model: add dots3-note - #27060
Merged
Merged
Conversation
github-actions
Bot
added
model
ngxson
commented
Aug 14, 2026
Comment on lines
+32
to
+49
| const bool has_swa = swa_type != LLAMA_SWA_TYPE_NONE; | ||
|
|
||
| const layer_filter_cb filter_mla_full = [&](int32_t il) { | ||
| if (filter_mla && !filter_mla(il)) { | ||
| return false; | ||
| } | ||
|
|
||
| return !(has_swa && model.hparams.is_swa(il)); | ||
| }; | ||
|
|
||
| const layer_filter_cb filter_mla_swa = [&](int32_t il) { | ||
| if (filter_mla && !filter_mla(il)) { | ||
| return false; | ||
| } | ||
|
|
||
| return model.hparams.is_swa(il); | ||
| }; | ||
|
|
Collaborator
Author
There was a problem hiding this comment.
CC @fairydreaming @ggerganov , I think adding SWA to DSA cache is not too invasive, do you think it's ok to keep this change as-is?
Contributor
There was a problem hiding this comment.
@ngxson From my understanding SWA in Dots3 is for non-DSA layers (they are ordinary MLA layers that don't use lightning indexer at all), so the cleanest solution that follows other memory classes would be probably to create llama_kv_cache_dsa_iswa class that would combine llama_kv_cache_dsa instance for DSA layers and llama_kv_cache instance for SWA layers.
Collaborator
Author
There was a problem hiding this comment.
that makes sense, I moved the logic to a new llama_kv_cache_dsa_iswa, PTAL: d9de2b2
ngxson
commented
Aug 14, 2026
4 tasks
Collaborator
Author
|
I think this PR is ready to be reviewed as-is 2 follow-up PRs planned: multimodal (see xsn/dots3-note...xsn/dots3-note-mtmd) and MTP support (TODO) |
ngxson
marked this pull request as ready for review
August 17, 2026 19:46
ngxson
requested review from
CISC,
JohannesGaessler and
ggerganov
as code owners
August 17, 2026 19:46
CISC
reviewed
Aug 18, 2026
CISC
reviewed
Aug 19, 2026
CISC
approved these changes
Aug 19, 2026
fairydreaming
approved these changes
Aug 19, 2026
Collaborator
Author
|
tested the Q8_0 weight on a big mac 512GB RAM 
|
therealkenc
pushed a commit
to therealkenc/llama.cpp
that referenced
this pull request
Aug 24, 2026
* text: conversion * init impl * address review comments * fix rope * move to a new llama_kv_cache_dsa_iswa
thecodacus
pushed a commit
to thecodacus/llama.cpp
that referenced
this pull request
Sep 7, 2026
* text: conversion * init impl * address review comments * fix rope * move to a new llama_kv_cache_dsa_iswa
JenySadadia
pushed a commit
to JenySadadia/qvac-fabric-llm.cpp
that referenced
this pull request
Sep 7, 2026
Implement `seq_token_count` for the new upstream `llama_kv_cache_dsa_iswa` wrapper. Fixes: 5a32f7b ("model: add dots3-note (ggml-org#27060)") Signed-off-by: JenySadadia <jeny.sadadia@collabora.com>
JenySadadia
pushed a commit
to JenySadadia/qvac-fabric-llm.cpp
that referenced
this pull request
Sep 7, 2026
Implement `seq_token_count` for the new upstream `llama_kv_cache_dsa_iswa` wrapper. Fixes: 5a32f7b ("model: add dots3-note (ggml-org#27060)") Signed-off-by: JenySadadia <jeny.sadadia@collabora.com>
JenySadadia
pushed a commit
to JenySadadia/qvac-fabric-llm.cpp
that referenced
this pull request
Sep 7, 2026
Implement `seq_token_count` for the new upstream `llama_kv_cache_dsa_iswa` wrapper. Fixes: 5a32f7b ("model: add dots3-note (ggml-org#27060)") Signed-off-by: JenySadadia <jeny.sadadia@collabora.com>
gagallo7
pushed a commit
to gagallo7/qvac-fabric-llm.cpp
that referenced
this pull request
Sep 8, 2026
Implement `seq_token_count` for the new upstream `llama_kv_cache_dsa_iswa` wrapper. Fixes: 5a32f7b ("model: add dots3-note (ggml-org#27060)") Signed-off-by: JenySadadia <jeny.sadadia@collabora.com>
gagallo7
pushed a commit
to gagallo7/qvac-fabric-llm.cpp
that referenced
this pull request
Sep 10, 2026
Implement `seq_token_count` for the new upstream `llama_kv_cache_dsa_iswa` wrapper. Fixes: 5a32f7b ("model: add dots3-note (ggml-org#27060)") Signed-off-by: JenySadadia <jeny.sadadia@collabora.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Overview
Support https://huggingface.co/dots-studio/dots3-note-prev
Ref HF impl: huggingface/transformers#47844
So far, the biggest change is DSA + SWA, so the
llama-kv-cache-dsahad to be extendedAdditional information
This PR was tested with a slice (3 layers, 16 experts) from the real weight. Other configs and tokenizer stay unchanged
Logits validation (passed --> good/acceptable NMSE):
TODO:
llama-kv-cache-dsais okRequirements