🦙 Newer Llamas support #129
Merged
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
|
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. |
The original Llama model has MLP bias disabled, but it might happen that it is enabled for finetuned models: in this case an error will appear when serving with Jetstream Pytorch.
If the config is enabled, on the Jetstream Pt TGI we create an alias, that will eventually result in untied weights (but at least it will work).
This does not seem to be necessary anymore.
Otherwise it only uses the last marker.
tengomucho
force-pushed
the
test_llama3.1
branch
from
8219b67 to
da1b9b8
Compare
dacorvo
approved these changes
Dec 13, 2024
I don't like monkey patching models either, but hopefully I will be able to remove it once AI-Hypercomputer/jetstream-pytorch#205 lands into Jetstream/Pytorch |
| ], | ||
| ids=["Mixtral-8x7B", "Meta-Llama-3-8B" ,"Meta-Llama-3-70B"], | ||
| ids=["Mixtral-8x7B", "Meta-Llama-3-8B" ,"Meta-Llama-3-70B", "Llama-3.3-70B-Instruct"], |
There was a problem hiding this comment.
Only question should we add test for 3.1 and 3.2 or that's unnecessary?
There was a problem hiding this comment.
We could, but 3.2 is tested with the 1B variant and 3.1 architecture is identical, so I do not think testing it would add any value.
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Outdated suggestions cannot be applied.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 PR mainly adds support for RoPE scaling to Jetstream Pytorch Llama models, that is required to support newer Llama models on TGI, such as Llama 3.1, Llama 3.2 and Llama 3.3.
Note that Llama3.3 is a 70B model, so quantization should be enabled to serve it.
Before submitting