IQ4_XS: a 4.25 bpw quantization (#5747)
* Try IQ4_NL with blocks of 64 - does not look good * iq4_xs: go to super-blocks of 256 and 6-bit scales for blocks of 32 * iq4_xs: CUDA works - 133.2 t/s * iq4_xs: AVX2 dot product * iq4_xs: ARM_NEON dot product * iq4_nl: Metal implementation As usual, Metal / Apple Silicon don't like my quants. * iq3_xs: minor fix * iq4_xs: shrink by using IQ3_S for attn_k and attn_q * iq4_xs: revert using IQ3_S for attn_k and attn_v PPL vs size is good, but CPU performance suffers: on M2 Max TG-128 drops to 21.7 t/s from 28.8, and on a Ryzen-7950X to 14.5 t/s from 15.8 t/s. On CUDA we have 135 t/s when using IQ3_S vs 133 t/s with pure IQ4_XS. * Fix CI * iq4_xs: Added forgotten check for 256 divisibility --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
GitHub
parent
c24a2a6e60
commit
0becb22ac0
11 changed files with 668 additions and 38 deletions
119
ggml-cuda.cu
119
ggml-cuda.cu
|
|
@ -571,6 +571,18 @@ typedef struct {
|
|||
} block_iq4_nl;
|
||||
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
|
||||
|
||||
// QR4_XS = 8 is very slightly faster than QR4_XS = 4
|
||||
#define QR4_XS 8
|
||||
#define QI4_XS (QK_K / (4*QR4_XS))
|
||||
typedef struct {
|
||||
half d;
|
||||
uint16_t scales_h;
|
||||
uint8_t scales_l[QK_K/64];
|
||||
uint8_t qs[QK_K/2];
|
||||
} block_iq4_xs;
|
||||
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
|
||||
|
||||
|
||||
#define WARP_SIZE 32
|
||||
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
|
||||
|
||||
|
|
@ -2427,6 +2439,25 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
|
|||
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const int i = blockIdx.x;
|
||||
const block_iq4_xs * x = (const block_iq4_xs *)vx;
|
||||
|
||||
const int tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 4*il;
|
||||
const uint8_t * q4 = x[i].qs + 16*ib + 4*il;
|
||||
const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
|
||||
for (int j = 0; j < 4; ++j) {
|
||||
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
|
||||
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
|
||||
|
||||
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
|
||||
|
|
@ -5286,6 +5317,76 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
|
|||
return d * (sumi1 + sumi2);
|
||||
}
|
||||
|
||||
static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
|
||||
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
||||
|
||||
#if QK_K == 256
|
||||
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
||||
|
||||
const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
|
||||
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
|
||||
|
||||
//// iqs is 0...7
|
||||
//const int ib64 = iqs/2;
|
||||
//const int il = iqs%2;
|
||||
//const int32_t * q8_1 = (const int *)bq8_1[2*ib64+0].qs + 2*il;
|
||||
//const int32_t * q8_2 = (const int *)bq8_1[2*ib64+1].qs + 2*il;
|
||||
//const uint32_t * q4_1 = (const uint32_t *)bq4->qs + 8*ib64 + 2*il;
|
||||
//const uint32_t * q4_2 = q4_1 + 4;
|
||||
//const int8_t ls1 = (bq4->scales_l[ib64] & 0xf) | (((bq4->scales_h >> (4*ib64+0)) & 3) << 4);
|
||||
//const int8_t ls2 = (bq4->scales_l[ib64] >> 4) | (((bq4->scales_h >> (4*ib64+2)) & 3) << 4);
|
||||
//const float d1 = (float)bq4->d * (ls1 - 32) * __low2float(bq8_1[2*ib64+0].ds);
|
||||
//const float d2 = (float)bq4->d * (ls2 - 32) * __low2float(bq8_1[2*ib64+1].ds);
|
||||
//int v1, v2;
|
||||
//int sumi1 = 0, sumi2 = 0;
|
||||
//for (int j = 0; j < 2; ++j) {
|
||||
// get_int_from_table_16(q4_1[j], values, v1, v2);
|
||||
// sumi1 = __dp4a(v2, q8_1[j+4], __dp4a(v1, q8_1[j+0], sumi1));
|
||||
// get_int_from_table_16(q4_2[j], values, v1, v2);
|
||||
// sumi2 = __dp4a(v2, q8_2[j+4], __dp4a(v1, q8_2[j+0], sumi2));
|
||||
//}
|
||||
//return d1 * sumi1 + d2 * sumi2;
|
||||
|
||||
// iqs is 0...7
|
||||
const int ib32 = iqs;
|
||||
const int32_t * q8 = (const int *)bq8_1[ib32].qs;
|
||||
const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32;
|
||||
const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
|
||||
const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
|
||||
int v1, v2;
|
||||
int sumi1 = 0, sumi2 = 0;
|
||||
for (int j = 0; j < 4; ++j) {
|
||||
get_int_from_table_16(q4[j], values, v1, v2);
|
||||
sumi1 = __dp4a(v1, q8[j+0], sumi1);
|
||||
sumi2 = __dp4a(v2, q8[j+4], sumi2);
|
||||
}
|
||||
return d * (sumi1 + sumi2);
|
||||
|
||||
//// iqs is 0...15
|
||||
//const int ib32 = iqs/2;
|
||||
//const int il = iqs%2;
|
||||
//const int32_t * q8 = (const int *)bq8_1[ib32].qs + 2*il;
|
||||
//const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32 + 2*il;
|
||||
//const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
|
||||
//const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
|
||||
//int v1, v2;
|
||||
//int sumi1 = 0, sumi2 = 0;
|
||||
//for (int j = 0; j < 2; ++j) {
|
||||
// get_int_from_table_16(q4[j], values, v1, v2);
|
||||
// sumi1 = __dp4a(v1, q8[j+0], sumi1);
|
||||
// sumi2 = __dp4a(v2, q8[j+4], sumi2);
|
||||
//}
|
||||
//return d * (sumi1 + sumi2);
|
||||
#else
|
||||
assert(false);
|
||||
return 0.f;
|
||||
#endif
|
||||
#else
|
||||
assert(false);
|
||||
return 0.f;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
|
||||
allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
|
||||
static __device__ __forceinline__ void mul_mat_q(
|
||||
|
|
@ -7340,6 +7441,12 @@ static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k,
|
|||
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = (k + QK_K - 1) / QK_K;
|
||||
dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template <typename src_t, typename dst_t>
|
||||
static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
|
||||
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
|
||||
|
|
@ -7385,6 +7492,8 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
|
|||
return dequantize_row_iq1_s_cuda;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
return dequantize_row_iq4_nl_cuda;
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
return dequantize_row_iq4_xs_cuda;
|
||||
case GGML_TYPE_IQ3_S:
|
||||
return dequantize_row_iq3_s_cuda;
|
||||
case GGML_TYPE_F32:
|
||||
|
|
@ -7428,6 +7537,8 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
|
|||
return dequantize_row_iq1_s_cuda;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
return dequantize_row_iq4_nl_cuda;
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
return dequantize_row_iq4_xs_cuda;
|
||||
case GGML_TYPE_IQ3_S:
|
||||
return dequantize_row_iq3_s_cuda;
|
||||
case GGML_TYPE_F16:
|
||||
|
|
@ -9176,6 +9287,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
|
|||
case GGML_TYPE_IQ3_XXS:
|
||||
case GGML_TYPE_IQ1_S:
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
case GGML_TYPE_IQ3_S:
|
||||
return max_compute_capability >= CC_RDNA2 ? 128 : 64;
|
||||
default:
|
||||
|
|
@ -9203,6 +9315,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
|
|||
case GGML_TYPE_IQ3_XXS:
|
||||
case GGML_TYPE_IQ1_S:
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
case GGML_TYPE_IQ3_S:
|
||||
return max_compute_capability >= CC_VOLTA ? 128 : 64;
|
||||
case GGML_TYPE_Q6_K:
|
||||
|
|
@ -9313,6 +9426,10 @@ static void ggml_cuda_op_mul_mat_vec_q(
|
|||
mul_mat_vec_q_cuda<QK4_NL, QI4_NL, block_iq4_nl, VDR_Q4_0_Q8_1_MMVQ, vec_dot_iq4_nl_q8_1>
|
||||
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
mul_mat_vec_q_cuda<QK_K, QI4_XS, block_iq4_xs, 1, vec_dot_iq4_xs_q8_1>
|
||||
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ3_S:
|
||||
mul_mat_vec_q_cuda<QK_K, QI3_XS, block_iq3_s, 1, vec_dot_iq3_s_q8_1>
|
||||
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
|
|
@ -12041,7 +12158,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
|
|||
ggml_type a_type = a->type;
|
||||
if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
|
||||
a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S ||
|
||||
a_type == GGML_TYPE_IQ2_S) {
|
||||
a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) {
|
||||
if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
|
||||
return false;
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue