kleidiai: add support for get_rows #14676
Conversation
github-actions
bot
added
the
ggml
| static inline float compute_fp16_to_fp32(ggml_fp16_t h) { | ||
| static_assert(sizeof(ggml_fp16_t) == sizeof(__fp16), "ggml_fp16_t and __fp16 must be the same size"); | ||
| __fp16 tmp; | ||
| memcpy(&tmp, &h, sizeof(ggml_fp16_t)); | ||
| return (float)tmp; | ||
| } |
There was a problem hiding this comment.
Can't we use ggml_fp16_to_fp32() instead introducing this function?
There was a problem hiding this comment.
Yes, good point — I'll update the patch to use ggml_fp16_to_fp32() instead.
There was a problem hiding this comment.
Since this is in the CPU backend, it could also use the potentially more efficient ggml_cpu_fp16_to_fp32.
There was a problem hiding this comment.
General advice is to try to keep the implementation more generic - it seems to focus a lot on Q4_0. Adding more asserts for the current underlying assumptions will help long term in case we add support for other types.
Another important thing we should improve soon is to add support for testing extra buffer types in test-backend-ops (see ggml-org/whisper.cpp#3223 (comment)). Without such tests it is very difficult to verify that these changes do not break something.
|
I've updated the patch to address all review comments. However, I noticed that three CI tests are currently failing due to what appear to be unrelated infrastructure issues. |
|
The build failures are unrelated. |
|
All the review comments have been addressed—just checking if there’s anything else needed from my side. Thanks! |
This patch adds support for KleidiAI acceleration of the Q4_0 matrix multiplication operation in cases where the weight tensor is shared with the get_rows operator. A typical use case is in whisper.cpp, where such weight sharing occurs between get_rows and matmul.