ggml-org/llama.cpp

Pull Requests Commits

metal : switch to execution barriers + fix one of the barriers

ggerganov committed 2 years ago

Verified e1241d9b

metal : limit kernels to not use more than the allowed threads

ggerganov committed 2 years ago

Verified 109e7aa8

metal : fix soft_max kernels

ggerganov committed 2 years ago

Verified ab558ac2

convert-hf : support for mixtral-instruct (#4428)

Mrkvak committed 2 years ago

Verified 82e4f645

ggml : do not use BLAS with ggml_mul_mat_id

ggerganov committed 2 years ago

Verified 90c12e6b

test-backend-ops : add one more sum_rows test

ggerganov committed 2 years ago

Verified ea4402bb

metal : fix binary ops for ne10 % 4 != 0

ggerganov committed 2 years ago

Verified a51bc0c1

metal : add cpy f16 -> f32 kernel

ggerganov committed 2 years ago

Verified 08eb9917

gguf-py : bump version

slaren committed 2 years ago

a742d9f9

convert : support safetensors format

ggerganov committed 2 years ago

Verified 6a419f4d

convert : fix style

slaren committed 2 years ago

f1cbfabd

convert : use 1e6 rope_freq_base for mixtral

slaren committed 2 years ago

7dc75e39

cuda : fix mul_mat_id with multi gpu

slaren committed 2 years ago

296c945d

test-backend-ops : disable MOE test with thread sanitizer

slaren committed 2 years ago

33e50f1b

test-backend-ops : simplify and disable slow tests to avoid CI timeout

slaren committed 2 years ago

ffda94c8

llama : fix hard-coded number of experts

ggerganov committed 2 years ago

Verified 8cbaed1d

test-backend-ops : fix dequantize block offset

slaren committed 2 years ago

b0029815

test-backend-ops : add cpy from f32 -> all types test

slaren committed 2 years ago

f1380d78

test-backend-ops : cleanup, add moe test for batches

slaren committed 2 years ago

54d254bb

test-backend-ops : make experts more evenly probable (test_moe)

ggerganov committed 2 years ago

Verified 54ba2634

metal : fix ggml_mul_mat_id for F32

ggerganov committed 2 years ago

Verified b0b83dd9

convert : determine n_ctx correctly

ggerganov committed 2 years ago

Verified 65923a8e

cuda : fix get_rows when ncols is odd

slaren committed 2 years ago

8614aa73

test-backend-ops : add moe test

slaren committed 2 years ago

cefebb36

llama : add n_expert and n_expert_used to hparams + change quants

ggerganov committed 2 years ago

Verified e640cbe0

llama : do not quantize expert gating tensors

ggerganov committed 2 years ago

Verified d1259b7b

metal : add indirect mat-vec kernels for all quantization types

ggerganov committed 2 years ago

Verified 6cfb31f9

metal : fix ggml_get_rows to work with non-cont src1

ggerganov committed 2 years ago

Verified 016f9bb5

llama : offload missing ffn_moe_silu

slaren committed 2 years ago

0710b0f7

cuda : support non-contiguous src1 in get_rows

slaren committed 2 years ago

62b95f93

Older