llama_cpp 0.15.0 → 0.15.2
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +12 -0
- data/ext/llama_cpp/llama_cpp.cpp +6 -0
- data/lib/llama_cpp/version.rb +2 -2
- data/sig/llama_cpp.rbs +6 -0
- data/vendor/tmp/llama.cpp/Makefile +6 -7
- data/vendor/tmp/llama.cpp/ggml-backend.c +2 -3
- data/vendor/tmp/llama.cpp/ggml-cuda.cu +303 -23
- data/vendor/tmp/llama.cpp/ggml-impl.h +84 -0
- data/vendor/tmp/llama.cpp/ggml-kompute.cpp +9 -3
- data/vendor/tmp/llama.cpp/ggml-metal.m +137 -133
- data/vendor/tmp/llama.cpp/ggml-metal.metal +87 -110
- data/vendor/tmp/llama.cpp/ggml-opencl.cpp +1 -0
- data/vendor/tmp/llama.cpp/ggml-quants.c +2220 -28
- data/vendor/tmp/llama.cpp/ggml-rpc.cpp +1032 -0
- data/vendor/tmp/llama.cpp/ggml-rpc.h +24 -0
- data/vendor/tmp/llama.cpp/ggml-sycl.cpp +35 -152
- data/vendor/tmp/llama.cpp/ggml-vulkan-shaders.hpp +46843 -39205
- data/vendor/tmp/llama.cpp/ggml-vulkan.cpp +953 -268
- data/vendor/tmp/llama.cpp/ggml.c +1762 -681
- data/vendor/tmp/llama.cpp/ggml.h +43 -24
- data/vendor/tmp/llama.cpp/llama.cpp +533 -296
- data/vendor/tmp/llama.cpp/llama.h +10 -1
- data/vendor/tmp/llama.cpp/sgemm.cpp +56 -21
- data/vendor/tmp/llama.cpp/unicode-data.cpp +6969 -1637
- data/vendor/tmp/llama.cpp/unicode-data.h +15 -11
- data/vendor/tmp/llama.cpp/unicode.cpp +286 -176
- data/vendor/tmp/llama.cpp/unicode.h +44 -10
- metadata +4 -2
@@ -0,0 +1,24 @@
|
|
1
|
+
#pragma once
|
2
|
+
|
3
|
+
#include "ggml.h"
|
4
|
+
#include "ggml-backend.h"
|
5
|
+
|
6
|
+
#ifdef __cplusplus
|
7
|
+
extern "C" {
|
8
|
+
#endif
|
9
|
+
|
10
|
+
#define GGML_RPC_MAX_SERVERS 16
|
11
|
+
|
12
|
+
// backend API
|
13
|
+
GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
|
14
|
+
GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend);
|
15
|
+
|
16
|
+
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
|
17
|
+
|
18
|
+
GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
|
19
|
+
|
20
|
+
GGML_API GGML_CALL void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
|
21
|
+
|
22
|
+
#ifdef __cplusplus
|
23
|
+
}
|
24
|
+
#endif
|
@@ -3154,7 +3154,6 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)(
|
|
3154
3154
|
#define SYCL_SCALE_BLOCK_SIZE 256
|
3155
3155
|
#define SYCL_CLAMP_BLOCK_SIZE 256
|
3156
3156
|
#define SYCL_ROPE_BLOCK_SIZE 256
|
3157
|
-
#define SYCL_ALIBI_BLOCK_SIZE 32
|
3158
3157
|
#define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32
|
3159
3158
|
#define SYCL_QUANTIZE_BLOCK_SIZE 256
|
3160
3159
|
#define SYCL_DEQUANTIZE_BLOCK_SIZE 256
|
@@ -8330,24 +8329,26 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_
|
|
8330
8329
|
const int blocks_per_row = ncols / qk;
|
8331
8330
|
const int blocks_per_warp = vdr * WARP_SIZE / qi;
|
8332
8331
|
|
8333
|
-
//
|
8332
|
+
const int qi_vdr = (qi / vdr); // N_threads processing 1 qk block
|
8333
|
+
|
8334
|
+
// partial sum for each thread
|
8334
8335
|
float tmp = 0.0f;
|
8335
8336
|
|
8336
8337
|
const block_q_t * x = (const block_q_t *) vx;
|
8337
8338
|
const block_q8_1 * y = (const block_q8_1 *) vy;
|
8338
8339
|
|
8339
|
-
for (int i = item_ct1.get_local_id(2) /
|
8340
|
+
for (int i = item_ct1.get_local_id(2) / qi_vdr; i < blocks_per_row;
|
8340
8341
|
i += blocks_per_warp) {
|
8341
|
-
|
8342
|
+
const int ibx = row * blocks_per_row + i; // x block index
|
8342
8343
|
|
8343
|
-
|
8344
|
+
const int iby = i * (qk / QK8_1); // y block index that aligns with ibx
|
8344
8345
|
|
8345
|
-
|
8346
|
-
|
8347
|
-
|
8348
|
-
|
8346
|
+
const int iqs =
|
8347
|
+
vdr *
|
8348
|
+
(item_ct1.get_local_id(2) -
|
8349
|
+
i * qi_vdr); // x block quant index when casting the quants to int
|
8349
8350
|
|
8350
|
-
|
8351
|
+
tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs);
|
8351
8352
|
}
|
8352
8353
|
|
8353
8354
|
// sum up partial sums and write back result
|
@@ -9314,32 +9315,6 @@ static void rope_glm_f32(
|
|
9314
9315
|
dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta;
|
9315
9316
|
}
|
9316
9317
|
|
9317
|
-
static void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
|
9318
|
-
const int n_heads_log2_floor, const float m0, const float m1,
|
9319
|
-
const sycl::nd_item<3> &item_ct1) {
|
9320
|
-
const int col = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
|
9321
|
-
item_ct1.get_local_id(2);
|
9322
|
-
|
9323
|
-
if (col >= ncols) {
|
9324
|
-
return;
|
9325
|
-
}
|
9326
|
-
|
9327
|
-
const int row = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
|
9328
|
-
item_ct1.get_local_id(1);
|
9329
|
-
const int i = row*ncols + col;
|
9330
|
-
|
9331
|
-
const int k = row/k_rows;
|
9332
|
-
|
9333
|
-
float m_k;
|
9334
|
-
if (k < n_heads_log2_floor) {
|
9335
|
-
m_k = dpct::pow(m0, k + 1);
|
9336
|
-
} else {
|
9337
|
-
m_k = dpct::pow(m1, 2 * (k - n_heads_log2_floor) + 1);
|
9338
|
-
}
|
9339
|
-
|
9340
|
-
dst[i] = col * m_k + x[i];
|
9341
|
-
}
|
9342
|
-
|
9343
9318
|
static void k_sum_rows_f32(const float * x, float * dst, const int ncols,
|
9344
9319
|
const sycl::nd_item<3> &item_ct1) {
|
9345
9320
|
const int row = item_ct1.get_group(1);
|
@@ -9441,7 +9416,7 @@ static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, con
|
|
9441
9416
|
|
9442
9417
|
|
9443
9418
|
template <bool vals_smem, int ncols_template, int block_size_template>
|
9444
|
-
static void soft_max_f32(const float * x, const float * mask,
|
9419
|
+
static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par,
|
9445
9420
|
const int nrows_y, const float scale, const float max_bias, const float m0,
|
9446
9421
|
const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) {
|
9447
9422
|
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
|
@@ -9455,7 +9430,7 @@ static void soft_max_f32(const float * x, const float * mask, const float *pos,
|
|
9455
9430
|
const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
|
9456
9431
|
const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
|
9457
9432
|
|
9458
|
-
float slope =
|
9433
|
+
float slope = 1.0f;
|
9459
9434
|
|
9460
9435
|
// ALiBi
|
9461
9436
|
if (max_bias > 0.0f) {
|
@@ -9480,7 +9455,7 @@ static void soft_max_f32(const float * x, const float * mask, const float *pos,
|
|
9480
9455
|
const int ix = rowx*ncols + col;
|
9481
9456
|
const int iy = rowy*ncols + col;
|
9482
9457
|
|
9483
|
-
const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f)
|
9458
|
+
const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f);
|
9484
9459
|
|
9485
9460
|
vals[col] = val;
|
9486
9461
|
max_val = sycl::max(max_val, val);
|
@@ -12962,20 +12937,6 @@ static void rope_glm_f32_sycl(const float *x, float *dst, int ncols, int nrows,
|
|
12962
12937
|
});
|
12963
12938
|
}
|
12964
12939
|
|
12965
|
-
static void alibi_f32_sycl(const float *x, float *dst, const int ncols,
|
12966
|
-
const int nrows, const int k_rows,
|
12967
|
-
const int n_heads_log2_floor, const float m0,
|
12968
|
-
const float m1, dpct::queue_ptr stream) {
|
12969
|
-
const sycl::range<3> block_dims(1, 1, SYCL_ALIBI_BLOCK_SIZE);
|
12970
|
-
const int num_blocks_x = (ncols + SYCL_ALIBI_BLOCK_SIZE - 1) / (SYCL_ALIBI_BLOCK_SIZE);
|
12971
|
-
const sycl::range<3> block_nums(1, nrows, num_blocks_x);
|
12972
|
-
stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
12973
|
-
[=](sycl::nd_item<3> item_ct1) {
|
12974
|
-
alibi_f32(x, dst, ncols, k_rows,
|
12975
|
-
n_heads_log2_floor, m0, m1, item_ct1);
|
12976
|
-
});
|
12977
|
-
}
|
12978
|
-
|
12979
12940
|
static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols,
|
12980
12941
|
const int nrows, dpct::queue_ptr stream) {
|
12981
12942
|
const sycl::range<3> block_dims(1, 1, WARP_SIZE);
|
@@ -13056,7 +13017,7 @@ static void diag_mask_inf_f32_sycl(const float *x, float *dst,
|
|
13056
13017
|
}
|
13057
13018
|
|
13058
13019
|
template <bool vals_smem, int ncols_template, int block_size_template>
|
13059
|
-
static void soft_max_f32_submitter(const float * x, const float * mask,
|
13020
|
+
static void soft_max_f32_submitter(const float * x, const float * mask, float * dst, const int ncols_par,
|
13060
13021
|
const int nrows_y, const float scale, const float max_bias, const float m0,
|
13061
13022
|
const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
|
13062
13023
|
const size_t n_local_scratch, dpct::queue_ptr stream) {
|
@@ -13066,7 +13027,7 @@ static void soft_max_f32_submitter(const float * x, const float * mask, const fl
|
|
13066
13027
|
cgh.parallel_for(
|
13067
13028
|
sycl::nd_range<3>(block_nums * block_dims, block_dims),
|
13068
13029
|
[=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
|
13069
|
-
soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask,
|
13030
|
+
soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, dst, ncols_par,
|
13070
13031
|
nrows_y, scale, max_bias, m0,
|
13071
13032
|
m1, n_head_log2, item_ct1,
|
13072
13033
|
local_buf_acc.get_pointer());
|
@@ -13074,7 +13035,7 @@ static void soft_max_f32_submitter(const float * x, const float * mask, const fl
|
|
13074
13035
|
});
|
13075
13036
|
}
|
13076
13037
|
|
13077
|
-
static void soft_max_f32_sycl(const float * x, const float * mask,
|
13038
|
+
static void soft_max_f32_sycl(const float * x, const float * mask,
|
13078
13039
|
float * dst, const int ncols_x, const int nrows_x,
|
13079
13040
|
const int nrows_y, const float scale, const float max_bias,
|
13080
13041
|
dpct::queue_ptr stream) {
|
@@ -13096,60 +13057,60 @@ static void soft_max_f32_sycl(const float * x, const float * mask, const float *
|
|
13096
13057
|
const size_t local_mem_size = stream->get_device().get_info<sycl::info::device::local_mem_size>();
|
13097
13058
|
if (n_local_scratch*sizeof(float) < local_mem_size) {
|
13098
13059
|
if (ncols_x > max_block_size) {
|
13099
|
-
soft_max_f32_submitter<true, 0, 0>(x, mask,
|
13060
|
+
soft_max_f32_submitter<true, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
|
13100
13061
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13101
13062
|
block_dims, n_local_scratch, stream);
|
13102
13063
|
return;
|
13103
13064
|
}
|
13104
13065
|
switch (ncols_x) {
|
13105
13066
|
case 32:
|
13106
|
-
soft_max_f32_submitter<true, 32, 32>(x, mask,
|
13067
|
+
soft_max_f32_submitter<true, 32, 32>(x, mask, dst, ncols_x, nrows_y, scale,
|
13107
13068
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13108
13069
|
block_dims, n_local_scratch, stream);
|
13109
13070
|
break;
|
13110
13071
|
case 64:
|
13111
|
-
soft_max_f32_submitter<true, 64, 64>(x, mask,
|
13072
|
+
soft_max_f32_submitter<true, 64, 64>(x, mask, dst, ncols_x, nrows_y, scale,
|
13112
13073
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13113
13074
|
block_dims, n_local_scratch, stream);
|
13114
13075
|
break;
|
13115
13076
|
case 128:
|
13116
|
-
soft_max_f32_submitter<true, 128, 128>(x, mask,
|
13077
|
+
soft_max_f32_submitter<true, 128, 128>(x, mask, dst, ncols_x, nrows_y, scale,
|
13117
13078
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13118
13079
|
block_dims, n_local_scratch, stream);
|
13119
13080
|
break;
|
13120
13081
|
case 256:
|
13121
|
-
soft_max_f32_submitter<true, 256, 256>(x, mask,
|
13082
|
+
soft_max_f32_submitter<true, 256, 256>(x, mask, dst, ncols_x, nrows_y, scale,
|
13122
13083
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13123
13084
|
block_dims, n_local_scratch, stream);
|
13124
13085
|
break;
|
13125
13086
|
case 512:
|
13126
|
-
soft_max_f32_submitter<true, 512, 512>(x, mask,
|
13087
|
+
soft_max_f32_submitter<true, 512, 512>(x, mask, dst, ncols_x, nrows_y, scale,
|
13127
13088
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13128
13089
|
block_dims, n_local_scratch, stream);
|
13129
13090
|
break;
|
13130
13091
|
case 1024:
|
13131
|
-
soft_max_f32_submitter<true, 1024, 1024>(x, mask,
|
13092
|
+
soft_max_f32_submitter<true, 1024, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
|
13132
13093
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13133
13094
|
block_dims, n_local_scratch, stream);
|
13134
13095
|
break;
|
13135
13096
|
case 2048:
|
13136
|
-
soft_max_f32_submitter<true, 2048, 1024>(x, mask,
|
13097
|
+
soft_max_f32_submitter<true, 2048, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
|
13137
13098
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13138
13099
|
block_dims, n_local_scratch, stream);
|
13139
13100
|
break;
|
13140
13101
|
case 4096:
|
13141
|
-
soft_max_f32_submitter<true, 4096, 1024>(x, mask,
|
13102
|
+
soft_max_f32_submitter<true, 4096, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
|
13142
13103
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13143
13104
|
block_dims, n_local_scratch, stream);
|
13144
13105
|
break;
|
13145
13106
|
default:
|
13146
|
-
soft_max_f32_submitter<true, 0, 0>(x, mask,
|
13107
|
+
soft_max_f32_submitter<true, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
|
13147
13108
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13148
13109
|
block_dims, n_local_scratch, stream);
|
13149
13110
|
break;
|
13150
13111
|
}
|
13151
13112
|
} else {
|
13152
|
-
soft_max_f32_submitter<false, 0, 0>(x, mask,
|
13113
|
+
soft_max_f32_submitter<false, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
|
13153
13114
|
max_bias, m0, m1, n_head_log2, block_nums,
|
13154
13115
|
block_dims, WARP_SIZE, stream);
|
13155
13116
|
}
|
@@ -14026,6 +13987,10 @@ inline void ggml_sycl_op_upscale(const ggml_tensor *src0,
|
|
14026
13987
|
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
14027
13988
|
GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
|
14028
13989
|
|
13990
|
+
#pragma message("TODO: generalize upscale operator")
|
13991
|
+
#pragma message(" https://github.com/ggerganov/ggml/pull/814")
|
13992
|
+
GGML_ASSERT(false && "TODO: generalize upscale operator");
|
13993
|
+
|
14029
13994
|
const int scale_factor = dst->op_params[0];
|
14030
13995
|
|
14031
13996
|
upscale_f32_sycl(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream);
|
@@ -14560,36 +14525,6 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
|
|
14560
14525
|
(void) src1_dd;
|
14561
14526
|
}
|
14562
14527
|
|
14563
|
-
inline void ggml_sycl_op_alibi(const ggml_tensor *src0, const ggml_tensor *src1,
|
14564
|
-
ggml_tensor *dst, const float *src0_dd,
|
14565
|
-
const float *src1_dd, float *dst_dd,
|
14566
|
-
const dpct::queue_ptr &main_stream) {
|
14567
|
-
|
14568
|
-
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
14569
|
-
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
14570
|
-
|
14571
|
-
GGML_TENSOR_LOCALS_3(int64_t, ne0, src0, ne);
|
14572
|
-
const int64_t nrows = ggml_nrows(src0);
|
14573
|
-
|
14574
|
-
//const int n_past = ((int32_t *) dst->op_params)[0];
|
14575
|
-
const int n_head = ((int32_t *) dst->op_params)[1];
|
14576
|
-
float max_bias;
|
14577
|
-
memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
|
14578
|
-
|
14579
|
-
//GGML_ASSERT(ne01 + n_past == ne00);
|
14580
|
-
GGML_ASSERT(n_head == ne02);
|
14581
|
-
|
14582
|
-
const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
|
14583
|
-
|
14584
|
-
const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
|
14585
|
-
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
|
14586
|
-
|
14587
|
-
alibi_f32_sycl(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream);
|
14588
|
-
|
14589
|
-
(void) src1;
|
14590
|
-
(void) src1_dd;
|
14591
|
-
}
|
14592
|
-
|
14593
14528
|
static void ggml_sycl_op_pool2d(const ggml_tensor *src0,
|
14594
14529
|
const ggml_tensor *src1, ggml_tensor *dst,
|
14595
14530
|
const float *src0_dd, const float *src1_dd,
|
@@ -14744,12 +14679,9 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0,
|
|
14744
14679
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
14745
14680
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
14746
14681
|
|
14747
|
-
|
14748
|
-
|
14749
|
-
#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 and src2 support")
|
14682
|
+
#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 support")
|
14750
14683
|
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021")
|
14751
14684
|
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
|
14752
|
-
GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32); // src2 contains positions and it is optional
|
14753
14685
|
|
14754
14686
|
const int64_t ne00 = src0->ne[0];
|
14755
14687
|
const int64_t nrows_x = ggml_nrows(src0);
|
@@ -14761,25 +14693,7 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0,
|
|
14761
14693
|
memcpy(&scale, dst->op_params + 0, sizeof(float));
|
14762
14694
|
memcpy(&max_bias, dst->op_params + 1, sizeof(float));
|
14763
14695
|
|
14764
|
-
|
14765
|
-
float * src2_dd = nullptr;
|
14766
|
-
sycl_pool_alloc<float> src2_f;
|
14767
|
-
|
14768
|
-
const bool use_src2 = src2 != nullptr;
|
14769
|
-
|
14770
|
-
if (use_src2) {
|
14771
|
-
const bool src2_on_device = src2->backend == GGML_BACKEND_TYPE_GPU;
|
14772
|
-
|
14773
|
-
if (src2_on_device) {
|
14774
|
-
ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
|
14775
|
-
src2_dd = (float *) src2_extra->data_device[g_main_device];
|
14776
|
-
} else {
|
14777
|
-
src2_dd = src2_f.alloc(ggml_nelements(src2));
|
14778
|
-
SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
|
14779
|
-
}
|
14780
|
-
}
|
14781
|
-
|
14782
|
-
soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00,
|
14696
|
+
soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00,
|
14783
14697
|
nrows_x, nrows_y, scale, max_bias, main_stream);
|
14784
14698
|
}
|
14785
14699
|
|
@@ -15654,26 +15568,6 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
|
|
15654
15568
|
const int64_t r2 = ne12/ne02;
|
15655
15569
|
const int64_t r3 = ne13/ne03;
|
15656
15570
|
|
15657
|
-
#if 0
|
15658
|
-
// use syclGemmEx
|
15659
|
-
{
|
15660
|
-
for (int i13 = 0; i13 < ne13; ++i13) {
|
15661
|
-
for (int i12 = 0; i12 < ne12; ++i12) {
|
15662
|
-
int i03 = i13 / r3;
|
15663
|
-
int i02 = i12 / r2;
|
15664
|
-
|
15665
|
-
SYCL_CHECK(
|
15666
|
-
syclGemmEx(g_sycl_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
|
15667
|
-
ne01, ne11, ne10,
|
15668
|
-
alpha, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , SYCL_R_16F, nb01/sizeof(half),
|
15669
|
-
(const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, SYCL_R_16F, nb11/sizeof(float),
|
15670
|
-
beta, ( char *) dst_t + i12*nbd2 + i13*nbd3, cu_data_type, ne01,
|
15671
|
-
cu_compute_type,
|
15672
|
-
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
|
15673
|
-
}
|
15674
|
-
}
|
15675
|
-
}
|
15676
|
-
#else
|
15677
15571
|
if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
|
15678
15572
|
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
|
15679
15573
|
SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
|
@@ -15685,7 +15579,6 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
|
|
15685
15579
|
nb11 / nb10, nb12 / nb10, beta,
|
15686
15580
|
(char *)dst_t, cu_data_type, ne01, nb2 / nb0,
|
15687
15581
|
ne12 * ne13, cu_compute_type)));
|
15688
|
-
g_sycl_handles[g_main_device]->wait();
|
15689
15582
|
} else {
|
15690
15583
|
const int ne23 = ne12*ne13;
|
15691
15584
|
|
@@ -15716,7 +15609,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
|
|
15716
15609
|
nb02, nb03, nb12_scaled, nb13_scaled,
|
15717
15610
|
nbd2, nbd3, r2, r3, item_ct1);
|
15718
15611
|
});
|
15719
|
-
})
|
15612
|
+
});
|
15720
15613
|
}
|
15721
15614
|
SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
|
15722
15615
|
*g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans,
|
@@ -15727,9 +15620,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
|
|
15727
15620
|
dpct::library_data_t::real_half, nb11 / nb10, beta,
|
15728
15621
|
(void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23,
|
15729
15622
|
cu_compute_type)));
|
15730
|
-
g_sycl_handles[g_main_device]->wait();
|
15731
15623
|
}
|
15732
|
-
#endif
|
15733
15624
|
|
15734
15625
|
if (no_mixed_dtypes) {
|
15735
15626
|
const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
|
@@ -16230,10 +16121,6 @@ static void ggml_sycl_rope(const ggml_tensor * src0, const ggml_tensor * src1, g
|
|
16230
16121
|
ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_rope);
|
16231
16122
|
}
|
16232
16123
|
|
16233
|
-
static void ggml_sycl_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
16234
|
-
ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_alibi);
|
16235
|
-
}
|
16236
|
-
|
16237
16124
|
static void ggml_sycl_pool2d(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
16238
16125
|
ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_pool2d);
|
16239
16126
|
}
|
@@ -16610,9 +16497,6 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_
|
|
16610
16497
|
case GGML_OP_ROPE:
|
16611
16498
|
func = ggml_sycl_rope;
|
16612
16499
|
break;
|
16613
|
-
case GGML_OP_ALIBI:
|
16614
|
-
func = ggml_sycl_alibi;
|
16615
|
-
break;
|
16616
16500
|
case GGML_OP_IM2COL:
|
16617
16501
|
func = ggml_sycl_im2col;
|
16618
16502
|
break;
|
@@ -17742,7 +17626,6 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
|
|
17742
17626
|
case GGML_OP_DIAG_MASK_INF:
|
17743
17627
|
case GGML_OP_SOFT_MAX:
|
17744
17628
|
case GGML_OP_ROPE:
|
17745
|
-
case GGML_OP_ALIBI:
|
17746
17629
|
case GGML_OP_IM2COL:
|
17747
17630
|
case GGML_OP_POOL_2D:
|
17748
17631
|
case GGML_OP_SUM_ROWS:
|