ggml-cuda : use graph allocator (#2684)

use a different function for no_alloc to avoid breaking backwards compat, fixes lora

remove 512 n_batch limit

fixed 2048 batch size

cleanup

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

1 files changed, 5 insertions, 0 deletions

diff --git a/ggml-cuda.h b/ggml-cuda.h
index cad05f5f..f66bb167 100644
--- a/ggml-cuda.h
+++ b/ggml-cuda.h
@@ -16,9 +16,14 @@ GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const str
 GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);
 GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+
 GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
+
+GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
+GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
+
 GGML_API void   ggml_cuda_set_main_device(int main_device);
 GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
 GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);