Commit 12b56c98 authored by dongcl's avatar dongcl
Browse files

support a2a overlap

parent 8551c38e
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
import contextlib
import weakref
from typing import Optional
import torch
from torch import Tensor
from megatron.core.pipeline_parallel.combined_1f1b import (
AbstractSchedulePlan,
FakeScheduleNode,
FreeInputsMemoryStrategy,
NoOpMemoryStrategy,
ScheduleNode,
get_com_stream,
get_comp_stream,
make_viewless,
)
from megatron.core.transformer import transformer_layer
from megatron.core.transformer.module import float16_to_fp32
def weak_method(method):
"""Creates a weak reference to a method to prevent circular references.
This function creates a weak reference to a method and returns a wrapper function
that calls the method when invoked. This helps prevent memory leaks from circular
references.
"""
method_ref = weakref.WeakMethod(method)
del method
def wrapped_func(*args, **kwarg):
# nonlocal object_ref
return method_ref()(*args, **kwarg)
return wrapped_func
class MemoryStrategyRegistry:
"""Registry for memory management strategies based on node names.
This class centralizes the definition of which memory strategy
should be used for each type of node in the computation graph.
"""
@classmethod
def get_strategy_by_name(cls, name, is_moe, is_deepep):
"""Gets the appropriate memory strategy for a node based on its name and MoE status.
Args:
name: The name of the node, which determines which strategy to use.
is_moe: Whether the node is part of a Mixture of Experts model.
Returns:
The memory strategy to use for the node.
"""
strategies = {
"default": NoOpMemoryStrategy(),
"attn": NoOpMemoryStrategy(), # Attention nodes keep their inputs
"dispatch": (
FreeInputsMemoryStrategy() if not is_deepep else NoOpMemoryStrategy()
), # deepep dispatch inputs share same storage with moe inputs
"mlp": FreeInputsMemoryStrategy(), # MLP nodes free inputs after use
"combine": FreeInputsMemoryStrategy(), # Combine nodes free inputs after use
}
if is_moe:
return strategies.get(name, strategies["default"])
# For dense layers [attn, fake, mlp, fake], the inputs of mlp are required for backward
return NoOpMemoryStrategy()
class PreProcessNode(ScheduleNode):
"""Node responsible for preprocessing operations in the model.
This node handles embedding and rotary positional embedding computations
before the main transformer layers.
"""
def __init__(self, gpt_model, model_chunk_state, event, stream):
"""Initializes a preprocessing node.
Args:
gpt_model: The GPT model instance.
model_chunk_state: State shared across the model chunk.
event: CUDA event for synchronization.
stream: CUDA stream for execution.
"""
super().__init__(weak_method(self.forward_impl), stream, event, name="pre_process")
self.gpt_model = gpt_model
self.model_chunk_state = model_chunk_state
def forward_impl(self):
"""Implements the forward pass for preprocessing.
This method handles:
1. Decoder embedding computation
2. Rotary positional embedding computation
3. Sequence length offset computation for flash decoding
Returns:
The processed decoder input tensor.
"""
gpt_model = self.gpt_model
decoder_input = self.model_chunk_state.decoder_input
input_ids = self.model_chunk_state.input_ids
position_ids = self.model_chunk_state.position_ids
inference_params = self.model_chunk_state.inference_params
packed_seq_params = self.model_chunk_state.packed_seq_params
# Decoder embedding.
if decoder_input is not None:
pass
elif gpt_model.pre_process:
decoder_input = gpt_model.embedding(input_ids=input_ids, position_ids=position_ids)
else:
# intermediate stage of pipeline
# decoder will get hidden_states from encoder.input_tensor
decoder_input = gpt_model.decoder.input_tensor
# Rotary positional embeddings (embedding is None for PP intermediate devices)
rotary_pos_emb = None
rotary_pos_cos = None
rotary_pos_sin = None
if (
gpt_model.position_embedding_type == 'rope'
and not gpt_model.config.multi_latent_attention
):
if not gpt_model.training and gpt_model.config.flash_decode and inference_params:
# Flash decoding uses precomputed cos and sin for RoPE
rotary_pos_cos, rotary_pos_sin = gpt_model.rotary_pos_emb_cache.setdefault(
inference_params.max_sequence_length,
gpt_model.rotary_pos_emb.get_cos_sin(inference_params.max_sequence_length),
)
else:
rotary_seq_len = gpt_model.rotary_pos_emb.get_rotary_seq_len(
inference_params,
gpt_model.decoder,
decoder_input,
gpt_model.config,
packed_seq_params,
)
rotary_pos_emb = gpt_model.rotary_pos_emb(
rotary_seq_len,
packed_seq=packed_seq_params is not None
and packed_seq_params.qkv_format == 'thd',
)
if (
(gpt_model.config.enable_cuda_graph or gpt_model.config.flash_decode)
and rotary_pos_cos is not None
and inference_params
):
sequence_len_offset = torch.tensor(
[inference_params.sequence_len_offset] * inference_params.current_batch_size,
dtype=torch.int32,
device=rotary_pos_cos.device, # Co-locate this with the rotary tensors
)
else:
sequence_len_offset = None
# saved for later use
self.model_chunk_state.rotary_pos_emb = rotary_pos_emb
self.model_chunk_state.rotary_pos_cos = rotary_pos_cos
self.model_chunk_state.rotary_pos_sin = rotary_pos_sin
self.model_chunk_state.sequence_len_offset = sequence_len_offset
return decoder_input
class PostProcessNode(ScheduleNode):
"""Node responsible for postprocessing operations in the model.
This node handles final layer normalization and output layer computation
after the main transformer layers.
"""
def __init__(self, gpt_model, model_chunk_state, event, stream):
"""Initializes a postprocessing node.
Args:
gpt_model: The GPT model instance.
model_chunk_state: State shared across the model chunk.
event: CUDA event for synchronization.
stream: CUDA stream for execution.
"""
super().__init__(weak_method(self.forward_impl), stream, event, name="post_process")
self.gpt_model = gpt_model
self.model_chunk_state = model_chunk_state
def forward_impl(self, hidden_states):
"""Implements the forward pass for postprocessing.
This method handles:
1. Final layer normalization
2. Output layer computation
3. Loss computation if labels are provided
Args:
hidden_states: The hidden states from the transformer layers.
Returns:
The logits or loss depending on whether labels are provided.
"""
# Final layer norm.
if self.gpt_model.decoder.final_layernorm is not None:
hidden_states = self.gpt_model.decoder.final_layernorm(hidden_states)
# TENorm produces a "viewed" tensor. This will result in schedule.py's
# deallocate_output_tensor() throwing an error, so a viewless tensor is
# created to prevent this.
hidden_states = transformer_layer.make_viewless_tensor(
inp=hidden_states, requires_grad=True, keep_graph=True
)
gpt_model = self.gpt_model
runtime_gather_output = self.model_chunk_state.runtime_gather_output
labels = self.model_chunk_state.labels
output_weight = None
if gpt_model.share_embeddings_and_output_weights:
output_weight = gpt_model.shared_embedding_or_output_weight()
logits, _ = gpt_model.output_layer(
hidden_states, weight=output_weight, runtime_gather_output=runtime_gather_output
)
if labels is None:
# [s b h] => [b s h]
return float16_to_fp32(logits.transpose(0, 1).contiguous())
loss = float16_to_fp32(gpt_model.compute_language_model_loss(labels, logits))
return loss
class TransformerLayerNode(ScheduleNode):
"""Base class for transformer layer computation nodes.
This class provides common functionality for different types of
transformer layer nodes (attention, MLP, etc.)
"""
def __init__(self, stream, event, state, callables, name="default"):
"""Initialize a transformer layer node.
Args:
stream (torch.cuda.Stream): CUDA stream for execution
event (torch.cuda.Event): Synchronization event
common_state (TransformerLayerState): State shared within a transformer layer
callables (Callable): The callables contain forward and dw function
it's the per_batch_state_context, o.w. nullcontext
name (str): Node name, also used to determine memory strategy
"""
# Get memory strategy based on node name
memory_strategy = MemoryStrategyRegistry.get_strategy_by_name(
name, callables.is_moe, callables.is_deepep
)
super().__init__(
weak_method(self.forward_impl),
stream,
event,
weak_method(self.backward_impl),
memory_strategy=memory_strategy,
name=name,
)
self.common_state = state
self.callables = callables
self.detached = tuple()
self.before_detached = tuple()
def detach(self, t):
"""Detaches a tensor and stores it for backward computation."""
detached = make_viewless(t).detach()
detached.requires_grad = t.requires_grad
self.before_detached = self.before_detached + (t,)
self.detached = self.detached + (detached,)
return detached
def forward_impl(self, *args):
"""Implements the forward pass for the transformer layer node."""
return self.callables.forward(self, *args)
def backward_impl(self, outputs, output_grad):
"""Implements the backward pass for the transformer layer node."""
detached_grad = tuple([e.grad for e in self.detached])
grads = output_grad + detached_grad
self.default_backward_func(outputs + self.before_detached, grads)
self.before_detached = None
self.detached = None
# return grads for record stream
return grads
def dw(self):
"""Computes the weight gradients for the transformer layer node."""
with torch.cuda.nvtx.range(f"{self.name} wgrad"):
self.callables.dw()
class TransformerLayerState:
"""State shared within a transformer layer.
This class holds state that is shared between different nodes
within a transformer layer.
"""
pass
class ModelChunkSate:
"""State shared across a model chunk.
This class holds state that is shared between different components
of a model chunk, such as input tensors, parameters, and configuration.
"""
pass
class TransformerLayerSchedulePlan:
"""Schedule plan for a transformer layer.
This class organizes the computation nodes for a transformer layer,
including attention, MLP, dispatch, and combine nodes.
"""
def __init__(self, layer, event, chunk_state, comp_stream, com_stream):
"""Initializes a transformer layer schedule plan.
Args:
layer (TransformerLayer): The transformer layer to schedule.
event (torch.cuda.Event): CUDA event for synchronization.
chunk_state (ModelChunkState): State shared across the model chunk.
comp_stream (torch.cuda.Stream): CUDA stream for computation.
com_stream (torch.cuda.Stream): CUDA stream for communication.
"""
self.common_state = TransformerLayerState()
# get callables for transformer layer
attn_callable, dispatch_callable, mlp_callable, combine_callable = (
layer.get_submodule_callables(chunk_state).as_array()
)
# Create nodes for different operations in the layer
# Each node type has a predefined name that determines its memory strategy
self.attn = TransformerLayerNode(
comp_stream, event, self.common_state, attn_callable, name="attn"
)
self.mlp = TransformerLayerNode(
comp_stream, event, self.common_state, mlp_callable, name="mlp"
)
if attn_callable.is_moe:
self.dispatch = TransformerLayerNode(
com_stream, event, self.common_state, dispatch_callable, name="dispatch"
)
self.combine = TransformerLayerNode(
com_stream, event, self.common_state, combine_callable, name="combine"
)
else:
self.dispatch = FakeScheduleNode()
self.combine = FakeScheduleNode()
class ModelChunkSchedulePlan(AbstractSchedulePlan):
"""Schedule plan for a model chunk.
This class organizes the computation nodes for a model chunk,
including preprocessing, transformer layers, and postprocessing.
"""
def __init__(self):
"""Initializes a model chunk schedule plan."""
super().__init__()
self._pre_process = None
self._post_process = None
self._model_chunk_state = ModelChunkSate()
self._transformer_layers = []
self._event = torch.cuda.Event()
@classmethod
def forward_backward(
cls,
f_schedule_plan,
b_schedule_plan,
grad=None,
f_context=None,
b_context=None,
pre_forward=None,
pre_backward=None,
post_forward=None,
post_backward=None,
):
"""Schedules forward and backward passes for model chunks.
Args:
f_schedule_plan (ModelChunkSchedulePlan): Forward schedule plan.
b_schedule_plan (ModelChunkSchedulePlan): Backward schedule plan.
grad (Tensor): Gradient for backward computation.
f_context (VppContextManager or None): The VppContextManager for the forward pass.
b_context (VppContextManager or None): The VppContextManager for the backward pass
pre_forward (Callable): Callback for preprocessing in forward pass.
pre_backward (Callable): Callback for preprocessing in backward pass.
post_forward (Callable): Callback for postprocessing in forward pass.
post_backward (Callable): Callback for postprocessing in backward pass.
Returns:
The output of the forward pass.
"""
return schedule_chunk_1f1b(
f_schedule_plan,
b_schedule_plan,
grad=grad,
f_context=f_context,
b_context=b_context,
pre_forward=pre_forward,
pre_backward=pre_backward,
post_forward=post_forward,
post_backward=post_backward,
)
@property
def event(self):
"""Gets the CUDA event for synchronization."""
return self._event
def record_current_stream(self):
"""Records the current CUDA stream in the event."""
stream = torch.cuda.current_stream()
self.event.record(stream)
def wait_current_stream(self):
"""Waits for the event to complete on the current CUDA stream."""
stream = torch.cuda.current_stream()
self.event.wait(stream)
@property
def pre_process(self):
"""Gets the preprocessing node."""
return self._pre_process
@pre_process.setter
def pre_process(self, value):
"""Sets the preprocessing node."""
self._pre_process = value
@property
def post_process(self):
"""Gets the postprocessing node."""
return self._post_process
@post_process.setter
def post_process(self, value):
"""Sets the postprocessing node."""
self._post_process = value
def get_layer(self, i):
"""Gets the transformer layer at the specified index."""
assert i < self.num_layers()
return self._transformer_layers[i]
def num_layers(self):
"""Gets the number of transformer layers."""
return len(self._transformer_layers)
def add_layer(self, layer):
"""Adds a transformer layer to the schedule plan."""
self._transformer_layers.append(layer)
@property
def state(self):
"""Gets the model chunk state."""
return self._model_chunk_state
def schedule_layer_1f1b(
f_layer,
b_layer,
f_input=None,
b_grad=None,
pre_forward=None,
pre_backward=None,
pre_backward_dw=None,
f_context=None,
b_context=None,
):
"""Schedule one-forward-one-backward operations for a single layer.
This function interleaves forward and backward operations to maximize
parallelism and efficiency.
Args:
f_layer (TransformerLayerSchedulePlan): Forward layer (for current microbatch)
b_layer (TransformerLayerSchedulePlan): Backward layer (for previous microbatch)
f_input (Tensor): Input for forward computation
b_grad (Tensor): Gradient for backward computation
pre_forward (Callable): Callback to get forward input if not provided
pre_backward (Callable): Callback to get backward gradient if not provided
pre_backward_dw (Callable): Callback for weight gradient computation
f_context (VppContextManager or None): The VppContextManager for the forward pass.
b_context (VppContextManager or None): The VppContextManager for the backward pass
Returns:
Functions or values for next iteration's computation
"""
f_context = f_context if f_context is not None else contextlib.nullcontext()
b_context = b_context if b_context is not None else contextlib.nullcontext()
if pre_forward is not None:
assert f_input is None
# combine from last iter
f_input = pre_forward()
del pre_forward
if pre_backward is not None:
# attn backward from last iter
assert b_grad is None
b_grad = pre_backward()
del pre_backward
if b_layer is not None:
with b_context:
b_grad = b_layer.combine.backward(b_grad)
if pre_backward_dw is not None:
pre_backward_dw()
del pre_backward_dw
if f_layer is not None:
with f_context:
f_input = f_layer.attn.forward(f_input)
if f_layer is not None:
with f_context:
f_input = f_layer.dispatch.forward(f_input)
if b_layer is not None:
with b_context:
b_grad = b_layer.mlp.backward(b_grad)
b_grad = b_layer.dispatch.backward(b_grad)
b_layer.mlp.dw()
if f_layer is not None:
with f_context:
f_input = f_layer.mlp.forward(f_input)
def next_iter_pre_forward():
if f_layer is not None:
with f_context:
output = f_layer.combine.forward(f_input)
return output
def next_iter_pre_backward():
if b_layer is not None:
with b_context:
grad = b_layer.attn.backward(b_grad)
return grad
def next_iter_pre_backward_dw():
if b_layer is not None:
with b_context:
b_layer.attn.dw()
if f_layer and b_layer:
return next_iter_pre_forward, next_iter_pre_backward, next_iter_pre_backward_dw
else:
return next_iter_pre_forward(), next_iter_pre_backward(), next_iter_pre_backward_dw()
def schedule_chunk_1f1b(
f_schedule_plan,
b_schedule_plan,
grad=None,
f_context=None,
b_context=None,
pre_forward=None,
pre_backward=None,
post_forward=None,
post_backward=None,
):
"""Schedules one-forward-one-backward operations for a model chunk.
This function interleaves forward and backward operations across multiple layers
to maximize parallelism and efficiency.
Args:
f_schedule_plan: Forward schedule plan.
b_schedule_plan: Backward schedule plan.
grad: Gradient for backward computation.
f_context: Context for forward computation.
b_context: Context for backward computation.
pre_forward: Callback for preprocessing in forward pass.
pre_backward: Callback for preprocessing in backward pass.
post_forward: Callback for postprocessing in forward pass.
post_backward: Callback for postprocessing in backward pass.
Returns:
The output of the forward pass.
"""
f_context = f_context if f_context is not None else contextlib.nullcontext()
b_context = b_context if b_context is not None else contextlib.nullcontext()
if f_schedule_plan:
# pp output send/receive sync
if pre_forward is not None:
with f_context: # virtual pipeline parallel context
pre_forward()
f_schedule_plan.record_current_stream()
if b_schedule_plan:
b_schedule_plan.record_current_stream()
f_input = None
def layer_pre_forward():
tmp = f_input
if f_schedule_plan is not None:
tmp = f_schedule_plan.pre_process.forward()
return tmp
def layer_pre_backward():
tmp = grad
if b_schedule_plan is not None:
assert grad is not None
if b_schedule_plan.post_process is not None:
with b_context: # virtual pipeline parallel context
tmp = b_schedule_plan.post_process.backward(grad)
if pre_backward is not None:
# pp grad send receive sync here, safe for now, maybe not safe in the future
with torch.cuda.stream(get_com_stream()):
b_schedule_plan.wait_current_stream()
with b_context: # virtual pipeline parallel context
pre_backward()
b_schedule_plan.record_current_stream()
return tmp
def layer_pre_backward_dw():
pass
f_num_layers = f_schedule_plan.num_layers() if f_schedule_plan is not None else 0
b_num_layers = b_schedule_plan.num_layers() if b_schedule_plan is not None else 0
overlaped_layers = min(f_num_layers, b_num_layers)
for i in range(overlaped_layers):
f_layer = f_schedule_plan.get_layer(i)
b_layer = b_schedule_plan.get_layer(b_num_layers - 1 - i)
torch.cuda.nvtx.range_push(f"layer_{i}f-layer_{b_num_layers - 1 - i}b")
layer_pre_forward, layer_pre_backward, layer_pre_backward_dw = schedule_layer_1f1b(
f_layer,
b_layer,
pre_forward=layer_pre_forward,
pre_backward=layer_pre_backward,
pre_backward_dw=layer_pre_backward_dw,
f_context=f_context,
b_context=b_context,
)
torch.cuda.nvtx.range_pop()
# tail forward
f_input = layer_pre_forward()
del layer_pre_forward
# tail backward
grad = layer_pre_backward()
del layer_pre_backward
with b_context:
for i in range(overlaped_layers, b_num_layers):
b_layer = b_schedule_plan.get_layer(b_num_layers - 1 - i)
torch.cuda.nvtx.range_push(f"layer_{b_num_layers - 1 - i}b")
tmp, grad, _ = schedule_layer_1f1b(None, b_layer, b_grad=grad)
torch.cuda.nvtx.range_pop()
# if b_schedule_plan is not None:
# b_schedule_plan.pre_process.backward(grad)
# # tail forward
# f_input = layer_pre_forward()
# del layer_pre_forward
with f_context:
for i in range(overlaped_layers, f_num_layers):
f_layer = f_schedule_plan.get_layer(i)
torch.cuda.nvtx.range_push(f"layer_{i}f")
f_input, tmp, _ = schedule_layer_1f1b(f_layer, None, f_input=f_input)
torch.cuda.nvtx.range_pop()
# if f_schedule_plan is not None and f_schedule_plan.post_process is not None:
# f_input = f_schedule_plan.post_process.forward(f_input)
# output pp send receive, overlapped with attn backward
if f_schedule_plan is not None and post_forward is not None:
with f_context:
f_schedule_plan.wait_current_stream()
post_forward(f_input)
# pp grad send / receive, overlapped with attn dw of cur micro-batch
# and forward attn of next micro-batch
if b_schedule_plan is not None and post_backward is not None:
with b_context:
b_schedule_plan.wait_current_stream()
post_backward(grad)
# The last wgrad of attention
layer_pre_backward_dw()
del layer_pre_backward_dw
with f_context:
if f_schedule_plan is not None and f_schedule_plan.post_process is not None:
f_input = f_schedule_plan.post_process.forward(f_input)
with b_context:
if b_schedule_plan is not None:
b_schedule_plan.pre_process.backward(grad)
if f_schedule_plan:
f_schedule_plan.wait_current_stream()
if b_schedule_plan:
b_schedule_plan.wait_current_stream()
return f_input
def build_model_chunk_schedule_plan(
model,
input_ids: Tensor,
position_ids: Tensor,
attention_mask: Tensor,
decoder_input: Tensor = None,
labels: Tensor = None,
inference_params=None,
packed_seq_params=None,
extra_block_kwargs=None,
runtime_gather_output: Optional[bool] = None,
):
"""Builds a schedule plan for a model chunk.
This function creates a schedule plan for a model chunk, including
preprocessing, transformer layers, and postprocessing.
Args:
model: The model to build a schedule plan for.
input_ids: Input token IDs.
position_ids: Position IDs.
attention_mask: Attention mask.
decoder_input: Decoder input tensor.
labels: Labels for loss computation.
inference_params: Parameters for inference.
packed_seq_params: Parameters for packed sequences.
extra_block_kwargs: Additional keyword arguments for blocks.
runtime_gather_output: Whether to gather output at runtime.
Returns:
The model chunk schedule plan.
"""
comp_stream = get_comp_stream()
com_stream = get_com_stream()
model_chunk_schedule_plan = ModelChunkSchedulePlan()
event = model_chunk_schedule_plan.event
state = model_chunk_schedule_plan.state
# save for later use
state.input_ids = input_ids
state.position_ids = position_ids
state.attention_mask = attention_mask
state.decoder_input = decoder_input
state.labels = labels
state.inference_params = inference_params
state.packed_seq_params = packed_seq_params
state.extra_block_kwargs = extra_block_kwargs
state.runtime_gather_output = runtime_gather_output
state.context = None
state.context_mask = None
state.attention_bias = None
# build preprocess
model_chunk_schedule_plan.pre_process = PreProcessNode(model, state, event, comp_stream)
# build for layers
for layer_idx in range(model.decoder.num_layers_per_pipeline_rank):
layer = model.decoder._get_layer(layer_idx)
layer_plan = TransformerLayerSchedulePlan(layer, event, state, comp_stream, com_stream)
model_chunk_schedule_plan.add_layer(layer_plan)
# build post process
if model.post_process:
model_chunk_schedule_plan.post_process = PostProcessNode(model, state, event, comp_stream)
return model_chunk_schedule_plan
......@@ -10,6 +10,7 @@ from torch import Tensor
from megatron.core import InferenceParams, tensor_parallel
from megatron.core.config_logger import has_config_logger_enabled, log_config_to_disk
from megatron.core.packed_seq_params import PackedSeqParams
from megatron.core.models.gpt import GPTModel as MegatronCoreGPTModel
from dcu_megatron.core.tensor_parallel import FluxColumnParallelLinear
......@@ -45,100 +46,143 @@ def gpt_model_init_wrapper(fn):
return wrapper
def gpt_model_forward(
self,
input_ids: Tensor,
position_ids: Tensor,
attention_mask: Tensor,
decoder_input: Tensor = None,
labels: Tensor = None,
inference_params: InferenceParams = None,
packed_seq_params: PackedSeqParams = None,
extra_block_kwargs: dict = None,
runtime_gather_output: Optional[bool] = None,
loss_mask: Optional[Tensor] = None,
) -> Tensor:
"""Forward function of the GPT Model This function passes the input tensors
through the embedding layer, and then the decoder and finally into the post
processing layer (optional).
It either returns the Loss values if labels are given or the final hidden units
Args:
runtime_gather_output (bool): Gather output at runtime. Default None means
`parallel_output` arg in the constructor will be used.
class GPTModel(MegatronCoreGPTModel):
"""
# If decoder_input is provided (not None), then input_ids and position_ids are ignored.
# Otherwise, apply embedding layer on input_ids and position_ids to get decoder_input.
# Decoder embedding.
if decoder_input is not None:
pass
elif self.pre_process:
decoder_input = self.embedding(input_ids=input_ids, position_ids=position_ids)
else:
# intermediate stage of pipeline
# decoder will get hidden_states from encoder.input_tensor
decoder_input = None
# Rotary positional embeddings (embedding is None for PP intermediate devices)
rotary_pos_emb = None
rotary_pos_cos = None
rotary_pos_sin = None
if self.position_embedding_type == 'rope' and not self.config.multi_latent_attention:
if not self.training and self.config.flash_decode and inference_params:
# Flash decoding uses precomputed cos and sin for RoPE
rotary_pos_cos, rotary_pos_sin = self.rotary_pos_emb_cache.setdefault(
inference_params.max_sequence_length,
self.rotary_pos_emb.get_cos_sin(inference_params.max_sequence_length),
)
else:
rotary_seq_len = self.rotary_pos_emb.get_rotary_seq_len(
inference_params, self.decoder, decoder_input, self.config, packed_seq_params
)
rotary_pos_emb = self.rotary_pos_emb(
rotary_seq_len,
packed_seq=packed_seq_params is not None
and packed_seq_params.qkv_format == 'thd',
)
if (
(self.config.enable_cuda_graph or self.config.flash_decode)
and rotary_pos_cos is not None
and inference_params
patch megatron GPTModel
"""
def get_transformer_callables_by_layer(self, layer_number: int):
"""
Get the callables for the layer at the given transformer layer number.
"""
return self.decoder.get_layer_callables(layer_number)
def build_schedule_plan(
self,
input_ids: Tensor,
position_ids: Tensor,
attention_mask: Tensor,
decoder_input: Tensor = None,
labels: Tensor = None,
inference_params: InferenceParams = None,
packed_seq_params: PackedSeqParams = None,
extra_block_kwargs: dict = None,
runtime_gather_output: Optional[bool] = None,
loss_mask: Optional[Tensor] = None,
):
sequence_len_offset = torch.tensor(
[inference_params.sequence_len_offset] * inference_params.current_batch_size,
dtype=torch.int32,
device=rotary_pos_cos.device, # Co-locate this with the rotary tensors
)
else:
sequence_len_offset = None
# Run decoder.
hidden_states = self.decoder(
hidden_states=decoder_input,
attention_mask=attention_mask,
inference_params=inference_params,
rotary_pos_emb=rotary_pos_emb,
rotary_pos_cos=rotary_pos_cos,
rotary_pos_sin=rotary_pos_sin,
packed_seq_params=packed_seq_params,
sequence_len_offset=sequence_len_offset,
**(extra_block_kwargs or {}),
)
# logits and loss
output_weight = None
if self.share_embeddings_and_output_weights:
output_weight = self.shared_embedding_or_output_weight()
if self.mtp_process:
hidden_states = self.mtp(
input_ids=input_ids,
position_ids=position_ids,
"""Builds a computation schedule plan for the model.
This function creates a schedule plan for a model chunk, including
preprocessing, transformer layers, and postprocessing.
The schedule plan is used to optimize computation and memory usage
in distributed environments.
Args:
input_ids (Tensor): Input token IDs.
position_ids (Tensor): Position IDs.
attention_mask (Tensor): Attention mask.
decoder_input (Tensor, optional): Decoder input tensor. Defaults to None.
labels (Tensor, optional): Labels for loss computation. Defaults to None.
inference_params (InferenceParams, optional):
Parameters for inference. Defaults to None.
packed_seq_params (PackedSeqParams, optional):
Parameters for packed sequences. Defaults to None.
extra_block_kwargs (dict, optional):
Additional keyword arguments for blocks. Defaults to None.
runtime_gather_output (Optional[bool], optional):
Whether to gather output at runtime. Defaults to None.
loss_mask (Optional[Tensor], optional): Loss mask. Defaults to None.
Returns:
ModelChunkSchedulePlan: The model chunk schedule plan.
"""
from .fine_grained_schedule import build_model_chunk_schedule_plan
return build_model_chunk_schedule_plan(
self,
input_ids,
position_ids,
attention_mask,
decoder_input=decoder_input,
labels=labels,
loss_mask=loss_mask,
hidden_states=hidden_states,
inference_params=inference_params,
packed_seq_params=packed_seq_params,
extra_block_kwargs=extra_block_kwargs,
runtime_gather_output=runtime_gather_output,
)
def forward(
self,
input_ids: Tensor,
position_ids: Tensor,
attention_mask: Tensor,
decoder_input: Tensor = None,
labels: Tensor = None,
inference_params: InferenceParams = None,
packed_seq_params: PackedSeqParams = None,
extra_block_kwargs: dict = None,
runtime_gather_output: Optional[bool] = None,
loss_mask: Optional[Tensor] = None,
) -> Tensor:
"""Forward function of the GPT Model This function passes the input tensors
through the embedding layer, and then the decoder and finally into the post
processing layer (optional).
It either returns the Loss values if labels are given or the final hidden units
Args:
runtime_gather_output (bool): Gather output at runtime. Default None means
`parallel_output` arg in the constructor will be used.
"""
# If decoder_input is provided (not None), then input_ids and position_ids are ignored.
# Otherwise, apply embedding layer on input_ids and position_ids to get decoder_input.
# Decoder embedding.
if decoder_input is not None:
pass
elif self.pre_process:
decoder_input = self.embedding(input_ids=input_ids, position_ids=position_ids)
else:
# intermediate stage of pipeline
# decoder will get hidden_states from encoder.input_tensor
decoder_input = None
# Rotary positional embeddings (embedding is None for PP intermediate devices)
rotary_pos_emb = None
rotary_pos_cos = None
rotary_pos_sin = None
if self.position_embedding_type == 'rope' and not self.config.multi_latent_attention:
if not self.training and self.config.flash_decode and inference_params:
# Flash decoding uses precomputed cos and sin for RoPE
rotary_pos_cos, rotary_pos_sin = self.rotary_pos_emb_cache.setdefault(
inference_params.max_sequence_length,
self.rotary_pos_emb.get_cos_sin(inference_params.max_sequence_length),
)
else:
rotary_seq_len = self.rotary_pos_emb.get_rotary_seq_len(
inference_params, self.decoder, decoder_input, self.config, packed_seq_params
)
rotary_pos_emb = self.rotary_pos_emb(
rotary_seq_len,
packed_seq=packed_seq_params is not None
and packed_seq_params.qkv_format == 'thd',
)
if (
(self.config.enable_cuda_graph or self.config.flash_decode)
and rotary_pos_cos is not None
and inference_params
):
sequence_len_offset = torch.tensor(
[inference_params.sequence_len_offset] * inference_params.current_batch_size,
dtype=torch.int32,
device=rotary_pos_cos.device, # Co-locate this with the rotary tensors
)
else:
sequence_len_offset = None
# Run decoder.
hidden_states = self.decoder(
hidden_states=decoder_input,
attention_mask=attention_mask,
inference_params=inference_params,
rotary_pos_emb=rotary_pos_emb,
......@@ -146,44 +190,66 @@ def gpt_model_forward(
rotary_pos_sin=rotary_pos_sin,
packed_seq_params=packed_seq_params,
sequence_len_offset=sequence_len_offset,
embedding=self.embedding,
output_layer=self.output_layer,
output_weight=output_weight,
runtime_gather_output=runtime_gather_output,
compute_language_model_loss=self.compute_language_model_loss,
**(extra_block_kwargs or {}),
)
if (
self.mtp_process is not None
and getattr(self.decoder, "main_final_layernorm", None) is not None
):
# move block main model final norms here
hidden_states = self.decoder.main_final_layernorm(hidden_states)
if not self.post_process:
return hidden_states
logits, _ = self.output_layer(
hidden_states, weight=output_weight, runtime_gather_output=runtime_gather_output
)
if has_config_logger_enabled(self.config):
payload = OrderedDict(
{
'input_ids': input_ids,
'position_ids': position_ids,
'attention_mask': attention_mask,
'decoder_input': decoder_input,
'logits': logits,
}
# logits and loss
output_weight = None
if self.share_embeddings_and_output_weights:
output_weight = self.shared_embedding_or_output_weight()
if self.mtp_process:
hidden_states = self.mtp(
input_ids=input_ids,
position_ids=position_ids,
labels=labels,
loss_mask=loss_mask,
hidden_states=hidden_states,
attention_mask=attention_mask,
inference_params=inference_params,
rotary_pos_emb=rotary_pos_emb,
rotary_pos_cos=rotary_pos_cos,
rotary_pos_sin=rotary_pos_sin,
packed_seq_params=packed_seq_params,
sequence_len_offset=sequence_len_offset,
embedding=self.embedding,
output_layer=self.output_layer,
output_weight=output_weight,
runtime_gather_output=runtime_gather_output,
compute_language_model_loss=self.compute_language_model_loss,
**(extra_block_kwargs or {}),
)
if (
self.mtp_process is not None
and getattr(self.decoder, "main_final_layernorm", None) is not None
):
# move block main model final norms here
hidden_states = self.decoder.main_final_layernorm(hidden_states)
if not self.post_process:
return hidden_states
logits, _ = self.output_layer(
hidden_states, weight=output_weight, runtime_gather_output=runtime_gather_output
)
log_config_to_disk(self.config, payload, prefix='input_and_logits')
if labels is None:
# [s b h] => [b s h]
return logits.transpose(0, 1).contiguous()
if has_config_logger_enabled(self.config):
payload = OrderedDict(
{
'input_ids': input_ids,
'position_ids': position_ids,
'attention_mask': attention_mask,
'decoder_input': decoder_input,
'logits': logits,
}
)
log_config_to_disk(self.config, payload, prefix='input_and_logits')
if labels is None:
# [s b h] => [b s h]
return logits.transpose(0, 1).contiguous()
loss = self.compute_language_model_loss(labels, logits)
loss = self.compute_language_model_loss(labels, logits)
return loss
return loss
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
import contextlib
from abc import ABC, abstractmethod
from contextlib import contextmanager
from typing import List, Union
import torch
from torch import Tensor
from torch.autograd.variable import Variable
from megatron.core import parallel_state
from megatron.core.distributed import DistributedDataParallel
from megatron.core.models.gpt.gpt_model import GPTModel
from megatron.core.transformer.module import Float16Module
from megatron.core.transformer.moe.router import MoEAuxLossAutoScaler
from megatron.core.utils import get_attr_wrapped_model, make_viewless_tensor
# Types
Shape = Union[List[int], torch.Size]
def make_viewless(e):
"""Make_viewless util func"""
e = make_viewless_tensor(inp=e, requires_grad=e.requires_grad, keep_graph=True)
return e
@contextmanager
def stream_acquire_context(stream, event):
"""Stream acquire context"""
event.wait(stream)
try:
yield
finally:
event.record(stream)
class FakeScheduleNode:
"""A placeholder node in the computation graph that simply passes through inputs and outputs.
This class is used as a no-op node in the scheduling system when a real computation node
is not needed but the interface must be maintained. It simply returns its inputs unchanged
in both forward and backward passes.
"""
def forward(self, inputs):
"""Passes through inputs unchanged in the forward pass."""
return inputs
def backward(self, outgrads):
"""Passes through gradients unchanged in the backward pass."""
return outgrads
class ScheduleNode:
"""Base node for fine-grained scheduling.
This class represents a computational node in the pipeline schedule.
It handles the execution of forward and backward operations on a stream.
"""
def __init__(
self,
forward_func,
stream,
event,
backward_func=None,
memory_strategy=None,
name="schedule_node",
):
"""Initialize a schedule node.
Args:
forward_func (callable): Function to execute during forward pass
stream (torch.cuda.Stream): CUDA stream for computation
event (torch.cuda.Event): Event for synchronization
backward_func (callable, optional): Function for backward pass
memory_strategy (MemoryManagementStrategy, optional): Strategy for memory management
name (str): Name of the node for debugging
"""
self.name = name
self.forward_func = forward_func
self.backward_func = backward_func if backward_func else self.default_backward_func
self.stream = stream
self.event = event
self.memory_strategy = memory_strategy or NoOpMemoryStrategy()
self.inputs = None
self.outputs = None
def default_backward_func(self, outputs, output_grad):
"""Default backward function"""
Variable._execution_engine.run_backward(
tensors=outputs,
grad_tensors=output_grad,
keep_graph=False,
create_graph=False,
inputs=tuple(),
allow_unreachable=True,
accumulate_grad=True,
)
return output_grad
def forward(self, inputs=()):
"""Schedule node forward"""
if not isinstance(inputs, tuple):
inputs = (inputs,)
return self._forward(*inputs)
def _forward(self, *inputs):
with stream_acquire_context(self.stream, self.event):
torch.cuda.nvtx.range_push(f"{self.name} forward")
with torch.cuda.stream(self.stream):
self.inputs = [make_viewless(e).detach() if e is not None else None for e in inputs]
for i, input in enumerate(self.inputs):
if input is not None:
input.requires_grad = inputs[i].requires_grad
data = tuple(self.inputs)
data = self.forward_func(*data)
if not isinstance(data, tuple):
data = make_viewless(data)
else:
data = tuple([make_viewless(e) if isinstance(e, Tensor) else e for e in data])
self.output = data
torch.cuda.nvtx.range_pop()
# Handle inputs using the memory strategy
self.memory_strategy.handle_inputs(inputs, self.stream)
return self.output
def get_output(self):
"""Get the forward output"""
return self.output
def backward(self, output_grad):
"""Schedule node backward"""
if not isinstance(output_grad, tuple):
output_grad = (output_grad,)
return self._backward(*output_grad)
def _backward(self, *output_grad):
with stream_acquire_context(self.stream, self.event):
torch.cuda.nvtx.range_push(f"{self.name} backward")
with torch.cuda.stream(self.stream):
outputs = self.output
if not isinstance(outputs, tuple):
outputs = (outputs,)
assert len(outputs) == len(output_grad), (
f"{len(outputs)} of {type(outputs[0])} is not equal to "
f"{len(output_grad)} of {type(output_grad[0])}"
)
output_grad = self.backward_func(outputs, output_grad)
torch.cuda.nvtx.range_pop()
# output_grad maybe from another stream
for g in output_grad:
g.record_stream(self.stream)
return self.get_grad()
def get_grad(self):
"""Get the grad of inputs"""
grad = tuple([e.grad if e is not None else None for e in self.inputs])
# clear state
self.inputs = None
self.output = None
# multiple in, multiple out
if len(grad) == 1:
grad = grad[0]
return grad
class AbstractSchedulePlan(ABC):
"""To use combined 1f1b, model must implement build_schedule_plan while take the same
signature as model forward but return an instance of AbstractSchedulePlan"""
@classmethod
@abstractmethod
def forward_backward(
cls,
f_schedule_plan,
b_schedule_plan,
grad=None,
f_context=None,
b_context=None,
pre_forward=None,
pre_backward=None,
post_forward=None,
post_backward=None,
):
"""forward_backward is the protocol between our schedule logic and model"""
...
def schedule_chunk_1f1b(
f_schedule_plan,
b_schedule_plan,
grad=None,
f_context=None,
b_context=None,
pre_forward=None,
pre_backward=None,
post_forward=None,
post_backward=None,
):
"""Model level 1f1b fine-grained schedule
This function schedules the forward and backward passes for a chunk of the model.
It takes in the forward schedule plan, backward schedule plan, gradient, and optional
context managers for the forward and backward passes.
Args:
f_schedule_plan (subclass of AbstractSchedulePlan): The forward schedule plan
b_schedule_plan (subclass of AbstractSchedulePlan): The backward schedule plan
grad (Tensor or None): The gradient of the loss function
f_context (VppContextManager or None): The VppContextManager for the forward pass
b_context (VppContextManager or None): The VppContextManager for the backward pass
pre_forward (callable or None): The function to call before the forward pass
pre_backward (callable or None): The function to call before the backward pass
post_forward (callable or None): The function to call after the forward pass
post_backward (callable or None): The function to call after the backward pass
Returns:
The output of the forward pass
"""
# Calls fine_grained_schedule.py::ModelChunkSchedulePlan.forward_backward(),
# which calls fine_grained_schedule.py::schedule_chunk_1f1b()
return type(f_schedule_plan or b_schedule_plan).forward_backward(
f_schedule_plan,
b_schedule_plan,
grad=grad,
f_context=f_context,
b_context=b_context,
pre_forward=pre_forward,
pre_backward=pre_backward,
post_forward=post_forward,
post_backward=post_backward,
)
_COMP_STREAM = None
_COM_STREAM = None
def set_streams(comp_stream=None, com_stream=None):
"""Set the streams for communication and computation"""
global _COMP_STREAM
global _COM_STREAM
if _COMP_STREAM is not None:
return
if comp_stream is None:
comp_stream = torch.cuda.current_stream()
if com_stream is None:
com_stream = torch.cuda.Stream(device="cuda")
assert _COMP_STREAM is None
assert _COM_STREAM is None
_COMP_STREAM = comp_stream
_COM_STREAM = com_stream
def get_comp_stream():
"""Get the stream for computation"""
global _COMP_STREAM
return _COMP_STREAM
def get_com_stream():
"""Get the stream for communication"""
global _COM_STREAM
return _COM_STREAM
class VppContextManager:
"""A reusable context manager for switch vpp stage"""
def __init__(self, vpp_rank):
self.vpp_rank = vpp_rank
def __enter__(self):
self.origin_vpp_rank = parallel_state.get_virtual_pipeline_model_parallel_rank()
parallel_state.set_virtual_pipeline_model_parallel_rank(self.vpp_rank)
return self
def __exit__(self, exc_type, exc_val, exc_tb):
parallel_state.set_virtual_pipeline_model_parallel_rank(self.origin_vpp_rank)
def forward_backward_step(
forward_step_func,
data_iterator,
f_model,
num_microbatches,
input_tensor,
forward_data_store,
b_model,
b_input_tensor,
b_output_tensor,
b_output_tensor_grad,
config,
f_context=None,
b_context=None,
pre_forward=None,
pre_backward=None,
post_forward=None,
post_backward=None,
collect_non_loss_data=False,
checkpoint_activations_microbatch=None,
is_first_microbatch=False,
current_microbatch=None,
encoder_decoder_xattn=False,
):
"""Merged forward and backward step for combined_1f1b.
Args:
Need to accept the argument of both forward_step() and backward_step().
forward_step_func (callable): is wrapped by wrap_forward_func() which is now returning
a forward schedule plan which is an input of schedule_chunk_1f1b function.
f_context (VppContextManager or nullcontext): The context manager for setting vpp ranks.
b_context (VppContextManager or nullcontext): The context manager for setting vpp ranks.
Only exists in 1f1b steady state with p2p overlap.
pre_forward (callable): The function to call before the forward_step.
pre_backward (callable): The function to call before the backward_step.
post_forward (callable): The function to call after the forward_step.
post_backward (callable): The function to call after the backward_step.
Returns:
forward_output_tensor (Tensor or list[Tensor]): The output object(s) from the forward step.
forward_num_tokens (Tensor): The number of tokens.
backward_input_tensor_grad (Tensor): The grad of the input tensor.
Descriptions:
This method merges the forward_step() and backward_step() methods in the schedules.py file.
Assuming that:
def forward_step():
# forward_preprocess()
# forward_compute()
# forward_postprocess()
def backward_step():
# backward_preprocess()
# backward_compute()
# backward_postprocess()
Then the forward_backward_step() method will be:
def forward_backward_step():
# forward_preprocess() // the same as the forward_step()
# GENERATE f_schedule_plan // schedule happens in schedule_chunk_1f1b()
# backward_preprocess() // the same as the backward_step()
# COMBINED_FORWARD_BACKWARD_COMPUTE() // by calling schedule_chunk_1f1b()
# forward_postprocess() // the same as the forward_step()
# backward_postprocess() // the same as the backward_step()
"""
assert (
checkpoint_activations_microbatch is None
), "checkpoint_activations_microbatch is not supported for combined_1f1b"
if config.combined_1f1b_recipe != "ep_a2a":
raise NotImplementedError(
f"combined_1f1b_recipe {config.combined_1f1b_recipe} not supported yet"
)
from .schedules import set_current_microbatch
if f_model is not None and config.timers is not None:
config.timers('forward-compute', log_level=2).start()
if config.enable_autocast:
context_manager = torch.autocast("cuda", dtype=config.autocast_dtype)
else:
context_manager = contextlib.nullcontext()
# forward preprocess
unwrap_output_tensor = False
f_schedule_plan = None
if f_model is not None:
with f_context:
if is_first_microbatch and hasattr(f_model, 'set_is_first_microbatch'):
f_model.set_is_first_microbatch()
if current_microbatch is not None:
set_current_microbatch(f_model, current_microbatch)
if not isinstance(input_tensor, list):
input_tensor = [input_tensor]
unwrap_output_tensor = True
set_input_tensor = get_attr_wrapped_model(f_model, "set_input_tensor")
set_input_tensor(input_tensor)
with context_manager: # autocast context
f_schedule_plan, loss_func = forward_step_func(data_iterator, f_model)
assert isinstance(
f_schedule_plan, AbstractSchedulePlan
), "first output of forward_step_func must be one instance of AbstractSchedulePlan"
# backward preprocess
unwrap_input_tensor_grad = False
b_schedule_plan = None
if b_model is not None:
# Retain the grad on the input_tensor.
if not isinstance(b_input_tensor, list):
b_input_tensor = [b_input_tensor]
unwrap_input_tensor_grad = True
for x in b_input_tensor:
if x is not None:
x.retain_grad()
if not isinstance(b_output_tensor, list):
b_output_tensor = [b_output_tensor]
if not isinstance(b_output_tensor_grad, list):
b_output_tensor_grad = [b_output_tensor_grad]
# Backward pass for loss function
b_schedule_plan = b_output_tensor[0].schedule_plan
b_output_tensor[0].schedule_plan = None
if b_output_tensor_grad[0] is None and config.grad_scale_func is not None:
# backward schedule plan
loss_node = b_output_tensor[0].loss_func
b_output_tensor[0].loss_func = None
b_output_tensor[0] = config.grad_scale_func(b_output_tensor[0])
torch.autograd.backward(b_output_tensor[0], grad_tensors=b_output_tensor_grad[0])
b_output_tensor_grad[0] = loss_node.get_grad()
grad = b_output_tensor_grad[0] if b_model else None
with context_manager: # autocast context
# schedule forward and backward
output_tensor = schedule_chunk_1f1b(
f_schedule_plan,
b_schedule_plan,
grad,
f_context=f_context,
b_context=b_context,
pre_forward=pre_forward,
pre_backward=pre_backward,
post_forward=post_forward,
post_backward=post_backward,
)
# forward post process
num_tokens = None
if f_model is not None:
with f_context:
num_tokens = torch.tensor(0, dtype=torch.int)
if parallel_state.is_pipeline_last_stage():
if not collect_non_loss_data:
loss_node = ScheduleNode(
loss_func,
torch.cuda.current_stream(),
f_schedule_plan.event,
name="loss_func",
)
loss_func = loss_node.forward
outputs = loss_func(output_tensor)
if len(outputs) == 3:
output_tensor, num_tokens, loss_reduced = outputs
if not config.calculate_per_token_loss:
output_tensor /= num_tokens
output_tensor /= num_microbatches
else:
# preserve legacy loss averaging behavior
# (ie, over the number of microbatches)
assert len(outputs) == 2
output_tensor, loss_reduced = outputs
output_tensor = output_tensor / num_microbatches
forward_data_store.append(loss_reduced)
# attach loss_func on output_tensor
output_tensor.loss_func = loss_node
else:
data = loss_func(output_tensor, non_loss_data=True)
forward_data_store.append(data)
# attach schedule plan on output tensor
output_tensor.schedule_plan = f_schedule_plan
if config.timers is not None:
config.timers('forward-compute').stop()
# Set the loss scale for the auxiliary loss of the MoE layer.
# Since we use a trick to do backward on the auxiliary loss, we need to set the scale
# explicitly.
if hasattr(config, 'num_moe_experts') and config.num_moe_experts is not None:
# Calculate the loss scale based on the grad_scale_func if available,
# else default to 1.
loss_scale = (
config.grad_scale_func(torch.ones(1, device=output_tensor.device))
if config.grad_scale_func is not None
else torch.tensor(1.0)
)
# Set the loss scale
MoEAuxLossAutoScaler.set_loss_scale(loss_scale / num_microbatches)
if not unwrap_output_tensor:
output_tensor, num_tokens = [output_tensor], num_tokens
# backward post process
input_tensor_grad = None
if b_model is not None:
input_tensor_grad = [None]
if b_input_tensor is not None:
input_tensor_grad = []
for x in b_input_tensor:
if x is None:
input_tensor_grad.append(None)
else:
input_tensor_grad.append(x.grad)
if unwrap_input_tensor_grad:
input_tensor_grad = input_tensor_grad[0]
return output_tensor, num_tokens, input_tensor_grad
def get_default_cls_for_unwrap():
"""Returns the default classes to unwrap from a model.
This function provides a tuple of classes that should be unwrapped from a model
to access the underlying GPTModel instance. It includes DistributedDataParallel
and Float16Module by default, and also attempts to include LegacyFloat16Module
if available for backward compatibility.
Returns:
tuple: A tuple of classes to unwrap from a model.
"""
cls = (DistributedDataParallel, Float16Module)
try:
# legacy should not be used in core, but for backward compatibility, we support it here
from megatron.legacy.model import Float16Module as LegacyFloat16Module
cls = cls + (LegacyFloat16Module,)
except:
pass
return cls
def unwrap_model(model, module_instances=get_default_cls_for_unwrap()):
"""Unwrap_model DistributedDataParallel and Float16Module wrapped model
to return GPTModel instance
"""
return_list = True
if not isinstance(model, list):
model = [model]
return_list = False
unwrapped_model = []
for model_module in model:
while isinstance(model_module, module_instances):
model_module = model_module.module
assert isinstance(
model_module, GPTModel
), "The final unwrapped model must be a GPTModel instance"
unwrapped_model.append(model_module)
if not return_list:
return unwrapped_model[0]
return unwrapped_model
def wrap_forward_func(forward_step_func):
"""Wrap the input to forward_step_func.
The wrapped function will return forward_schedule_plan and the loss_function.
"""
def wrapped_func(data_iterator, model):
# Model is unwrapped to get GPTModel instance.
# GPTModel.build_schedule_plan(model_forward_inputs) is called in the forward_step.
# The return value becomes (forward_schedule_plan, loss_function),
# which is used to be (forward_output_tensor, loss_function).
return forward_step_func(data_iterator, unwrap_model(model).build_schedule_plan)
return wrapped_func
class MemoryManagementStrategy:
"""Base class for memory management strategies.
Different memory management strategies can be implemented by subclassing this class.
These strategies control how tensors are handled in memory during the computation.
"""
def handle_inputs(self, inputs, stream):
"""Process input tensors after computation.
Args:
inputs (tuple): Input tensors that have been used
stream (torch.cuda.Stream): Current CUDA stream
"""
pass
def handle_outputs(self, outputs, stream):
"""Process output tensors after computation.
Args:
outputs (tuple): Output tensors produced by the computation
stream (torch.cuda.Stream): Current CUDA stream
"""
pass
class NoOpMemoryStrategy(MemoryManagementStrategy):
"""Strategy that performs no memory management operations.
This is the default strategy - it doesn't free any memory.
"""
pass
class FreeInputsMemoryStrategy(MemoryManagementStrategy):
"""Strategy that immediately frees input tensors after they are used.
This strategy is useful for nodes where inputs are no longer needed
after computation, helping to reduce memory usage.
"""
def handle_inputs(self, inputs, stream):
"""Free input tensors by resizing their storage to zero.
Args:
inputs (tuple): Input tensors to be freed
stream (torch.cuda.Stream): Current CUDA stream
"""
for input in inputs:
if input is not None:
input.record_stream(stream)
input.untyped_storage().resize_(0)
from megatron.core.transformer.moe.token_dispatcher import _DeepepManager as MegatronCoreDeepepManager
class MoEAlltoAllTokenDispatcher(MoETokenDispatcher):
def token_permutation(
self, hidden_states: torch.Tensor, probs: torch.Tensor, routing_map: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Dispatch tokens to local experts using AlltoAll communication.
This method performs the following steps:
1. Preprocess the routing map to get metadata for communication and permutation.
2. Permute input tokens for AlltoAll communication.
3. Perform expert parallel AlltoAll communication.
4. Sort tokens by local expert (if multiple local experts exist).
Args:
hidden_states (torch.Tensor): Input token embeddings.
probs (torch.Tensor): The probabilities of token to experts assignment.
routing_map (torch.Tensor): The mapping of token to experts assignment.
Returns:
Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
- Permuted token embeddings for local experts.
- Number of tokens per expert.
- Permuted probs of each token produced by the router.
"""
# Preprocess: Get the metadata for communication, permutation and computation operations.
self.hidden_shape = hidden_states.shape
self.probs = probs
self.routing_map = routing_map
assert probs.dim() == 2, "Expected 2D tensor for probs"
assert routing_map.dim() == 2, "Expected 2D tensor for token2expert mask"
assert routing_map.dtype == torch.bool, "Expected bool tensor for mask"
hidden_states = hidden_states.view(-1, self.hidden_shape[-1])
tokens_per_expert = self.preprocess(self.routing_map)
if self.shared_experts is not None:
self.shared_experts.pre_forward_comm(hidden_states.view(self.hidden_shape))
# Permutation 1: input to AlltoAll input
tokens_per_expert = self._maybe_dtoh_and_synchronize(
"before_permutation_1", tokens_per_expert
)
self.hidden_shape_before_permute = hidden_states.shape
(
permutated_local_input_tokens,
permuted_probs,
self.reversed_local_input_permutation_mapping,
) = permute(
hidden_states,
routing_map,
probs=probs,
num_out_tokens=self.num_out_tokens,
fused=self.config.moe_permute_fusion,
drop_and_pad=self.drop_and_pad,
)
# Perform expert parallel AlltoAll communication
tokens_per_expert = self._maybe_dtoh_and_synchronize(
"before_ep_alltoall", tokens_per_expert
)
global_input_tokens = all_to_all(
self.ep_group, permutated_local_input_tokens, self.output_splits, self.input_splits
)
global_probs = all_to_all(
self.ep_group, permuted_probs, self.output_splits, self.input_splits
)
if self.shared_experts is not None:
self.shared_experts.linear_fc1_forward_and_act(global_input_tokens)
if self.tp_size > 1:
if self.output_splits_tp is None:
output_split_sizes = None
else:
output_split_sizes = self.output_splits_tp.tolist()
global_input_tokens = gather_from_sequence_parallel_region(
global_input_tokens, group=self.tp_group, output_split_sizes=output_split_sizes
)
global_probs = gather_from_sequence_parallel_region(
global_probs, group=self.tp_group, output_split_sizes=output_split_sizes
)
# Permutation 2: Sort tokens by local expert.
tokens_per_expert = self._maybe_dtoh_and_synchronize(
"before_permutation_2", tokens_per_expert
)
if self.num_local_experts > 1:
if self.drop_and_pad:
global_input_tokens = (
global_input_tokens.view(
self.tp_size * self.ep_size,
self.num_local_experts,
self.capacity,
*global_input_tokens.size()[1:],
)
.transpose(0, 1)
.contiguous()
.flatten(start_dim=0, end_dim=2)
)
global_probs = (
global_probs.view(
self.tp_size * self.ep_size,
self.num_local_experts,
self.capacity,
*global_probs.size()[1:],
)
.transpose(0, 1)
.contiguous()
.flatten(start_dim=0, end_dim=2)
)
else:
global_input_tokens, global_probs = sort_chunks_by_idxs(
global_input_tokens,
self.num_global_tokens_per_local_expert.ravel(),
self.sort_input_by_local_experts,
probs=global_probs,
fused=self.config.moe_permute_fusion,
)
tokens_per_expert = self._maybe_dtoh_and_synchronize("before_finish", tokens_per_expert)
return global_input_tokens, tokens_per_expert, global_probs
class _DeepepManager(MegatronCoreDeepepManager):
"""
patch megatron _DeepepManager. async
"""
def dispatch(
self,
hidden_states: torch.Tensor,
async_finish: bool = False,
allocate_on_comm_stream: bool = False,
) -> torch.Tensor:
# DeepEP only supports float32 probs
if self.token_probs.dtype != torch.float32:
if self.token_probs.dtype in [torch.bfloat16, torch.float16]:
print("DeepEP only supports float32 probs, please set --moe-router-dtype=fp32")
self.token_probs = self.token_probs.float() # downcast or upcast
hidden_states, dispatched_indices, dispatched_probs, num_tokens_per_expert, handle = (
fused_dispatch(
hidden_states,
self.token_indices,
self.token_probs,
self.num_experts,
self.group,
async_finish=async_finish,
allocate_on_comm_stream=allocate_on_comm_stream,
)
)
self.handle = handle
self.tokens_per_expert = num_tokens_per_expert
self.dispatched_indices = dispatched_indices
self.dispatched_probs = dispatched_probs
return hidden_states
def combine(
self,
hidden_states: torch.Tensor,
async_finish: bool = False,
allocate_on_comm_stream: bool = False,
) -> torch.Tensor:
hidden_states, _ = fused_combine(
hidden_states,
self.group,
self.handle,
async_finish=async_finish,
allocate_on_comm_stream=allocate_on_comm_stream,
)
# Release the handle after combine operation
self.handle = None
return hidden_states
class MoEFlexTokenDispatcher(MoETokenDispatcher):
"""
Flex token dispatcher using DeepEP.
"""
def dispatch_preprocess(
self, hidden_states: torch.Tensor, routing_map: torch.Tensor, probs: torch.Tensor
):
"""
Preprocesses the hidden states and routing information before dispatching tokens to experts.
Args:
hidden_states (torch.Tensor): Input hidden states to be processed
routing_map (torch.Tensor): Map indicating which expert each token should be routed to
probs (torch.Tensor): Routing probabilities for each token-expert pair
Returns:
Tuple containing:
- torch.Tensor: Reshaped hidden states
- torch.Tensor: Token probabilities from the communication manager
- None: Placeholder for compatibility
"""
self.hidden_shape = hidden_states.shape
hidden_states = hidden_states.view(-1, self.hidden_shape[-1])
# Initialize metadata
routing_map, probs = self._initialize_metadata(routing_map, probs)
self._comm_manager.setup_metadata(routing_map, probs)
return hidden_states, self._comm_manager.token_probs, None
def dispatch_all_to_all(
self,
hidden_states: torch.Tensor,
probs: torch.Tensor = None,
async_finish: bool = True,
allocate_on_comm_stream: bool = True,
):
"""
Performs all-to-all communication to dispatch tokens across expert parallel ranks.
"""
return (
self._comm_manager.dispatch(hidden_states, async_finish, allocate_on_comm_stream),
self._comm_manager.dispatched_probs,
)
def dispatch_postprocess(self, hidden_states: torch.Tensor):
"""
Post-processes the dispatched hidden states after all-to-all communication.
This method retrieves the permuted hidden states by experts, calculates the number of tokens
per expert, and returns the processed data ready for expert processing.
"""
global_input_tokens, permuted_probs = (
self._comm_manager.get_permuted_hidden_states_by_experts(hidden_states)
)
tokens_per_expert = self._comm_manager.get_number_of_tokens_per_expert()
return global_input_tokens, tokens_per_expert, permuted_probs
def token_permutation(
self, hidden_states: torch.Tensor, probs: torch.Tensor, routing_map: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Permutes tokens according to the routing map and dispatches them to experts.
This method implements the token permutation process in three steps:
1. Preprocess the hidden states and routing information
2. Perform all-to-all communication to dispatch tokens
3. Post-process the dispatched tokens for expert processing
"""
hidden_states, _, _ = self.dispatch_preprocess(hidden_states, routing_map, probs)
hidden_states, _ = self.dispatch_all_to_all(
hidden_states, async_finish=False, allocate_on_comm_stream=False
)
global_input_tokens, tokens_per_expert, permuted_probs = self.dispatch_postprocess(
hidden_states
)
return global_input_tokens, tokens_per_expert, permuted_probs
def combine_preprocess(self, hidden_states: torch.Tensor):
"""
Pre-processes the hidden states before combining them after expert processing.
This method restores the hidden states to their original ordering before expert processing
by using the communication manager's restoration function.
"""
hidden_states = self._comm_manager.get_restored_hidden_states_by_experts(hidden_states)
return hidden_states
def combine_all_to_all(
self,
hidden_states: torch.Tensor,
async_finish: bool = True,
allocate_on_comm_stream: bool = True,
):
"""
Performs all-to-all communication to combine tokens after expert processing.
"""
return self._comm_manager.combine(hidden_states, async_finish, allocate_on_comm_stream)
def combine_postprocess(self, hidden_states: torch.Tensor):
"""
Post-processes the combined hidden states after all-to-all communication.
This method reshapes the combined hidden states to match the original input shape.
"""
return hidden_states.view(self.hidden_shape)
def token_unpermutation(
self, hidden_states: torch.Tensor, bias: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
"""
Reverses the token permutation process to restore the original token order.
This method implements the token unpermutation process in three steps:
1. Pre-process the hidden states to restore their original ordering
2. Perform all-to-all communication to combine tokens
3. Post-process the combined tokens to match the original input shape
"""
assert bias is None, "Bias is not supported in MoEFlexTokenDispatcher"
hidden_states = self.combine_preprocess(hidden_states)
hidden_states = self.combine_all_to_all(hidden_states, False, False)
hidden_states = self.combine_postprocess(hidden_states)
return hidden_states, None
from functools import wraps
from megatron.core.transformer.transformer_block import TransformerBlock as MegatronCoreTransformerBlock
def transformer_block_init_wrapper(fn):
@wraps(fn)
......@@ -13,3 +14,22 @@ def transformer_block_init_wrapper(fn):
self.final_layernorm = None
return wrapper
class TransformerBlock(MegatronCoreTransformerBlock):
def __init__(
self, *args, **kwargs
):
super().__init__(*args, **kwargs)
# mtp require seperate layernorms for main model and mtp modules, thus move finalnorm out of block
config = args[0] if len(args) > 1 else kwargs['config']
if getattr(config, "mtp_num_layers", 0) > 0:
self.main_final_layernorm = self.final_layernorm
self.final_layernorm = None
def get_layer_callables(self, layer_number: int):
"""
Get the callables for the layer at the given layer number.
"""
return self.layers[layer_number].get_submodule_callables()
from megatron.core.transformer.transformer_layer import TransformerLayer as MegatronCoreTransformerLayer
class TransformerLayer(MegatronCoreTransformerLayer):
def _submodule_attn_router_forward(
self,
hidden_states,
attention_mask=None,
inference_params=None,
rotary_pos_emb=None,
rotary_pos_cos=None,
rotary_pos_sin=None,
attention_bias=None,
packed_seq_params=None,
sequence_len_offset=None,
state=None,
):
"""
Performs a combined forward pass that includes self-attention and MLP routing logic.
"""
hidden_states, _ = self._forward_attention(
hidden_states=hidden_states,
attention_mask=attention_mask,
rotary_pos_emb=rotary_pos_emb,
rotary_pos_cos=rotary_pos_cos,
rotary_pos_sin=rotary_pos_sin,
attention_bias=attention_bias,
packed_seq_params=packed_seq_params,
sequence_len_offset=sequence_len_offset,
inference_params=inference_params,
)
pre_mlp_layernorm_output = self.pre_mlp_layernorm(hidden_states)
probs, routing_map = self.mlp.router(pre_mlp_layernorm_output)
local_tokens, probs, tokens_per_expert = self.mlp.token_dispatcher.dispatch_preprocess(
pre_mlp_layernorm_output, routing_map, probs
)
return (local_tokens, probs, hidden_states, pre_mlp_layernorm_output, tokens_per_expert)
def _submodule_dispatch_forward(self, local_tokens, probs, state=None):
"""
Dispatches tokens to the appropriate experts based on the router output.
"""
token_dispatcher = self.mlp.token_dispatcher
if self.is_deepep:
token_dispatcher._comm_manager.token_probs = probs
return token_dispatcher.dispatch_all_to_all(local_tokens, probs)
def _submodule_moe_forward(self, dispatched_tokens, probs=None, state=None):
"""
Performs a forward pass for the MLP submodule, including both expert-based
and optional shared-expert computations.
"""
shared_expert_output = None
token_dispatcher = self.mlp.token_dispatcher
if self.is_deepep:
token_dispatcher._comm_manager.dispatched_probs = state.dispatched_probs
dispatched_tokens, tokens_per_expert, permuted_probs = (
token_dispatcher.dispatch_postprocess(dispatched_tokens)
)
else:
dispatched_tokens, permuted_probs = token_dispatcher.dispatch_postprocess(
dispatched_tokens, probs
)
tokens_per_expert = state.tokens_per_expert
expert_output, mlp_bias = self.mlp.experts(
dispatched_tokens, tokens_per_expert, permuted_probs
)
assert mlp_bias is None, f"Bias is not supported in {token_dispatcher.__class__.__name__}"
if self.mlp.use_shared_expert and not self.mlp.shared_expert_overlap:
shared_expert_output = self.mlp.shared_experts(state.pre_mlp_layernorm_output)
expert_output = self.mlp.token_dispatcher.combine_preprocess(expert_output)
return expert_output, shared_expert_output, mlp_bias
def _submodule_combine_forward(self, output, shared_expert_output=None, state=None):
residual = state.residual
token_dispatcher = self.mlp.token_dispatcher
output = token_dispatcher.combine_all_to_all(output)
output = token_dispatcher.combine_postprocess(output)
if shared_expert_output is not None:
output = output + shared_expert_output
mlp_output_with_bias = (output, None)
with self.bias_dropout_add_exec_handler():
hidden_states = self.mlp_bda(self.training, self.config.bias_dropout_fusion)(
mlp_output_with_bias, residual, self.hidden_dropout
)
output = make_viewless_tensor(
inp=hidden_states, requires_grad=hidden_states.requires_grad, keep_graph=True
)
return output
def _submodule_attn_router_dw(self):
self.self_attention.backward_dw()
def _submodule_mlp_dw(self):
self.mlp.backward_dw()
def _submodule_attn_router_postprocess(
self, node, local_tokens, probs, residual, pre_mlp_layernorm_output, tokens_per_expert
):
node.common_state.residual = node.detach(residual)
if self.mlp.use_shared_expert:
node.common_state.pre_mlp_layernorm_output = node.detach(pre_mlp_layernorm_output)
if not self.is_deepep:
node.common_state.tokens_per_expert = tokens_per_expert
return local_tokens, probs
def _submodule_dispatch_postprocess(self, node, dispatched_tokens, probs):
if self.is_deepep:
node.common_state.dispatched_probs = node.detach(probs)
return dispatched_tokens
else:
return dispatched_tokens, probs
def _submodule_mlp_postprocess(self, node, expert_output, shared_expert_output, mlp_bias):
assert mlp_bias is None
node.common_state.pre_mlp_layernorm_output = None
if shared_expert_output is None:
return expert_output
return expert_output, shared_expert_output
def _submodule_combine_postprocess(self, node, output):
cur_stream = torch.cuda.current_stream()
node.common_state.residual.record_stream(cur_stream)
node.common_state.residual = None
return output
def _submodule_attn_postprocess(self, node, hidden_states, context):
return hidden_states
def _submodule_dense_postprocess(self, node, hidden_states):
return hidden_states
def _submodule_not_implemented(self, *args):
raise NotImplementedError("This callable is not implemented.")
def get_submodule_callables(self, chunk_state):
"""
The forward callables take 2 parts of inputs:
1. The ScheduleNode object.
2. The input tensors.
"""
from megatron.core.transformer.moe.moe_layer import MoELayer
from megatron.core.transformer.moe.token_dispatcher import MoEFlexTokenDispatcher
self.is_moe = isinstance(self.mlp, MoELayer)
self.is_deepep = False
if self.is_moe:
self.is_deepep = isinstance(self.mlp.token_dispatcher, MoEFlexTokenDispatcher)
def get_func_with_default(func, default_func):
if self.is_moe:
return func
return default_func
def callable_wrapper(forward_func, postprocess_func, node, *args):
state = getattr(node, 'common_state', None)
callable_outputs = forward_func(*args, state=state)
if isinstance(callable_outputs, tuple):
outputs = postprocess_func(node, *callable_outputs)
else:
outputs = postprocess_func(node, callable_outputs)
return outputs
attn_func = get_func_with_default(
self._submodule_attn_router_forward, self._forward_attention
)
def attn_wrapper(hidden_states, state=None):
return attn_func(
hidden_states=hidden_states,
attention_mask=chunk_state.attention_mask,
attention_bias=chunk_state.attention_bias,
inference_params=chunk_state.inference_params,
packed_seq_params=chunk_state.packed_seq_params,
sequence_len_offset=chunk_state.sequence_len_offset,
rotary_pos_emb=chunk_state.rotary_pos_emb,
rotary_pos_cos=chunk_state.rotary_pos_cos,
rotary_pos_sin=chunk_state.rotary_pos_sin,
state=state,
)
attn_postprocess_func = get_func_with_default(
self._submodule_attn_router_postprocess, self._submodule_attn_postprocess
)
dispatch_func = get_func_with_default(
self._submodule_dispatch_forward, self._submodule_not_implemented
)
dispatch_postprocess_func = get_func_with_default(
self._submodule_dispatch_postprocess, self._submodule_not_implemented
)
mlp_func = get_func_with_default(self._submodule_moe_forward, self._forward_mlp)
mlp_postprocess_func = get_func_with_default(
self._submodule_mlp_postprocess, self._submodule_dense_postprocess
)
combine_func = get_func_with_default(
self._submodule_combine_forward, self._submodule_not_implemented
)
combine_postprocess_func = get_func_with_default(
self._submodule_combine_postprocess, self._submodule_not_implemented
)
attn_forward = partial(callable_wrapper, attn_wrapper, attn_postprocess_func)
dispatch_forward = partial(callable_wrapper, dispatch_func, dispatch_postprocess_func)
mlp_forward = partial(callable_wrapper, mlp_func, mlp_postprocess_func)
combine_forward = partial(callable_wrapper, combine_func, combine_postprocess_func)
callables = TransformerLayerSubmoduleCallables(
attention=SubmoduleCallables(forward=attn_forward, dw=self._submodule_attn_router_dw),
dispatch=SubmoduleCallables(forward=dispatch_forward),
mlp=SubmoduleCallables(forward=mlp_forward, dw=self._submodule_mlp_dw),
combine=SubmoduleCallables(forward=combine_forward),
is_moe=self.is_moe,
is_deepep=self.is_deepep,
)
return callables
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment