utils.ctx_managers.sequence_parallel

utils.ctx_managers.sequence_parallel

Module for Axolotl trainer sequence parallelism manager and utilities

Classes

Name	Description
AllGatherWithGrad	Custom autograd function for all-gather to preserve gradients.
SequenceParallelContextManager	Context manager for sequence parallelism operations.

AllGatherWithGrad

utils.ctx_managers.sequence_parallel.AllGatherWithGrad()

Custom autograd function for all-gather to preserve gradients.

Methods

Name	Description
backward	Backward pass for all-gather operation.
forward	Forward pass of all-gather of data with sequence dimension.

backward

utils.ctx_managers.sequence_parallel.AllGatherWithGrad.backward(
    ctx,
    grad_output,
)

Backward pass for all-gather operation.

Extracts the gradient slice corresponding to this rank’s original input from the full gradient tensor.

Parameters

Name	Type	Description	Default
ctx	torch.autograd.function.FunctionCtx	torch.autograd function context.	required
grad_output	torch.Tensor	Gradient from subsequent layers with respect to the concatenated output tensor.	required

Returns

Name	Type	Description
	tuple[torch.Tensor, None]	Tuple containing the gradient slice for this rank’s input tensor and None for the process group parameter which doesn’t require gradients.

forward

utils.ctx_managers.sequence_parallel.AllGatherWithGrad.forward(
    ctx,
    input_tensor,
    group,
)

Forward pass of all-gather of data with sequence dimension.

Parameters

Name	Type	Description	Default
ctx	torch.autograd.function.FunctionCtx	torch.autograd function context.	required
input_tensor	torch.Tensor	Tensor from model output with sequence dimension.	required
group	dist.ProcessGroup	torch.distributed process group.	required

Returns

Name	Type	Description
	torch.Tensor	Tensor from gathering the input_tensor from across the process group and concatenating along the sequence dimension.

SequenceParallelContextManager

utils.ctx_managers.sequence_parallel.SequenceParallelContextManager(
    models,
    sequence_parallel_degree,
    gradient_accumulation_steps,
    ring_attn_func,
    heads_k_stride,
)

Context manager for sequence parallelism operations.

This class provides a context that will automatically apply sequence parallelism during model forward passes using a pre-forward hook, and gather outputs from across the sequence parallelism group using a post-forward hook.

Parameters

Name	Type	Description	Default
models	list[nn.Module]	List of models to apply sequence parallelism to pre- and post- forward hooks.	required
sequence_parallel_degree	int	Number of processes to split sequences over.	required
gradient_accumulation_steps	int	Number of steps to accumulate gradients over.	required
ring_attn_func	RingAttnFunc	Which ring attention function to use. Currently unused.	required
heads_k_stride	int | None	Sequence parallelism K head stride size. Passed through to varlen_llama3 ring_flash_attn implementation.	required

Functions

Name	Description
apply_sequence_parallelism	Apply sequence parallelism slicing to a batch.

apply_sequence_parallelism

utils.ctx_managers.sequence_parallel.apply_sequence_parallelism(
    batch,
    local_rank,
    local_world_size,
    gradient_accumulation_steps,
    ring_attn_func,
)

Apply sequence parallelism slicing to a batch.

Special handling is implemented for integer logits_to_keep, which indicates to only keep the last N tokens in the sequence during generation.

Parameters

Name	Type	Description	Default
batch	dict[str, torch.Tensor]	Batch dictionary (e.g., input_ids, attention_mask, etc.).	required
local_rank	int	Local rank in the sequence parallel group.	required
local_world_size	int	World size of the sequence parallel group.	required
gradient_accumulation_steps	int	Number of steps to accumulate gradients over.	required
ring_attn_func	RingAttnFunc	Which ring attention function to use. Currently unused, but related to above TODO.	required

Returns

Name	Type	Description
	tuple[dict[str, torch.Tensor], int, int]	tuple of: - Batch dictionary with sliced tensors. - The original sequence length before padding. - The number of padding tokens added.