vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source
This commit is contained in:
Concedo 2025-03-22 09:42:33 +08:00
commit 27bd7b95f5
7 changed files with 33 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -2285,13 +2285,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
if (device->float_controls_rte_fp16) {
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
} else {
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
}
ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);

5
ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp
View file

@ -1,5 +1,10 @@
#version 450
#if RTE16
#extension GL_EXT_spirv_intrinsics : enable
spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
#endif // RTE16
#include "types.comp"
#include "generic_unary_head.comp"

4
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
View file

@ -82,8 +82,8 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1]));
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const i8vec2 v0 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2]);
const i8vec2 v1 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2 + 1]);
const i8vec2 v0 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2])).xy; // vec4 used due to #12147
const i8vec2 v1 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2 + 1])).xy;
return vec4(v0.x, v0.y, v1.x, v1.y);
}
#endif

4
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp
View file

@ -19,8 +19,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
const float db = d * (0.5 + scale) * 0.25;
const uint qh = data_a[ibi].qh[ib32];
const u8vec2 qs16 = unpack8(data_a_packed16[ibi].qs[itid]);
const u8vec2 sign16 = unpack8(data_a_packed16[ibi].qs[QUANT_K / 16 + itid]);
const u8vec2 qs16 = unpack8(uint32_t(data_a_packed16[ibi].qs[itid])).xy; // vec4 used due to #12147
const u8vec2 sign16 = unpack8(uint32_t(data_a_packed16[ibi].qs[QUANT_K / 16 + itid])).xy;
[[unroll]] for (uint l = 0; l < 2; ++l) {
const uint8_t sign = sign16[l];
const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300);

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp
View file

@ -21,7 +21,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
sum[j] = 0.0;
}
[[unroll]] for (uint l = 0; l < 4; ++l) {
const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]);
const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147
const uint sign = data_a[ibi].signs[4 * ib32 + l];
const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)]));

14
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View file

@ -336,8 +336,8 @@ void main() {
const uint iqs = idx & 0x07;
const float d = float(data_a_packed16[ib].d);
const i8vec2 v0 = unpack8(data_a_packed16[ib].qs[2*iqs]);
const i8vec2 v1 = unpack8(data_a_packed16[ib].qs[2*iqs + 1]);
const i8vec2 v0 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs])).xy; // vec4 used due to #12147
const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy;
const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
@ -544,7 +544,7 @@ void main() {
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@ -564,7 +564,7 @@ void main() {
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@ -586,7 +586,7 @@ void main() {
const float db = d * 0.25 * (0.5 + scale);
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
const vec2 v = db * vec2(sign01) * vec2(unpack8(uint32_t(grid)).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@ -611,7 +611,7 @@ void main() {
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@ -631,7 +631,7 @@ void main() {
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);

1
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View file

@ -451,6 +451,7 @@ void process_shaders() {
for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
}