CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: add RTE variants of exp shader (#16165)

Browse Source 
This fixes some failures on Turing where "round to zero" rounds to the max f16
value but the CPU reference value is infinite.
This commit is contained in:
Jeff Bolz
2025-09-22 00:37:17 -05:00
committed by GitHub
parent 4d0a7cbc61
commit a20d810d79
3 changed files with 17 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -3391,7 +3391,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
CREATE_UNARY(exp)
CREATE_UNARY(gelu)
CREATE_UNARY(gelu_erf)
CREATE_UNARY(gelu_quick)
@@ -3403,6 +3402,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
CREATE_UNARY(hardswish)
#undef CREATE_UNARY
#define CREATE_UNARY_RTE(name) \
if (device->float_controls_rte_fp16) { \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16_rte", name ## _f16_rte_len, name ## _f16_rte_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \
} else { \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \
}
CREATE_UNARY_RTE(exp)
#undef CREATE_UNARY_RTE
#define CREATE_GLU(name) \
if (device->float_controls_rte_fp16) { \
ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true); \