nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
ik_llama_opt/examples
History
Kawrakow c03c2d7cc6
Merge ffn_up and ffn_gate experts tensors (#1137) 
* WIP - not working

* WIP - not working

* WIP - GPT-OSS working

However, extremely stupid. The only way I could correctly repack the
up/gate experts is to copy up and gate into host buffers, repack
into another host buffer, copy back into the ffn_up_gate_exps tensor.
This is going to be very slow for giant 500 GB models.

My attempts to do this via a compute graph on the backend holding
the tensors was unsuccessful.

For GPT-OSS-20B I see ~6-7% better PP when using the original
ik_llama.cpp fused_up_gate CUDA implementation, and ~10% when
using the small batch size implementation.

Other models are not working yet on CUDA as I need to fix the
fused mul-unary implementation.

* WIP

* WIP - Qwen3-MoE (and hopefully all others) working

But when I say here and in the previous commit "working",
I mean PP is working. TG is still broken.

* WIP: TG seems to be working

* Minor

* Add command line option to merge experts up/gate

* Add merge up/gate command line parameter to llama-bench

* Turn off merge_up_gate_exps if split mode graph

It is not yet implemented

* When no bias, allow merging up/gate with tensor overrides

* Arghh, we need to increase the context size again

* Cleanup
 		2026-01-12 18:30:53 +02:00
..
baby-llama Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
batched Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
batched-bench MoE fix for R4 quants (#170) 2025-01-12 13:19:14 +02:00
batched.swift Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
benchmark `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
convert-llama2c-to-ggml `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
cvector-generator CUDA: set compute parameters via command line arguments (#910) 2025-11-07 07:11:23 +02:00
deprecation-warning Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
embedding Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
eval-callback Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
export-lora Merge vulkan code from mainline up to commit of 6/28/2025 (#563) 2025-07-02 08:49:42 +02:00
gbnf-validator Update grammar (#1023) 2025-11-30 18:45:38 +01:00
gguf Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
gguf-hash Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
gguf-split gguf-split : update (#444) 2025-05-23 08:07:42 +03:00
gritlm llama : allow pooled embeddings on any model (#7477) 2024-06-21 08:38:22 +03:00
imatrix Fix imatrix calculation for MLA models (#411) 2025-05-13 17:53:38 +03:00
infill Tool calls support from mainline (#723) 2025-09-01 08:38:49 +03:00
jeopardy `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
llama-bench Merge ffn_up and ffn_gate experts tensors (#1137) 2026-01-12 18:30:53 +02:00
llama.android Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
llama.swiftui Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
llava add dry sampler (#513) 2025-06-19 10:24:53 +03:00
lookahead add dry sampler (#513) 2025-06-19 10:24:53 +03:00
lookup add dry sampler (#513) 2025-06-19 10:24:53 +03:00
main Port mdmd from mainline + Qwen2/2.5-VL support (#798) 2025-09-27 08:45:29 +02:00
main-cmake-pkg Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
mtmd Update mtmd to improve accuracy of M-RoPE (#993) 2025-11-29 07:27:15 +01:00
parallel Tool calls support from mainline (#723) 2025-09-01 08:38:49 +03:00
passkey Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
perplexity More informative PPL readout line (#914) 2025-11-07 16:41:24 +02:00
quantize Allow quantization of ffn_gate_inp (#896) 2025-11-05 10:44:32 +02:00
quantize-stats Disable experimental code that causes issues with MSVC (#707) 2025-08-19 18:09:49 +03:00
retrieval Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
rpc Refactor chat and server file (#1062) 2025-12-15 08:27:20 +01:00
save-load-state Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
server server: improve speed of speculative decoding (#1119) 2026-01-10 08:01:22 +02:00
simple Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
speculative Support --device and --device-draft parameter (#866) 2025-10-27 18:13:28 +02:00
sweep-bench sweep-bench: be able to set TG tokens via -n (#897) 2025-11-04 14:39:30 +02:00
sycl Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
tokenize Merge mainline - Aug 12 2024 (#17) 2024-08-12 15:14:32 +02:00
CMakeLists.txt Port mdmd from mainline + Qwen2/2.5-VL support (#798) 2025-09-27 08:45:29 +02:00
Miku.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
base-translate.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
chat-13B.bat
…
chat-13B.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
chat-persistent.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
chat-vicuna.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
chat.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
convert_legacy_llama.py Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
json_schema_pydantic_example.py Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
json_schema_to_grammar.py Update grammar (#1023) 2025-11-30 18:45:38 +01:00
llama.vim llama.vim : added api key support (#5090) 2024-01-23 08:51:27 +02:00
llm.vim llm.vim : stop generation at multiple linebreaks, bind to <F2> (#2879) 2023-08-30 09:50:55 +03:00
pydantic_models_to_grammar.py Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
pydantic_models_to_grammar_examples.py Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
reason-act.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
regex_to_grammar.py Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
server-llama2-13B.sh `build`: rename main → llama-cli, server → llama-server, llava-cli → llama-llava-cli, etc... (#7809) 2024-06-13 00:41:52 +01:00
server_embd.py Merge mainline llama.cpp (#3) 2024-07-27 07:55:01 +02:00
ts-type-to-grammar.sh JSON schema conversion: ⚡️ faster repetitions, min/maxLength for strings, cap number length (#6555) 2024-04-12 19:43:38 +01:00