Sampling

Advanced sampling techniques for discrete generative models.

medlatents.sampling.sample_nucleus(logits, p=0.9)[source][source]

Nucleus (top-p) sampling: keep smallest set of tokens with cumulative probability >= p.

More dynamic than top-k: adapts to the confidence of the model. Useful for discrete sampling in any domain including medical imaging.

Parameters:

  • logits (Tensor) – [batch, vocab_size] unnormalized logits

  • p (float, default: 0.9) – cumulative probability threshold (0 < p <= 1)

Returns:

[batch, vocab_size] with low-probability logits set to -inf

Return type:

filtered_logits

medlatents.sampling.sample_min_p(logits, min_p=0.05, base_top_p=1.0)[source][source]

Min-p sampling: Remove tokens with probability < min_p * max_probability.

Better than top-p: scales with model confidence instead of using fixed threshold. When model is confident, it’s more selective. When uncertain, it’s more diverse.

Works well for discrete latent codes across domains.

Reference: https://github.com/ggerganov/llama.cpp/pull/3841

Parameters:

  • logits (Tensor) – [batch, vocab_size] unnormalized logits

  • min_p (float, default: 0.05) – minimum probability threshold (relative to max)

  • base_top_p (float, default: 1.0) – optional top-p to apply first

Returns:

[batch, vocab_size] with low-probability logits set to -inf

Return type:

filtered_logits

medlatents.sampling.sample_autoregressive(logits, temperature=1.0, top_k=None, top_p=None, min_p=None)[source][source]

Sample from autoregressive model with temperature and filtering.

Order of operations: 1. Apply temperature 2. Apply filtering (top-k, top-p, or min-p) 3. Sample from filtered distribution

Parameters:

  • logits (Tensor) – [batch, vocab_size] unnormalized logits

  • temperature (float, default: 1.0) – sampling temperature (higher = more random)

  • top_k (int | None, default: None) – top-k filtering (None = disabled)

  • top_p (float | None, default: None) – nucleus sampling threshold (None = disabled)

  • min_p (float | None, default: None) – min-p sampling threshold (None = disabled, recommended: 0.05)

Returns:

[batch] sampled token indices

Return type:

sampled_tokens

medlatents.sampling.sample_with_cfg(conditional_logits, unconditional_logits, guidance_scale=1.5)[source][source]

Classifier-Free Guidance: Interpolate between conditional and unconditional predictions.

Useful for conditional generation tasks such as: - Generating specific modalities or conditions - Conditioning on metadata - Guided inpainting/reconstruction

Formula: logits_cfg = unconditional + guidance_scale * (conditional - unconditional)

Parameters:

  • conditional_logits (Tensor) – [batch, vocab_size] logits with conditioning

  • unconditional_logits (Tensor) – [batch, vocab_size] logits without conditioning

  • guidance_scale (float, default: 1.5) – strength of guidance (1.0 = no guidance, >1.0 = stronger conditioning)

Returns:

[batch, vocab_size]

Return type:

guided_logits

class medlatents.sampling.TemperatureScheduler(schedule='constant', start_temp=1.0, end_temp=0.7, num_steps=None)[source][source]

Bases: object

Dynamic temperature scheduling during generation.

Start with high temperature (exploration), end with low (refinement): - High temperature early: Explore different structures - Low temperature late: Refine details with high confidence

Supports multiple schedules: - constant: fixed temperature - linear: linearly decrease temperature - cosine: smooth annealing (recommended) - exponential: exponential decay - adaptive: adjust based on entropy (experimental)

Parameters:

  • schedule (str, default: 'constant')

  • start_temp (float, default: 1.0)

  • end_temp (float, default: 0.7)

  • num_steps (int | None, default: None)

__init__(schedule='constant', start_temp=1.0, end_temp=0.7, num_steps=None)[source][source]
Parameters:

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

  • start_temp (float, default: 1.0) – initial temperature (higher = more exploration)

  • end_temp (float, default: 0.7) – final temperature (lower = more confident)

  • num_steps (int | None, default: None) – number of generation steps (required for non-constant schedules)

get_temperature(step, logits=None)[source][source]

Get temperature for current step.

Parameters:
Return type:

float

medlatents.sampling.confidence_based_schedule(logits, current_tokens, mask_token, target_mask_ratio, temperature=1.0)[source][source]

Adaptive masking based on model confidence.

Instead of using a fixed schedule, mask tokens where the model is least confident. This allows the model to focus on difficult regions.

Parameters:

  • logits (Tensor) – [batch, seq_len, vocab_size] model predictions

  • current_tokens (Tensor) – [batch, seq_len] current token assignments

  • mask_token (int) – index of mask token

  • target_mask_ratio (float) – desired fraction of tokens to mask

  • temperature (float, default: 1.0) – temperature for confidence scoring

Returns:

[batch, seq_len] boolean mask (True = keep masked)

Return type:

mask

medlatents.sampling.adaptive_masking(logits, current_tokens, mask_token, step, total_steps, base_schedule='cosine', confidence_weight=0.5, temperature=1.0, schedule_power=2.0, halton_base=2)[source][source]

Hybrid masking: combine fixed schedule with confidence-based adjustment.

Parameters:

  • logits (Tensor) – [batch, seq_len, vocab_size] model predictions

  • current_tokens (Tensor) – [batch, seq_len] current token assignments

  • mask_token (int) – index of mask token

  • step (int) – current iteration

  • total_steps (int) – total iterations

  • base_schedule (str, default: 'cosine') – base schedule type (‘cosine’, ‘linear’, ‘sqrt’, ‘power’, ‘halton’)

  • confidence_weight (float, default: 0.5) – weight for confidence-based adjustment (0-1)

  • temperature (float, default: 1.0) – temperature for confidence scoring

  • schedule_power (float, default: 2.0) – power for power schedule

  • halton_base (int, default: 2) – base for halton schedule

Returns:

[batch, seq_len] boolean mask

Return type:

mask

medlatents.sampling.gumbel_max_sampling(logits, temperature=1.0, hard=True)[source][source]

Gumbel-max sampling for categorical distributions.

Adds Gumbel noise and takes argmax (or softmax for soft samples). This is equivalent to sampling but differentiable when soft=True.

Parameters:

  • logits (Tensor) – [batch, seq_len, vocab_size] unnormalized logits

  • temperature (float, default: 1.0) – sampling temperature

  • hard (bool, default: True) – if True, return discrete samples, if False, return soft samples

Returns:

[batch, seq_len] if hard, [batch, seq_len, vocab_size] if soft

Return type:

samples

class medlatents.sampling.MaskGITScheduler(num_steps=None, mask_schedule=None, temp_schedule='cosine', start_temp=1.0, end_temp=0.7, confidence_weight=0.3, schedule_power=2.0, halton_base=2, max_unmask_per_step=0.25, num_iterations=None, schedule=None)[source][source]

Bases: object

Advanced scheduler for MaskGIT-style generation.

Combines: - Flexible masking schedules (cosine, linear, sqrt, power, halton) - Temperature annealing - Confidence-based adaptation

Parameters:

  • num_steps (int | None, default: None)

  • mask_schedule (str | None, default: None)

  • temp_schedule (str, default: 'cosine')

  • start_temp (float, default: 1.0)

  • end_temp (float, default: 0.7)

  • confidence_weight (float, default: 0.3)

  • schedule_power (float, default: 2.0)

  • halton_base (int, default: 2)

  • max_unmask_per_step (float | None, default: 0.25)

  • num_iterations (int | None, default: None)

  • schedule (str | None, default: None)

__init__(num_steps=None, mask_schedule=None, temp_schedule='cosine', start_temp=1.0, end_temp=0.7, confidence_weight=0.3, schedule_power=2.0, halton_base=2, max_unmask_per_step=0.25, num_iterations=None, schedule=None)[source][source]
Parameters:

  • num_steps (int | None, default: None)

  • mask_schedule (str | None, default: None)

  • temp_schedule (str, default: 'cosine')

  • start_temp (float, default: 1.0)

  • end_temp (float, default: 0.7)

  • confidence_weight (float, default: 0.3)

  • schedule_power (float, default: 2.0)

  • halton_base (int, default: 2)

  • max_unmask_per_step (float | None, default: 0.25)

  • num_iterations (int | None, default: None)

  • schedule (str | None, default: None)

property num_iterations: int
step(logits, tokens, mask, step, temperature=None)[source][source]

Perform one step of MaskGIT generation.

Uses a fixed K_MAX for topk to enable torch.compile(fullgraph=True). Final step uses a separate codepath that unmasks all remaining tokens without calling topk.

Parameters:

  • logits (Tensor) – [batch, seq_len, vocab_size] model output

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

  • mask (Tensor) – [batch, seq_len] True for positions to update

  • step (int) – current step number

  • temperature (float | None, default: None) – optional override for temperature

Returns:

updated token sequence mask: updated mask (fewer positions still masked)

Return type:

tokens

get_mask_ratio(step)[source][source]

Get mask ratio for a given step.

CONTRACT: - mask_ratio is the fraction of tokens that should REMAIN masked - mask_ratio must decay from ~1.0 (step=0) to ~0.0 (step=T-1) - All schedules must return values in [0, 1] - progress is defined as step/(num_steps-1) so final step has progress=1.0

The unmask_ratio = 1 - mask_ratio represents the fraction of currently-masked tokens to unmask at this step.

Parameters:

step (int) – Current step (0-indexed, must be < num_steps)

Return type:

float

Returns:

mask_ratio in [0, 1], decaying from ~1 at step=0 to ~0 at step=T-1

get_temperature(step)[source][source]
Parameters:

step (int)

Return type:

float

get_mask(logits, current_tokens, mask_token, step)[source][source]
Parameters:
Return type:

Tensor

class medlatents.sampling.RunningConfidenceRemasker(decay=0.9, remask_threshold=0.1, min_steps_before_remask=2, max_remask_per_step=0.1)[source][source]

Bases: object

Running Confidence Remasking for improved MaskGIT generation.

Instead of using only the current step’s confidence, this tracks confidence across multiple steps and uses running statistics to make more stable unmasking decisions.

Reference: “Improving MaskGIT with Running Confidence” (2024)

Parameters:

  • decay (float, default: 0.9) – Exponential decay for running confidence (0.9 = 90% old, 10% new)

  • remask_threshold (float, default: 0.1) – Tokens below this percentile of running confidence get remasked

  • min_steps_before_remask (int, default: 2) – Don’t remask until this many steps have passed

  • max_remask_per_step (float, default: 0.1) – Maximum fraction of tokens to remask per step

__init__(decay=0.9, remask_threshold=0.1, min_steps_before_remask=2, max_remask_per_step=0.1)[source][source]
Parameters:

  • decay (float, default: 0.9)

  • remask_threshold (float, default: 0.1)

  • min_steps_before_remask (int, default: 2)

  • max_remask_per_step (float, default: 0.1)

reset()[source][source]

Reset running statistics for new generation.

update(confidence, mask)[source][source]

Update running confidence and compute remasking decisions.

Parameters:
Return type:

Tensor

class medlatents.sampling.RunningConfidenceScheduler(remasker=None, **kwargs)[source][source]

Bases: medlatents.sampling.maskgit.sampling.MaskGITScheduler

MaskGIT scheduler with running confidence remasking.

Parameters:

remasker (RunningConfidenceRemasker | None, default: None)

__init__(remasker=None, **kwargs)[source][source]
Parameters:

remasker (RunningConfidenceRemasker | None, default: None)

step(logits, tokens, mask, step, temperature=None)[source][source]

Perform one step with running confidence remasking.

Uses fixed K_MAX for topk to enable torch.compile(fullgraph=True). Final step unmasks all remaining positions without topk.

Parameters:
Return type:

tuple[Tensor, Tensor]

reset()[source][source]

Reset remasker state for new generation.

get_mask(logits, current_tokens, mask_token, step)[source]
Parameters:
Return type:

Tensor

get_mask_ratio(step)[source]

Get mask ratio for a given step.

CONTRACT: - mask_ratio is the fraction of tokens that should REMAIN masked - mask_ratio must decay from ~1.0 (step=0) to ~0.0 (step=T-1) - All schedules must return values in [0, 1] - progress is defined as step/(num_steps-1) so final step has progress=1.0

The unmask_ratio = 1 - mask_ratio represents the fraction of currently-masked tokens to unmask at this step.

Parameters:

step (int) – Current step (0-indexed, must be < num_steps)

Return type:

float

Returns:

mask_ratio in [0, 1], decaying from ~1 at step=0 to ~0 at step=T-1

get_temperature(step)[source]
Parameters:

step (int)

Return type:

float

property num_iterations: int
class medlatents.sampling.RCRGenerator(scheduler, mask_token)[source][source]

Bases: object

Running Confidence Remasking Generator for MaskGIT.

Parameters:
__init__(scheduler, mask_token)[source][source]
Parameters:
generate(model, batch_size, seq_len, condition=None, condition_mask=None)[source][source]

Generate tokens with running confidence remasking.

Uses fixed-iteration loop (no Python breaks for compile-friendliness).

Parameters:
Return type:

tuple[Tensor, dict]

class medlatents.sampling.KLASSGenerator(scheduler, kl_threshold=0.01, min_steps=3, max_steps=None, window_size=2)[source][source]

Bases: object

KL-Adaptive Stability Sampling (KLASS) for MaskGIT and masked diffusion.

Dynamically stops generation when predictions stabilize, rather than using a fixed number of steps. Can significantly speed up inference.

Reference: “KL-Guided Fast Inference in Masked Diffusion Models” (arXiv:2511.05664)

Parameters:

  • scheduler (MaskGITScheduler) – Base MaskGITScheduler for masking and temperature

  • kl_threshold (float, default: 0.01) – Stop when KL divergence falls below this threshold

  • min_steps (int, default: 3) – Minimum number of steps before early stopping

  • max_steps (int | None, default: None) – Maximum number of steps (hard limit)

  • window_size (int, default: 2) – Number of steps to average KL over

__init__(scheduler, kl_threshold=0.01, min_steps=3, max_steps=None, window_size=2)[source][source]
Parameters:
should_stop(step, kl_divergence)[source][source]

Determine if generation should stop based on KL divergence.

Uses tensor operations to avoid GPU->CPU sync in the hot sampling loop.

Parameters:

  • step (int) – Current step number

  • kl_divergence (Tensor) – KL divergence tensor (scalar)

Return type:

bool

Returns:

True if generation should stop, False otherwise

generate(model, initial_tokens, mask_token)[source][source]

Generate tokens with KL-guided early stopping.

Uses fixed-iteration loop (no Python breaks for compile-friendliness).

Parameters:

  • model (Module) – Model that takes (tokens, mask_ratio) and returns logits

  • initial_tokens (Tensor) – [batch, seq_len] initial tokens (typically all masked)

  • mask_token (int) – Mask token ID

Returns:

[batch, seq_len] generated tokens metrics: Dict with generation statistics

Return type:

tokens

class medlatents.sampling.DDIMScheduler(eta=0.0, **kwargs)[source][source]

Bases: medlatents.sampling.diffusion.DiscreteScheduler

DDIM scheduler for faster sampling.

Reference: “Denoising Diffusion Implicit Models” (Song et al., 2020) https://arxiv.org/abs/2010.02502

Allows deterministic sampling and can skip timesteps for faster generation.

Parameters:

eta (float, default: 0.0)

__init__(eta=0.0, **kwargs)[source][source]
Parameters:

eta (float, default: 0.0) – stochasticity parameter (0 = deterministic DDIM, 1 = DDPM)

step(model_output, timestep, sample, prev_timestep=None)[source][source]

DDIM sampling step for discrete data.

Parameters:

  • model_output (Tensor) – [batch, seq_len, vocab_size] predicted probabilities

  • timestep (int) – current timestep

  • sample (Tensor) – [batch, seq_len] current discrete sample

  • prev_timestep (int | None, default: None) – previous timestep (for skipping)

Returns:

[batch, seq_len] denoised sample

Return type:

prev_sample

set_timesteps(num_inference_steps)[source]

Set the number of inference steps.

Parameters:

num_inference_steps (int)

class medlatents.sampling.DPMSolverScheduler(solver_order=2, prediction_type='epsilon', **kwargs)[source][source]

Bases: medlatents.sampling.diffusion.DiscreteScheduler

DPM-Solver scheduler for high-quality fast sampling.

Reference: “DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling” https://arxiv.org/abs/2206.00927

Achieves high quality with very few steps (10-20).

Parameters:

  • solver_order (int, default: 2)

  • prediction_type (str, default: 'epsilon')

__init__(solver_order=2, prediction_type='epsilon', **kwargs)[source][source]
Parameters:

  • solver_order (int, default: 2) – order of DPM-Solver (1, 2, or 3)

  • prediction_type (str, default: 'epsilon') – ‘epsilon’ or ‘sample’

step(model_output, timestep, sample)[source][source]

DPM-Solver sampling step adapted for discrete data.

Uses multi-step method for improved quality.

Parameters:
Return type:

Tensor

set_timesteps(num_inference_steps)[source]

Set the number of inference steps.

Parameters:

num_inference_steps (int)

class medlatents.sampling.EulerDiscreteScheduler(**kwargs)[source][source]

Bases: medlatents.sampling.diffusion.DiscreteScheduler

Euler discrete scheduler - simple ODE solver.

Good balance of speed and quality. Similar to DDPM but can skip steps.

__init__(**kwargs)[source][source]
step(model_output, timestep, sample)[source][source]

Euler method step for discrete sampling.

Parameters:

  • model_output (Tensor) – [batch, seq_len, vocab_size] predicted probabilities

  • timestep (int) – current timestep

  • sample (Tensor) – [batch, seq_len] current sample

Returns:

[batch, seq_len] denoised sample

Return type:

prev_sample

set_timesteps(num_inference_steps)[source]

Set the number of inference steps.

Parameters:

num_inference_steps (int)

class medlatents.sampling.AncestralSamplingScheduler(noise_scale=1.0, **kwargs)[source][source]

Bases: medlatents.sampling.diffusion.DiscreteScheduler

Ancestral sampling (stochastic) scheduler.

Adds noise at each step for more diverse samples. Good for generation tasks where diversity is important.

Parameters:

noise_scale (float, default: 1.0)

__init__(noise_scale=1.0, **kwargs)[source][source]
Parameters:

noise_scale (float, default: 1.0) – scale of added noise (1.0 = standard, >1.0 = more noise)

step(model_output, timestep, sample)[source][source]

Ancestral sampling step with added noise.

Parameters:

  • model_output (Tensor) – [batch, seq_len, vocab_size] predicted probabilities

  • timestep (int) – current timestep

  • sample (Tensor) – [batch, seq_len] current sample

Returns:

[batch, seq_len] noisy denoised sample

Return type:

prev_sample

set_timesteps(num_inference_steps)[source]

Set the number of inference steps.

Parameters:

num_inference_steps (int)

medlatents.sampling.sample_with_cfg_flow(model, x_t, t, condition=None, guidance_scale=1.0, null_condition=None)[source][source]

Classifier-Free Guidance for flow matching models.

Parameters:

  • model (Module) – Flow matching model

  • x_t (Tensor) – Current state [batch, seq_len] or [batch, seq_len, dim]

  • t (Tensor) – Time values [batch] or [batch, 1]

  • condition (Tensor | None, default: None) – Conditioning tensor (class labels, text embeddings, etc.)

  • guidance_scale (float, default: 1.0) – CFG scale (1.0 = no guidance, >1.0 = stronger conditioning)

  • null_condition (Tensor | None, default: None) – Null/unconditional token for CFG

Return type:

Tensor

Returns:

Guided model output (logits or probabilities)

medlatents.sampling.sample_with_cfg_diffusion(model, x_t, t, condition=None, guidance_scale=1.0, null_condition=None, model_kwargs=None)[source][source]

Classifier-Free Guidance for diffusion models.

Parameters:

  • model (Module) – Diffusion model (predicts noise or x_0)

  • x_t (Tensor) – Noisy sample [batch, channels, …] or [batch, seq_len, dim]

  • t (Tensor) – Timestep values [batch] or [batch, 1]

  • condition (Tensor | None, default: None) – Conditioning tensor (class labels, text embeddings, etc.)

  • guidance_scale (float, default: 1.0) – CFG scale (1.0 = no guidance, >1.0 = stronger conditioning)

  • null_condition (Tensor | None, default: None) – Null/unconditional token for CFG

  • model_kwargs (dict | None, default: None) – Additional model arguments

Return type:

Tensor

Returns:

Guided model prediction (noise or x_0)

class medlatents.sampling.CFGFlowMatcher(model, guidance_scale=1.5, null_condition=None)[source][source]

Bases: object

Wrapper for flow matching models with built-in CFG support.

Simplifies CFG inference by handling conditional/unconditional splits.

Parameters:
__init__(model, guidance_scale=1.5, null_condition=None)[source][source]
Parameters:
__call__(x_t, t, condition=None)[source][source]

Forward pass with CFG.

Parameters:
Return type:

Tensor

class medlatents.sampling.CFGDiffuser(model, guidance_scale=1.5, null_condition=None)[source][source]

Bases: object

Wrapper for diffusion models with built-in CFG support.

Simplifies CFG inference by handling conditional/unconditional splits.

Parameters:
__init__(model, guidance_scale=1.5, null_condition=None)[source][source]
Parameters:
__call__(x_t, t, condition=None, **model_kwargs)[source][source]

Forward pass with CFG.

Parameters:
Return type:

Tensor

medlatents.sampling.generate_with_cfg(model, solver, x_init, condition=None, guidance_scale=1.0, null_condition=None, model_type='flow', **solver_kwargs)[source][source]

Unified generation with CFG for flow matching or diffusion.

Parameters:

  • model (Module) – Model (flow matching or diffusion)

  • solver (Callable) – Solver function (e.g., MixtureDiscreteEulerSolver or scheduler.step)

  • x_init (Tensor) – Initial noise/state

  • condition (Tensor | None, default: None) – Conditioning information

  • guidance_scale (float, default: 1.0) – CFG scale

  • null_condition (Tensor | None, default: None) – Unconditional token

  • model_type (str, default: 'flow') – ‘flow’ or ‘diffusion’

  • **solver_kwargs – Additional solver arguments

Return type:

Tensor

Returns:

Generated samples

medlatents.sampling.rescale_cfg(pred_cond, pred_uncond, guidance_scale, rescale_factor=0.7)[source][source]

CFG with rescaling to prevent oversaturation.

From “Common Diffusion Noise Schedules and Sample Steps are Flawed” https://arxiv.org/abs/2305.08891

Parameters:

  • pred_cond (Tensor) – Conditional prediction

  • pred_uncond (Tensor) – Unconditional prediction

  • guidance_scale (float) – CFG scale

  • rescale_factor (float, default: 0.7) – Rescaling factor (0.7 is recommended)

Return type:

Tensor

Returns:

Rescaled CFG output

medlatents.sampling.sample_with_dynamic_cfg(model, x_t, t, condition=None, guidance_schedule=<function <lambda>>, null_condition=None, model_type='flow')[source][source]

CFG with dynamic guidance scale based on timestep.

Useful for adaptive guidance (e.g., stronger at early steps, weaker at late steps).

Parameters:

  • model (Module) – Model (flow matching or diffusion)

  • x_t (Tensor) – Current state

  • t (Tensor) – Time/timestep values [batch]

  • condition (Tensor | None, default: None) – Conditioning tensor

  • guidance_schedule (Callable[[float | Tensor], float | Tensor], default: <function <lambda> at 0x7f9d459084a0>) – Function t -> guidance_scale (supports float or tensor input)

  • null_condition (Tensor | None, default: None) – Unconditional token

  • model_type (str, default: 'flow') – ‘flow’ or ‘diffusion’

Return type:

Tensor

Returns:

Guided model output

medlatents.sampling.sample_bayesian_flow(model, batch_size, seq_len, num_steps=None, temperature=1.0, temperature_schedule=None, condition=None, guidance_scale=None, null_condition=None, **model_kwargs)[source][source]
Parameters:
Return type:

Int[Tensor, 'batch seq']

medlatents.sampling.sample_with_early_stopping(model, batch_size, seq_len, confidence_threshold=0.95, min_steps=10, max_steps=None, temperature=1.0, **model_kwargs)[source][source]
Parameters:
Return type:

tuple[Int[Tensor, 'batch seq'], int]

medlatents.sampling.get_confidence(params)[source][source]
Parameters:

params (Float[Tensor, 'batch seq classes'])

Return type:

Float[Tensor, 'batch seq']

class medlatents.sampling.BayesianFlowSampler(model, default_temperature=1.0, default_num_steps=None)[source][source]

Bases: object

Convenience wrapper for Bayesian Flow sampling strategies.

Parameters:
__init__(model, default_temperature=1.0, default_num_steps=None)[source][source]
Parameters:
sample(batch_size, seq_len, num_steps=None, temperature=None, **model_kwargs)[source][source]
Parameters:

  • batch_size (int)

  • seq_len (int)

  • num_steps (int | None, default: None)

  • temperature (float | None, default: None)

Return type:

Int[Tensor, 'batch seq']

sample_with_cfg(batch_size, seq_len, condition, guidance_scale=1.5, null_condition=None, num_steps=None, temperature=None, **model_kwargs)[source][source]
Parameters:
Return type:

Int[Tensor, 'batch seq']

sample_with_schedule(batch_size, seq_len, temperature_schedule, **model_kwargs)[source][source]
Parameters:
Return type:

Int[Tensor, 'batch seq']

sample_with_early_stopping(batch_size, seq_len, confidence_threshold=0.95, min_steps=10, max_steps=None, temperature=None, **model_kwargs)[source][source]
Parameters:

  • batch_size (int)

  • seq_len (int)

  • confidence_threshold (float, default: 0.95)

  • min_steps (int, default: 10)

  • max_steps (int | None, default: None)

  • temperature (float | None, default: None)

Return type:

tuple[Int[Tensor, 'batch seq'], int]

class medlatents.sampling.DecoupledSTGumbelSoftmax(forward_temp=1.0, backward_temp=0.5, hard=True, dim=-1)[source][source]

Bases: torch.nn.modules.module.Module

Decoupled Straight-Through Gumbel-Softmax estimator.

Uses different temperatures for forward (sampling) and backward (gradient) passes. This provides better gradient estimates while maintaining sharp samples.

Reference: “Improving Discrete Optimisation Via Decoupled Straight-Through

Gumbel-Softmax” (arXiv:2410.13331)

Parameters:

  • forward_temp (float, default: 1.0) – Temperature for forward pass (sampling)

  • backward_temp (float, default: 0.5) – Temperature for backward pass (gradients)

  • hard (bool, default: True) – If True, use straight-through estimator

  • dim (int, default: -1) – Dimension to softmax over

__init__(forward_temp=1.0, backward_temp=0.5, hard=True, dim=-1)[source][source]

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

Parameters:

  • forward_temp (float, default: 1.0)

  • backward_temp (float, default: 0.5)

  • hard (bool, default: True)

  • dim (int, default: -1)

forward(logits)[source][source]

Apply decoupled ST-Gumbel-Softmax.

Parameters:

logits (Tensor) – [batch, seq_len, vocab_size] unnormalized logits

Returns:

[batch, seq_len, vocab_size] one-hot if hard=True, else soft samples

Return type:

samples

T_destination = ~T_destination
add_module(name, module)[source]

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)[source]

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()[source]

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)[source]

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()[source]

Return an iterator over immediate children modules.

Yields:

Module – a child module

Return type:

Iterator[Module]

compile(*args, **kwargs)[source]

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()[source]

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)[source]

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()[source]

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()[source]

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()[source]

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()[source]

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)[source]

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()[source]

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)[source]

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)[source]

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()[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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()[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

Alias for add_module().

Parameters:
Return type:

None

register_parameter(name, param)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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()[source]

See torch.Tensor.share_memory_().

Return type:

Self

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

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)[source]

Move and/or cast the parameters and buffers.

This can be called as

to(device=None, dtype=None, non_blocking=False)[source]
to(dtype, non_blocking=False)[source]
to(tensor, non_blocking=False)[source]
to(memory_format=torch.channels_last)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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.sampling.ReinMax(temperature=1.0, dim=-1)[source][source]

Bases: torch.nn.modules.module.Module

ReinMax: Second-order accurate straight-through estimator.

Provides better gradient estimates than standard ST-Gumbel.

Reference: “Bridging Discrete and Backpropagation: Straight-Through…=” (NeurIPS 2023)

Parameters:

  • temperature (float, default: 1.0) – Sampling temperature

  • dim (int, default: -1) – Dimension to softmax over

__init__(temperature=1.0, dim=-1)[source][source]

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

Parameters:

  • temperature (float, default: 1.0)

  • dim (int, default: -1)

forward(logits)[source][source]

Apply ReinMax estimator.

Parameters:

logits (Tensor) – [batch, seq_len, vocab_size] unnormalized logits

Returns:

[batch, seq_len, vocab_size]

Return type:

samples

T_destination = ~T_destination
add_module(name, module)[source]

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)[source]

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()[source]

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)[source]

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()[source]

Return an iterator over immediate children modules.

Yields:

Module – a child module

Return type:

Iterator[Module]

compile(*args, **kwargs)[source]

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()[source]

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)[source]

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()[source]

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()[source]

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()[source]

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()[source]

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)[source]

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()[source]

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)[source]

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)[source]

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()[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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()[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

Alias for add_module().

Parameters:
Return type:

None

register_parameter(name, param)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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()[source]

See torch.Tensor.share_memory_().

Return type:

Self

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

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)[source]

Move and/or cast the parameters and buffers.

This can be called as

to(device=None, dtype=None, non_blocking=False)[source]
to(dtype, non_blocking=False)[source]
to(tensor, non_blocking=False)[source]
to(memory_format=torch.channels_last)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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)[source]

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.sampling.GuidedSampler(model, uncond_model=None, guidance_schedule='constant', base_scale=5.0)[source][source]

Bases: object

Applies classifier-free guidance with time-dependent schedules.

Reference: - “Theory-Informed Improvements to Classifier-Free Guidance” (arXiv:2507.08965) - “What Does Guidance Do in Masked Discrete Diffusion?” (arXiv:2506.10971)

Parameters:

  • model (Module) – Model that takes (x, t, conditioning) and returns logits

  • uncond_model (Module | None, default: None) – Optional unconditioned model (if None, uses empty conditioning)

  • guidance_schedule (str | Callable[[float, float], float], default: 'constant') – Schedule function or name

  • base_scale (float, default: 5.0) – Base guidance scale

__init__(model, uncond_model=None, guidance_schedule='constant', base_scale=5.0)[source][source]
Parameters:
__call__(x, t, conditioning=None, progress=0.5)[source][source]

Get guided predictions.

Parameters:

  • x (Tensor) – Input tensor

  • t (Tensor) – Time tensor

  • conditioning (dict[str, Tensor] | None, default: None) – Conditioning dict (class labels, text, etc.)

  • progress (float, default: 0.5) – Sampling progress (0 to 1) for guidance schedule

Return type:

Tensor

Returns:

Guided logits

medlatents.sampling.get_guidance_schedule(schedule)[source][source]

Get a guidance schedule function by name.

Parameters:

schedule (str) – Schedule type name. One of: constant, linear, cosine, cosine_decay, triangular, cfg_zero, cfg_zero_star, dynamic.

Return type:

Callable[[float, float], float]

Returns:

Function that takes (progress, base_scale) and returns a guidance scale.

Examples

>>> schedule = get_guidance_schedule("cosine")
>>> scale = schedule(0.5, 5.0)  # guidance scale at 50% progress
medlatents.sampling.constant_guidance(progress, base_scale)[source][source]

Constant guidance weight base_scale for the whole trajectory.

Parameters:
Return type:

float

medlatents.sampling.linear_guidance(progress, base_scale)[source][source]

Linear ramp of the guidance weight from 1.0 to base_scale.

Parameters:
Return type:

float

medlatents.sampling.cosine_guidance(progress, base_scale)[source][source]

Cosine ramp of the guidance weight from 1.0 (at t=0) to base_scale (at t=1).

Parameters:
Return type:

float

medlatents.sampling.cosine_decay_guidance(progress, base_scale)[source][source]

Cosine decay of the guidance weight from base_scale (at t=0) to 1.0 (at t=1).

Parameters:
Return type:

float

medlatents.sampling.triangular_guidance(progress, base_scale)[source][source]

Triangular schedule peaking at base_scale at the midpoint (t=0.5).

Parameters:
Return type:

float

medlatents.sampling.cfg_zero_guidance(progress, base_scale)[source][source]

CFG-Zero time-dependent guidance weighting.

Uses parabolic weighting: w(t) = 4*t*(1-t) which peaks at t=0.5 and is zero at the boundaries. This prevents over-guidance at the start and end of sampling.

Reference:

“Classifier-Free Guidance is a Predictor-Corrector” and related work on time-dependent CFG weighting.

Parameters:

  • progress (float) – Progress from 0 to 1 (corresponds to time t)

  • base_scale (float) – Base guidance scale

Returns:

1 + (base_scale - 1) * 4*t*(1-t)

Return type:

Time-weighted guidance scale

medlatents.sampling.cfg_zero_star_guidance(progress, base_scale, zero_init_steps=0.1, scale_factor=1.0)[source][source]

CFG-Zero* improved guidance (2025).

Adds two improvements over CFG-Zero: 1. Zero-init: Zero out guidance for the first few steps (pure noise region) 2. Optimized scale: Apply a learned/tuned scale factor

Reference: “CFG-Zero*: Improved Classifier-Free Guidance for Flow Matching Models”

(arXiv:2503.18886)

Parameters:

  • progress (float) – Progress from 0 to 1 (corresponds to time t)

  • base_scale (float) – Base guidance scale

  • zero_init_steps (float, default: 0.1) – Fraction of steps at start to zero out guidance (default 0.1 = 10%)

  • scale_factor (float, default: 1.0) – Scale factor for guidance (tune per model, default 1.0)

Return type:

float

Returns:

Optimized guidance scale

medlatents.sampling.dynamic_guidance(progress, base_scale, warmup=0.2, cooldown=0.1)[source][source]

Dynamic guidance schedule with warmup and cooldown.

Provides weak guidance at start (noise region) and end (fine details), with strong guidance in the middle where structure forms.

Based on findings from discrete diffusion guidance research (2024-2025).

Parameters:

  • progress (float) – Progress from 0 to 1

  • base_scale (float) – Maximum guidance scale

  • warmup (float, default: 0.2) – Fraction of steps for warmup (weak guidance)

  • cooldown (float, default: 0.1) – Fraction of steps for cooldown (weak guidance)

Return type:

float

Returns:

Guidance scale at this progress

medlatents.sampling.is_compile_available()[source][source]

Check if torch.compile is available (PyTorch 2.0+).

Return type:

bool

medlatents.sampling.compile_model(model, mode='reduce-overhead', fullgraph=False, dynamic=True, disable=False)[source][source]

Compile a model for faster inference using torch.compile.

This is a no-op on PyTorch < 2.0.

Parameters:

  • model (Module) – Model to compile

  • mode (Literal['default', 'reduce-overhead', 'max-autotune'], default: 'reduce-overhead') – Compilation mode: - “default”: Good balance of compile time and runtime - “reduce-overhead”: Minimize overhead, best for small models - “max-autotune”: Maximum optimization, longer compile time

  • fullgraph (bool, default: False) – If True, require entire function to be captured as single graph

  • dynamic (bool, default: True) – If True, enable dynamic shape support (recommended)

  • disable (bool, default: False) – If True, skip compilation (useful for debugging)

Return type:

Module

Returns:

Compiled model (or original if torch.compile unavailable)

Example

>>> model = MaskGIT(vocab_size=8192, seq_length=1024)
>>> model = compile_model(model, mode="reduce-overhead")
>>> # Now generate() calls will be faster after warmup
>>> samples = model.generate(initial_tokens)
medlatents.sampling.compile_function(fn, mode='reduce-overhead', fullgraph=False, dynamic=True, disable=False)[source][source]

Compile a function for faster execution using torch.compile.

This is a no-op on PyTorch < 2.0.

Parameters:

  • fn (Callable) – Function to compile

  • mode (Literal['default', 'reduce-overhead', 'max-autotune'], default: 'reduce-overhead') – Compilation mode (see compile_model)

  • fullgraph (bool, default: False) – If True, require entire function to be captured

  • dynamic (bool, default: True) – If True, enable dynamic shape support

  • disable (bool, default: False) – If True, skip compilation

Return type:

Callable

Returns:

Compiled function (or original if torch.compile unavailable)

Example

>>> @compile_function
... def sample_step(logits, mask):
...     probs = F.softmax(logits, dim=-1)
...     return torch.multinomial(probs, 1)
class medlatents.sampling.CompiledSampler(model, mode='reduce-overhead', warmup_steps=3)[source][source]

Bases: object

Wrapper that provides compiled sampling for any discrete generative model.

This wrapper compiles the model’s forward pass and provides optimized sampling methods. Use this for maximum inference throughput.

Parameters:

  • model (Module) – Model to wrap (MaskGIT, AR transformer, etc.)

  • mode (Literal['default', 'reduce-overhead', 'max-autotune'], default: 'reduce-overhead') – Compilation mode

  • warmup_steps (int, default: 3) – Number of warmup calls before measuring (compilation happens here)

Example

>>> model = MaskGIT(vocab_size=8192, seq_length=1024)
>>> sampler = CompiledSampler(model)
>>> # First call triggers compilation (slow)
>>> samples = sampler.generate(initial_tokens)
>>> # Subsequent calls are fast
>>> samples = sampler.generate(initial_tokens)
__init__(model, mode='reduce-overhead', warmup_steps=3)[source][source]
Parameters:

  • model (Module)

  • mode (Literal['default', 'reduce-overhead', 'max-autotune'], default: 'reduce-overhead')

  • warmup_steps (int, default: 3)

property is_compiled: bool

Check if the model is compiled.

forward(*args, **kwargs)[source][source]

Forward pass through the (potentially compiled) model.

__call__(*args, **kwargs)[source][source]

Call the model.

generate(*args, **kwargs)[source][source]

Generate samples using the model’s generate method.

Delegates to the underlying model’s generate() method.

warmup(sample_input, num_steps=None)[source][source]

Warmup the compiled model with sample inputs.

This triggers compilation and optimizes the compute graph.

Parameters:

  • sample_input (Tensor) – Sample input tensor for warmup

  • num_steps (int | None, default: None) – Number of warmup steps (default: self.warmup_steps)

Return type:

None