Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: fix mmq out of bounds reads (#17108)

Browse Source 
* vulkan: fix mmq out of bounds reads, streamline outdated matmul host code

* fix mul_mat_id quantization call

* Fix compiler warnings
This commit is contained in:
Ruben Ortlam
2025-11-09 09:52:57 +01:00
committed by GitHub
parent 80a6cf6347
commit 8a3519b708
5 changed files with 113 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
+4 -2
View File
@@ -211,7 +211,9 @@ void main() {
const uint iqs = loadr_a;
[[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
block_a_to_shmem(k_step * BM + buf_ib, ib + k_step, iqs);
if (block + k_step * BK < end_k) {
block_a_to_shmem(k_step * BM + buf_ib, ib + k_step, iqs);
}
}
}
[[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) {
@@ -226,7 +228,7 @@ void main() {
const uint iqs = loadr_b;
[[unroll]] for (uint k_step = 0; k_step < BK_STEP; k_step++) {
block_b_to_shmem(k_step * BN + buf_ib, ib + k_step, iqs);
block_b_to_shmem(k_step * BN + buf_ib, ib + k_step, iqs, block + k_step * BK < end_k);
}
}
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl
+22 -11
View File
@@ -469,19 +469,30 @@ ACC_TYPE mmq_dot_product(const uint ib_a) {
#endif
#ifdef MMQ_SHMEM
void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs) {
const uint ib_outer = ib / 4;
const uint ib_inner = ib % 4;
void block_b_to_shmem(const uint buf_ib, const uint ib, const uint iqs, const bool is_in_bounds) {
if (is_in_bounds) {
const uint ib_outer = ib / 4;
const uint ib_inner = ib % 4;
if (iqs == 0) {
buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
if (iqs == 0) {
buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
}
const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
buf_b[buf_ib].qs[iqs * 4 ] = values.x;
buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
} else {
if (iqs == 0) {
buf_b[buf_ib].ds = FLOAT_TYPE_VEC2(0.0f);
}
buf_b[buf_ib].qs[iqs * 4 ] = 0;
buf_b[buf_ib].qs[iqs * 4 + 1] = 0;
buf_b[buf_ib].qs[iqs * 4 + 2] = 0;
buf_b[buf_ib].qs[iqs * 4 + 3] = 0;
}
const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
buf_b[buf_ib].qs[iqs * 4 ] = values.x;
buf_b[buf_ib].qs[iqs * 4 + 1] = values.y;
buf_b[buf_ib].qs[iqs * 4 + 2] = values.z;
buf_b[buf_ib].qs[iqs * 4 + 3] = values.w;
}
void block_b_to_registers(const uint ib) {
ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
+1 -1
View File
@@ -61,7 +61,7 @@ void quantize() {
const uint a_idx = ib * 8 + iqs;
vec4 vals = a_idx < p.ne ? data_a[a_idx] : vec4(0.0f);
vec4 vals = a_idx < p.ne / 4 ? data_a[a_idx] : vec4(0.0f);
const vec4 abs_vals = abs(vals);
// Find absolute max for each block