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

ggml : add new Q4_2 quantization (ARM only) (#1046)

Browse source 
* ggml : Q4_2 ARM

* ggml : add ggml_is_quantized()

* llama : update llama_type_name() with Q4_2 entry

* ggml : speed-up q4_2

- 4 threads: ~100ms -> ~90ms
- 8 threads:  ~55ms -> ~50ms

* ggml : optimize q4_2 using vmlaq_n_f32 + vmulq_n_f32
This commit is contained in:
Georgi Gerganov 2023-04-18 23:54:57 +03:00 • committed by GitHub
parent 50a8a2af97
commit 77a73403ca
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
5 changed files with 287 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

10
llama.cpp
View file

@ -478,6 +478,7 @@ struct llama_file_loader {
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q4_2:
break;
default: {
throw format("unrecognized tensor type %u\n", shard.type);
@ -550,6 +551,7 @@ struct llama_file_saver {
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q4_2:
break;
default: LLAMA_ASSERT(false);
}
@ -838,6 +840,7 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1";
case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
return "mostly Q4_1, some F16";
case LLAMA_FTYPE_MOSTLY_Q4_2: return "mostly Q4_2";
default: return "unknown, may not work";
}
}
@ -1571,6 +1574,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
switch (ftype) {
case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break;
case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break;
case LLAMA_FTYPE_MOSTLY_Q4_2: quantized_type = GGML_TYPE_Q4_2; break;
default: throw format("invalid output file type %d\n", ftype);
};
@ -1644,6 +1648,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
{
new_size = ggml_quantize_q4_1(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
} break;
case GGML_TYPE_Q4_2:
{
new_size = ggml_quantize_q4_2(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
} break;
default:
LLAMA_ASSERT(false);
}
@ -1955,7 +1963,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
base_t = dest_t;
}
if (base_t->type == GGML_TYPE_Q4_0 || base_t->type == GGML_TYPE_Q4_1) {
if (base_t->type == GGML_TYPE_Q4_0 || base_t->type == GGML_TYPE_Q4_1 || base_t->type == GGML_TYPE_Q4_2) {
if (!warned) {
fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, "
"use a f16 or f32 base model with --lora-base\n", __func__);