Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-15 14:50:51 +01:00
Code Issues Packages Projects Releases Wiki Activity

Scale the workgroup count down to allow correct generation for falcon with

Browse Source 
AMD radeon cards with lower workgroup count limit

Partially fixes #1581
This commit is contained in:
Adam Treat 2023-10-27 18:32:29 -04:00 committed by cebtenzzre
parent 89b71278ff
commit e006d377dd
7 changed files with 27 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
ggml-vulkan.cpp
View File

@ -1356,7 +1356,7 @@ void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph
{ {
if (ggml_nelements(src1) == ne10) { if (ggml_nelements(src1) == ne10) {
// src1 is a row // src1 is a row
ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst), ne00); ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
} else { } else {
ggml_vk_add(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4); ggml_vk_add(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4);
} }
@ -1365,7 +1365,7 @@ void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph
{ {
if (ggml_nelements(src1) == ne10) { if (ggml_nelements(src1) == ne10) {
// src1 is a row // src1 is a row
ggml_vk_mulrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst), ne00); ggml_vk_mulrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
} else { } else {
ggml_vk_mul(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4); ggml_vk_mul(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4);
} }
@ -1373,7 +1373,7 @@ void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph
case GGML_OP_SCALE: case GGML_OP_SCALE:
{ {
const float scale = *(const float *) src1->data; const float scale = *(const float *) src1->data;
ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale); ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst)/8, scale);
} break; } break;
case GGML_OP_UNARY: case GGML_OP_UNARY:
switch (ggml_get_unary_op(gf->nodes[i])) { switch (ggml_get_unary_op(gf->nodes[i])) {
@ -1387,7 +1387,7 @@ void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph
} break; } break;
case GGML_UNARY_OP_GELU: case GGML_UNARY_OP_GELU:
{ {
ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst)/4); ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst)/8);
} break; } break;
default: default:
{ {

7
kompute/op_addrow.comp
View File

@ -24,7 +24,10 @@ layout(push_constant) uniform PushConstants {
} pcs; } pcs;
void main() { void main() {
const uint i = gl_WorkGroupID.x; const uint baseIndex = gl_WorkGroupID.x * 4;
out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff]; for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
}
} }

4
kompute/op_gelu.comp
View File

@ -20,9 +20,9 @@ layout(push_constant) uniform PushConstants {
} pcs; } pcs;
void main() { void main() {
const uint baseIndex = gl_WorkGroupID.x * 4; const uint baseIndex = gl_WorkGroupID.x * 8;
for (uint x = 0; x < 4; x++) { for (uint x = 0; x < 8; x++) {
const uint i = baseIndex + x; const uint i = baseIndex + x;
const float y = in_[i + pcs.inOff]; const float y = in_[i + pcs.inOff];
out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y))); out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y)));

7
kompute/op_mulrow.comp
View File

@ -24,7 +24,10 @@ layout(push_constant) uniform PushConstants {
} pcs; } pcs;
void main() { void main() {
const uint i = gl_WorkGroupID.x; const uint baseIndex = gl_WorkGroupID.x * 4;
out_[i + pcs.outOff] = inA[i + pcs.inAOff] * inB[(i % pcs.row) + pcs.inBOff]; for (uint x = 0; x < 4; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = inA[i + pcs.inAOff] * inB[(i % pcs.row) + pcs.inBOff];
}
} }

7
kompute/op_scale.comp
View File

@ -22,7 +22,10 @@ layout(push_constant) uniform PushConstants {
} pcs; } pcs;
void main() { void main() {
const uint i = gl_WorkGroupID.x; const uint baseIndex = gl_WorkGroupID.x * 8;
out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale; for (uint x = 0; x < 8; x++) {
const uint i = baseIndex + x;
out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
}
} }

2
kompute/op_silu.comp
View File

@ -18,8 +18,8 @@ layout(push_constant) uniform PushConstants {
uint inOff; uint inOff;
uint outOff; uint outOff;
} pcs; } pcs;
void main() {
void main() {
const uint baseIndex = gl_WorkGroupID.x * 4; const uint baseIndex = gl_WorkGroupID.x * 4;
for (uint x = 0; x < 4; x++) { for (uint x = 0; x < 4; x++) {

4
kompute/src/Algorithm.cpp
View File

@ -387,6 +387,10 @@ Algorithm::recordDispatch(const vk::CommandBuffer& commandBuffer)
void void
Algorithm::setWorkgroup(const Workgroup& workgroup, uint32_t minSize) Algorithm::setWorkgroup(const Workgroup& workgroup, uint32_t minSize)
{ {
if (workgroup[0] > 65535) {
fprintf(stderr, "workgroup size is %d\n", workgroup[0]);
fflush(stderr);
}
KP_LOG_INFO("Kompute OpAlgoCreate setting dispatch size"); KP_LOG_INFO("Kompute OpAlgoCreate setting dispatch size");