LLM-Inference/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-08-12 22:31:22 +04:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: fix l2_norm epsilon handling (llama/20350)

Browse Source
This commit is contained in:
Jeff Bolz
2026-03-16 13:10:15 +02:00
committed by Georgi Gerganov
parent 6c5e3aac3e
commit 86e312d61d
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+1 -1
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) {
ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp
+1 -1
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]));