nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
ik_llama_opt/ggml/src
History
Kawrakow 676f0e71b4 FlashMLA-2 (CPU): faster and smaller compute buffer size (#253) 
* FlashMLA-2: eliminate intermediate f32 tensors

This works on the CPU. PP performance is ~13% better for 16k tokens
and compute buffer is quite a bit smaller.

* FlashMLA-2: enable fast path only on the CPU for now

I did implement the necessary ops on CUDA, but something is
still wrong there, so for now we only use it when running
CPU-only.

* FlashMLA-2: slightly smaller computer buffer size

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
 		2025-03-13 12:07:43 +02:00
..
ggml-cann Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-cuda MLA-2: Allow usage of q8_0 for KV cache on CUDA (#252) 2025-03-12 07:21:46 +02:00
ggml-sycl Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
iqk FlashMLA-2 (CPU): faster and smaller compute buffer size (#253) 2025-03-13 12:07:43 +02:00
kompute@4565194ed7 Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
kompute-shaders Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
llamafile Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
vulkan-shaders Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
CMakeLists.txt Better FlashMLA (#243) 2025-03-07 09:46:58 +02:00
ggml-aarch64.c Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-aarch64.h Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
ggml-alloc.c Give the user the option to override where model weights are stored (#232) 2025-02-25 17:55:58 +02:00
ggml-backend-impl.h Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
ggml-backend.c FlashMLA-2 (CPU): faster and smaller compute buffer size (#253) 2025-03-13 12:07:43 +02:00
ggml-blas.cpp Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-cann.cpp Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-common.h Use Q8_K_128 for IQ1_S_R4 and IQ1_M_R4 matrix multiplications (#194) 2025-02-09 09:14:52 +02:00
ggml-cuda.cu MLA-2: Allow usage of q8_0 for KV cache on CUDA (#252) 2025-03-12 07:21:46 +02:00
ggml-impl.h Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-kompute.cpp Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-metal.m Faster MoE inference (#112) 2024-10-31 12:05:27 +01:00
ggml-metal.metal Faster MoE inference (#112) 2024-10-31 12:05:27 +01:00
ggml-quants.c Flash MLA (CPU only) (#240) 2025-03-03 15:17:51 +02:00
ggml-quants.h IQ1_M_R4: better 1.75 bpw quants (#187) 2025-02-06 14:08:52 +02:00
ggml-rpc.cpp Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-sycl.cpp Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml-vulkan.cpp Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
ggml.c FlashMLA-2 (CPU): faster and smaller compute buffer size (#253) 2025-03-13 12:07:43 +02:00