CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-18 11:46:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: set_rows + cpy.cu refactor (#14712)

Browse Source
This commit is contained in:
Aman Gupta
2025-07-18 14:54:18 +08:00
committed by GitHub
parent 8f974bc1e9
commit f9a31eea06
4 changed files with 396 additions and 244 deletions
Download Patch File Download Diff File Expand all files Collapse all files

239
ggml/src/ggml-cuda/cpy.cu
View File

@@ -1,46 +1,12 @@
#include "cpy.cuh"
#include "dequantize.cuh"
#include "cpy-utils.cuh"
#ifdef GGML_USE_MUSA
#include "ggml-musa/mudnn.cuh"
#endif // GGML_USE_MUSA
typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
float * dsti = (float *) cdsti;
*dsti = *xi;
}
static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti;
*dsti = *xi;
}
static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
half * dsti = (half *) cdsti;
*dsti = __float2half(*xi);
}
static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
const half * xi = (const half *) cxi;
half * dsti = (half *) cdsti;
*dsti = *xi;
}
static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
const half * xi = (const half *) cxi;
float * dsti = (float *) cdsti;
*dsti = *xi;
}
template <cpy_kernel_t cpy_1>
static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne,
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@@ -71,29 +37,6 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const in
cpy_1(cx + x_offset, cdst + dst_offset);
}
static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
block_q8_0 * dsti = (block_q8_0 *) cdsti;
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_0; j++) {
const float v = xi[j];
amax = fmaxf(amax, fabsf(v));
}
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
dsti->d = d;
for (int j = 0; j < QK8_0; ++j) {
const float x0 = xi[j]*id;
dsti->qs[j] = roundf(x0);
}
}
static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
float * cdstf = (float *)(cdsti);
@@ -106,139 +49,6 @@ static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
}
}
static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
block_q4_0 * dsti = (block_q4_0 *) cdsti;
float amax = 0.0f;
float vmax = 0.0f;
for (int j = 0; j < QK4_0; ++j) {
const float v = xi[j];
if (amax < fabsf(v)) {
amax = fabsf(v);
vmax = v;
}
}
const float d = vmax / -8;
const float id = d ? 1.0f/d : 0.0f;
dsti->d = d;
for (int j = 0; j < QK4_0/2; ++j) {
const float x0 = xi[0 + j]*id;
const float x1 = xi[QK4_0/2 + j]*id;
const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
dsti->qs[j] = xi0;
dsti->qs[j] |= xi1 << 4;
}
}
static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
block_q4_1 * dsti = (block_q4_1 *) cdsti;
float vmin = FLT_MAX;
float vmax = -FLT_MAX;
for (int j = 0; j < QK4_1; ++j) {
const float v = xi[j];
if (v < vmin) vmin = v;
if (v > vmax) vmax = v;
}
const float d = (vmax - vmin) / ((1 << 4) - 1);
const float id = d ? 1.0f/d : 0.0f;
dsti->dm.x = d;
dsti->dm.y = vmin;
for (int j = 0; j < QK4_1/2; ++j) {
const float x0 = (xi[0 + j] - vmin)*id;
const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
dsti->qs[j] = xi0;
dsti->qs[j] |= xi1 << 4;
}
}
static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
block_q5_0 * dsti = (block_q5_0 *) cdsti;
float amax = 0.0f;
float vmax = 0.0f;
for (int j = 0; j < QK5_0; ++j) {
const float v = xi[j];
if (amax < fabsf(v)) {
amax = fabsf(v);
vmax = v;
}
}
const float d = vmax / -16;
const float id = d ? 1.0f/d : 0.0f;
dsti->d = d;
uint32_t qh = 0;
for (int j = 0; j < QK5_0/2; ++j) {
const float x0 = xi[0 + j]*id;
const float x1 = xi[QK5_0/2 + j]*id;
const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
}
memcpy(dsti->qh, &qh, sizeof(qh));
}
static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
block_q5_1 * dsti = (block_q5_1 *) cdsti;
float min = xi[0];
float max = xi[0];
for (int j = 1; j < QK5_1; ++j) {
const float v = xi[j];
min = v < min ? v : min;
max = v > max ? v : max;
}
const float d = (max - min) / 31;
const float id = d ? 1.0f/d : 0.0f;
dsti->dm.x = d;
dsti->dm.y = min;
uint32_t qh = 0;
for (int j = 0; j < QK5_1/2; ++j) {
const float x0 = (xi[0 + j] - min)*id;
const float x1 = (xi[QK5_1/2 + j] - min)*id;
const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
}
memcpy(dsti->qh, &qh, sizeof(qh));
}
template<dequantize_kernel_t dequant, int qk>
static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
float * cdstf = (float *)(cdsti);
@@ -252,53 +62,6 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
}
}
static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
if (x <= val[0]) return 0;
if (x >= val[n-1]) return n-1;
int ml = 0, mu = n-1;
while (mu-ml > 1) {
int mav = (ml+mu)/2;
if (x < val[mav]) mu = mav; else ml = mav;
}
return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
}
static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
const float * xi = (const float *) cxi;
block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
float amax = 0.0f;
float vmax = 0.0f;
for (int j = 0; j < QK4_NL; ++j) {
const float v = xi[j];
if (amax < fabsf(v)) {
amax = fabsf(v);
vmax = v;
}
}
float d = vmax / kvalues_iq4nl[0];
const float id = d ? 1.0f/d : 0.0f;
float sumqx = 0, sumq2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) {
const float x0 = xi[0 + j]*id;
const float x1 = xi[QK4_NL/2 + j]*id;
const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
dsti->qs[j] = xi0 | (xi1 << 4);
const float v0 = kvalues_iq4nl[xi0];
const float v1 = kvalues_iq4nl[xi1];
const float w0 = xi[0 + j]*xi[0 + j];
const float w1 = xi[QK4_NL/2 + j]*xi[QK4_NL/2 + j];
sumqx += w0*v0*xi[j] + w1*v1*xi[QK4_NL/2 + j];
sumq2 += w0*v0*v0 + w1*v1*v1;
}
dsti->d = sumq2 > 0 ? sumqx/sumq2 : d;
}
template <cpy_kernel_t cpy_blck, int qk>
static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,