Implementation:Ollama Ollama Llama Arch

Knowledge Sources	Ollama
Domains	Model Architecture, GGUF
Last Updated	2025-02-15 00:00 GMT

Overview

Defines the mapping tables between model architecture enums, their string names, GGUF key-value identifiers, and tensor name patterns for all supported LLM architectures.

Description

Contains large static std::map tables: LLM_ARCH_NAMES maps llm_arch enum values to string names (e.g., LLM_ARCH_LLAMA to "llama"), LLM_KV_NAMES maps key-value identifiers to GGUF metadata key strings, and per-architecture tensor name maps that associate llm_tensor enum values with their expected tensor name patterns in GGUF files. Each architecture (LLaMA, Falcon, GPT-2, Qwen, Gemma, DeepSeek, Mamba, etc.) has its own tensor mapping entry. Also implements the LLM_KV and LLM_TN helper classes for constructing architecture-specific key/tensor name strings.

Usage

This file is the core registry that enables the model loader to work with dozens of different model architectures. It is used whenever loading, saving, or interpreting GGUF model files.

Code Reference

Source Location

• Repository: Ollama
• File: llama/llama.cpp/src/llama-arch.cpp
• Lines: 1-2499

Signature

static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_LLAMA,    "llama"    },
    { LLM_ARCH_FALCON,   "falcon"   },
    { LLM_ARCH_GPT2,     "gpt2"     },
    { LLM_ARCH_QWEN2,    "qwen2"    },
    { LLM_ARCH_GEMMA,    "gemma"    },
    { LLM_ARCH_DEEPSEEK2,"deepseek2"},
    { LLM_ARCH_MAMBA,    "mamba"    },
    // ... 100+ architectures
};

static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" },
    { LLM_KV_CONTEXT_LENGTH,       "%s.context_length"    },
    { LLM_KV_EMBEDDING_LENGTH,     "%s.embedding_length"  },
    // ... 200+ keys
};

Import

#include "llama-arch.h"

I/O Contract

Inputs

Name	Type	Required	Description
arch	llm_arch	Yes	Architecture enum to look up
kv	llm_kv	Yes	Key-value enum to look up

Outputs

Name	Type	Description
name	const char *	Architecture or key name string
tensor_names	map	Tensor name patterns for a specific architecture

Usage Examples

#include "llama-arch.h"

// Look up architecture name
const char * arch_name = LLM_ARCH_NAMES.at(LLM_ARCH_LLAMA); // "llama"

// Construct architecture-prefixed key
LLM_KV kv(LLM_ARCH_LLAMA);
std::string key = kv(LLM_KV_CONTEXT_LENGTH); // "llama.context_length"

// Construct tensor name with layer index
LLM_TN tn(LLM_ARCH_LLAMA);
std::string tname = tn(LLM_TENSOR_ATTN_Q, "weight", 0); // "blk.0.attn_q.weight"

Related Pages

• Principle:Ollama_Ollama_Model_Architecture_Support