happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: fix failing dequant shaders (#15862) 

* vulkan: fix failing dequant shaders

* add missing const
This commit is contained in:
Jeff Bolz 2025-09-13 16:29:43 +01:00 committed by GitHub
parent b9c9c9f789
commit aa0c461efe
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 13 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp
View File

@ -29,7 +29,7 @@ void main() {
uint qs = data_a[ib].qs[4 * ib32 + l]; uint qs = data_a[ib].qs[4 * ib32 + l];
const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l]; const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l];
qs |= (qh << (8 - 2 * l)) & 0x300; qs |= (qh << (8 - 2 * l)) & 0x300;
const uvec2 grid = iq2s_grid[qs & 511]; const uvec2 grid = iq2s_grid[qs];
const u8vec4 grid0 = unpack8(grid.x); const u8vec4 grid0 = unpack8(grid.x);
const u8vec4 grid1 = unpack8(grid.y); const u8vec4 grid1 = unpack8(grid.y);
data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0)); data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0));

3
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp
View File

@ -33,7 +33,8 @@ void main() {
[[unroll]] for (uint l = 0; l < 4; ++l) { [[unroll]] for (uint l = 0; l < 4; ++l) {
const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7); const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit
const uvec2 grid = iq2xxs_grid[data_a[ib].qs[8 * is + l]]; const uint qs = data_a[ib].qs[8 * is + l];
const uvec2 grid = iq2xxs_grid[qs];
const u8vec4 grid0 = unpack8(grid.x); const u8vec4 grid0 = unpack8(grid.x);
const u8vec4 grid1 = unpack8(grid.y); const u8vec4 grid1 = unpack8(grid.y);
data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0)); data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));

11
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp
View File

@ -22,15 +22,16 @@ void main() {
const uint b_idx = 256 * ib + 32 * is; const uint b_idx = 256 * ib + 32 * is;
const float d = float(data_a[ib].d); const float d = float(data_a[ib].d);
const float db = d * (1 + 2 * ((data_a[ib].scales[is] >> (4 * (is % 2))) & 0xf)); const float db = d * (1 + 2 * ((data_a[ib].scales[is / 2] >> (4 * (is % 2))) & 0xf));
// We must produce 32 values using 4 sign bytes, 1 qh byte, 8 qs bytes. // We must produce 32 values using 4 sign bytes, 1 qh byte, 8 qs bytes.
uint qh = data_a[ib].qh[is]; uint qh = data_a[ib].qh[is];
[[unroll]] for (uint l = 0; l < 8; ++l) { [[unroll]] for (uint l = 0; l < 8; ++l) {
uint qs = data_a[ib].qs[8 * is + l]; const uint iqs = 8 * is + l;
uint gidx = qs | ((qh << (8 - l)) & 256); const uint qs = data_a[ib].qs[iqs];
uint8_t signs = data_a[ib].signs[8 * is + l / 2] >> (4 * (l & 1)); const uint gidx = qs | ((qh << (8 - l)) & 256);
u8vec4 grid = unpack8(iq3s_grid[gidx]); const uint8_t signs = data_a[ib].signs[iqs / 2] >> (4 * (l & 1));
const u8vec4 grid = unpack8(iq3s_grid[gidx]);
data_b[b_idx + 4 * l + 0] = D_TYPE(db * grid.x * ((signs & 1) != 0 ? -1.0 : 1.0)); data_b[b_idx + 4 * l + 0] = D_TYPE(db * grid.x * ((signs & 1) != 0 ? -1.0 : 1.0));
data_b[b_idx + 4 * l + 1] = D_TYPE(db * grid.y * ((signs & 2) != 0 ? -1.0 : 1.0)); data_b[b_idx + 4 * l + 1] = D_TYPE(db * grid.y * ((signs & 2) != 0 ? -1.0 : 1.0));
data_b[b_idx + 4 * l + 2] = D_TYPE(db * grid.z * ((signs & 4) != 0 ? -1.0 : 1.0)); data_b[b_idx + 4 * l + 2] = D_TYPE(db * grid.z * ((signs & 4) != 0 ? -1.0 : 1.0));

6
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp
View File

@ -35,8 +35,10 @@ void main() {
const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7); const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
// Restore parity bit. // Restore parity bit.
const uint sign8 = sign7 | (bitCount(sign7) << 7); const uint sign8 = sign7 | (bitCount(sign7) << 7);
const u8vec4 grid0 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l]]); const uint qs0 = data_a[ib].qs[8 * is + 2 * l];
const u8vec4 grid1 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l + 1]]); const uint qs1 = data_a[ib].qs[8 * is + 2 * l + 1];
const u8vec4 grid0 = unpack8(iq3xxs_grid[qs0]);
const u8vec4 grid1 = unpack8(iq3xxs_grid[qs1]);
data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0)); data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0)); data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0)); data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0));