Packages

erllama

0.4.0
0.8.0 retired 0.7.0 retired 0.6.2 retired 0.6.1 retired 0.6.0 retired 0.5.1 retired 0.5.0 retired 0.4.0 retired 0.3.0 retired 0.2.0 retired 0.1.2 retired 0.1.1 retired 0.1.0 retired

Native Erlang/OTP wrapper around llama.cpp with token-exact tiered KV cache; cache more warm state than fits in RAM

Retired package: Renamed - Renamed to barrel_inference: https://github.com/barrel-platform/barrel_inference
HexDocs

Current section

Files

Jump to
Readme
13 Versions
0 Dependencies
Dependants
Files
Activity
erllama c_src llama.cpp ggml src ggml-cuda softmax.cuh
Raw

c_src/llama.cpp/ggml/src/ggml-cuda/softmax.cuh 

#include "common.cuh"



#define CUDA_SOFT_MAX_BLOCK_SIZE 1024



void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst);



void ggml_cuda_op_soft_max_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);