Fix cache config inconsistency between shortfin and sharktank #406
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
renxida
force-pushed
the
cache-config-consistency
branch
from
754ae09 to
93c0d19
Compare
renxida
force-pushed
the
cache-config-consistency
branch
from
93c0d19 to
cd309ac
Compare
|
Proof it works. Tested with script: https://gist.github.com/renxida/2023ec7f9931c697879762c9330de001 |
| "block_seq_stride": llama_config.block_seq_stride, | ||
| "paged_kv_cache": { | ||
| "block_seq_stride": llama_config.block_seq_stride, | ||
| "attn_head_count_kv": hp.attention_head_count_kv, |
There was a problem hiding this comment.
Just FYI, in sharktank's llm_configs.py, hp.attention_head_count_kv==hp.attention_head_count.
There was a problem hiding this comment.
If you req vmfb and irpa for shortfin, is there a reason we rely on generate_params_json params instead of the irpa params? There are ways to load & access irpa params directly in sharktank. Do we want/have something similar in shortfin?
There was a problem hiding this comment.
Oh ty! i should probably take a deeper look then to figure this part out.
|
Replaced by #487 |
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Outdated suggestions cannot be applied.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.
Fixes #405 and #401
Up till now we have been patching the configs to reorganize it after exporting from sharktank, but for llama3 8b (which has grouped attention), the config doesn't specify grouping, so this has to be fixed.
Might be a good idea to rename all attn to attention for consistency. But I want this PR to just fix the issue at hand.
After this fix, the only required bit of editing is setting config.paged_kv_cache.device_block_count