Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml : update ggml_prec specification (#26675)

Browse Source 
* ggml : update ggml_prec specification

[no ci]

* cont : add GGML_PREC_BF16

* cont : rework API

* cont : use new API

* cont : swap arg order

* cont : support for MUL_MAT_ID

* cont : fix accidental remove of "break;"

* cont : return bools, add doc TAG_GGML_PREC, clean-up

* cont : add search tag

* cont : ws
This commit is contained in:
Georgi Gerganov
2026-09-08 09:06:24 +03:00
committed by GitHub
parent 9dcf84e5ae
commit 5a6caa05fc
10 changed files with 134 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/include/ggml.h
+56 -7
View File
@@ -433,10 +433,21 @@ extern "C" {
GGML_TYPE_COUNT = 43,
};
// precision
// [TAG_GGML_PREC]
// this enum is used to declare the allowed numerical precision/data-types types that can be used during the compute of an op
// the declared types can be:
// - result accumulation type
// - source tensor data representation type
// - etc.
// the precision parameters are stored as ggml_tensor.op_params to the respective ops
enum ggml_prec {
GGML_PREC_DEFAULT = 0, // stored as ggml_tensor.op_params, 0 by default
GGML_PREC_F32 = 10,
GGML_PREC_UNDEFINED = 0,
GGML_PREC_DEFAULT = 0, // note: deprecated, use GGML_PREC_UNDEFINED
GGML_PREC_F32 = 10,
GGML_PREC_BF16 = 15,
GGML_PREC_F16 = 20,
GGML_PREC_Q8 = 30,
GGML_PREC_Q4 = 40,
};
// op hint
@@ -1429,6 +1440,42 @@ extern "C" {
struct ggml_tensor * b,
float eps);
// [TAG_GGML_PREC]
// set the minimum required accumulator type for the implementation to use during the compute
// for example:
// - GGML_PREC_F32 - requires accumulation of the results in F32
// - GGML_PREC_BF16 - can accumulate the results in BF16, F32
// - GGML_PREC_F16 - can accumulate the results in F16, F32
// - GGML_PREC_Q8 - not allowed
// - GGML_PREC_Q4 - not allowed
//
// return false on faliure
GGML_API bool ggml_prec_set_acc(
struct ggml_tensor * a,
enum ggml_prec prec);
// [TAG_GGML_PREC]
// set the smallest rank that the implementation can use to internally convert the src[idx] data to
// ranks in decreasing order:
// - GGML_PREC_F32 - GGML_TYPE_F32
// - GGML_PREC_BF16 - GGML_TYPE_BF16
// - GGML_PREC_F16 - GGML_TYPE_F16,
// - GGML_PREC_Q8 - GGML_TYPE_Q8_0, GGML_TYPE_Q8_1, GGML_TYPE_Q8_K, etc.
// - GGML_PREC_Q4 - GGML_TYPE_Q4_0, GGML_TYPE_Q4_1, GGML_TYPE_Q4_K, GGML_TYPE_NVFP4, GGML_TYPE_MXFP4, etc.
//
// for example:
// - ggml_prec_set_src(a, GGML_PREC_Q8, 1):
// - allows the implementation to quantize F32, BF16, F16 data of src[1] down to GGML_TYPE_Q8_0
// - cannot quantize it down to GGML_TYPE_Q4_0 or GGML_TYPE_NVFP4
// - ggml_prec_set_src(a, GGML_PREC_Q4, 1):
// - allows the implementation to quantize F32, BF16, F16 data of src[1] down to 4-bit datatypes such as GGML_TYPE_Q4_K, GGML_TYPE_NVFP4 etc.
//
// return false on faliure
GGML_API bool ggml_prec_set_src(
struct ggml_tensor * a,
enum ggml_prec prec,
int idx);
// A: k columns, n rows => [ne03, ne02, n, k]
// B: k columns, m rows (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
// result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
@@ -1439,9 +1486,10 @@ extern "C" {
// change the precision of a matrix multiplication
// set to GGML_PREC_F32 for higher precision (useful for phi-2)
GGML_API void ggml_mul_mat_set_prec(
GGML_DEPRECATED(GGML_API void ggml_mul_mat_set_prec(
struct ggml_tensor * a,
enum ggml_prec prec);
enum ggml_prec prec),
"use ggml_prec_set_acc() instead");
// change the hint of a matrix multiplication
GGML_API void ggml_mul_mat_set_hint(
@@ -2446,9 +2494,10 @@ extern "C" {
float max_bias,
float logit_softcap);
GGML_API void ggml_flash_attn_ext_set_prec(
GGML_DEPRECATED(GGML_API void ggml_flash_attn_ext_set_prec(
struct ggml_tensor * a,
enum ggml_prec prec);
enum ggml_prec prec),
"use ggml_prec_set_acc() instead");
GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
const struct ggml_tensor * a);
ggml/src/ggml-impl.h
+12
View File
@@ -160,6 +160,18 @@ static float ggml_get_op_params_f32(const struct ggml_tensor * tensor, uint32_t
return ((const float *)(tensor->op_params))[i];
}
// [TAG_GGML_PREC]
// - GGML_OP_MUL_MAT
// 0 - acc
// 1 - hint
// 2 - src0 precision
// 3 - src1 precision
//
// - GGML_OP_MUL_MAT_ID
// 0 - acc
// 1 - hint
// 2 - src0 precision
// 3 - src1 precision
static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
((int32_t *)(tensor->op_params))[i] = value;
ggml/src/ggml.c
+51
View File
@@ -3277,6 +3277,57 @@ struct ggml_tensor * ggml_l2_norm_inplace(
return ggml_l2_norm_impl(ctx, a, eps, true);
}
// ggml_prec
bool ggml_prec_set_acc(
struct ggml_tensor * a,
enum ggml_prec prec) {
switch (a->op) {
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
{
const int32_t prec_i32 = (int32_t) prec;
ggml_set_op_params_i32(a, 0, prec_i32);
}
break;
case GGML_OP_FLASH_ATTN_EXT:
{
const int32_t prec_i32 = (int32_t) prec;
ggml_set_op_params_i32(a, 3, prec_i32);
}
break;
default:
return false;
};
return true;
}
bool ggml_prec_set_src(
struct ggml_tensor * a,
enum ggml_prec prec,
int idx) {
GGML_ASSERT(idx >= 0 && idx < GGML_MAX_SRC);
switch (a->op) {
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
{
if (idx != 1) {
return false;
}
const int32_t prec_i32 = (int32_t) prec;
ggml_set_op_params_i32(a, 2 + idx, prec_i32);
}
break;
default:
return false;
};
return true;
}
// ggml_mul_mat
static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {