Add an HLO backend for LLM models #775
Open
+7,151
−1,394
Conversation
This file contains 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
This backend is a backport of features previously implemented in the AWS Neuron SDK transformers-neuronx package.
When using the new HLO backend, the graphs will be slightly modified, so we bump the dev version to avoid trying to reuse the cached test artifacts from the previous dev version.
The name of the class is confusing as there are already NeuronConfig classes in AWS Neuron SDK (both in NxDI and TnX).
dacorvo
force-pushed
the
add_hlo_backend
branch
from
370867d to
e415a8a
Compare
These tests are taking some time, so it is better to have them separated.
dacorvo
force-pushed
the
add_hlo_backend
branch
from
e415a8a to
c510221
Compare
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
tengomucho
approved these changes
Feb 5, 2025
optimum/exporters/neuron/__main__.py
Outdated
| @@ -724,7 +730,7 @@ def main(): | |||
| submodels = None | |||
| else: | |||
| input_shapes, neuron_config_class = get_input_shapes_and_config_class(task, args) | |||
| if NeuronDecoderConfig in inspect.getmro(neuron_config_class): | |||
| if NeuronDecoderExportConfig in inspect.getmro(neuron_config_class): | |||
There was a problem hiding this comment.
if !is_transformers_neuronx_available() then NeuronDecoderExportConfig will not be defined.
The HF_ENDPOINT variable is not always taken into account when using the huggingface_hub client depending on the order of imports. This modifies the tests to create temporary dorectories under the testing user account instead.
dacorvo
force-pushed
the
add_hlo_backend
branch
from
836271c to
9299f94
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
What does this PR do?
This backend is a backport of features first implemented in the transformers-neuronx package from the
AWS Neuron SDK.
As the original transformers-neuronx implementation, it relies on XLA High Level Operations (HLO)
as the compiled language for implementing Neuron optimized transformer decoder classes.
More specifically, it uses a syntax called “PyHLO”, name of a Neuron internal tool for writing/compiling the HLO language in Python.
See
backends\hlo\README.mdfor details.The
Llama,GraniteandQwen2models previously usingtransformers-neuronxare now using this new backend directly.