kv-cache : simplify the interface #13660
Merged
+89
−153
Conversation
This file contains hidden or 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
ggerganov
force-pushed
the
gg/kv-cache-simplify-part1
branch
from
70321a1 to
ef880b3
Compare
ggerganov
force-pushed
the
gg/kv-cache-simplify-part1
branch
from
ef880b3 to
a91b15f
Compare
slaren
reviewed
May 21, 2025
src/llama-batch.cpp
Outdated
| @@ -283,7 +283,7 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0 | |||
| if (!batch.pos) { | |||
| pos.resize(batch.n_tokens); | |||
| for (int32_t i = 0; i < batch.n_tokens; i++) { | |||
| pos[i] = i + p0; | |||
| pos[i] = p0 + i + 1; | |||
There was a problem hiding this comment.
This change is a bit confusing to me. With p0 I would understand "position zero", but now this parameter seem to mean "previous max pos" instead.
There was a problem hiding this comment.
I was focused on simplifying the call site by absorbing the + 1 into the function, but you are right that this makes the parameter meaning more confusing. Changed back to the previous version and also added assert for p0.
ggerganov
force-pushed
the
gg/kv-cache-simplify-part1
branch
from
0e096d4 to
e987482
Compare
slaren
approved these changes
May 21, 2025
infil00p
pushed a commit
to baseweight/llama.cpp
that referenced
this pull request
May 22, 2025
* kv-cache : simplify the interface ggml-ci * context : revert llama_batch_allocr position change ggml-ci
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Outdated suggestions cannot be applied.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.
First part of some KV cache interface refactoring and simplification.
Public API changes
llama_kv_self_n_tokensllama_kv_self_used_cellsInternal
llama_kv_cachechangesllama_kv_cache::get_n_tokens()llama_kv_cache::get_used_cells()llama_kv_cache::get_pos_max()n_seq_maxto the KV cache objects. Will be needed later for improving the data structures for tracking the per-sequence information.type_kandtype_vmemberspadding->n_padfor consistencyOther changes
llama_decode()now verifies that if the input batch haspos == nullit should also haveseq_id == nullso that we can automatically assign all tokens toseq_id == 0starting from the max position currently in the cache. This fixes/prevents an edge case where a batch withpos == nullthat also has tokens withseq_id != 0would be assigned incorrect positions by thellama_batch_allocr./metricsendpoint. These are too internal and implementation-specific and should not be exposed to the public.