haraldwolff/llama.cpp
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity
llama.cpp/examples
History
Kawrakow f4d7e54974
SOTA 3-bit quants (#5196) 
* iq3_xxs: quantize/dequantize

RMSE seems a bit high-ish at about half-way between q2_K and
q3_K, so need to check more.

* iq3_xxs: CUDA dequantize works

* iq2_xxs: tuning quantization

* iq3_xxs: starting to look better

PPL on wiki.test.raw
LLaMA-v1-7B: 6.4218
LLaMA-v2-7B: 6.3560
Mistral-7B : 6.0717

This is better than Q3_K_XS, with a 5% reduction in quantized model
size.

* iq3_xxs: CUDA dot product

We have
PP-512: 5891 t/s
TG-128: 143.9 t/s

* iq3_xxs: scalar and AVX2 dot products

* iq3_xxs: ARM_NEON and Metal

Metal performance is decent, ARM_NEON is pathetic

* iq3_xxs: slightly better grid points

* Faster iq3_xxs and iq2_xs dot products on CUDA

* iq3_xxs: add some quant mix

* iq3_xxs: fix failing quantization test

Dot product still fails. Is this real?

* iq3_xxs: hopefully fix ROCm

* iq3_xxs: failing tests

This time the dot product accuracy did find an actual bug
in the AVX2 implementation.

* Add IQ3_XXS to test-backend-ops

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-01-30 15:14:12 +02:00
..
baby-llama ggml : change ggml_scale to take a float instead of tensor (#4573) 2023-12-21 23:20:49 +02:00
batched examples : add passkey test (#3856) 2024-01-08 11:14:04 +02:00
batched-bench llama : ggml-backend integration (#4766) 2024-01-12 20:07:38 +01:00
batched.swift swift : fix prompt tokenization logic (#4321) 2023-12-04 15:43:45 +02:00
beam-search llama : remove token functions with context args in favor of model (#3720) 2023-10-23 22:40:03 +03:00
benchmark 2-bit quantizations (#4897) 2024-01-14 09:45:56 +02:00
convert-llama2c-to-ggml ggml : remove n_dims from ggml_tensor (#4469) 2023-12-14 16:52:08 +01:00
embedding build : link against build info instead of compiling against it (#3879) 2023-11-02 08:50:16 +02:00
export-lora export-lora : use LLAMA_FILE_MAGIC_GGLA (#4894) 2024-01-12 19:54:53 +02:00
finetune finetune : print sample-start/include-sample-start (#5072) 2024-01-22 13:11:01 +02:00
gguf gguf : simplify example dependencies 2023-12-21 23:08:14 +02:00
imatrix imatrix : keep intermediate imatrix results (#5077) 2024-01-22 14:18:43 +02:00
infill Remove unused data and add fixes (#5154) 2024-01-27 15:25:55 +01:00
jeopardy parallel : add option to load external prompt file (#3416) 2023-10-06 16:16:38 +03:00
llama-bench ggml : add Vulkan backend (#2059) 2024-01-28 19:03:59 +02:00
llama.android android : use release cmake build type by default (#5123) 2024-01-25 19:05:51 +02:00
llama.swiftui llama.swiftui : update models layout (#4826) 2024-01-12 14:48:00 +02:00
llava llava : support for Yi-VL and fix for mobileVLM (#5093) 2024-01-27 17:09:18 +02:00
lookahead english : use typos to fix comments and logs (#4354) 2023-12-12 11:53:36 +02:00
lookup lookup : add prompt lookup decoding example (#4484) 2023-12-22 18:05:56 +02:00
main main : allow empty --prompt-cache file (#5176) 2024-01-30 11:18:02 +02:00
main-cmake-pkg main-cmake-pkg : fix build issue (#4665) 2023-12-29 16:18:20 +02:00
parallel llama : KV cache view API + better KV cache management (#4170) 2023-11-23 19:07:56 +02:00
passkey examples : add passkey test (#3856) 2024-01-08 11:14:04 +02:00
perplexity Additional KL-divergence statistics (#5081) 2024-01-23 15:17:20 +02:00
quantize SOTA 3-bit quants (#5196) 2024-01-30 15:14:12 +02:00
quantize-stats SOTA 3-bit quants (#5196) 2024-01-30 15:14:12 +02:00
save-load-state llama : minimize size used for state save/load (#4820) 2024-01-13 18:29:43 +02:00
server server : improve README (#5209) 2024-01-30 11:11:46 +02:00
simple simple : update error message for KV cache check (#4324) 2023-12-04 18:04:21 +02:00
speculative speculative : threading options (#4959) 2024-01-16 13:04:32 +02:00
sycl ggml : add unified SYCL backend for Intel GPUs (#2690) 2024-01-28 17:56:23 +02:00
tokenize tokenize example: Respect normal add BOS token behavior (#4126) 2023-11-18 14:48:17 -07:00
train-text-from-scratch finetune : fix ggml_allocr lifetimes (tmp workaround) (#5033) 2024-01-19 20:20:50 +02:00
alpaca.sh alpaca.sh : update model file name (#2074) 2023-07-06 19:17:50 +03:00
base-translate.sh examples : improve base-translate.sh script (#4783) 2024-01-06 11:40:24 +02:00
chat-13B.bat Create chat-13B.bat (#592) 2023-03-29 20:21:09 +03:00
chat-13B.sh examples : read chat prompts from a template file (#1196) 2023-05-03 20:58:11 +03:00
chat-persistent.sh llama : fix session saving/loading (#3400) 2023-10-03 21:04:01 +03:00
chat-vicuna.sh examples : add chat-vicuna.sh (#1854) 2023-06-15 21:05:53 +03:00
chat.sh main : log file (#2748) 2023-08-30 09:29:32 +03:00
CMakeLists.txt ggml : add unified SYCL backend for Intel GPUs (#2690) 2024-01-28 17:56:23 +02:00
gpt4all.sh examples : add -n to alpaca and gpt4all scripts (#706) 2023-04-13 16:03:39 +03:00
json-schema-to-grammar.py chmod : make scripts executable (#2675) 2023-08-23 17:29:09 +03:00
llama.vim llama.vim : added api key support (#5090) 2024-01-23 08:51:27 +02:00
llama2-13b.sh gitignore : changes for Poetry users + chat examples (#2284) 2023-07-21 13:53:27 +03:00
llama2.sh gitignore : changes for Poetry users + chat examples (#2284) 2023-07-21 13:53:27 +03:00
llm.vim llm.vim : stop generation at multiple linebreaks, bind to <F2> (#2879) 2023-08-30 09:50:55 +03:00
make-ggml.py make-ggml.py : compatibility with more models and GGUF (#3290) 2023-09-27 19:25:12 +03:00
Miku.sh MIKU MAYHEM: Upgrading the Default Model for Maximum Fun 🎉 (#2287) 2023-07-21 11:13:18 +03:00
pydantic-models-to-grammar-examples.py examples : make pydantic scripts pass mypy and support py3.8 (#5099) 2024-01-25 14:51:24 -05:00
pydantic_models_to_grammar.py examples : make pydantic scripts pass mypy and support py3.8 (#5099) 2024-01-25 14:51:24 -05:00
reason-act.sh chmod : make scripts executable (#2675) 2023-08-23 17:29:09 +03:00
server-llama2-13B.sh chmod : make scripts executable (#2675) 2023-08-23 17:29:09 +03:00