Fix Mllama model placement#2125
Merged
astachowiczhabana merged 1 commit intoJul 14, 2025
Merged
Conversation
Collaborator
|
Why not targeting the main branch? |
Collaborator
|
Hi @regisss |
pbielak
force-pushed
the
dev/pbielak/fix-mllama-model-placement
branch
from
a840dd0 to
0e255c3
Compare
pbielak
force-pushed
the
dev/pbielak/fix-mllama-model-placement
branch
from
0e255c3 to
0f03cc1
Compare
astachowiczhabana
force-pushed
the
v1.19-release
branch
from
79dc01d to
62b45d7
Compare
With Llama 4 support in Transformers 4.51, there was a change in the `Pipeline` class [1], which causes the pipeline device to be set to `self.model.device`. In the case of Mllama, DeepSpeed is used to create the `.language_model` on HPU, whereas the rest of the model stays on CPU [2]. Hence, always `self.model.device = CPU`, which causes the whole model to be placed back on CPU. This commit explicitly moves the model to HPU, so the pipeline will be also placed on HPU. [1] https://github.com/huggingface/transformers/pull/37307/files#diff-441f558737166b045444da9c4be81f566b3d69054e8f20e288aed746a691fa61 [2] https://github.com/huggingface/optimum-habana/blob/v1.18.0/examples/image-to-text/run_pipeline.py#L360
pbielak
force-pushed
the
dev/pbielak/fix-mllama-model-placement
branch
from
0f03cc1 to
86537e5
Compare
Collaborator
|
This PR looks like a solution for blockers of release 1.19, as well as 1.18.1. @regisss If merged, please cherry pick it to v1.18-release branch. |
astachowiczhabana
approved these changes
Jul 14, 2025
astachowiczhabana
pushed a commit
that referenced
this pull request
Jul 14, 2025
With Llama 4 support in Transformers 4.51, there was a change in the `Pipeline` class [1], which causes the pipeline device to be set to `self.model.device`. In the case of Mllama, DeepSpeed is used to create the `.language_model` on HPU, whereas the rest of the model stays on CPU [2]. Hence, always `self.model.device = CPU`, which causes the whole model to be placed back on CPU. This commit explicitly moves the model to HPU, so the pipeline will be also placed on HPU. [1] https://github.com/huggingface/transformers/pull/37307/files#diff-441f558737166b045444da9c4be81f566b3d69054e8f20e288aed746a691fa61 [2] https://github.com/huggingface/optimum-habana/blob/v1.18.0/examples/image-to-text/run_pipeline.py#L360
astachowiczhabana
pushed a commit
that referenced
this pull request
Aug 22, 2025
With Llama 4 support in Transformers 4.51, there was a change in the `Pipeline` class [1], which causes the pipeline device to be set to `self.model.device`. In the case of Mllama, DeepSpeed is used to create the `.language_model` on HPU, whereas the rest of the model stays on CPU [2]. Hence, always `self.model.device = CPU`, which causes the whole model to be placed back on CPU. This commit explicitly moves the model to HPU, so the pipeline will be also placed on HPU. [1] https://github.com/huggingface/transformers/pull/37307/files#diff-441f558737166b045444da9c4be81f566b3d69054e8f20e288aed746a691fa61 [2] https://github.com/huggingface/optimum-habana/blob/v1.18.0/examples/image-to-text/run_pipeline.py#L360
astachowiczhabana
pushed a commit
that referenced
this pull request
Aug 25, 2025
With Llama 4 support in Transformers 4.51, there was a change in the `Pipeline` class [1], which causes the pipeline device to be set to `self.model.device`. In the case of Mllama, DeepSpeed is used to create the `.language_model` on HPU, whereas the rest of the model stays on CPU [2]. Hence, always `self.model.device = CPU`, which causes the whole model to be placed back on CPU. This commit explicitly moves the model to HPU, so the pipeline will be also placed on HPU. [1] https://github.com/huggingface/transformers/pull/37307/files#diff-441f558737166b045444da9c4be81f566b3d69054e8f20e288aed746a691fa61 [2] https://github.com/huggingface/optimum-habana/blob/v1.18.0/examples/image-to-text/run_pipeline.py#L360
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.
What does this PR do?
With Llama 4 support in Transformers 4.51, there was a change in the
Pipelineclass [1], which causes the pipeline device to be set toself.model.device. In the case of Mllama, DeepSpeed is used to create the.language_modelon HPU, whereas the rest of the model stays on CPU [2]. Hence, alwaysself.model.device = CPU, which causes the whole model to be placed back on CPU. This commit explicitly moves the model to HPU, so the pipeline will be also placed on HPU.[1] https://github.com/huggingface/transformers/pull/37307/files#diff-441f558737166b045444da9c4be81f566b3d69054e8f20e288aed746a691fa61
[2] https://github.com/huggingface/optimum-habana/blob/v1.18.0/examples/image-to-text/run_pipeline.py#L360