Llama3 hybrid implementation using submeshes #18777
Draft
+898
−455
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
yieldthought
requested changes
Mar 7, 2025
There was a problem hiding this comment.
Clean 👌
To do:
- Add at least one CI test that will exercise DP. I suggest adding a demo to the t3k tests.
cfjchu
approved these changes
Mar 8, 2025
Comment on lines
+452
to
+453
| if is_ci_env and num_devices == 8 and data_parallel > 1 and not ("3.2-1B" in llama_dir or "3.1-8B" in llama_dir): | ||
| pytest.skip("CI runs only hybrid Llama3 1b and 8b on T3K") |
There was a problem hiding this comment.
Wanted to avoid burdening the CI with additional tests. 1B and 8B seemed okay to cover perf regression checks as the smallest and biggest variants of the smaller Llama3 models. Should we add 3B anyway?
| return data_parallel, mesh_device.create_submeshes(ttnn.MeshShape(1, num_devices // data_parallel)) | ||
|
|
||
|
|
||
| def allocate_kv_cache(kv_cache_shape, dtype, num_layers, mesh_device): |
There was a problem hiding this comment.
TODO (@ipotkonjak-tt and/or @skhorasganiTT) Modify KV creation in vLLM to use this function and test with DP
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Outdated suggestions cannot be applied.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.
Problem description
Missing support for data / hybrid parallelism for Llama3 models.
What's changed
Addition of hybrid parallelism within llama code base with concept of submeshes. Implementation is mainly based at the LlamaGenerator level. MeshDevice is partitioned into submeshes where each subset of devices has an independent model. Models remain implemented in the tensor parallel manner.
Checklist