ggerganov/llama.cpp

Pull Requests Commits

ggerganov committed 1 year ago

Verified 06c2d0d1

metal : parallel reduce across heads

ggerganov committed 1 year ago

Verified 17720fad

metal : parallelize across KV size

ggerganov committed 1 year ago

Verified 77d08f32

wip : template for rows per warp

ggerganov committed 1 year ago

Verified a4b6341c

wip : 4 rows per simd group

ggerganov committed 1 year ago

Verified f31955f5

wip : 8 rows per simd group

ggerganov committed 1 year ago

Verified 8cde449b

metal : specialize for head size

ggerganov committed 1 year ago

Verified b9732580

metal : reduce branches

ggerganov committed 1 year ago

Verified 52ae0857

metal : f16 precision

ggerganov committed 1 year ago

Verified 528da751

metal : initial implementation

ggerganov committed 1 year ago

Verified 1173f49c

ggml : online attention (CPU)

ggerganov committed 1 year ago

Verified a9681feb

Merge branch 'master' into gg/flash-attn

ggerganov committed 1 year ago

Verified c3cdfffa

cuda : fix compile error in jetson platform (#4975)

KyL0N committed 1 year ago

Verified cca894f1

finetune : fix ggml_allocr lifetimes (tmp workaround) (#5033)

bzuzo committed 1 year ago

Verified 381ee195

ggml : fix GQA support in ggml_flash_attn_ext

ggerganov committed 1 year ago

fa7ebcca

imatrix : add README.md

ggerganov committed 1 year ago

Verified a5cacb22

llama : support upcoming Qwen2 (#5037)

simonJJJ committed 1 year ago

Verified 9b75cb2b

py : fix flake8 lint

ggerganov committed 1 year ago

de9a147d

winogrande: evaluate log-probs in parallel (#5036)

ikawrakow committed 1 year ago

Verified 7051aacf

llama : add CodeShell support (#5016)

chiranko committed 1 year ago

Verified 2b3b999c

perplexity: avoid unnecessary alloocations and logit copies (#5035)

ikawrakow committed 1 year ago

Verified 993fba81

perplexity : faster Winogrande via batching (#5024)

ggerganov committed 1 year ago

Verified 8b20858e

llama : fix falcon arch for tied output embeddings (#4978)

cmp-nct committed 1 year ago

Verified 57e2a7a5

cmake : add ggml public headers (#5011)

ggerganov committed 1 year ago

Verified 9b6ea426

server : defer tasks when "slot unavailable" (#5018)

ngxson committed 1 year ago

Verified 821f0a27

llama : fix mlock with no-mmap with Metal (#5025)

slaren committed 1 year ago

Verified 96d7f56d

imatrix : fix assert for src0 non-cont check

ggerganov committed 1 year ago

Verified 2d5419d0

perplexity : fix winogrande N tasks option

ggerganov committed 1 year ago

Verified d391ae9b

scripts : add get-winogrande.sh

ggerganov committed 1 year ago

Verified e9240cdf

convert.py : fix llama/llama2 conversion due to vocab_size=-1 (#5019)

databyte committed 1 year ago

Verified b4675773

Older