vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

iq1_m: CUDA dequantize works

Browse source 
Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B.
This commit is contained in:
Iwan Kawrakow 2024-03-22 19:46:35 +02:00
parent ac8b3dd2eb
commit 1df37b654b
1 changed files with 3 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml-cuda.cu
View file

@ -615,6 +615,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@ -643,6 +644,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@ -2560,7 +2562,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
ggml_type a_type = a->type;
if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S ||
a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) {
a_type == GGML_TYPE_IQ1_M || a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) {
if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
return false;
}