Packages

erllama

0.8.0
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-cpu spacemit spine_mem_pool.h
Raw

c_src/llama.cpp/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h 

#pragma once



#include <cstddef>

#include <cstdint>



namespace ggml::cpu::riscv64_spacemit {



enum class spine_mem_pool_backend : uint8_t {

    none,

    posix_memalign,

    transparent_hugepage,

    hugetlb_1g,

};



struct spine_mem_pool_tcm_info {

    bool   available{ false };

    size_t blk_size{ 0 };

    size_t blk_num{ 0 };

    bool   is_fake_tcm{ false };

};



bool   spine_mem_pool_tcm_init(spine_mem_pool_tcm_info * info) noexcept;

void * spine_mem_pool_tcm_mem_get(int cpu_id) noexcept;

void * spine_mem_pool_tcm_mem_wait(int cpu_id) noexcept;

int    spine_mem_pool_tcm_mem_release(int cpu_id) noexcept;



void * spine_mem_pool_alloc(size_t size, size_t alignment) noexcept;

void * spine_mem_pool_shared_mem_alloc(size_t size, size_t alignment) noexcept;

void   spine_mem_pool_free(void * base) noexcept;

void   spine_mem_pool_shared_mem_free(void * base) noexcept;



}  // namespace ggml::cpu::riscv64_spacemit