Add Splitwise implementation to vLLM #1
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
Initialize MSCCL++ communication group if sep_prompt_token is set in ParallelConfig. Also add documentation for MSCCL++ installation.
- Add worker_type to differentiate prompt, token, and mixed workers - Set a driver for each prompt machines and token machines - Allow broadcasts to take a group - Setup KV Cache communication using MSCCL++ - Add test for KV Cache communication
- Obtain `blocks_to_nw` when creating batches in scheduler. Coalesce blocks where possible for fast network transfers. - Use a Sequence to Semaphore Mapper to allow for fine-grained waiting for kv-cache transfer per sequence - Separately run prompt and token workers using the `_run_stage_workers` helper - Populate KVCacheCommunicator for all PagedAttention modules, which allows implementation of layer-wise sends from within `attention.py` - Populate destination rank for Sampler, which will be used as root in `gather` operations. - Fix `tensor_model_parallel_gather` - use global rank instead of group local rank.
goiri
reviewed
Feb 3, 2024
| @@ -0,0 +1,41 @@ | |||
| """Test the KV cache communication operators. | |||
|
|
|||
| Run `python test_kvcache_comm.py`. | |||
There was a problem hiding this comment.
I'm guessing there's no CI infra or test running?
There was a problem hiding this comment.
Actually, they do have one. I will have to figure out how to add this test to the CI infra. Multiple things to figure out - how to use pytest with argparse for LLMEngine and setting up MSCCL++ environment in the CI infra.
goiri
approved these changes
Feb 4, 2024
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Outdated suggestions cannot be applied.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.
No description provided.