Skip to content

Commit 3abf024

Browse files
committed
- remove expert-profile sources from Cmakelist
- fix metal and vulkan backend to support gclip parameter
1 parent 298fae0 commit 3abf024

5 files changed

Lines changed: 7 additions & 5 deletions

File tree

ggml/src/ggml-metal/ggml-metal.metal

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10526,8 +10526,9 @@ kernel void kernel_opt_step_adamw_f32(
1052610526
const float wd = pars[4];
1052710527
const float beta1h = pars[5];
1052810528
const float beta2h = pars[6];
10529+
const float gclip = pars[7];
1052910530

10530-
const float gi = g[gid];
10531+
const float gi = (gclip > 0.0f) ? clamp(g[gid], -gclip, gclip) : g[gid];
1053110532
const float gmi = g_m[gid] * beta1 + gi * (1.0f - beta1);
1053210533
const float gvi = g_v[gid] * beta2 + gi * gi * (1.0f - beta2);
1053310534

ggml/src/ggml-vulkan/ggml-vulkan.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10813,7 +10813,7 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
1081310813
GGML_ASSERT(ggml_are_same_shape(x, g));
1081410814
GGML_ASSERT(ggml_are_same_shape(x, gm));
1081510815
GGML_ASSERT(ggml_are_same_shape(x, gv));
10816-
GGML_ASSERT(ggml_nelements(p) == 7);
10816+
GGML_ASSERT(ggml_nelements(p) == 8);
1081710817

1081810818
vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, g, gm, gv, dst, GGML_OP_OPT_STEP_ADAMW);
1081910819
GGML_ASSERT(pipeline != nullptr);

ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -11,7 +11,7 @@ layout (binding = 0) buffer X {A_TYPE x[];};
1111
layout (binding = 1) readonly buffer G {A_TYPE grad[];};
1212
layout (binding = 2) buffer GM {A_TYPE gradm[];};
1313
layout (binding = 3) buffer GV {A_TYPE gradv[];};
14-
layout (binding = 4) readonly buffer P {float params[7];};
14+
layout (binding = 4) readonly buffer P {float params[8];};
1515

1616
void main() {
1717
const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
@@ -27,8 +27,9 @@ void main() {
2727
const float wd = params[4];
2828
const float beta1h = params[5];
2929
const float beta2h = params[6];
30+
const float gclip = params[7];
3031

31-
const float gi = grad[i];
32+
const float gi = (gclip > 0.0f) ? clamp(grad[i], -gclip, gclip) : grad[i];
3233
const float gmi = gradm[i]*beta1 + gi*(1.0f - beta1);
3334
const float gvi = gradv[i]*beta2 + gi*gi*(1.0f - beta2);
3435

tools/CMakeLists.txt

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -39,5 +39,4 @@ else()
3939
endif()
4040
add_subdirectory(fit-params)
4141
add_subdirectory(results)
42-
add_subdirectory(expert-profile)
4342
endif()

tools/tts/OuteTTS-0.2-500M

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1 @@
1+
Subproject commit 71d6994ffa92c066628d83db99f1298516ece592

0 commit comments

Comments
 (0)