Files
LocalAI/backend
Ettore Di Giacinto e983919516 feat(paged): route GQA-grouped tile kernel by default for paged decode (patch 0011)
Increment 3 attention lever. In the paged in-kernel decode dispatch, route the
common grouped-query F16 case to the tile kernel and keep the inc-1 vec kernel
for everything else. Tile groups the q-heads that share a kv-head (ncols2) so
each K/V row is loaded once per group instead of once per q-head, and runs at
higher occupancy (108-128 regs vs vec 168 -> 25%). On GB10 (Qwen3-32B NVFP4,
F16 cache, gqa 8, batch 32, 1024 ctx, same build, env-toggled) this cuts the
decode step from 186.3 to 177.9 ms/step (-4.5%), within 1.8% of stock (174.8).
The win grows with context (tile vs vec decode step, npl=8): 1024 -2.3%, 4096
-3.3%, 8192 -4.1%, 16384 -6.1%, as attention takes a larger share of the step.

Routing guard: tile has no K/V type template (loads half2), so a non-F16 cache
would be converted to a contiguous F16 copy by launch_fattn, breaking the
in-kernel block-table read. So tile is correct only for an F16 cache, and the
grouping only helps at gqa>=2. tile is used only for {F16 K and V, gqa_ratio>=2};
everything else falls back to the inc-1 vec path, exactly as before this change.
LLAMA_KV_PAGED_VEC=1 forces vec for A/B. The inc-2 phys(j) tile read (patch 0010)
was already plumbed; this only adds the default route. (Paged decode currently
needs an F16 cache; quantized + paged is a pre-existing limitation unaffected by
this change: stock+q8_0 works, paged+q8_0 aborts both before and after.)

Split-K was ruled out: the vec decode grid is already block-saturated (~43 waves
over 144 resident on 48 SM), so more parallel_blocks adds no SM fill; the
under-saturation is intra-SM occupancy + 8x KV re-streaming, which GQA grouping
attacks directly.

Validated (greedy): CPU plumbing gate (0.6B, build-cpu, paged-on vs off)
byte-identical; GPU 0.6B gqa=2 tile token-coherent with the inc-1 vec path
(7/8 sequences identical, 8th in the same kernel-noise band where vec also
drifts from stock); 32B gqa=8 tile tracks stock at least as well as vec. Stock
(no block table) is byte-identical: the dispatch guard only diverts on src[5].
Full rationale and numbers in the patch header.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:opus-4.8 [Claude Code]
2026-06-22 22:38:28 +00:00
..

LocalAI Backend Architecture

This directory contains the core backend infrastructure for LocalAI, including the gRPC protocol definition, multi-language Dockerfiles, and language-specific backend implementations.

Overview

LocalAI uses a unified gRPC-based architecture that allows different programming languages to implement AI backends while maintaining consistent interfaces and capabilities. The backend system supports multiple hardware acceleration targets and provides a standardized way to integrate various AI models and frameworks.

Architecture Components

1. Protocol Definition (backend.proto)

The backend.proto file defines the gRPC service interface that all backends must implement. This ensures consistency across different language implementations and provides a contract for communication between LocalAI core and backend services.

Core Services

  • Text Generation: Predict, PredictStream for LLM inference
  • Embeddings: Embedding for text vectorization
  • Image Generation: GenerateImage for stable diffusion and image models
  • Audio Processing: AudioTranscription, TTS, SoundGeneration
  • Video Generation: GenerateVideo for video synthesis
  • Object Detection: Detect for computer vision tasks
  • Vector Storage: StoresSet, StoresGet, StoresFind for RAG operations
  • Reranking: Rerank for document relevance scoring
  • Voice Activity Detection: VAD for audio segmentation

Key Message Types

  • PredictOptions: Comprehensive configuration for text generation
  • ModelOptions: Model loading and configuration parameters
  • Result: Standardized response format
  • StatusResponse: Backend health and memory usage information

2. Multi-Language Dockerfiles

The backend system provides language-specific Dockerfiles that handle the build environment and dependencies for different programming languages:

  • Dockerfile.python
  • Dockerfile.golang
  • Dockerfile.llama-cpp

3. Language-Specific Implementations

Python Backends (python/)

  • transformers: Hugging Face Transformers framework
  • vllm: High-performance LLM inference
  • mlx: Apple Silicon optimization
  • diffusers: Stable Diffusion models
  • Audio: coqui, faster-whisper, kitten-tts
  • Vision: mlx-vlm, rfdetr
  • Specialized: rerankers, chatterbox, kokoro

Go Backends (go/)

  • whisper: OpenAI Whisper speech recognition in Go with GGML cpp backend (whisper.cpp)
  • stablediffusion-ggml: Stable Diffusion in Go with GGML Cpp backend
  • piper: Text-to-speech synthesis Golang with C bindings using rhaspy/piper
  • local-store: Vector storage backend

C++ Backends (cpp/)

  • llama-cpp: Llama.cpp integration
  • grpc: GRPC utilities and helpers

Hardware Acceleration Support

CUDA (NVIDIA)

  • Versions: CUDA 12.x, 13.x
  • Features: cuBLAS, cuDNN, TensorRT optimization
  • Targets: x86_64, ARM64 (Jetson)

ROCm (AMD)

  • Features: HIP, rocBLAS, MIOpen
  • Targets: AMD GPUs with ROCm support

Intel

  • Features: oneAPI, Intel Extension for PyTorch
  • Targets: Intel GPUs, XPUs, CPUs

Vulkan

  • Features: Cross-platform GPU acceleration
  • Targets: Windows, Linux, Android, macOS

Apple Silicon

  • Features: MLX framework, Metal Performance Shaders
  • Targets: M1/M2/M3 Macs

Backend Registry (index.yaml)

The index.yaml file serves as a central registry for all available backends, providing:

  • Metadata: Name, description, license, icons
  • Capabilities: Hardware targets and optimization profiles
  • Tags: Categorization for discovery
  • URLs: Source code and documentation links

Building Backends

Prerequisites

  • Docker with multi-architecture support
  • Appropriate hardware drivers (CUDA, ROCm, etc.)
  • Build tools (make, cmake, compilers)

Build Commands

Example of build commands with Docker

# Build Python backend
docker build -f backend/Dockerfile.python \
  --build-arg BACKEND=transformers \
  --build-arg BUILD_TYPE=cublas12 \
  --build-arg CUDA_MAJOR_VERSION=12 \
  --build-arg CUDA_MINOR_VERSION=0 \
  -t localai-backend-transformers .

# Build Go backend
docker build -f backend/Dockerfile.golang \
  --build-arg BACKEND=whisper \
  --build-arg BUILD_TYPE=cpu \
  -t localai-backend-whisper .

# Build C++ backend
docker build -f backend/Dockerfile.llama-cpp \
  --build-arg BACKEND=llama-cpp \
  --build-arg BUILD_TYPE=cublas12 \
  -t localai-backend-llama-cpp .

For ARM64/Mac builds, docker can't be used, and the makefile in the respective backend has to be used.

Build Types

  • cpu: CPU-only optimization
  • cublas12, cublas13: CUDA 12.x, 13.x with cuBLAS
  • hipblas: ROCm with rocBLAS
  • intel: Intel oneAPI optimization
  • vulkan: Vulkan-based acceleration
  • metal: Apple Metal optimization

Backend Development

Creating a New Backend

  1. Choose Language: Select Python, Go, or C++ based on requirements
  2. Implement Interface: Implement the gRPC service defined in backend.proto
  3. Add Dependencies: Create appropriate requirements files
  4. Configure Build: Set up Dockerfile and build scripts
  5. Register Backend: Add entry to index.yaml
  6. Test Integration: Verify gRPC communication and functionality

Backend Structure

backend-name/
├── backend.py/go/cpp    # Main implementation
├── requirements.txt      # Dependencies
├── Dockerfile           # Build configuration
├── install.sh           # Installation script
├── run.sh              # Execution script
├── test.sh             # Test script
└── README.md           # Backend documentation

Required gRPC Methods

At minimum, backends must implement:

  • Health() - Service health check
  • LoadModel() - Model loading and initialization
  • Predict() - Main inference endpoint
  • Status() - Backend status and metrics

Integration with LocalAI Core

Backends communicate with LocalAI core through gRPC:

  1. Service Discovery: Core discovers available backends
  2. Model Loading: Core requests model loading via LoadModel
  3. Inference: Core sends requests via Predict or specialized endpoints
  4. Streaming: Core handles streaming responses for real-time generation
  5. Monitoring: Core tracks backend health and performance

Performance Optimization

Memory Management

  • Model Caching: Efficient model loading and caching
  • Batch Processing: Optimize for multiple concurrent requests
  • Memory Pinning: GPU memory optimization for CUDA/ROCm

Hardware Utilization

  • Multi-GPU: Support for tensor parallelism
  • Mixed Precision: FP16/BF16 for memory efficiency
  • Kernel Fusion: Optimized CUDA/ROCm kernels

Troubleshooting

Common Issues

  1. GRPC Connection: Verify backend service is running and accessible
  2. Model Loading: Check model paths and dependencies
  3. Hardware Detection: Ensure appropriate drivers and libraries
  4. Memory Issues: Monitor GPU memory usage and model sizes

Contributing

When contributing to the backend system:

  1. Follow Protocol: Implement the exact gRPC interface
  2. Add Tests: Include comprehensive test coverage
  3. Document: Provide clear usage examples
  4. Optimize: Consider performance and resource usage
  5. Validate: Test across different hardware targets