ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-28 10:21:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Vulkan: a fresh start (#608)

Browse Source 
* It compiles

* Seems to be working with coopmat

* Vulkan needs f32 precision for flash attention

* Vulkan: fix u_batch > 4096/n_active_experts

for coopmat1. Without this fix we get an assert.
We get the same assert in mainline too.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-07-15 08:03:13 +02:00
committed by GitHub
parent f375799f17
commit 19c57dbe1d
35 changed files with 2342 additions and 596 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/include/ggml-vulkan.h
View File

@@ -10,7 +10,7 @@ extern "C" {
#define GGML_VK_NAME "Vulkan"
#define GGML_VK_MAX_DEVICES 16
GGML_API GGML_CALL void ggml_vk_instance_init(void);
//GGML_API GGML_CALL void ggml_vk_instance_init(void);
// backend API
GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);

9
ggml/include/ggml.h
View File

@@ -884,6 +884,15 @@ extern "C" {
GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
// returns whether the tensor elements are allocated as one contiguous block of memory (no gaps, but permutation ok)
GGML_API bool ggml_is_contiguously_allocated(const struct ggml_tensor * tensor);
// true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
// true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);