vulkan: handle rope with large number of rows (#18306)

This commit is contained in:
Jeff Bolz
2025-12-26 09:53:46 -06:00
committed by GitHub
parent 4893cc07bb
commit 10dc500bdb
7 changed files with 34 additions and 7 deletions
@@ -6,6 +6,9 @@
void main() {
const uint i0 = 2*gl_GlobalInvocationID.y;
// i1 is actually i2*nb2+i1, but the rows are contiguous
const uint i1 = gl_GlobalInvocationID.x;
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
if (i1 >= pc.nrows) {
return;
}
rope_multi(i0, i1, pc);
}
@@ -6,6 +6,9 @@
void main() {
const uint i0 = 2*gl_GlobalInvocationID.y;
// i1 is actually i2*nb2+i1, but the rows are contiguous
const uint i1 = gl_GlobalInvocationID.x;
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
if (i1 >= pc.nrows) {
return;
}
rope_neox(i0, i1, pc);
}
@@ -6,6 +6,9 @@
void main() {
const uint i0 = 2*gl_GlobalInvocationID.y;
// i1 is actually i2*nb2+i1, but the rows are contiguous
const uint i1 = gl_GlobalInvocationID.x;
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
if (i1 >= pc.nrows) {
return;
}
rope_norm(i0, i1, pc);
}
@@ -6,6 +6,7 @@
struct rope_params {
uint rope_mode;
uint ncols;
uint nrows;
uint n_dims;
float freq_scale;
uint p_delta_rows;
@@ -6,6 +6,9 @@
void main() {
const uint i0 = 2*gl_GlobalInvocationID.y;
// i1 is actually i2*nb2+i1, but the rows are contiguous
const uint i1 = gl_GlobalInvocationID.x;
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
if (i1 >= pc.nrows) {
return;
}
rope_vision(i0, i1, pc);
}