[Compiler Toolkit] Add option for full inductor.#2150
Merged
Conversation
meta-cla
bot
added
the
CLA Signed
SherlockNoMad
requested changes
Dec 15, 2025
Contributor
SherlockNoMad
left a comment
SherlockNoMad
left a comment
There was a problem hiding this comment.
lgtm with comments.
yiming0416
reviewed
Dec 15, 2025
aditvenk
force-pushed
the
ps/rr/_compiler_toolkit__add_option_for_full_inductor_
branch
from
bd66bbd to
82834ba
Compare
Contributor
Author
|
aot_eager for llama3: full inductor: |
yiming0416
requested changes
Jan 5, 2026
Contributor
yiming0416
left a comment
yiming0416
left a comment
There was a problem hiding this comment.
Can we add commands to run this to compiler_toolkit/README?
Also you can add 8-GPU CI following existing ones in compiler_toolkit/tests/integration_tests.py
aditvenk
force-pushed
the
ps/rr/_compiler_toolkit__add_option_for_full_inductor_
branch
from
82834ba to
0083214
Compare
Contributor
Author
Added to README and integration test. |
aditvenk
force-pushed
the
ps/rr/_compiler_toolkit__add_option_for_full_inductor_
branch
from
0083214 to
f8e7294
Compare
yiming0416
reviewed
Jan 6, 2026
yiming0416
reviewed
Jan 6, 2026
| """ | ||
|
|
||
| joint_passes: list[str] = field(default_factory=list) | ||
| passes: list[str] = field(default_factory=list) |
Contributor
There was a problem hiding this comment.
non-blocking. I wonder if we should have a better naming here to distinguish joint_passes and passes applied on partitioned graphs. Do we consider having fwd_passes and bwd_passes (maybe an overkill for now)
Contributor
Author
There was a problem hiding this comment.
I think joint_passes is pretty descriptive for what it is.
I think passes could be renamed to post_partition_passes, as we don't yet have an example for fwd-only or bwd-only pass, so it is overkill to split to two options for fwd and bwd.
- Being able to compile fw/bw graphs using compile_fx_inner could help with establishing perf rooflines. - Full inductor compilation is achieved using compile_fx_inner, however, it requires the graph to have been decomposed using Inductor's default decomposition table. We apply this decomposition as a pass on the joint graph. We need to be careful to suitably unwrap the primals/tangents before running this decomposition. <!-- ps-id: 861b7e8c-8917-4806-8fec-25b0e5288f29 -->
aditvenk
force-pushed
the
ps/rr/_compiler_toolkit__add_option_for_full_inductor_
branch
from
f8e7294 to
2f8bfb2
Compare
yiming0416
approved these changes
Jan 7, 2026
SherlockNoMad
approved these changes
Jan 7, 2026
wwwjn
pushed a commit
to wwwjn/torchtitan
that referenced
this pull request
Jan 15, 2026
Being able to compile fw/bw graphs using compile_fx_inner could help with establishing perf rooflines. Full inductor compilation is achieved using `compile_fx_inner`, however, it requires the graph to have been decomposed using Inductor's default decomposition table. We apply this decomposition as a pass on the joint graph. We need to be careful to suitably unwrap the primals/tangents before running this decomposition. Manual testing: NGPU=4 \ CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml \ TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train \ ./run_train.sh \ --model.name $MODEL_NAME \ --parallelism.data_parallel_shard_degree=2 \ --parallelism.tensor_parallel_degree=2 \ --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config \ --compile.joint_passes inductor_decomposition \ --compile.passes full_inductor_compilation <!-- ps-id: 5d590700-6d1f-44fe-8f70-4d2ea39106f4 -->
xrsrke
pushed a commit
to NousResearch/torchtitan
that referenced
this pull request
Feb 13, 2026
Being able to compile fw/bw graphs using compile_fx_inner could help with establishing perf rooflines. Full inductor compilation is achieved using `compile_fx_inner`, however, it requires the graph to have been decomposed using Inductor's default decomposition table. We apply this decomposition as a pass on the joint graph. We need to be careful to suitably unwrap the primals/tangents before running this decomposition. Manual testing: NGPU=4 \ CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml \ TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train \ ./run_train.sh \ --model.name $MODEL_NAME \ --parallelism.data_parallel_shard_degree=2 \ --parallelism.tensor_parallel_degree=2 \ --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config \ --compile.joint_passes inductor_decomposition \ --compile.passes full_inductor_compilation <!-- ps-id: 5d590700-6d1f-44fe-8f70-4d2ea39106f4 -->
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Being able to compile fw/bw graphs using compile_fx_inner could help with establishing perf rooflines.
Full inductor compilation is achieved using
compile_fx_inner, however, it requires the graph to have been decomposed using Inductor's default decomposition table. We apply this decomposition as a pass on the joint graph. We need to be careful to suitably unwrap the primals/tangents before running this decomposition.Manual testing:
NGPU=4
CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml
TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train
./run_train.sh
--model.name $MODEL_NAME
--parallelism.data_parallel_shard_degree=2
--parallelism.tensor_parallel_degree=2
--job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config
--compile.joint_passes inductor_decomposition
--compile.passes full_inductor_compilation