Skip to content

llama4 distributed: compile optimizer #2659

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 6, 2025
Merged

llama4 distributed: compile optimizer #2659

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion recipes/configs/llama4/scout_17B_16E_full.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,7 +69,14 @@ device: cuda
enable_activation_checkpointing: True
enable_activation_offloading: False
fsdp_cpu_offload: True
compile: False # torch.compile, set to true for perf/memory improvement
# compile True means use torch.compile for all components
# compile False means no torch.compile
# compile Dictionary with keys: "model", "loss", "optimizer_step"
# enables torch.compile only for specified components.
compile: False
# model: True
# loss: True
# optimizer_step: False

# Reduced precision
dtype: bf16
Expand Down
23 changes: 20 additions & 3 deletions recipes/full_finetune_distributed.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

import os
import sys
import time

Expand Down Expand Up @@ -306,7 +307,18 @@ def setup(self, cfg: DictConfig) -> None:
# Load the base model
checkpoint_dict = self._checkpoint_client.load_base_checkpoint()

self._compile = cfg.get("compile", False)
compile = cfg.get("compile")
compile_bool = bool(compile)
self._compile_backend = os.environ.get("TORCH_COMPILE_BACKEND", "inductor")

self._compile_model = compile_bool
self._compile_loss = compile_bool
self._compile_optimizer_step = compile_bool
if isinstance(compile, DictConfig):
self._compile_model = compile.get("model", True)
self._compile_loss = compile.get("loss", True)
Comment on lines +318 to +319
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A minor thing, but I wonder whether we should default to False here just to match the default when the user is not passing a dict. Not a huge deal since users will likely need to opt-in by passing the dict anyways

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I look at this values as a default pieces that we recommend user to compile if user specified compilation as dict and did not specify exactly model, loss.
I think this will be similar alignment what was before for "compile: True", where it was compiling model and loss.

self._compile_optimizer_step = compile.get("optimizer_step", False)

self._model = self._setup_model(
cfg_model=cfg.model,
enable_activation_checkpointing=self._enable_activation_checkpointing,
Expand All @@ -329,6 +341,11 @@ def setup(self, cfg: DictConfig) -> None:
else None
),
)
if self._compile_optimizer_step:
self._optimizer.step = torch.compile(
self._optimizer.step,
backend=self._compile_backend,
)

if self._resume_from_checkpoint:
# If async checkpointing is enabled, intermediate checkpoints are saved asynchronously
Expand Down Expand Up @@ -358,7 +375,7 @@ def setup(self, cfg: DictConfig) -> None:
# initialize loss
self._loss_fn = config.instantiate(cfg.loss)

if self._compile:
if self._compile_loss:
training.compile_loss(self._loss_fn, verbose=self._is_rank_zero)

# The loss may handle the output projection. If true, the model should skip it.
Expand Down Expand Up @@ -569,7 +586,7 @@ def _setup_model(
with training.set_default_dtype(self._dtype), torch.device("meta"):
model = config.instantiate(cfg_model)

if self._compile:
if self._compile_model:
training.compile_model(model, verbose=self._is_rank_zero)

if self._enable_fp8_training:
Expand Down
Loading