Skip to content

Conversation

CISC
Copy link
Collaborator

@CISC CISC commented Sep 9, 2025

Change reshape_4d to view_4d, fixes non-contiguous assertion error in build_attn_mha.

Tested with jina-reranker-v2, starcoder, bloomz (and Qwen3MoE for good measure), all with and without FA.

Fixes #15895

@CISC CISC requested a review from ggerganov September 9, 2025 20:11
@CISC
Copy link
Collaborator Author

CISC commented Sep 9, 2025

Hmmm, looks like tinyllama fails... :(

ggml-ci

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
@CISC
Copy link
Collaborator Author

CISC commented Sep 10, 2025

I will run the tests again later today before merging, just in case.

@CISC CISC merged commit 6ab397e into master Sep 10, 2025
54 of 55 checks passed
@CISC CISC deleted the cisc/non-contiguous-q-build-attn-mha branch September 10, 2025 17:09
njsyw1997 pushed a commit to aizip/llama.cpp that referenced this pull request Sep 10, 2025
* support non-contiguous Q in build_attn_mha

* Update src/llama-graph.cpp

ggml-ci

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Eval bug: GGML_ASSERT(ggml_is_contiguous(a)) with Jina reranker model

2 participants