ggml-org/llama.cpp

Pull Requests Commits

py : fix padded dummy tokens (I hope)

ggerganov committed 2 years ago

Verified 23742deb

py : fix BPE vocab conversion

ggerganov committed 2 years ago

Verified d92351e2

py : pad with unknown tokens when data is missing

ggerganov committed 2 years ago

Verified a1372737

py : fix missing added_tokens_dict for SPM vocab

ggerganov committed 2 years ago

Verified 9b464b4e

ggml : introduce GGML_CALL function annotation (#4850)

jart committed 2 years ago

Verified a0b3ac8c

finetune : use LLAMA_FILE_MAGIC_GGLA (#4961)

danbev committed 2 years ago

Verified d75c232e

speculative : threading options (#4959)

stduhpf committed 2 years ago

Verified e0324285

pass cpu-architecture arguments only to host code (C;C++) (#4943)

ngc92 committed 2 years ago

Verified 3e5ca793

llama : apply classifier-free guidance to logits directly (#4951)

dfriehs committed 2 years ago

Verified 44833967

awq-py : fix typo in awq-py/README.md (#4947)

ziliangpeng committed 2 years ago

Verified d9aa4ffa

cuda : fix dequantize kernel names (#4938)

ggerganov committed 2 years ago

Verified ddb008d8

llama : check for 256 divisibility for IQ2_XS, IQ2_XXS (#4950)

ikawrakow committed 2 years ago

Verified 2faaef39

CUDA: faster dequantize kernels for Q4_0 and Q4_1 (#4938)

ikawrakow committed 2 years ago

Verified 4a3156de

llama : fix missing quotes (#4937)

David Pflug committed 2 years ago

Verified a836c8f5

Add ability to use importance matrix for all k-quants (#4930)

ikawrakow committed 2 years ago

Verified 467a882f

llama : check LLAMA_TRACE env for extra logging (#4929)

ggerganov committed 2 years ago

Verified bb0c1392

scripts : sync-ggml-am.sh option to skip commits

ggerganov committed 2 years ago

Verified 9408cfda

llama : use LLAMA_LOG_ macros for logging

ggerganov committed 2 years ago

Verified 03c52674

Fix ffn_down quantization mix for MoE models (#4927)

ikawrakow committed 2 years ago

Verified a128c38d

metal : correctly set SIMD support flags on iOS (#4923)

azarovalex committed 2 years ago

Verified 5f5fe1bd

llama : support WinXP build with MinGW 8.1.0 (#3419)

guilt committed 2 years ago

Verified ac32902a

2-bit quantizations (#4897)

ikawrakow committed 2 years ago

Verified 147b17ac

Make Q3_K_S be the same as olf Q3_K_L for Mixtral-8x7B (#4906)

ikawrakow committed 2 years ago

Verified 807179ec

ggerganov committed 2 years ago

Verified 76484fbf

ggml: cache sin/cos for RoPE (#4908)

JohannesGaessler committed 2 years ago

Verified c71d608c

metal : remove old API (#4919)

ggerganov committed 2 years ago

Verified 4be5ef55

server : fix prompt caching with system prompt (#4914)

ggerganov committed 2 years ago

Verified 0ea069b8

llama : fix detokenization of non-special added-tokens (#4916)

ggerganov committed 2 years ago

Verified f172de03

metal : disable log for loaded kernels (#4794)

ggerganov committed 2 years ago

Verified 2d57de52

llama : minimize size used for state save/load (#4820)

dfriehs committed 2 years ago

Verified df845cc9

Older