llama.cpp/ggml-cuda.cu

958 lines
35 KiB
Text
Raw Normal View History

#include <cstddef>
#include <cstdint>
#include <stdint.h>
#include <stdio.h>
#include <atomic>
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <cuda_fp16.h>
#include "ggml-cuda.h"
#include "ggml.h"
static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
#define CUDA_CHECK(err) \
do { \
cudaError_t err_ = (err); \
if (err_ != cudaSuccess) { \
fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \
cudaGetErrorString(err_)); \
exit(1); \
} \
} while (0)
#define CUBLAS_CHECK(err) \
do { \
cublasStatus_t err_ = (err); \
if (err_ != CUBLAS_STATUS_SUCCESS) { \
fprintf(stderr, "cuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, float & v0, float & v1);
typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream);
// QK = number of values after dequantization
// QR = QK / number of values before dequantization
#define QK4_0 32
#define QR4_0 2
typedef struct {
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
#define QR4_1 2
typedef struct {
half d; // delta
half m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
#define QR5_0 2
typedef struct {
half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
#define QR5_1 2
typedef struct {
half d; // delta
half m; // min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
#define QR8_0 1
typedef struct {
half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define WARP_SIZE 32
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
#define CUDA_MUL_BLOCK_SIZE 256
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
// dmmv = dequantize_mul_mat_vec
#ifndef GGML_CUDA_DMMV_X
#define GGML_CUDA_DMMV_X 32
#endif
#ifndef GGML_CUDA_DMMV_Y
#define GGML_CUDA_DMMV_Y 1
#endif
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= kx) {
return;
}
dst[i] = x[i] * y[i%ky];
}
static __device__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
const block_q4_0 * x = (const block_q4_0 *) vx;
const float d = x[ib].d;
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
v0 = (vi0 - 8)*d;
v1 = (vi1 - 8)*d;
}
static __device__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
const block_q4_1 * x = (const block_q4_1 *) vx;
const float d = x[ib].d;
const float m = x[ib].m;
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
v0 = vi0*d + m;
v1 = vi1*d + m;
}
static __device__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
const block_q5_0 * x = (const block_q5_0 *) vx;
const float d = x[ib].d;
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1) - 16;
v0 = x0*d;
v1 = x1*d;
}
static __device__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
const block_q5_1 * x = (const block_q5_1 *) vx;
const float d = x[ib].d;
const float m = x[ib].m;
uint32_t qh;
memcpy(&qh, x[ib].qh, sizeof(qh));
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1);
v0 = x0*d + m;
v1 = x1*d + m;
}
static __device__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
const block_q8_0 * x = (const block_q8_0 *) vx;
const float d = x[ib].d;
const int8_t vi0 = x[ib].qs[iqs + 0];
const int8_t vi1 = x[ib].qs[iqs + 1];
v0 = vi0*d;
v1 = vi1*d;
}
static __device__ void convert_f16(const void * vx, const int ib, const int iqs, float & v0, float & v1){
const half * x = (const half *) vx;
v0 = __half2float(x[ib + 0]);
v1 = __half2float(x[ib + 1]);
}
template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
static __global__ void dequantize_block(const void * vx, float * y, const int k) {
const int i = blockDim.x*blockIdx.x + 2*threadIdx.x;
ggml : remove bit shuffling (#1405) * ggml : remove Q4_0 bit shufling (ARM NEON) * ggml : remove Q4_1 bit shuffling (ARM NEON + reference) * ggml : nibbles_from_floats() + bytes_from_nibbles() (ARM NEON) * ggml : remove Q4_2 bit shuffling (WIP, BROKEN) * ggml : remove Q5_0 bit shuffling (ARM NEON) * ggml : 2x faster scalar implementations * ggml : remove Q5_1 bit shuffling (ARM NEON + scalar) * ggml : simplify scalar dot * ggml : remove WASM SIMD bit shuffling + remove vzip for ARM 32-bit * ggml : fix Q4_1 quantization * ggml : update cuBLAS + normalize variable names * ggml : remove Q4_2 mode * ggml : minor formatting * ggml : fix Q5_0 quantization * scripts : add script for measuring the time per token * AVX implementations (#1370) * ggml : uniform 5th bit extraction * llama : produce error upon loading old model files * llama : fix model magic/version write * ggml : speed-up Q5_0 + Q5_1 at 4 threads * ggml : preserve old Q4 and Q5 formats * ggml : simplify Q8_1 - no need for low / high sums anymore * ggml : fix Q8_0 and Q8_1 rounding * Revert "AVX implementations (#1370)" This reverts commit 948d124837f9d287d8490f41338e0e4cceb0814f. * ggml : fix AVX2 implementation * sha : update hashes for 7B and 13B * readme : update timings + remove warning banner * llama : update v2 PR number to 1405 * ggml : fix WASM comments * ggml : back to original bit order * readme : add note that Q4 and Q5 have been changed * llama : fix return for unknown version --------- Co-authored-by: Stephan Walter <stephan@walter.name>
2023-05-11 21:23:08 +00:00
if (i >= k) {
return;
}
const int ib = i/qk; // block index
const int iqs = (i%qk)/qr; // quant index
const int iybs = i - i%qk; // y block start index
const int y_offset = qr == 1 ? 1 : qk/2;
// dequantize
float & v0 = y[iybs + iqs + 0];
float & v1 = y[iybs + iqs + y_offset];
dequantize_kernel(vx, ib, iqs, v0, v1);
}
template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y, float * dst, const int ncols) {
// qk = quantized weights per x block
// qr = number of quantized weights per data value in x block
const int row = blockIdx.x*blockDim.y + threadIdx.y;
const int tid = threadIdx.x;
const int iter_stride = 2*GGML_CUDA_DMMV_X;
const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
const int y_offset = qr == 1 ? 1 : qk/2;
float tmp = 0; // partial sum for thread in warp
for (int i = 0; i < ncols; i += iter_stride) {
const int col = i + vals_per_iter*tid;
const int ib = (row*ncols + col)/qk; // x block index
const int iqs = (col%qk)/qr; // x quant index
const int iybs = col - col%qk; // y block start index
// processing >2 values per i iter is faster for fast GPUs
#pragma unroll
for (int j = 0; j < vals_per_iter; j += 2) {
// process 2 vals per j iter
// dequantize
float v0, v1;
dequantize_kernel(vx, ib, iqs + j/qr, v0, v1);
// for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
// matrix multiplication
tmp += v0 * y[iybs + iqs + j/qr + 0];
tmp += v1 * y[iybs + iqs + j/qr + y_offset];
// for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
}
}
// sum up partial sums and write back result
__syncthreads();
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
}
if (tid == 0) {
dst[row] = tmp;
}
}
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE;
mul_f32<<<num_blocks, CUDA_MUL_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
}
static void dequantize_row_q4_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
dequantize_block<QK4_0, QR4_0, dequantize_q4_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void dequantize_row_q4_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
dequantize_block<QK4_1, QR4_1, dequantize_q4_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void dequantize_row_q5_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
dequantize_block<QK5_0, QR5_0, dequantize_q5_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void dequantize_row_q5_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
dequantize_block<QK5_1, QR5_1, dequantize_q5_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void dequantize_row_q8_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
dequantize_block<QK8_0, QR8_0, dequantize_q8_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
<<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
<<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
<<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
<<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
<<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
dequantize_block<32, 1, convert_f16><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void convert_mul_mat_vec_f16_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
const dim3 block_dims(WARP_SIZE, GGML_CUDA_DMMV_Y, 1);
dequantize_mul_mat_vec<1, 1, convert_f16>
<<<nrows/GGML_CUDA_DMMV_Y, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return dequantize_row_q4_0_cuda;
case GGML_TYPE_Q4_1:
return dequantize_row_q4_1_cuda;
case GGML_TYPE_Q5_0:
return dequantize_row_q5_0_cuda;
case GGML_TYPE_Q5_1:
return dequantize_row_q5_1_cuda;
case GGML_TYPE_Q8_0:
return dequantize_row_q8_0_cuda;
case GGML_TYPE_F16:
return convert_fp16_to_fp32_cuda;
default:
return nullptr;
}
}
static dequantize_mul_mat_vec_cuda_t ggml_get_dequantize_mul_mat_vec_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return dequantize_mul_mat_vec_q4_0_cuda;
case GGML_TYPE_Q4_1:
return dequantize_mul_mat_vec_q4_1_cuda;
case GGML_TYPE_Q5_0:
return dequantize_mul_mat_vec_q5_0_cuda;
case GGML_TYPE_Q5_1:
return dequantize_mul_mat_vec_q5_1_cuda;
case GGML_TYPE_Q8_0:
return dequantize_mul_mat_vec_q8_0_cuda;
case GGML_TYPE_F16:
2023-05-13 14:40:58 +00:00
return convert_mul_mat_vec_f16_cuda;
default:
return nullptr;
}
}
// buffer pool for cuda
#define MAX_CUDA_BUFFERS 256
struct scoped_spin_lock {
std::atomic_flag& lock;
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
while (lock.test_and_set(std::memory_order_acquire)) {
; // spin
}
}
~scoped_spin_lock() {
lock.clear(std::memory_order_release);
}
scoped_spin_lock(const scoped_spin_lock&) = delete;
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
};
struct cuda_buffer {
void * ptr = nullptr;
size_t size = 0;
};
static cuda_buffer g_cuda_buffer_pool[MAX_CUDA_BUFFERS];
static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT;
static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
scoped_spin_lock lock(g_cuda_pool_lock);
for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
cuda_buffer& b = g_cuda_buffer_pool[i];
if (b.size >= size && b.ptr != nullptr) {
void * ptr = b.ptr;
*actual_size = b.size;
b.ptr = nullptr;
b.size = 0;
return ptr;
}
}
void * ptr;
CUDA_CHECK(cudaMalloc((void **) &ptr, size));
*actual_size = size;
return ptr;
}
static void ggml_cuda_pool_free(void * ptr, size_t size) {
scoped_spin_lock lock(g_cuda_pool_lock);
for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
cuda_buffer& b = g_cuda_buffer_pool[i];
if (b.ptr == nullptr) {
b.ptr = ptr;
b.size = size;
return;
}
}
fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
CUDA_CHECK(cudaFree(ptr));
}
#define GGML_CUDA_MAX_STREAMS 8 // Set this to 1 for reproducible matrix multiplication.
#define GGML_CUDA_MAX_EVENTS 64
static cublasHandle_t g_cublasH = nullptr;
static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_STREAMS] = { nullptr };
static cudaStream_t g_cudaStreams2[GGML_CUDA_MAX_STREAMS] = { nullptr };
static cudaEvent_t g_cudaEvents[GGML_CUDA_MAX_EVENTS] = { nullptr };
void ggml_init_cublas() {
if (g_cublasH == nullptr) {
// create streams
for (int i = 0; i < GGML_CUDA_MAX_STREAMS; ++i) {
CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[i], cudaStreamNonBlocking));
CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams2[i], cudaStreamNonBlocking));
}
// create events
for (int i = 0; i < GGML_CUDA_MAX_EVENTS; ++i) {
CUDA_CHECK(cudaEventCreateWithFlags(&g_cudaEvents[i], cudaEventDisableTiming));
}
2023-04-20 18:49:53 +00:00
// create cublas handle
CUBLAS_CHECK(cublasCreate(&g_cublasH));
CUBLAS_CHECK(cublasSetMathMode(g_cublasH, CUBLAS_TF32_TENSOR_OP_MATH));
// configure logging to stdout
// CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
}
}
void * ggml_cuda_host_malloc(size_t size) {
if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
return nullptr;
2023-04-20 18:49:53 +00:00
}
void * ptr = nullptr;
cudaError_t err = cudaMallocHost((void **) &ptr, size);
if (err != cudaSuccess) {
fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
size/1024.0/1024.0, cudaGetErrorString(err));
return nullptr;
}
return ptr;
}
void ggml_cuda_host_free(void * ptr) {
CUDA_CHECK(cudaFreeHost(ptr));
}
static cudaError_t ggml_cuda_h2d_tensor_2d(void * dst, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cudaStream_t stream) {
const uint64_t ne0 = src->ne[0];
const uint64_t ne1 = src->ne[1];
const uint64_t nb0 = src->nb[0];
const uint64_t nb1 = src->nb[1];
const uint64_t nb2 = src->nb[2];
const uint64_t nb3 = src->nb[3];
const enum ggml_type type = src->type;
const size_t ts = ggml_type_size(type);
const size_t bs = ggml_blck_size(type);
const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
if (nb0 == ts && nb1 == ts*ne0/bs) {
return cudaMemcpyAsync(dst, x, ne1*nb1, cudaMemcpyHostToDevice, stream);
} else if (nb0 == ts) {
return cudaMemcpy2DAsync(dst, ts*ne0/bs, x, nb1, ts*ne0/bs, ne1, cudaMemcpyHostToDevice, stream);
} else {
for (uint64_t i1 = 0; i1 < ne1; i1++) {
const void * rx = (const void *) ((const char *) x + i1*nb1);
void * rd = (void *) ((char *) dst + i1*ts*ne0/bs);
// pretend the row is a matrix with cols=1
cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, cudaMemcpyHostToDevice, stream);
if (r != cudaSuccess) return r;
}
return cudaSuccess;
}
}
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
static void ggml_cuda_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->backend == GGML_BACKEND_CUDA);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[2];
const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
size_t x_size, d_size;
float * d_X = (float *) ggml_cuda_pool_malloc(ne0 * sizeof(float), &x_size); // src0
float * d_Y = (float *) src1->data; // src1 is already on device, broadcasted.
float * d_D = (float *) ggml_cuda_pool_malloc(ne0 * sizeof(float), &d_size); // dst
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
const int i0 = i03*ne02 + i02;
float * c_X2 = d_X + i0*ne01*ne00;
float * c_D2 = d_D + i0*ne01*ne00;
cudaStream_t cudaStream = g_cudaStreams[i0 % GGML_CUDA_MAX_STREAMS];
cudaStream_t cudaStream2 = g_cudaStreams2[i0 % GGML_CUDA_MAX_STREAMS];
cudaEvent_t cudaEvent = g_cudaEvents[i0 % GGML_CUDA_MAX_EVENTS];
// copy src0 to device
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X2, src0, i03, i02, cudaStream2));
CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
// wait for data
CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
for (int64_t i01 = 0; i01 < ne01; i01++) {
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int64_t i11 = i01%ne11;
const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
float * c_X1 = c_X2 + i01*ne00;
float * c_Y = d_Y + i1*ne10;
float * c_D1 = c_D2 + i01*ne00;
// compute
mul_f32_cuda(c_X1, c_Y, c_D1, ne00, ne10, cudaStream);
CUDA_CHECK(cudaGetLastError());
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CUDA_CHECK(cudaMemcpyAsync(d, c_D2, sizeof(float)*ne00*ne01, cudaMemcpyDeviceToHost, cudaStream));
}
}
CUDA_CHECK(cudaDeviceSynchronize());
ggml_cuda_pool_free(d_X, x_size);
ggml_cuda_pool_free(d_D, d_size);
}
static void ggml_cuda_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
const int n_mm = ne03 * ne02;
size_t x_size, y_size, d_size;
float * d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size);
float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size);
float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
int i = i03*ne02 + i02;
cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS];
float * c_X = d_X + i * x_ne;
float * c_Y = d_Y + i * y_ne;
float * c_D = d_D + i * d_ne;
// copy data to device
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X, src0, i03, i02, cudaStream));
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
// compute
CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
CUBLAS_CHECK(
cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
&alpha, c_X, ne00,
c_Y, ne10,
&beta, c_D, ne01));
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
}
}
CUDA_CHECK(cudaDeviceSynchronize());
ggml_cuda_pool_free(d_X, x_size);
ggml_cuda_pool_free(d_Y, y_size);
ggml_cuda_pool_free(d_D, d_size);
}
static void ggml_cuda_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb10 = src1->nb[0];
const int nb11 = src1->nb[1];
const int nb12 = src1->nb[2];
const int nb13 = src1->nb[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
const int n_mm = ne03 * ne02;
size_t x_size, y_size, d_size;
half * d_X = (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * x_ne, &x_size);
half * d_Y = (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * y_ne, &y_size);
float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
bool src1_cont_rows = nb10 == sizeof(float);
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
int i = i03*ne02 + i02;
cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS];
half * c_X = d_X + i * x_ne;
half * c_Y = d_Y + i * y_ne;
float * c_D = d_D + i * d_ne;
// copy src0 to device
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X, src0, i03, i02, cudaStream));
// convert src1 to fp16
// TODO: use multiple threads
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
if (src1_cont_rows) {
if (src1_cont_cols) {
ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
}
}
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
for (int64_t i00 = 0; i00 < ne10; i00++) {
// very slow due to no inlining
tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
}
}
}
// copy src1 to device
CUDA_CHECK(cudaMemcpyAsync(c_Y, tmp, sizeof(half) * y_ne, cudaMemcpyHostToDevice, cudaStream));
// compute
CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
CUBLAS_CHECK(
cublasGemmEx(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
&alpha, c_X, CUDA_R_16F, ne00,
c_Y, CUDA_R_16F, ne10,
&beta, c_D, CUDA_R_32F, ne01,
CUBLAS_COMPUTE_32F_FAST_16F,
CUBLAS_GEMM_DEFAULT));
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
}
}
CUDA_CHECK(cudaDeviceSynchronize());
ggml_cuda_pool_free(d_X, x_size);
ggml_cuda_pool_free(d_Y, y_size);
ggml_cuda_pool_free(d_D, d_size);
}
static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_type type = src0->type;
const bool mul_mat_vec = ne11 == 1;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
const int n_mm = ne03 * ne02;
const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
size_t x_size, y_size, d_size, q_size;
float * d_X = nullptr;
if (!mul_mat_vec) {
d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size);
}
float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size);
float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
char * d_Q = (char *) ggml_cuda_pool_malloc(n_mm * q_sz, &q_size);
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(type);
dequantize_mul_mat_vec_cuda_t dmmv = ggml_get_dequantize_mul_mat_vec_cuda(type);
GGML_ASSERT(to_fp32_cuda != nullptr);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
int i = i03*ne02 + i02;
cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS];
cudaStream_t cudaStream2 = g_cudaStreams2[i % GGML_CUDA_MAX_STREAMS];
cudaEvent_t cudaEvent = g_cudaEvents[i % GGML_CUDA_MAX_EVENTS];
float * c_Y = d_Y + i * y_ne;
float * c_D = d_D + i * d_ne;
char * c_Q = d_Q + i * q_sz;
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_CPU) {
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Q, src0, i03, i02, cudaStream2));
} else if (src0->backend == GGML_BACKEND_CUDA) {
c_Q = ((char *) src0->data) + i * q_sz;
} else {
GGML_ASSERT(false);
}
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
// copy src1 to device
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
// wait for data
CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
// compute
dmmv(c_Q, c_Y, c_D, ne00, ne01, cudaStream);
CUDA_CHECK(cudaGetLastError());
} else { // general dequantization kernel + cuBLAS matrix matrix multiplication
float * c_X = d_X + i * x_ne;
// convert src0 to fp32 on device
to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2);
CUDA_CHECK(cudaGetLastError());
CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
// copy src1 to device
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
// wait for conversion
CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
// compute
CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
CUBLAS_CHECK(
cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
ne01, ne11, ne10,
&alpha, c_X, ne00,
c_Y, ne10,
&beta, c_D, ne01));
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
}
}
CUDA_CHECK(cudaDeviceSynchronize());
if (!mul_mat_vec) {
ggml_cuda_pool_free(d_X, x_size);
}
ggml_cuda_pool_free(d_Y, y_size);
ggml_cuda_pool_free(d_D, d_size);
ggml_cuda_pool_free(d_Q, q_size);
}
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
void ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
ggml_cuda_mul_f32(src0, src1, dst);
}
bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
// TODO: find the optimal values for these
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
src1->type == GGML_TYPE_F32 &&
dst->type == GGML_TYPE_F32 &&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CUDA)) {
return true;
}
return false;
}
bool ggml_cuda_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
size_t src0_sz = ggml_nbytes(src0);
size_t src1_sz = ggml_nbytes(src1);
// mul_mat_q: src0 is converted to fp32 on device
size_t mul_mat_q_transfer = src0_sz + src1_sz;
// mul_mat_f16: src1 is converted to fp16 on cpu
size_t mul_mat_f16_transfer = src0_sz + sizeof(half) * ggml_nelements(src1);
// choose the smaller one to transfer to the device
// TODO: this is not always the best choice due to the overhead of converting to fp16
return mul_mat_f16_transfer < mul_mat_q_transfer;
}
void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
GGML_ASSERT(ggml_cuda_can_mul_mat(src0, src1, dst));
if (src0->type == GGML_TYPE_F32) {
ggml_cuda_mul_mat_f32(src0, src1, dst);
}
else if (src0->type == GGML_TYPE_F16) {
if (ggml_cuda_mul_mat_use_f16(src0, src1, dst)) {
ggml_cuda_mul_mat_f16(src0, src1, dst, wdata, wsize);
}
else {
ggml_cuda_mul_mat_q_f32(src0, src1, dst);
}
}
else if (ggml_is_quantized(src0->type)) {
ggml_cuda_mul_mat_q_f32(src0, src1, dst);
}
else {
GGML_ASSERT(false);
}
}
size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
if (ggml_cuda_mul_mat_use_f16(src0, src1, dst)) {
return ggml_nelements(src1) * sizeof(ggml_fp16_t);
}
else {
return 0;
}
}
void ggml_cuda_transform_tensor(ggml_tensor * tensor) {
const int64_t ne0 = tensor->ne[0];
const int64_t ne1 = tensor->ne[1];
const int64_t ne2 = tensor->ne[2];
const int64_t ne3 = tensor->ne[3];
const ggml_type type = tensor->type;
const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
size_t q_size;
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
char * dst = (char *) ggml_cuda_pool_malloc(q_sz, &q_size);
cudaStream_t cudaStream2 = g_cudaStreams2[0];
// copy tensor to device
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
for (int64_t i3 = 0; i3 < ne3; i3++) {
for (int64_t i2 = 0; i2 < ne2; i2++) {
int i = i3*ne2 + i2;
CUDA_CHECK(ggml_cuda_h2d_tensor_2d(dst + i*ne0*ne1, tensor, i3, i2, cudaStream2));
}
}
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
tensor->data = dst;
tensor->backend = GGML_BACKEND_CUDA;
}
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483) * Broadcasting for ggml_mul * CUDA kernel for ggml_mul, norms in VRAM * GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * fixup! GPU weights not in RAM, direct loading with cuFile * define default model path once, sync path with readme (#1366) * ~7% faster Q5_1 AVX2 code (#1477) * convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo * benchmark-matmul: Print the average of the test results (#1490) * Remove unused n_parts parameter (#1509) * Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw * make kv_f16 the default for api users (#1517) * minor : fix compile warnings * readme : adds WizardLM to the list of supported models (#1485) * main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 * examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * tests : add missing header * ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics * ggml : fix scalar implementation of Q4_1 dot * llama : fix compile warnings in llama_set_state_data() * llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Fix for mingw (#1462) * llama : add llama_init_backend() API (close #1527) * feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. * GPU weights not in RAM, direct loading with cuFile * llama : code style fixes + progress print fix * ggml : ggml_mul better broadcast support * cmake : workarounds for cufile when CMake version < 3.25 * gg rebase fixup * Loop in llama.cpp, fixed progress callback * Attempt clang-tidy fix * llama : fix vram size computation * Add forgotten fclose() --------- Co-authored-by: András Salamon <ott2@users.noreply.github.com> Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com> Co-authored-by: Stephan Walter <stephan@walter.name> Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com> Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: David Kennedy <dakennedyd@gmail.com> Co-authored-by: Jason McCartney <jmac@theroot.org> Co-authored-by: Evan Jones <evan.q.jones@gmail.com> Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
void ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensor, const size_t offset) {
FILE * fp = fopen(fname, "rb");
const size_t size = ggml_nbytes(tensor);
void * buf;
CUDA_CHECK(cudaMalloc(&buf, size));
void * buf_host = malloc(size);
#ifdef _WIN32
int ret = _fseeki64(fp, (__int64) offset, SEEK_SET);
#else
int ret = fseek(fp, (long) offset, SEEK_SET);
#endif
GGML_ASSERT(ret == 0); // same
size_t ret2 = fread(buf_host, size, 1, fp);
if (ret2 != 1) {
fprintf(stderr, "unexpectedly reached end of file");
exit(1);
}
cudaMemcpy(buf, buf_host, size, cudaMemcpyHostToDevice);
cudaDeviceSynchronize();
tensor->data = buf;
free(buf_host);
fclose(fp);
}