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

ggml : add AArch64 optimized GEMV and GEMM Q4 kernels (#5780)

Browse source 
* Arm AArch64: optimized GEMV and GEMM kernels for q4_0_q8_0, and q8_0_q8_0 quantization

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add copyright claim only to ggml-aarch64.cpp and ggml-aarch64.h files

* Arm AArch64: minor code refactoring for rebase

* Arm AArch64: minor code refactoring for resolving a build issue with cmake

* Arm AArch64: minor code refactoring to split the Q4_0_AARC64 type into three separate types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: minor code change for resolving a build issue with server-windows

* retrigger checks

* Arm AArch64: minor code changes for rebase

* Arm AArch64: minor changes to skip the pr#7433 vec_dot code for arm cpus with SVE VL not equal to 256 bits

* Arm AArch64: remove stale LLAMA_QKK_64 from CMakeLists.txt and delete build.zig

* Arm AArch64: add reference scalar gemm and gemv, and avoid dynamic memory allocations during quantization for Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: add multithreaded quantization support for the new types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: minor code refactoring

* Arm AArch64: simplify logic for calling gemm and gemv functions in ggml_compute_forward_mul_mat

* Arm AArch64: minimize changes in ggml_compute_forward_mul_mat

* Arm AArch64: minor code refactoring, and add reference scalar code to quantize routines for new quant types

* Arm AArch64: minor code refactoring

* Arm AArch64: minor code refactoring

* Arm AArch64: minor code refactoring

* rebase on the latest master commit 3fd62a6 and adapt to the new directory structure

* Arm AArch64: remove a redundant comment

* Arm AArch64: add pragma in ggml-aarch64.c to turn -Woverlength-strings warning off

* Arm AArch64: use __aarch64__ check to guard 64-bit neon kernels

* Arm AArch64: update docs/build.md README to include compile time flags for buiilding the Q4_0_4_4 quant type
This commit is contained in:
Dibakar Gope 2024-07-10 07:14:51 -05:00 • committed by GitHub
parent 83321c6958
commit 0f1a39f343
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
14 changed files with 2534 additions and 53 deletions
Download patch file Download diff file Expand all files Collapse all files

24
src/llama.cpp
View file

@ -3788,6 +3788,9 @@ struct llama_model_loader {
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break;
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break;
case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break;
case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@ -4481,6 +4484,9 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
default: return "unknown, may not work";
}
@ -17768,6 +17774,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_S;
}
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
new_type == GGML_TYPE_Q4_0_8_8) {
new_type = GGML_TYPE_Q4_0;
}
}
} else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
@ -18080,6 +18090,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
@ -18390,6 +18403,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
f32_data = (float *) f32_conv_buf.data();
}
int chunk_size_multiplier = 1;
if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
}
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);
@ -18402,7 +18423,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const int64_t nrows = tensor->ne[1];
static const int64_t min_chunk_size = 32 * 512;
const int64_t chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row);
const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
chunk_size_multiplier;
const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;