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 l2_norm epsilon handling (#20350)

This commit is contained in:
Jeff Bolz 2026-03-12 00:39:41 -05:00 committed by GitHub
parent aa429cf507
commit 246ffc4b05
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -16061,7 +16061,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
tensor_clone = ggml_arange(ggml_ctx, start, stop, step);
} else if (tensor->op == GGML_OP_FILL) {
const float value = ggml_get_op_params_f32(tensor, 0);
tensor_clone = ggml_fill(ggml_ctx, tensor_clone, value);
tensor_clone = ggml_fill(ggml_ctx, src_clone[0], value);
} else if (tensor->op == GGML_OP_SQR) {
tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
} else if (tensor->op == GGML_OP_SQRT) {

2
ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp
View File

@ -36,7 +36,7 @@ void main() {
barrier();
}
const FLOAT_TYPE scale = inversesqrt(max(sum[0], FLOAT_TYPE(p.param1)));
const FLOAT_TYPE scale = 1.0f / max(sqrt(sum[0]), FLOAT_TYPE(p.param1));
[[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]));

2
tests/test-backend-ops.cpp
View File

@ -7656,7 +7656,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_softcap(GGML_TYPE_F32, {10, 10, 10, 10}, 50.0f));
test_cases.emplace_back(new test_silu_back());
for (float eps : { 0.0f, 1e-6f, 1e-4f, 1e-1f }) {
for (float eps : { 0.0f, 1e-6f, 1e-4f, 1e-1f, 10.f }) {
for (uint32_t n : { 64, 1025 }) {
for (bool v : { false, true }) {
test_cases.emplace_back(new test_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));