Patch the _get_embeddings method of MultiTokenPredictionLayer

Patch the Megatron-Bridge provider to support QAT quantized layers.

Recover left padding from result return result

Postprocess bshd sequences

_summary_ For fused kernels, the output is a dictionary with keys like 'log_probs', 'entropy', etc. This function post-processes each tensor i

Gather local router maps from all PP ranks into a global router map. Args: local_layers_router_map (torch.Tensor): Local router map

Remove left padding from input_ids, attention_mask and position_ids return new_input_ids, new_attention_mask, new_position_ids

AllGather data from sp region This is because the data is first sharded along the FSDP dimension as we utilize the DP_COMPUTE

Converts a video dict into a [n_frames, 3, H, W] tensor Add video sample FPS in a future MR

Put information dictionary and extra lines on an image. Args: image: Input image info: Dictionary of key-value pairs to disp

Put a batch sample into the queue Args: sample: Sample data Returns: bool: Whether the sample was s

FP8 quantization based on parameter name using a memory-efficient generator. Args: weights: Generator, AsyncGenerator, or iterab

Streaming quantize: consume input layer by layer, yield (name, tensor) pairs.

Query the collect info for a given mesh name. Args: mesh_name (str): Name of the mesh to query collect info for.

Deduplicate consecutive image tokens in prompt_ids for Qwen2.5-VL, since vLLM will replicate the <|image_pad|> and <|video_pad|> token by image_da

Access underlying torch stream.

Block until the remote agent sends the metadata. Returns: dict: Metadata from the remote agent.

Rebuild GPU-resident buffers from CPU state. Call this before training resumes after release_gpu_buffers().

Release weights and kv cache in GPU memory.

Release all GPU-resident buffers to free GPU memory. Call this when the GPU is needed for other purposes (e.g., inference engine). Us

Remove items from the replay buffer. Args: partition_id (str): Partition of transfer queue, e.g. "train" or "val". ke

Reorder and merge per-VPP layer blocks into a contiguous layer dimension. Given a tensor shaped as [bs*vpp_size, max_token_len, layer_num_pe

This is the equivalent of torch.repeat_interleave(x, dim=2, repeats=n_rep). The hidden states go from (batch, seqlen, num_key_value_heads, he

Reset all statistics.

Resolve agent loop configuration file path. In multi-node Ray training, relative paths may not resolve correctly because the working director

Convert padded response tensor to nested tensor. Args: tensor: a tensor with shape (bsz, response_len) response_mask: a nested te

Restore base model weights from CPU backup. This function restores the base model weights from the CPU backup, effectively undoing any LoRA m

Restore the model state from CPU memory. Args: n: Identifier/Key for the saved model state to restore.

Context switch hybridengine to rollout mode.

Context switch hybridengine to rollout mode.

Execute the graph in topological order.

Run forward and backward pass, return logits and gradients.

Compute actions and their log probabilities from state features. Args: state_features: Any data structure representing the proces

Get the parameters of the critic head for optimization. Returns: A list of torch.nn.Parameter objects representing the critic hea

Repeat each row of the batch data a specified number of times. Args: repeat_times (torch.tensor, list, tuple, ndarray):

Generate Gaussian noise for the action trajectory. Args: shape: Desired output shape, typically (B, n_action_steps, action_dim).

Save an FSDP checkpoint for this rank. Writes: - model & optimizer shard files - extra state dict (scheduler + R

Save the current model state to CPU memory. Args: n: Identifier/Key for the saved model state.

Saves an MP4 replay of an episode. Args: rollout_images: List of images from the episode output_dir: Directory to save the v

Cast tensor from high precision to FP8 with blockwise quantization. This function automatically selects the best available implementation: 1.

Send the weights of the model. Args: weights: A generator that yields the name of the weight tensor and the tensor itself.

Kill the current process when the parent process exits.

Sets if the module should reshard parameters after forward. This can be used to change the ``reshard_after_forward`` FSDP arg at runtime. For

Scatter the packed router top-k indices back to sequence-parallel ranks and update each local RouterReplay instance with target indices for r

Sets the target topk indices for replay.

Set timesteps and sigmas on the scheduler and move them to *device*. Args: scheduler (SchedulerMixin): the scheduler used for the

Sleep all rollout replica instances.

Sort the placement groups by node ip, all bundles in a single placement group should be on the same node. FSDPCheckpointManager saves sharde

Split the elements in a tuple/interval, while handling well-formatted commas in large numbers

Block until all ops on this stream complete.

Copied from maniskill https://github.com/haosulab/ManiSkill Tile multiple images to a single image comprised of nrows and an appropriate

Move working copy to device.

Manual control of load/offload

Refresh shared-memory weight flats from CPU master params. Called by main process after optimizer.step().

Verify if the solution is correct. Args: solution_str: The solution string to verify answer: The ground truth answer stri

Wait for the broadcast operation to complete. Returns: list[ParameterMeta]: The bucket meta after broadcast.