Update model definition to support Flash-Decoding #177
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
masahi
commented
Jan 30, 2024
| @@ -402,6 +403,10 @@ class BuildArgs: | |||
| "action": "store_true", | |||
| }, | |||
| ) | |||
| paged_kv_cache_type: str = field( | |||
| default="vllm", | |||
| metadata={"help": "The type of paged KV cache, either vllm or flash-decoding"}, | |||
There was a problem hiding this comment.
This new option makes --use_vllm_attention obsolete. Since removing it is a breaking change, I'll do that later when I integrate Flash-Decoding into mlc-serve. @sunggg
vinx13
approved these changes
Jan 30, 2024
masahi
force-pushed
the
flash-decoding
branch
from
c9d7fac to
b9e41e1
Compare
Lunderberg
pushed a commit
to Lunderberg/mlc-llm
that referenced
this pull request
Jan 30, 2024
The repetition penalty (introduced in [CTRL](https://arxiv.org/abs/1909.05858)) can help prevent the LLM from generating repetitive tokens. This PR implements the repetition penalty. Note: Previous the logits softmax is performed on GPU, this PR moves it to CPU to accommodate the repetition penalty.
sunggg
approved these changes
Jan 30, 2024
This was referenced Jan 31, 2024
Merged
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Outdated suggestions cannot be applied.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.
This PR integrates Flash-Decoding support from apache/tvm#16474. This is a drop-in replacement for the vLLM kernel. The only difference with the vLLM-based build is the shape of KV cache blocks. In particular, the block size for vLLM is 16 while for Flash-Decoding it is 256.
In addition, it supports decoding with multiple, fixed length queries per request, which is necessary for speculative decoding.
evaluate_multi_queryfrom #156 can also be used for this purpose, but it supports variable-length queries per request and piggy-backs to the prefill attention, which is not efficient when the number of queries is fixed and small. The changes inrun_llama_batched_vllm.pydemonstrates that the new Relax function,decode_multi_query, can do exactly the same thing asevaluate_multi_querywhen the query length is fixed.This PR only updates the model definition and
run_llama_batched_vllm.pyexample script. I'll follow up with the integration intomlc-servenext.Need the latest https://github.com/octoml/tvm/tree/for-mlc-serve-jan12
@sunggg @yelite @vinx13