vulkan: fix l2_norm epsilon handling (llama/20350)

diff --git a/src/ggml-vulkan/ggml-vulkan.cpp b/src/ggml-vulkan/ggml-vulkan.cpp
index 6574955cf103dfa987c7cccb33380b7236efc8a7..ce3c85e7589a65a31e2431cbec4b090c17892f5d 100644 (file)
--- a/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/src/ggml-vulkan/ggml-vulkan.cpp
@@ -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) {

diff --git a/src/ggml-vulkan/vulkan-shaders/l2_norm.comp b/src/ggml-vulkan/vulkan-shaders/l2_norm.comp
index 7d0a1de0df9bd5ac5ba0a05921ae796816ed3a08..f9af46744df41759e40974288f7e3700dfbc6d9e 100644 (file)
--- a/src/ggml-vulkan/vulkan-shaders/l2_norm.comp
+++ b/src/ggml-vulkan/vulkan-shaders/l2_norm.comp
@@ -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]));

diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 58d67d97f8d439426d4952066495a4245a52a41a..a821655d10c7b28bd70d042ffd2b801d171a5dfc 100644 (file)
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -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));