char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Titaniumtownsycl: *glu flat path (#26354)6b5c2efb4

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
2.3 KiB79 linesraw
1//
2// MIT license
3// Copyright (C) 2024 Intel Corporation
4// SPDX-License-Identifier: MIT
5//
6
7//
8// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9// See https://llvm.org/LICENSE.txt for license information.
10// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11//
12
13#ifndef GGML_SYCL_PRESETS_HPP
14#define GGML_SYCL_PRESETS_HPP
15
16#define GGML_SYCL_MAX_STREAMS       8
17#define GGML_SYCL_MAX_BUFFERS       256
18
19#define WARP_SIZE GGML_SYCL_WARP_SIZE
20#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
21
22#define SYCL_COL2IM_1D_BLOCK_SIZE 256
23#define SYCL_TANH_BLOCK_SIZE 256
24#define SYCL_RELU_BLOCK_SIZE 256
25#define SYCL_HARDSIGMOID_BLOCK_SIZE 256
26#define SYCL_HARDSWISH_BLOCK_SIZE 256
27#define SYCL_EXP_BLOCK_SIZE 256
28#define SYCL_NEG_BLOCK_SIZE 256
29#define SYCL_SIGMOID_BLOCK_SIZE 256
30#define SYCL_SQRT_BLOCK_SIZE 256
31#define SYCL_SIN_BLOCK_SIZE 256
32#define SYCL_SQR_BLOCK_SIZE 256
33#define SYCL_SET_BLOCK_SIZE 256
34#define SYCL_CPY_BLOCK_SIZE 32
35#define SYCL_SCALE_BLOCK_SIZE 256
36#define SYCL_CLAMP_BLOCK_SIZE 256
37#define SYCL_ROPE_BLOCK_SIZE 256
38#define SYCL_ALIBI_BLOCK_SIZE 32
39#define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32
40#define SYCL_QUANTIZE_BLOCK_SIZE 256
41#define SYCL_DEQUANTIZE_BLOCK_SIZE 256
42#define SYCL_GET_ROWS_BLOCK_SIZE 256
43#define SYCL_UPSCALE_BLOCK_SIZE 256
44#define SYCL_CONCAT_BLOCK_SIZE 256
45#define SYCL_PAD_BLOCK_SIZE 256
46#define SYCL_ACC_BLOCK_SIZE 256
47#define SYCL_IM2COL_BLOCK_SIZE 256
48#define SYCL_POOL1D_BLOCK_SIZE 256
49#define SYCL_POOL2D_BLOCK_SIZE 256
50#define SYCL_ARGMAX_BLOCK_SIZE 256
51#define SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE 256
52#define SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
53#define SYCL_ARANGE_BLOCK_SIZE 256
54
55// dmmv = dequantize_mul_mat_vec
56#ifndef GGML_SYCL_DMMV_X
57#define GGML_SYCL_DMMV_X 32
58#endif
59#ifndef GGML_SYCL_MMV_Y
60#define GGML_SYCL_MMV_Y 1
61#endif
62
63#ifndef K_QUANTS_PER_ITERATION
64#define K_QUANTS_PER_ITERATION 1
65#else
66static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
67#endif
68
69#ifndef GGML_SYCL_PEER_MAX_BATCH_SIZE
70#define GGML_SYCL_PEER_MAX_BATCH_SIZE 128
71#endif // GGML_SYCL_PEER_MAX_BATCH_SIZE
72
73#define MUL_MAT_SRC1_COL_STRIDE 128
74
75#define QK_WARP_SIZE 32
76#define WARP_32_SIZE 32
77#define WARP_16_SIZE 16
78
79#endif // GGML_SYCL_PRESETS_HPP