Repository navigation
Model arch: Flint - #314
Open
danielhanchen wants to merge 2 commits into
Open
danielhanchen wants to merge 2 commits into
danielhanchen wants to merge 2 commits into
Conversation
Adds the Flint architecture: GGUF conversion (conversion/flint.py, pre-tokenizer hash), the llama.cpp model graph (src/models/flint.cpp), router and expert handling, adaptive F16 activation scaling for the expert down projection, KV cache shifting, tokenizer pre-type, and test-llama-archs / Metal fusion entries. Squash of recursive-org/llama.cpp-flint-ext#1 at 403d072bf7a9b7598300fc688fbe0c3cc005dda0, applied on ggml-org/llama.cpp 0a1ef30. Co-authored-by: Alexey <alexey@recursive.com> Co-authored-by: Tankred Saanum <62603648+tankred-saanum@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
The account paying for this security review has reached its Codex usage limits. The payer can check the Codex usage dashboard. For personal accounts, using credits requires enabling “Use credits for security reviews” in Code review settings. If you do not manage the paying account, contact this repository's admins. |
Inkling (ggml-org#25731) takes 62. The pre-type is resolved from the tokenizer.ggml.pre string at load time, so the number is not stored in GGUFs.
This was referenced Oct 21, 2026
Merged
This branch has not been deployed
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
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.
What it adds
conversion/flint.py: HF to GGUF conversion forFlintForCausalLM, with F32 router weights and the standardflintpre-tokenizer hash (also registered inconvert_hf_to_gguf_update.py).src/models/flint.cpp: the model graph (MoE with router, BF16 expert matmuls kept on CUDA).build_moe_ffn: Flint joins Mistral4 and Mellum onbuild_amax_src1_f16. The helper's target changes from 32768 to 1024, which leaves headroom for the F16 sums of 32 values in Q8 activation blocks. This also changes Mistral4 and Mellum: their expert down-projection inputs are now rescaled whenever the max absolute value exceeds 1024, instead of only above 32768.test-llama-archsplus Metal fusion count entries.Squash details
403d072bf7a9b7598300fc688fbe0c3cc005dda0(18 commits, including two merges of upstream master).0a1ef309c, the upstream commit the source branch last merged. The tree is identical to the source head except for that repository's own.github/workflowschanges, which are not part of the PR.Testing
test-llama-archspasses all 140 tests (flint OK), andtest-chatpasses.Pin set note
A second commit renumbers
LLAMA_VOCAB_PRE_TYPE_FLINTfrom 62 to 63, because Inkling (ggml-org#25731) takes 62 and both are pinned together. The pre-type is resolved from thetokenizer.ggml.prestring at load time, so the number is not stored in GGUF files and existing Flint GGUFs are unaffected.