huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 19:58:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Faster kernels for quantized matmul on cuda (#2060)

Browse Source 
* Hook the quantized matmul cuda kernels.

* Add a (currently broken) test.

* Kernel fixes.

* Fix by transposing the rhs matrix.

* Add the q4-1 kernels.

* Proper block sizes.

* More details in the tests.
This commit is contained in:
Laurent Mazare
2024-04-15 08:32:47 +02:00
committed by GitHub
parent c119600d6e
commit f7d5bf5b97
2 changed files with 255 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

129
candle-kernels/src/quantized.cu
View File

@ -71,8 +71,6 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t *
}
#define CUDA_USE_TENSOR_CORES
#define WARP_SIZE 32
#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
@ -103,6 +101,25 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t *
#define MMQ_Y_Q4_0_PASCAL 64
#define NWARPS_Q4_0_PASCAL 8
#define MMQ_X_Q4_1_RDNA2 64
#define MMQ_Y_Q4_1_RDNA2 128
#define NWARPS_Q4_1_RDNA2 8
#define MMQ_X_Q4_1_RDNA1 64
#define MMQ_Y_Q4_1_RDNA1 64
#define NWARPS_Q4_1_RDNA1 8
#if defined(CUDA_USE_TENSOR_CORES)
#define MMQ_X_Q4_1_AMPERE 4
#define MMQ_Y_Q4_1_AMPERE 32
#define NWARPS_Q4_1_AMPERE 4
#else
#define MMQ_X_Q4_1_AMPERE 64
#define MMQ_Y_Q4_1_AMPERE 128
#define NWARPS_Q4_1_AMPERE 4
#endif
#define MMQ_X_Q4_1_PASCAL 64
#define MMQ_Y_Q4_1_PASCAL 64
#define NWARPS_Q4_1_PASCAL 8
#define MMQ_X_Q5_0_RDNA2 64
#define MMQ_Y_Q5_0_RDNA2 128
#define NWARPS_Q5_0_RDNA2 8
@ -558,6 +575,52 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
}
}
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
GGML_CUDA_ASSUME(i_offset >= 0);
GGML_CUDA_ASSUME(i_offset < nwarps);
GGML_CUDA_ASSUME(k >= 0);
GGML_CUDA_ASSUME(k < WARP_SIZE);
const int kbx = k / QI4_1;
const int kqsx = k % QI4_1;
const block_q4_1 * bx0 = (const block_q4_1 *) vx;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
int i = i0 + i_offset;
if (need_check) {
i = min(i, i_max);
}
const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx;
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
}
const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
const int kbxd = k % blocks_per_tile_x_row;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row;
if (need_check) {
i = min(i, i_max);
}
const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd;
x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm;
}
}
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
(void)x_qh; (void)x_sc;
@ -568,6 +631,16 @@ template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_0(
*x_dm = (half2 *) tile_x_d;
}
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
__shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + + mmq_y];
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_1) + mmq_y/QI4_1];
*x_ql = tile_x_qs;
*x_dm = tile_x_dm;
}
static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
const block_q4_0 * x = (const block_q4_0 *) vx;
@ -3493,6 +3566,26 @@ static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat(
y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
}
static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat(
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
GGML_UNUSED(x_qh); GGML_UNUSED(x_sc);
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
int u[2*VDR_Q4_1_Q8_1_MMQ];
#pragma unroll
for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE];
u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE];
}
return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
(&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1],
y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
}
extern "C" __global__ void
mul_mat_q4_0(
@ -3503,10 +3596,24 @@ extern "C" __global__ void
const int nwarps = NWARPS_Q4_0_AMPERE;
mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
load_tiles_q4_0<mmq_y, nwarps, false>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
load_tiles_q4_0<mmq_y, nwarps, true>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
extern "C" __global__ void
mul_mat_q4_1(
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
const int mmq_x = MMQ_X_Q4_1_AMPERE;
const int mmq_y = MMQ_Y_Q4_1_AMPERE;
const int nwarps = NWARPS_Q4_1_AMPERE;
mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
load_tiles_q4_1<mmq_y, nwarps, true>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
extern "C" __global__ void
mul_mat_q5_0(
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
@ -3516,7 +3623,7 @@ extern "C" __global__ void
const int nwarps = NWARPS_Q5_0_AMPERE;
mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
load_tiles_q5_0<mmq_y, nwarps, false>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
load_tiles_q5_0<mmq_y, nwarps, true>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
@ -3529,7 +3636,7 @@ mul_mat_q5_1(
const int nwarps = NWARPS_Q5_1_AMPERE;
mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
load_tiles_q5_1<mmq_y, nwarps, false>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
load_tiles_q5_1<mmq_y, nwarps, true>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
@ -3542,7 +3649,7 @@ extern "C" __global__ void
const int nwarps = NWARPS_Q8_0_AMPERE;
mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
load_tiles_q8_0<mmq_y, nwarps, false>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
load_tiles_q8_0<mmq_y, nwarps, true>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
@ -3554,7 +3661,7 @@ mul_mat_q2_K(
const int mmq_y = MMQ_Y_Q2_K_AMPERE;
const int nwarps = NWARPS_Q2_K_AMPERE;
mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
load_tiles_q2_K<mmq_y, nwarps, false>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
load_tiles_q2_K<mmq_y, nwarps, true>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
@ -3566,7 +3673,7 @@ extern "C" __global__ void
const int mmq_y = MMQ_Y_Q3_K_AMPERE;
const int nwarps = NWARPS_Q3_K_AMPERE;
mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
load_tiles_q3_K<mmq_y, nwarps, false>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
load_tiles_q3_K<mmq_y, nwarps, true>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
@ -3578,7 +3685,7 @@ extern "C" __global__ void
const int mmq_y = MMQ_Y_Q4_K_AMPERE;
const int nwarps = NWARPS_Q4_K_AMPERE;
mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
load_tiles_q4_K<mmq_y, nwarps, false>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
load_tiles_q4_K<mmq_y, nwarps, true>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
@ -3590,7 +3697,7 @@ mul_mat_q5_K(
const int mmq_y = MMQ_Y_Q5_K_AMPERE;
const int nwarps = NWARPS_Q5_K_AMPERE;
mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
load_tiles_q5_K<mmq_y, nwarps, false>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
load_tiles_q5_K<mmq_y, nwarps, true>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}
@ -3602,6 +3709,6 @@ extern "C" __global__ void
const int mmq_y = MMQ_Y_Q6_K_AMPERE;
const int nwarps = NWARPS_Q6_K_AMPERE;
mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
load_tiles_q6_K<mmq_y, nwarps, false>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
load_tiles_q6_K<mmq_y, nwarps, true>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
}