[Snippets][CPU] Disable MHA tokenization in LLM #28601
Open
+14
−5
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
a-sidorova
force-pushed
the
feature/snippets/disable_mha_token_in_llm
branch
2 times, most recently
from
0e62943 to
08aeea7
Compare
IvanNovoselov
approved these changes
Jan 22, 2025
Comment on lines
1040
to
1042
| // Note: the variable `ops` should not exist during `SnippetsTokenization` execution. | ||
| // Otherwise, it will extend the life time of ops (since they're stored as shared ptrs) and | ||
| // they will be visible in the model during the tokenization passes even after removing or replacing. |
There was a problem hiding this comment.
Minor: this comment looks a bit confusing to be honest. Had to read it a few times, but still don't fully understand it.
It looks like you compare the present check with some other solution that could've been implemented (i.e. store ops somewhere and access them from the tokenization pass?). Not sure we need to do this.
There was a problem hiding this comment.
This comment is outdated now since I use ov::op::util::has_op_with_type<OP>(model) in cad0554. Thanks!
dmitry-gorokhov
approved these changes
Jan 23, 2025
a-sidorova
force-pushed
the
feature/snippets/disable_mha_token_in_llm
branch
from
cad0554 to
e432b62
Compare
akladiev
pushed a commit
that referenced
this pull request
Jan 23, 2025
### Details: - *The second inference in LLM is usually single token inference. It means that `M` dimension of MatMuls in SDPA pattern will have the value `1` (during compilation model this dimension is dynamic (unknown)). Snippets cannot provide efficient execution for single token inference. So we decided to disable MHA tokenization by Snippets in CPU Plugin on LLMs'. We consider the presence of `ScaledDotProductAttentionWithKVCache` op in the model as a sign that this model is LLM.* - *Cherry-picked from #28601 ### Tickets: - *160634* - *160978 (contains performance validation results)*
a-sidorova
force-pushed
the
feature/snippets/disable_mha_token_in_llm
branch
from
e432b62 to
8e13f0c
Compare
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.
Details:
Mdimension of MatMuls in SDPA pattern will have the value1(during compilation model this dimension is dynamic (unknown)). Snippets cannot provide efficient execution for single token inference. So we decided to disable MHA tokenization by Snippets in CPU Plugin on LLMs'. We consider the presence ofScaledDotProductAttentionWithKVCacheop in the model as a sign that this model is LLM.Tickets:
TODO: