attributes using given names.

Parameters:

• name (str) – name of the buffer. The buffer can be accessed from this module using the given name

• tensor (Tensor or None) – buffer to be registered. If None, then operations that run on buffers, such as cuda, are ignored. If None, the buffer is not included in the module’s state_dict.

• persistent (bool) – whether the buffer is part of this module’s state_dict.

Return type:

None

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> self.register_buffer('running_mean', torch.zeros(num_features))

register_forward_hook(hook, *, prepend=False, with_kwargs=False, always_call=False)

Register a forward hook on the module.

The hook will be called every time after forward() has computed an output.

If with_kwargs is False or not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to the forward. The hook can modify the output. It can modify the input inplace but it will not have effect on forward since this is called after forward() is called. The hook should have the following signature:

hook(module, args, output) -> None or modified output

If with_kwargs is True, the forward hook will be passed the kwargs given to the forward function and be expected to return the output possibly modified. The hook should have the following signature:

hook(module, args, kwargs, output) -> None or modified output

Parameters:

• hook (Callable) – The user defined hook to be registered.

• prepend (bool) – If True, the provided hook will be fired before all existing forward hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing forward hooks on this torch.nn.Module. Note that global forward hooks registered with register_module_forward_hook() will fire before all hooks registered by this method. Default: False

• with_kwargs (bool) – If True, the hook will be passed the kwargs given to the forward function. Default: False

• always_call (bool) – If True the hook will be run regardless of whether an exception is raised while calling the Module. Default: False

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_forward_pre_hook(hook, *, prepend=False, with_kwargs=False)

Register a forward pre-hook on the module.

The hook will be called every time before forward() is invoked.

If with_kwargs is false or not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to the forward. The hook can modify the input. User can either return a tuple or a single modified value in the hook. We will wrap the value into a tuple if a single value is returned (unless that value is already a tuple). The hook should have the following signature:

hook(module, args) -> None or modified input

If with_kwargs is true, the forward pre-hook will be passed the kwargs given to the forward function. And if the hook modifies the input, both the args and kwargs should be returned. The hook should have the following signature:

hook(module, args, kwargs) -> None or a tuple of modified input and kwargs

Parameters:

• hook (Callable) – The user defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing forward_pre hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing forward_pre hooks on this torch.nn.Module. Note that global forward_pre hooks registered with register_module_forward_pre_hook() will fire before all hooks registered by this method. Default: False

• with_kwargs (bool) – If true, the hook will be passed the kwargs given to the forward function. Default: False

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_full_backward_hook(hook, prepend=False)

Register a backward hook on the module.

The hook will be called every time the gradients with respect to a module are computed, and its firing rules are as follows:

1. Ordinarily, the hook fires when the gradients are computed with respect to the module inputs.

2. If none of the module inputs require gradients, the hook will fire when the gradients are computed with respect to module outputs.

3. If none of the module outputs require gradients, then the hooks will not fire.

The hook should have the following signature:

hook(module, grad_input, grad_output) -> tuple(Tensor) or None

The grad_input and grad_output are tuples that contain the gradients with respect to the inputs and outputs respectively. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the input that will be used in place of grad_input in subsequent computations. grad_input will only correspond to the inputs given as positional arguments and all kwarg arguments are ignored. Entries in grad_input and grad_output will be None for all non-Tensor arguments.

For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.

Warning

Modifying inputs or outputs inplace is not allowed when using backward hooks and will raise an error.

Parameters:

• hook (Callable) – The user-defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing backward hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing backward hooks on this torch.nn.Module. Note that global backward hooks registered with register_module_full_backward_hook() will fire before all hooks registered by this method.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_full_backward_pre_hook(hook, prepend=False)

Register a backward pre-hook on the module.

The hook will be called every time the gradients for the module are computed. The hook should have the following signature:

hook(module, grad_output) -> tuple[Tensor, ...], Tensor or None

The grad_output is a tuple. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the output that will be used in place of grad_output in subsequent computations. Entries in grad_output will be None for all non-Tensor arguments.

For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.

Warning

Modifying inputs inplace is not allowed when using backward hooks and will raise an error.

Parameters:

• hook (Callable) – The user-defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing backward_pre hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing backward_pre hooks on this torch.nn.Module. Note that global backward_pre hooks registered with register_module_full_backward_pre_hook() will fire before all hooks registered by this method.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_load_state_dict_post_hook(hook)

Register a post-hook to be run after module’s load_state_dict() is called.

It should have the following signature::

hook(module, incompatible_keys) -> None

The module argument is the current module that this hook is registered on, and the incompatible_keys argument is a NamedTuple consisting of attributes missing_keys and unexpected_keys. missing_keys is a list of str containing the missing keys and unexpected_keys is a list of str containing the unexpected keys.

The given incompatible_keys can be modified inplace if needed.

Note that the checks performed when calling load_state_dict() with strict=True are affected by modifications the hook makes to missing_keys or unexpected_keys, as expected. Additions to either set of keys will result in an error being thrown when strict=True, and clearing out both missing and unexpected keys will avoid an error.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_load_state_dict_pre_hook(hook)

Register a pre-hook to be run before module’s load_state_dict() is called.

It should have the following signature::

hook(module, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) -> None # noqa: B950

Parameters:

hook (Callable) – Callable hook that will be invoked before loading the state dict.

register_module(name, module)

Alias for add_module().

Parameters:

• name (str)

• module (Optional[Module])

Return type:

None

register_parameter(name, param)

Add a parameter to the module.

The parameter can be accessed as an attribute using given name.

Parameters:

• name (str) – name of the parameter. The parameter can be accessed from this module using the given name

• param (Parameter or None) – parameter to be added to the module. If None, then operations that run on parameters, such as cuda, are ignored. If None, the parameter is not included in the module’s state_dict.

Return type:

None

register_state_dict_post_hook(hook)

Register a post-hook for the state_dict() method.

It should have the following signature::

hook(module, state_dict, prefix, local_metadata) -> None

The registered hooks can modify the state_dict inplace.

register_state_dict_pre_hook(hook)

Register a pre-hook for the state_dict() method.

It should have the following signature::

hook(module, prefix, keep_vars) -> None

The registered hooks can be used to perform pre-processing before the state_dict call is made.

requires_grad_(requires_grad=True)

Change if autograd should record operations on parameters in this module.

This method sets the parameters’ requires_grad attributes in-place.

This method is helpful for freezing part of the module for finetuning or training parts of a model individually (e.g., GAN training).

See Locally disabling gradient computation for a comparison between .requires_grad_() and several similar mechanisms that may be confused with it.

Parameters:

requires_grad (bool) – whether autograd should record operations on parameters in this module. Default: True.

Returns:

self

Return type:

Module

set_extra_state(state)

Set extra state contained in the loaded state_dict.

This function is called from load_state_dict() to handle any extra state found within the state_dict. Implement this function and a corresponding get_extra_state() for your module if you need to store extra state within its state_dict.

Parameters:

state (dict) – Extra state from the state_dict

Return type:

None

set_submodule(target, module, strict=False)

Set the submodule given by target if it exists, otherwise throw an error.

Note

If strict is set to False (default), the method will replace an existing submodule or create a new submodule if the parent module exists. If strict is set to True, the method will only attempt to replace an existing submodule and throw an error if the submodule does not exist.

For example, let’s say you have an nn.Module A that looks like this:

A(
    (net_b): Module(
        (net_c): Module(
            (conv): Conv2d(3, 3, 3)
        )
        (linear): Linear(3, 3)
    )
)

(The diagram shows an nn.Module A. A has a nested submodule net_b, which itself has two submodules net_c and linear. net_c then has a submodule conv.)

To override the Conv2d with a new submodule Linear, you could call set_submodule("net_b.net_c.conv", nn.Linear(1, 1)) where strict could be True or False

To add a new submodule Conv2d to the existing net_b module, you would call set_submodule("net_b.conv", nn.Conv2d(1, 1, 1)).

In the above if you set strict=True and call set_submodule("net_b.conv", nn.Conv2d(1, 1, 1), strict=True), an AttributeError will be raised because net_b does not have a submodule named conv.

Parameters:

• target (str) – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)

• module (Module) – The module to set the submodule to.

• strict (bool, default: False) – If False, the method will replace an existing submodule or create a new submodule if the parent module exists. If True, the method will only attempt to replace an existing submodule and throw an error if the submodule doesn’t already exist.

Raises:

• ValueError – If the target string is empty or if module is not an instance of nn.Module.

• AttributeError – If at any point along the path resulting from the target string the (sub)path resolves to a non-existent attribute name or an object that is not an instance of nn.Module.

Return type:

None

share_memory()

See torch.Tensor.share_memory_().

Return type:

Self

state_dict(*args, destination=None, prefix='', keep_vars=False)

Return a dictionary containing references to the whole state of the module.

Both parameters and persistent buffers (e.g. running averages) are included. Keys are corresponding parameter and buffer names. Parameters and buffers set to None are not included.

Note

The returned object is a shallow copy. It contains references to the module’s parameters and buffers.

Warning

Currently state_dict() also accepts positional arguments for destination, prefix and keep_vars in order. However, this is being deprecated and keyword arguments will be enforced in future releases.

Warning

Please avoid the use of argument destination as it is not designed for end-users.

Parameters:

• destination (dict, optional) – If provided, the state of module will be updated into the dict and the same object is returned. Otherwise, an OrderedDict will be created and returned. Default: None.

• prefix (str, optional) – a prefix added to parameter and buffer names to compose the keys in state_dict. Default: ''.

• keep_vars (bool, optional) – by default the Tensor s returned in the state dict are detached from autograd. If it’s set to True, detaching will not be performed. Default: False.

Returns:

a dictionary containing a whole state of the module

Return type:

dict

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> module.state_dict().keys()
['bias', 'weight']

to(*args, **kwargs)

Move and/or cast the parameters and buffers.

This can be called as

to(device=None, dtype=None, non_blocking=False)

to(dtype, non_blocking=False)

to(tensor, non_blocking=False)

to(memory_format=torch.channels_last)

Its signature is similar to torch.Tensor.to(), but only accepts floating point or complex dtypes. In addition, this method will only cast the floating point or complex parameters and buffers to dtype (if given). The integral parameters and buffers will be moved device, if that is given, but with dtypes unchanged. When non_blocking is set, it tries to convert/move asynchronously with respect to the host if possible, e.g., moving CPU Tensors with pinned memory to CUDA devices.

See below for examples.

Note

This method modifies the module in-place.

Parameters:

• device (torch.device) – the desired device of the parameters and buffers in this module

• dtype (torch.dtype) – the desired floating point or complex dtype of the parameters and buffers in this module

• tensor (torch.Tensor) – Tensor whose dtype and device are the desired dtype and device for all parameters and buffers in this module

• memory_format (torch.memory_format) – the desired memory format for 4D parameters and buffers in this module (keyword only argument)

Returns:

self

Return type:

Module

Examples:

>>> # xdoctest: +IGNORE_WANT("non-deterministic")
>>> linear = nn.Linear(2, 2)
>>> linear.weight
Parameter containing:
tensor([[ 0.1913, -0.3420],
        [-0.5113, -0.2325]])
>>> linear.to(torch.double)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1913, -0.3420],
        [-0.5113, -0.2325]], dtype=torch.float64)
>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CUDA1)
>>> gpu1 = torch.device("cuda:1")
>>> linear.to(gpu1, dtype=torch.half, non_blocking=True)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1914, -0.3420],
        [-0.5112, -0.2324]], dtype=torch.float16, device='cuda:1')
>>> cpu = torch.device("cpu")
>>> linear.to(cpu)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1914, -0.3420],
        [-0.5112, -0.2324]], dtype=torch.float16)

>>> linear = nn.Linear(2, 2, bias=None).to(torch.cdouble)
>>> linear.weight
Parameter containing:
tensor([[ 0.3741+0.j,  0.2382+0.j],
        [ 0.5593+0.j, -0.4443+0.j]], dtype=torch.complex128)
>>> linear(torch.ones(3, 2, dtype=torch.cdouble))
tensor([[0.6122+0.j, 0.1150+0.j],
        [0.6122+0.j, 0.1150+0.j],
        [0.6122+0.j, 0.1150+0.j]], dtype=torch.complex128)

to_empty(*, device, recurse=True)

Move the parameters and buffers to the specified device without copying storage.

Parameters:

• device (torch.device) – The desired device of the parameters and buffers in this module.

• recurse (bool) – Whether parameters and buffers of submodules should be recursively moved to the specified device.

Returns:

self

Return type:

Module

train(mode=True)

Set the module in training mode.

This has an effect only on certain modules. See the documentation of particular modules for details of their behaviors in training/evaluation mode, i.e., whether they are affected, e.g. Dropout, BatchNorm, etc.

Parameters:

mode (bool) – whether to set training mode (True) or evaluation mode (False). Default: True.

Returns:

self

Return type:

Module

type(dst_type)

Casts all parameters and buffers to dst_type.

Note

This method modifies the module in-place.

Parameters:

dst_type (type or string) – the desired type

Returns:

self

Return type:

Module

xpu(device=None)

Move all model parameters and buffers to the XPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing optimizer if the module will live on XPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

zero_grad(set_to_none=True)

Reset gradients of all model parameters.

See similar function under torch.optim.Optimizer for more context.

Parameters:

set_to_none (bool) – instead of setting to zero, set the grads to None. See torch.optim.Optimizer.zero_grad() for details.

Return type:

None

training: bool

class medlatents.training.REPAProjection(hidden_dim, target_dim=768, proj_dim=256, num_layers=2, activation='gelu', dropout=0.0)

Bases: torch.nn.modules.module.Module

MLP projection head for REPA feature alignment.

Projects DiT hidden states to the same dimension as the target encoder for similarity computation. Applied after early-to-mid transformer blocks.

Parameters:

• hidden_dim (int) – DiT hidden dimension (input)

• target_dim (int, default: 768) – Target encoder dimension (e.g., 768 for DINOv2-B)

• proj_dim (int, default: 256) – Projection space dimension (for similarity computation)

• num_layers (int, default: 2) – Number of MLP layers

• activation (str, default: 'gelu') – Activation function

• dropout (float, default: 0.0) – Dropout probability

Usage:

proj = REPAProjection(hidden_dim=1024, target_dim=768, proj_dim=256)

# In forward pass, after block k: hidden = dit_block_k(x) proj_hidden = proj(hidden) # [batch, seq, proj_dim]

__init__(hidden_dim, target_dim=768, proj_dim=256, num_layers=2, activation='gelu', dropout=0.0)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

• hidden_dim (int)

• target_dim (int, default: 768)

• proj_dim (int, default: 256)

• num_layers (int, default: 2)

• activation (str, default: 'gelu')

• dropout (float, default: 0.0)

projection: torch.nn.modules.container.Sequential

target_proj: torch.nn.modules.module.Module

forward(hidden_states)

Project hidden states to alignment space.

Parameters:

hidden_states (Tensor) – DiT hidden states [batch, seq, hidden_dim]

Return type:

Tensor

Returns:

Projected features [batch, seq, proj_dim]

project_target(target_features)

Project target encoder features to alignment space.

Parameters:

target_features (Tensor) – Encoder features [batch, seq, target_dim]

Return type:

Tensor

Returns:

Projected features [batch, seq, proj_dim]

T_destination = ~T_destination

add_module(name, module)

Add a child module to the current module.

The module can be accessed as an attribute using the given name.

Parameters:

• name (str) – name of the child module. The child module can be accessed from this module using the given name

• module (Module) – child module to be added to the module.

Return type:

None

apply(fn)

Apply fn recursively to every submodule (as returned by .children()) as well as self.

Typical use includes initializing the parameters of a model (see also torch.nn.init).

Parameters:

fn (Module -> None) – function to be applied to each submodule

Returns:

self

Return type:

Module

Example:

>>> @torch.no_grad()
>>> def init_weights(m):
>>>     print(m)
>>>     if type(m) is nn.Linear:
>>>         m.weight.fill_(1.0)
>>>         print(m.weight)
>>> net = nn.Sequential(nn.Linear(2, 2), nn.Linear(2, 2))
>>> net.apply(init_weights)
Linear(in_features=2, out_features=2, bias=True)
Parameter containing:
tensor([[1., 1.],
        [1., 1.]], requires_grad=True)
Linear(in_features=2, out_features=2, bias=True)
Parameter containing:
tensor([[1., 1.],
        [1., 1.]], requires_grad=True)
Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
)

bfloat16()

Casts all floating point parameters and buffers to bfloat16 datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

buffers(recurse=True)

Return an iterator over module buffers.

Parameters:

recurse (bool) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module.

Yields:

torch.Tensor – module buffer

Return type:

Iterator[Tensor]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for buf in model.buffers():
>>>     print(type(buf), buf.size())
<class 'torch.Tensor'> (20L,)
<class 'torch.Tensor'> (20L, 1L, 5L, 5L)

call_super_init: bool = False

children()

Return an iterator over immediate children modules.

Yields:

Module – a child module

Return type:

Iterator[Module]

compile(*args, **kwargs)

Compile this Module’s forward using torch.compile().

This Module’s __call__ method is compiled and all arguments are passed as-is to torch.compile().

See torch.compile() for details on the arguments for this function.

Return type:

None

cpu()

Move all model parameters and buffers to the CPU.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

cuda(device=None)

Move all model parameters and buffers to the GPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on GPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

double()

Casts all floating point parameters and buffers to double datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

dump_patches: bool = False

eval()

Set the module in evaluation mode.

This has an effect only on certain modules. See the documentation of particular modules for details of their behaviors in training/evaluation mode, i.e. whether they are affected, e.g. Dropout, BatchNorm, etc.

This is equivalent with self.train(False).

See Locally disabling gradient computation for a comparison between .eval() and several similar mechanisms that may be confused with it.

Returns:

self

Return type:

Module

extra_repr()

Return the extra representation of the module.

To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.

Return type:

str

float()

Casts all floating point parameters and buffers to float datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

get_buffer(target)

Return the buffer given by target if it exists, otherwise throw an error.

See the docstring for get_submodule for a more detailed explanation of this method’s functionality as well as how to correctly specify target.

Parameters:

target (str) – The fully-qualified string name of the buffer to look for. (See get_submodule for how to specify a fully-qualified string.)

Returns:

The buffer referenced by target

Return type:

torch.Tensor

Raises:

AttributeError – If the target string references an invalid path or resolves to something that is not a buffer

get_extra_state()

Return any extra state to include in the module’s state_dict.

Implement this and a corresponding set_extra_state() for your module if you need to store extra state. This function is called when building the module’s state_dict().

Note that extra state should be picklable to ensure working serialization of the state_dict. We only provide backwards compatibility guarantees for serializing Tensors; other objects may break backwards compatibility if their serialized pickled form changes.

Returns:

Any extra state to store in the module’s state_dict

Return type:

object

get_parameter(target)

Return the parameter given by target if it exists, otherwise throw an error.

See the docstring for get_submodule for a more detailed explanation of this method’s functionality as well as how to correctly specify target.

Parameters:

target (str) – The fully-qualified string name of the Parameter to look for. (See get_submodule for how to specify a fully-qualified string.)

Returns:

The Parameter referenced by target

Return type:

torch.nn.Parameter

Raises:

AttributeError – If the target string references an invalid path or resolves to something that is not an nn.Parameter

get_submodule(target)

Return the submodule given by target if it exists, otherwise throw an error.

For example, let’s say you have an nn.Module A that looks like this:

A(
    (net_b): Module(
        (net_c): Module(
            (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2))
        )
        (linear): Linear(in_features=100, out_features=200, bias=True)
    )
)

(The diagram shows an nn.Module A. A which has a nested submodule net_b, which itself has two submodules net_c and linear. net_c then has a submodule conv.)

To check whether or not we have the linear submodule, we would call get_submodule("net_b.linear"). To check whether we have the conv submodule, we would call get_submodule("net_b.net_c.conv").

The runtime of get_submodule is bounded by the degree of module nesting in target. A query against named_modules achieves the same result, but it is O(N) in the number of transitive modules. So, for a simple check to see if some submodule exists, get_submodule should always be used.

Parameters:

target (str) – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)

Returns:

The submodule referenced by target

Return type:

torch.nn.Module

Raises:

AttributeError – If at any point along the path resulting from the target string the (sub)path resolves to a non-existent attribute name or an object that is not an instance of nn.Module.

half()

Casts all floating point parameters and buffers to half datatype.

Note

This method modifies the module in-place.

Returns:

self

Return type:

Module

ipu(device=None)

Move all model parameters and buffers to the IPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on IPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

load_state_dict(state_dict, strict=True, assign=False)

Copy parameters and buffers from state_dict into this module and its descendants.

If strict is True, then the keys of state_dict must exactly match the keys returned by this module’s state_dict() function.

Warning

If assign is True the optimizer must be created after the call to load_state_dict unless get_swap_module_params_on_conversion() is True.

Parameters:

• state_dict (dict) – a dict containing parameters and persistent buffers.

• strict (bool, optional) – whether to strictly enforce that the keys in state_dict match the keys returned by this module’s state_dict() function. Default: True

• assign (bool, optional) – When set to False, the properties of the tensors in the current module are preserved whereas setting it to True preserves properties of the Tensors in the state dict. The only exception is the requires_grad field of Parameter for which the value from the module is preserved. Default: False

Returns:

• missing_keys is a list of str containing any keys that are expected

  by this module but missing from the provided state_dict.

• unexpected_keys is a list of str containing the keys that are not

  expected by this module but present in the provided state_dict.

Return type:

NamedTuple with missing_keys and unexpected_keys fields

Note

If a parameter or buffer is registered as None and its corresponding key exists in state_dict, load_state_dict() will raise a RuntimeError.

modules(remove_duplicate=True)

Return an iterator over all modules in the network.

Parameters:

remove_duplicate (bool, default: True) – whether to remove the duplicated module instances in the result or not.

Yields:

Module – a module in the network

Return type:

Iterator[Module]

Note

Duplicate modules are returned only once by default. In the following example, l will be returned only once.

Example:

>>> l = nn.Linear(2, 2)
>>> net = nn.Sequential(l, l)
>>> for idx, m in enumerate(net.modules()):
...     print(idx, '->', m)

0 -> Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
)
1 -> Linear(in_features=2, out_features=2, bias=True)

mtia(device=None)

Move all model parameters and buffers to the MTIA.

This also makes associated parameters and buffers different objects. So it should be called before constructing the optimizer if the module will live on MTIA while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

named_buffers(prefix='', recurse=True, remove_duplicate=True)

Return an iterator over module buffers, yielding both the name of the buffer as well as the buffer itself.

Parameters:

• prefix (str) – prefix to prepend to all buffer names.

• recurse (bool, optional) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module. Defaults to True.

• remove_duplicate (bool, optional) – whether to remove the duplicated buffers in the result. Defaults to True.

Yields:

(str, torch.Tensor) – Tuple containing the name and buffer

Return type:

Iterator[tuple[str, Tensor]]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, buf in self.named_buffers():
>>>     if name in ['running_var']:
>>>         print(buf.size())

named_children()

Return an iterator over immediate children modules, yielding both the name of the module as well as the module itself.

Yields:

(str, Module) – Tuple containing a name and child module

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, module in model.named_children():
>>>     if name in ['conv4', 'conv5']:
>>>         print(module)

Return type:

Iterator[tuple[str, Module]]

named_modules(memo=None, prefix='', remove_duplicate=True)

Return an iterator over all modules in the network, yielding both the name of the module as well as the module itself.

Parameters:

• memo (set[Module] | None, default: None) – a memo to store the set of modules already added to the result

• prefix (str, default: '') – a prefix that will be added to the name of the module

• remove_duplicate (bool, default: True) – whether to remove the duplicated module instances in the result or not

Yields:

(str, Module) – Tuple of name and module

Note

Duplicate modules are returned only once. In the following example, l will be returned only once.

Example:

>>> l = nn.Linear(2, 2)
>>> net = nn.Sequential(l, l)
>>> for idx, m in enumerate(net.named_modules()):
...     print(idx, '->', m)

0 -> ('', Sequential(
  (0): Linear(in_features=2, out_features=2, bias=True)
  (1): Linear(in_features=2, out_features=2, bias=True)
))
1 -> ('0', Linear(in_features=2, out_features=2, bias=True))

named_parameters(prefix='', recurse=True, remove_duplicate=True)

Return an iterator over module parameters, yielding both the name of the parameter as well as the parameter itself.

Parameters:

• prefix (str) – prefix to prepend to all parameter names.

• recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.

• remove_duplicate (bool, optional) – whether to remove the duplicated parameters in the result. Defaults to True.

Yields:

(str, Parameter) – Tuple containing the name and parameter

Return type:

Iterator[tuple[str, Parameter]]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for name, param in self.named_parameters():
>>>     if name in ['bias']:
>>>         print(param.size())

parameters(recurse=True)

Return an iterator over module parameters.

This is typically passed to an optimizer.

Parameters:

recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.

Yields:

Parameter – module parameter

Return type:

Iterator[Parameter]

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> for param in model.parameters():
>>>     print(type(param), param.size())
<class 'torch.Tensor'> (20L,)
<class 'torch.Tensor'> (20L, 1L, 5L, 5L)

register_backward_hook(hook)

Register a backward hook on the module.

This function is deprecated in favor of register_full_backward_hook() and the behavior of this function will change in future versions.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

Parameters:

hook (Callable[[Module, tuple[Tensor, ...] | Tensor, tuple[Tensor, ...] | Tensor], tuple[Tensor, ...] | Tensor | None])

register_buffer(name, tensor, persistent=True)

Add a buffer to the module.

This is typically used to register a buffer that should not be considered a model parameter. For example, BatchNorm’s running_mean is not a parameter, but is part of the module’s state. Buffers, by default, are persistent and will be saved alongside parameters. This behavior can be changed by setting persistent to False. The only difference between a persistent buffer and a non-persistent buffer is that the latter will not be a part of this module’s state_dict.

Buffers can be accessed as attributes using given names.

Parameters:

• name (str) – name of the buffer. The buffer can be accessed from this module using the given name

• tensor (Tensor or None) – buffer to be registered. If None, then operations that run on buffers, such as cuda, are ignored. If None, the buffer is not included in the module’s state_dict.

• persistent (bool) – whether the buffer is part of this module’s state_dict.

Return type:

None

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> self.register_buffer('running_mean', torch.zeros(num_features))

register_forward_hook(hook, *, prepend=False, with_kwargs=False, always_call=False)

Register a forward hook on the module.

The hook will be called every time after forward() has computed an output.

If with_kwargs is False or not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to the forward. The hook can modify the output. It can modify the input inplace but it will not have effect on forward since this is called after forward() is called. The hook should have the following signature:

hook(module, args, output) -> None or modified output

If with_kwargs is True, the forward hook will be passed the kwargs given to the forward function and be expected to return the output possibly modified. The hook should have the following signature:

hook(module, args, kwargs, output) -> None or modified output

Parameters:

• hook (Callable) – The user defined hook to be registered.

• prepend (bool) – If True, the provided hook will be fired before all existing forward hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing forward hooks on this torch.nn.Module. Note that global forward hooks registered with register_module_forward_hook() will fire before all hooks registered by this method. Default: False

• with_kwargs (bool) – If True, the hook will be passed the kwargs given to the forward function. Default: False

• always_call (bool) – If True the hook will be run regardless of whether an exception is raised while calling the Module. Default: False

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_forward_pre_hook(hook, *, prepend=False, with_kwargs=False)

Register a forward pre-hook on the module.

The hook will be called every time before forward() is invoked.

If with_kwargs is false or not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to the forward. The hook can modify the input. User can either return a tuple or a single modified value in the hook. We will wrap the value into a tuple if a single value is returned (unless that value is already a tuple). The hook should have the following signature:

hook(module, args) -> None or modified input

If with_kwargs is true, the forward pre-hook will be passed the kwargs given to the forward function. And if the hook modifies the input, both the args and kwargs should be returned. The hook should have the following signature:

hook(module, args, kwargs) -> None or a tuple of modified input and kwargs

Parameters:

• hook (Callable) – The user defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing forward_pre hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing forward_pre hooks on this torch.nn.Module. Note that global forward_pre hooks registered with register_module_forward_pre_hook() will fire before all hooks registered by this method. Default: False

• with_kwargs (bool) – If true, the hook will be passed the kwargs given to the forward function. Default: False

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_full_backward_hook(hook, prepend=False)

Register a backward hook on the module.

The hook will be called every time the gradients with respect to a module are computed, and its firing rules are as follows:

1. Ordinarily, the hook fires when the gradients are computed with respect to the module inputs.

2. If none of the module inputs require gradients, the hook will fire when the gradients are computed with respect to module outputs.

3. If none of the module outputs require gradients, then the hooks will not fire.

The hook should have the following signature:

hook(module, grad_input, grad_output) -> tuple(Tensor) or None

The grad_input and grad_output are tuples that contain the gradients with respect to the inputs and outputs respectively. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the input that will be used in place of grad_input in subsequent computations. grad_input will only correspond to the inputs given as positional arguments and all kwarg arguments are ignored. Entries in grad_input and grad_output will be None for all non-Tensor arguments.

For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.

Warning

Modifying inputs or outputs inplace is not allowed when using backward hooks and will raise an error.

Parameters:

• hook (Callable) – The user-defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing backward hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing backward hooks on this torch.nn.Module. Note that global backward hooks registered with register_module_full_backward_hook() will fire before all hooks registered by this method.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_full_backward_pre_hook(hook, prepend=False)

Register a backward pre-hook on the module.

The hook will be called every time the gradients for the module are computed. The hook should have the following signature:

hook(module, grad_output) -> tuple[Tensor, ...], Tensor or None

The grad_output is a tuple. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the output that will be used in place of grad_output in subsequent computations. Entries in grad_output will be None for all non-Tensor arguments.

For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.

Warning

Modifying inputs inplace is not allowed when using backward hooks and will raise an error.

Parameters:

• hook (Callable) – The user-defined hook to be registered.

• prepend (bool) – If true, the provided hook will be fired before all existing backward_pre hooks on this torch.nn.Module. Otherwise, the provided hook will be fired after all existing backward_pre hooks on this torch.nn.Module. Note that global backward_pre hooks registered with register_module_full_backward_pre_hook() will fire before all hooks registered by this method.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_load_state_dict_post_hook(hook)

Register a post-hook to be run after module’s load_state_dict() is called.

It should have the following signature::

hook(module, incompatible_keys) -> None

The module argument is the current module that this hook is registered on, and the incompatible_keys argument is a NamedTuple consisting of attributes missing_keys and unexpected_keys. missing_keys is a list of str containing the missing keys and unexpected_keys is a list of str containing the unexpected keys.

The given incompatible_keys can be modified inplace if needed.

Note that the checks performed when calling load_state_dict() with strict=True are affected by modifications the hook makes to missing_keys or unexpected_keys, as expected. Additions to either set of keys will result in an error being thrown when strict=True, and clearing out both missing and unexpected keys will avoid an error.

Returns:

a handle that can be used to remove the added hook by calling handle.remove()

Return type:

torch.utils.hooks.RemovableHandle

register_load_state_dict_pre_hook(hook)

Register a pre-hook to be run before module’s load_state_dict() is called.

It should have the following signature::

hook(module, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) -> None # noqa: B950

Parameters:

hook (Callable) – Callable hook that will be invoked before loading the state dict.

register_module(name, module)

Alias for add_module().

Parameters:

• name (str)

• module (Optional[Module])

Return type:

None

register_parameter(name, param)

Add a parameter to the module.

The parameter can be accessed as an attribute using given name.

Parameters:

• name (str) – name of the parameter. The parameter can be accessed from this module using the given name

• param (Parameter or None) – parameter to be added to the module. If None, then operations that run on parameters, such as cuda, are ignored. If None, the parameter is not included in the module’s state_dict.

Return type:

None

register_state_dict_post_hook(hook)

Register a post-hook for the state_dict() method.

It should have the following signature::

hook(module, state_dict, prefix, local_metadata) -> None

The registered hooks can modify the state_dict inplace.

register_state_dict_pre_hook(hook)

Register a pre-hook for the state_dict() method.

It should have the following signature::

hook(module, prefix, keep_vars) -> None

The registered hooks can be used to perform pre-processing before the state_dict call is made.

requires_grad_(requires_grad=True)

Change if autograd should record operations on parameters in this module.

This method sets the parameters’ requires_grad attributes in-place.

This method is helpful for freezing part of the module for finetuning or training parts of a model individually (e.g., GAN training).

See Locally disabling gradient computation for a comparison between .requires_grad_() and several similar mechanisms that may be confused with it.

Parameters:

requires_grad (bool) – whether autograd should record operations on parameters in this module. Default: True.

Returns:

self

Return type:

Module

set_extra_state(state)

Set extra state contained in the loaded state_dict.

This function is called from load_state_dict() to handle any extra state found within the state_dict. Implement this function and a corresponding get_extra_state() for your module if you need to store extra state within its state_dict.

Parameters:

state (dict) – Extra state from the state_dict

Return type:

None

set_submodule(target, module, strict=False)

Set the submodule given by target if it exists, otherwise throw an error.

Note

If strict is set to False (default), the method will replace an existing submodule or create a new submodule if the parent module exists. If strict is set to True, the method will only attempt to replace an existing submodule and throw an error if the submodule does not exist.

For example, let’s say you have an nn.Module A that looks like this:

A(
    (net_b): Module(
        (net_c): Module(
            (conv): Conv2d(3, 3, 3)
        )
        (linear): Linear(3, 3)
    )
)

(The diagram shows an nn.Module A. A has a nested submodule net_b, which itself has two submodules net_c and linear. net_c then has a submodule conv.)

To override the Conv2d with a new submodule Linear, you could call set_submodule("net_b.net_c.conv", nn.Linear(1, 1)) where strict could be True or False

To add a new submodule Conv2d to the existing net_b module, you would call set_submodule("net_b.conv", nn.Conv2d(1, 1, 1)).

In the above if you set strict=True and call set_submodule("net_b.conv", nn.Conv2d(1, 1, 1), strict=True), an AttributeError will be raised because net_b does not have a submodule named conv.

Parameters:

• target (str) – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)

• module (Module) – The module to set the submodule to.

• strict (bool, default: False) – If False, the method will replace an existing submodule or create a new submodule if the parent module exists. If True, the method will only attempt to replace an existing submodule and throw an error if the submodule doesn’t already exist.

Raises:

• ValueError – If the target string is empty or if module is not an instance of nn.Module.

• AttributeError – If at any point along the path resulting from the target string the (sub)path resolves to a non-existent attribute name or an object that is not an instance of nn.Module.

Return type:

None

share_memory()

See torch.Tensor.share_memory_().

Return type:

Self

state_dict(*args, destination=None, prefix='', keep_vars=False)

Return a dictionary containing references to the whole state of the module.

Both parameters and persistent buffers (e.g. running averages) are included. Keys are corresponding parameter and buffer names. Parameters and buffers set to None are not included.

Note

The returned object is a shallow copy. It contains references to the module’s parameters and buffers.

Warning

Currently state_dict() also accepts positional arguments for destination, prefix and keep_vars in order. However, this is being deprecated and keyword arguments will be enforced in future releases.

Warning

Please avoid the use of argument destination as it is not designed for end-users.

Parameters:

• destination (dict, optional) – If provided, the state of module will be updated into the dict and the same object is returned. Otherwise, an OrderedDict will be created and returned. Default: None.

• prefix (str, optional) – a prefix added to parameter and buffer names to compose the keys in state_dict. Default: ''.

• keep_vars (bool, optional) – by default the Tensor s returned in the state dict are detached from autograd. If it’s set to True, detaching will not be performed. Default: False.

Returns:

a dictionary containing a whole state of the module

Return type:

dict

Example:

>>> # xdoctest: +SKIP("undefined vars")
>>> module.state_dict().keys()
['bias', 'weight']

to(*args, **kwargs)

Move and/or cast the parameters and buffers.

This can be called as

to(device=None, dtype=None, non_blocking=False)

to(dtype, non_blocking=False)

to(tensor, non_blocking=False)

to(memory_format=torch.channels_last)

Its signature is similar to torch.Tensor.to(), but only accepts floating point or complex dtypes. In addition, this method will only cast the floating point or complex parameters and buffers to dtype (if given). The integral parameters and buffers will be moved device, if that is given, but with dtypes unchanged. When non_blocking is set, it tries to convert/move asynchronously with respect to the host if possible, e.g., moving CPU Tensors with pinned memory to CUDA devices.

See below for examples.

Note

This method modifies the module in-place.

Parameters:

• device (torch.device) – the desired device of the parameters and buffers in this module

• dtype (torch.dtype) – the desired floating point or complex dtype of the parameters and buffers in this module

• tensor (torch.Tensor) – Tensor whose dtype and device are the desired dtype and device for all parameters and buffers in this module

• memory_format (torch.memory_format) – the desired memory format for 4D parameters and buffers in this module (keyword only argument)

Returns:

self

Return type:

Module

Examples:

>>> # xdoctest: +IGNORE_WANT("non-deterministic")
>>> linear = nn.Linear(2, 2)
>>> linear.weight
Parameter containing:
tensor([[ 0.1913, -0.3420],
        [-0.5113, -0.2325]])
>>> linear.to(torch.double)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1913, -0.3420],
        [-0.5113, -0.2325]], dtype=torch.float64)
>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CUDA1)
>>> gpu1 = torch.device("cuda:1")
>>> linear.to(gpu1, dtype=torch.half, non_blocking=True)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1914, -0.3420],
        [-0.5112, -0.2324]], dtype=torch.float16, device='cuda:1')
>>> cpu = torch.device("cpu")
>>> linear.to(cpu)
Linear(in_features=2, out_features=2, bias=True)
>>> linear.weight
Parameter containing:
tensor([[ 0.1914, -0.3420],
        [-0.5112, -0.2324]], dtype=torch.float16)

>>> linear = nn.Linear(2, 2, bias=None).to(torch.cdouble)
>>> linear.weight
Parameter containing:
tensor([[ 0.3741+0.j,  0.2382+0.j],
        [ 0.5593+0.j, -0.4443+0.j]], dtype=torch.complex128)
>>> linear(torch.ones(3, 2, dtype=torch.cdouble))
tensor([[0.6122+0.j, 0.1150+0.j],
        [0.6122+0.j, 0.1150+0.j],
        [0.6122+0.j, 0.1150+0.j]], dtype=torch.complex128)

to_empty(*, device, recurse=True)

Move the parameters and buffers to the specified device without copying storage.

Parameters:

• device (torch.device) – The desired device of the parameters and buffers in this module.

• recurse (bool) – Whether parameters and buffers of submodules should be recursively moved to the specified device.

Returns:

self

Return type:

Module

train(mode=True)

Set the module in training mode.

This has an effect only on certain modules. See the documentation of particular modules for details of their behaviors in training/evaluation mode, i.e., whether they are affected, e.g. Dropout, BatchNorm, etc.

Parameters:

mode (bool) – whether to set training mode (True) or evaluation mode (False). Default: True.

Returns:

self

Return type:

Module

type(dst_type)

Casts all parameters and buffers to dst_type.

Note

This method modifies the module in-place.

Parameters:

dst_type (type or string) – the desired type

Returns:

self

Return type:

Module

xpu(device=None)

Move all model parameters and buffers to the XPU.

This also makes associated parameters and buffers different objects. So it should be called before constructing optimizer if the module will live on XPU while being optimized.

Note

This method modifies the module in-place.

Parameters:

device (int, optional) – if specified, all parameters will be copied to that device

Returns:

self

Return type:

Module

zero_grad(set_to_none=True)

Reset gradients of all model parameters.

See similar function under torch.optim.Optimizer for more context.

Parameters:

set_to_none (bool) – instead of setting to zero, set the grads to None. See torch.optim.Optimizer.zero_grad() for details.

Return type:

None

training: bool

class medlatents.training.REPAETrainer(dit, vae, frozen_encoder, repa_projection, config=None, vae_type='continuous', noise_scheduler=None)

Bases: object

End-to-end trainer for VAE + DiT using REPA alignment.

This trainer enables joint optimization of VAE tokenizer and DiT by using REPA alignment loss instead of standard diffusion loss for VAE gradients.

The key insight is that diffusion loss is ineffective for E2E VAE training, but REPA loss provides meaningful gradients for both components.

Example

>>> trainer = REPAETrainer(
...     dit=dit_model,
...     vae=vae_model,
...     frozen_encoder=dino_encoder,
...     config=REPAEConfig(repa_weight=1.0),
... )
>>>
>>> for images in dataloader:
...     loss_dict = trainer.compute_loss(images, timesteps)
...     loss_dict["total"].backward()
...     optimizer.step()

Parameters:

• dit (Module)

• vae (Module)

• frozen_encoder (Module)

• repa_projection (REPAProjection)

• config (REPAEConfig | None, default: None)

• vae_type (Literal['continuous', 'discrete'], default: 'continuous')

• noise_scheduler (Module | None, default: None)

__init__(dit, vae, frozen_encoder, repa_projection, config=None, vae_type='continuous', noise_scheduler=None)

Parameters:

• dit (Module) – Diffusion transformer model

• vae (Module) – VAE or discrete tokenizer

• frozen_encoder (Module) – Pretrained encoder (DINOv2, SigLIP, etc.)

• repa_projection (REPAProjection) – Projection head for alignment

• config (REPAEConfig | None, default: None) – Training configuration

• vae_type (Literal['continuous', 'discrete'], default: 'continuous') – Type of VAE (‘continuous’ or ‘discrete’)

• noise_scheduler (Module | None, default: None) – Diffusion noise scheduler (for adding noise)

compute_loss(images, timesteps=None, return_intermediates=False)

Compute REPA-E loss for a batch of images.

Parameters:

• images (Tensor) – Input images [batch, channels, H, W]

• timesteps (Tensor | None, default: None) – Diffusion timesteps [batch]. If None, sampled uniformly.

• return_intermediates (bool, default: False) – Whether to return intermediate tensors

Return type:

dict[str, Tensor]

Returns:

Dict with loss components and optionally intermediates

get_vae_parameters()

Get VAE parameters for separate optimizer group.

Return type:

list[Parameter]

get_dit_parameters()

Get DiT parameters for separate optimizer group.

Return type:

list[Parameter]

create_optimizer_groups(dit_lr=0.0001, vae_lr=1e-05, weight_decay=0.01)

Create optimizer parameter groups with different learning rates.

VAE typically uses a lower learning rate to prevent catastrophic forgetting.

Parameters:

• dit_lr (float, default: 0.0001)

• vae_lr (float, default: 1e-05)

• weight_decay (float, default: 0.01)

Return type:

list[dict]

class medlatents.training.REPAEConfig(repa_weight=1.0, diffusion_weight=1.0, vae_recon_weight=0.0, kl_weight=0.0, timestep_threshold=0.5, vae_grad_scale=1.0, freeze_vae_encoder=False, freeze_vae_decoder=False, vae_warmup_steps=0, alignment_layers=None)

Bases: object

Configuration for REPA-E end-to-end training.

Variables:

• repa_weight – Weight for REPA alignment loss

• diffusion_weight – Weight for diffusion denoising loss

• vae_recon_weight – Weight for VAE reconstruction loss (optional)

• kl_weight – Weight for VAE KL divergence loss

• timestep_threshold – Only apply REPA at t > threshold

• vae_grad_scale – Scale factor for gradients flowing to VAE

• freeze_vae_encoder – Whether to freeze VAE encoder (only train decoder)

• freeze_vae_decoder – Whether to freeze VAE decoder (only train encoder)

• vae_warmup_steps – Steps to warmup VAE learning rate

• alignment_layers – Which DiT layers to use for alignment (list of indices)

Parameters:

• repa_weight (float, default: 1.0)

• diffusion_weight (float, default: 1.0)

• vae_recon_weight (float, default: 0.0)

• kl_weight (float, default: 0.0)

• timestep_threshold (float, default: 0.5)

• vae_grad_scale (float, default: 1.0)

• freeze_vae_encoder (bool, default: False)

• freeze_vae_decoder (bool, default: False)

• vae_warmup_steps (int, default: 0)

• alignment_layers (list[int] | None, default: None)

repa_weight: float = 1.0

diffusion_weight: float = 1.0

vae_recon_weight: float = 0.0

kl_weight: float = 0.0

timestep_threshold: float = 0.5

vae_grad_scale: float = 1.0

freeze_vae_encoder: bool = False

freeze_vae_decoder: bool = False

vae_warmup_steps: int = 0

alignment_layers: list[int] | None = None

__init__(repa_weight=1.0, diffusion_weight=1.0, vae_recon_weight=0.0, kl_weight=0.0, timestep_threshold=0.5, vae_grad_scale=1.0, freeze_vae_encoder=False, freeze_vae_decoder=False, vae_warmup_steps=0, alignment_layers=None)

Parameters:

• repa_weight (float, default: 1.0)

• diffusion_weight (float, default: 1.0)

• vae_recon_weight (float, default: 0.0)

• kl_weight (float, default: 0.0)

• timestep_threshold (float, default: 0.5)

• vae_grad_scale (float, default: 1.0)

• freeze_vae_encoder (bool, default: False)

• freeze_vae_decoder (bool, default: False)

• vae_warmup_steps (int, default: 0)

• alignment_layers (list[int] | None, default: None)

Checkpointing

medlatents.training.save_checkpoint(path, model, optimizer, epoch, metric, wandb_id, hparams=None)

Save a training checkpoint.

Parameters:

• path (str)

• model (Module)

• optimizer (Optimizer)

• epoch (int)

• metric (float)

• wandb_id (str)

• hparams (dict | None, default: None)

Return type:

None

medlatents.training.load_checkpoint(checkpoint_dir, name, best=False, weights_only=True)

Load a training checkpoint if it exists.

Parameters:

• checkpoint_dir (str)

• name (str)

• best (bool, default: False)

• weights_only (bool, default: True)

Return type:

dict | None

medlatents.training.resume_from_checkpoint(checkpoint, model, optimizer=None)

Resume training from a checkpoint.

Returns:

Epoch to resume from metric_best: Best metric value wandb_id: W&B run ID for resuming

Return type:

start_epoch

Parameters:

• checkpoint (dict)

• model (Module)

• optimizer (Optimizer | None, default: None)

Distributed Training

class medlatents.training.DistributedConfig(backend='nccl', world_size=1, rank=0, local_rank=0, master_addr='localhost', master_port='29500', use_fsdp=False, fsdp_sharding_strategy='full', mixed_precision='bf16', gradient_checkpointing=True)

Bases: object

Configuration for distributed training.

Variables:

• backend – Distributed backend (‘nccl’ for GPU, ‘gloo’ for CPU)

• world_size – Total number of processes

• rank – Rank of current process

• local_rank – Local rank on current node

• master_addr – Master node address

• master_port – Master node port

• use_fsdp – Use FSDP instead of DDP

• fsdp_sharding_strategy – FSDP sharding strategy

• mixed_precision – Mixed precision dtype (‘fp16’, ‘bf16’, or None)

Parameters:

• backend (str, default: 'nccl')

• world_size (int, default: 1)

• rank (int, default: 0)

• local_rank (int, default: 0)

• master_addr (str, default: 'localhost')

• master_port (str, default: '29500')

• use_fsdp (bool, default: False)

• fsdp_sharding_strategy (str, default: 'full')

• mixed_precision (str | None, default: 'bf16')

• gradient_checkpointing (bool, default: True)

backend: str = 'nccl'

world_size: int = 1

rank: int = 0

local_rank: int = 0

master_addr: str = 'localhost'

master_port: str = '29500'

use_fsdp: bool = False

fsdp_sharding_strategy: str = 'full'

mixed_precision: str | None = 'bf16'

gradient_checkpointing: bool = True

classmethod from_env()

Create config from environment variables (set by torchrun).

Return type:

DistributedConfig

property is_distributed: bool

Whether distributed training is enabled.

property is_main_process: bool

Whether this is the main (rank 0) process.

__init__(backend='nccl', world_size=1, rank=0, local_rank=0, master_addr='localhost', master_port='29500', use_fsdp=False, fsdp_sharding_strategy='full', mixed_precision='bf16', gradient_checkpointing=True)

Parameters:

• backend (str, default: 'nccl')

• world_size (int, default: 1)

• rank (int, default: 0)

• local_rank (int, default: 0)

• master_addr (str, default: 'localhost')

• master_port (str, default: '29500')

• use_fsdp (bool, default: False)

• fsdp_sharding_strategy (str, default: 'full')

• mixed_precision (str | None, default: 'bf16')

• gradient_checkpointing (bool, default: True)

medlatents.training.setup_distributed(config=None)

Initialize distributed training environment.

Parameters:

config (DistributedConfig | None, default: None) – Distributed configuration (auto-detects from env if None)

Return type:

DistributedConfig

Returns:

Initialized configuration

medlatents.training.wrap_model_fsdp(model, config, auto_wrap_policy=None, cpu_offload=False)

Wrap model with FullyShardedDataParallel.

FSDP shards model parameters across GPUs for memory efficiency.

Parameters:

• model (Module) – Model to wrap

• config (DistributedConfig) – Distributed configuration

• auto_wrap_policy (Any | None, default: None) – Policy for automatic module wrapping

• cpu_offload (bool, default: False) – Offload parameters to CPU (slower but lower GPU memory)

Return type:

Module

Returns:

FSDP-wrapped model

medlatents.training.wrap_model_ddp(model, config, find_unused_parameters=False, gradient_as_bucket_view=True)

Wrap model with DistributedDataParallel.

Parameters:

• model (Module) – Model to wrap

• config (DistributedConfig) – Distributed configuration

• find_unused_parameters (bool, default: False) – Enable for models with unused params

• gradient_as_bucket_view (bool, default: True) – Memory optimization

Return type:

Module

Returns:

DDP-wrapped model

Data Loading

class medlatents.training.DataLoaderConfig(batch_size=32, num_workers=None, pin_memory=None, prefetch_factor=2, persistent_workers=True, drop_last=True, shuffle=True)

Bases: object

Configuration for high-performance data loading.

Variables:

• batch_size – Batch size

• num_workers – Number of data loading workers

• pin_memory – Pin memory for faster GPU transfer

• prefetch_factor – Batches to prefetch per worker

• persistent_workers – Keep workers alive between epochs

• drop_last – Drop incomplete last batch (important for DDP)

• shuffle – Shuffle data

Parameters:

• batch_size (int, default: 32)

• num_workers (int | None, default: None)

• pin_memory (bool | None, default: None)

• prefetch_factor (int, default: 2)

• persistent_workers (bool, default: True)

• drop_last (bool, default: True)

• shuffle (bool, default: True)

batch_size: int = 32

num_workers: int | None = None

pin_memory: bool | None = None

prefetch_factor: int = 2

persistent_workers: bool = True

drop_last: bool = True

shuffle: bool = True

__init__(batch_size=32, num_workers=None, pin_memory=None, prefetch_factor=2, persistent_workers=True, drop_last=True, shuffle=True)

Parameters:

• batch_size (int, default: 32)

• num_workers (int | None, default: None)

• pin_memory (bool | None, default: None)

• prefetch_factor (int, default: 2)

• persistent_workers (bool, default: True)

• drop_last (bool, default: True)

• shuffle (bool, default: True)

medlatents.training.create_optimized_dataloader(dataset, config=None, distributed=False, prefetch=True, device=None)

Create an optimized DataLoader with optional prefetching.

Parameters:

• dataset (Dataset) – PyTorch Dataset

• config (DataLoaderConfig | None, default: None) – DataLoader configuration (uses defaults if None)

• distributed (bool, default: False) – Whether to use DistributedSampler

• prefetch (bool, default: True) – Whether to wrap with async prefetcher

• device (device | str | None, default: None) – Target device for prefetching

Return type:

DataLoader | AsyncPrefetcher | CUDAPrefetcher

Returns:

DataLoader or prefetcher wrapper

medlatents.training.collate_variable_length(batch, pad_value=0, max_length=None)

Collate function for variable-length sequences.

Pads sequences to the maximum length in the batch.

Parameters:

• batch (list[Tensor]) – List of 1D tensors

• pad_value (int, default: 0) – Value to use for padding

• max_length (int | None, default: None) – Optional maximum length (truncates if exceeded)

Return type:

Tensor

Returns:

Padded batch tensor [batch_size, max_seq_len]

Performance and Profiling

medlatents.training.enable_cuda_optimizations(benchmark=True, tf32=True, deterministic=False)

Enable CUDA performance optimizations.

Should be called at the start of training before model creation.

Parameters:

• benchmark (bool, default: True) – Enable cuDNN benchmarking. Best for consistent input sizes. Adds ~1-2min warmup but can give 10-30% speedup.

• tf32 (bool, default: True) – Enable TensorFloat-32 for Ampere+ GPUs. 2-3x faster matmuls with minimal precision loss.

• deterministic (bool, default: False) – Force deterministic operations. Disables some optimizations but ensures reproducibility.

Return type:

dict[str, bool]

Returns:

Dict of applied settings

medlatents.training.create_fused_adamw(model, lr=0.0001, weight_decay=0.01, betas=(0.9, 0.999), eps=1e-08, foreach=None)

Create AdamW optimizer with fused kernels when available.

Fused AdamW is 10-30% faster on CUDA by combining operations.

Parameters:

• model (Module) – Model to optimize

• lr (float, default: 0.0001) – Learning rate

• weight_decay (float, default: 0.01) – Weight decay coefficient

• betas (tuple[float, float], default: (0.9, 0.999)) – Adam beta parameters

• eps (float, default: 1e-08) – Epsilon for numerical stability

• foreach (bool | None, default: None) – Use foreach implementation (batched updates). None = auto-detect, True = force, False = disable.

Return type:

Optimizer

Returns:

Configured AdamW optimizer

medlatents.training.get_optimal_dtype(prefer_bf16=True, check_cuda_capability=True)

Get optimal dtype for mixed precision training.

Parameters:

• prefer_bf16 (bool, default: True) – Prefer bfloat16 over float16 when available. bf16 has better numerical range, fp16 has better hardware support.

• check_cuda_capability (bool, default: True) – Check GPU capability for bf16 support.

Return type:

dtype

Returns:

Recommended dtype for autocast

class medlatents.training.TrainingProfiler(enabled=True, sync_cuda=True, _stats=<factory>)

Bases: object

Profiler for training loop bottleneck detection.

Usage:

profiler = TrainingProfiler(enabled=True)

for batch in dataloader:

with profiler.section(“data_load”):

data = preprocess(batch)

with profiler.section(“forward”):

loss = model(data)

with profiler.section(“backward”):

loss.backward()

with profiler.section(“optimizer”):

optimizer.step()

profiler.print_report()

Parameters:

• enabled (bool, default: True)

• sync_cuda (bool, default: True)

• _stats (dict[str, TimingStats], default: <factory>)

enabled: bool = True

sync_cuda: bool = True

section(name)

Time a section of code.

Parameters:

name (str) – Name of the section (e.g., “forward”, “backward”, “data_load”)

Return type:

Iterator[None]

reset()

Reset all timing statistics.

Return type:

None

get_stats()

Get timing statistics for all sections.

Return type:

dict[str, dict[str, float]]

Returns:

Dict mapping section names to their timing statistics.

get_percentages()

Get percentage of time spent in each section.

Return type:

dict[str, float]

Returns:

Dict mapping section names to their percentage of total time.

print_report(top_n=None)

Print a formatted profiling report.

Parameters:

top_n (int | None, default: None) – Only show top N sections by time (None = show all)

Return type:

None

get_wandb_metrics(prefix='profiler/')

Get metrics formatted for wandb logging.

Parameters:

prefix (str, default: 'profiler/') – Prefix for metric names

Return type:

dict[str, float]

Returns:

Dict of metrics suitable for wandb.log()

__init__(enabled=True, sync_cuda=True, _stats=<factory>)

Parameters:

• enabled (bool, default: True)

• sync_cuda (bool, default: True)

• _stats (dict[str, TimingStats], default: <factory>)

class medlatents.training.MemoryProfiler(enabled=True)

Bases: object

Track GPU memory usage during training.

Parameters:

enabled (bool, default: True)

__init__(enabled=True)

Parameters:

enabled (bool, default: True)

track(name)

Track memory usage for a section of code.

Parameters:

name (str)

Return type:

Iterator[None]

get_stats()

Get memory statistics in MB.

Return type:

dict[str, dict[str, float]]

print_report()

Print memory usage report.

Return type:

None