API Reference Overview¶

Complete API documentation for llcuda v2.2.0.

Main Components¶

llcuda provides a simple, PyTorch-style API for GPU-accelerated LLM inference.

Choose Your Interface¶

Pick the workflow that fits your use case:

• OpenAI-compatible client → LlamaCppClient for chat/completions APIs
• Simple engine wrapper → InferenceEngine for quick local inference

Core Classes¶

Utility Functions¶

Quick API Reference¶

Basic Usage¶

import llcuda

# Create engine
engine = llcuda.InferenceEngine()

# Load model
engine.load_model(
    "unsloth/gemma-3-1b-it-GGUF:gemma-3-1b-it-Q4_K_M.gguf",
    silent=True
)

# Run inference
result = engine.infer("What is AI?", max_tokens=100)

# Access results
print(result.text)                    # Generated text
print(result.tokens_per_sec)          # Speed in tokens/sec
print(result.latency_ms)              # Latency in milliseconds
print(result.tokens_generated)        # Number of tokens generated

InferenceEngine Methods¶

__init__(server_url=None)¶

Create a new inference engine instance.

Parameters: - server_url (str, optional): Custom llama-server URL. Default: http://127.0.0.1:8090

load_model(model_path, silent=False, auto_start=True, **kwargs)¶

Load a GGUF model for inference.

Parameters: - model_path (str): Model identifier or path - HuggingFace: "unsloth/repo-name:filename.gguf" - Registry: "gemma-3-1b-Q4_K_M" - Local: "/path/to/model.gguf" - silent (bool): Suppress llama-server output. Default: False - auto_start (bool): Start server automatically. Default: True - **kwargs: Additional options (context_size, gpu_layers, etc.)

infer(prompt, max_tokens=512, temperature=0.7, **kwargs)¶

Run inference on a single prompt.

Parameters: - prompt (str): Input text - max_tokens (int): Maximum tokens to generate. Default: 512 - temperature (float): Sampling temperature. Default: 0.7 - top_p (float): Nucleus sampling threshold. Default: 0.9 - top_k (int): Top-k sampling. Default: 40 - stop_sequences (list): Stop generation at these sequences

Returns: - InferenceResult: Result object with text and metrics

batch_infer(prompts, max_tokens=512, **kwargs)¶

Run inference on multiple prompts.

Parameters: - prompts (list[str]): List of input texts - max_tokens (int): Maximum tokens per prompt - **kwargs: Same as infer()

Returns: - list[InferenceResult]: List of results

get_metrics()¶

Get aggregated performance metrics.

Returns: - dict: Metrics dictionary with throughput and latency stats

InferenceResult Attributes¶

Utility Functions¶

check_gpu_compatibility()¶

Check if current GPU is compatible with llcuda.

Returns:

{
    'gpu_name': str,          # e.g., "Tesla T4"
    'compute_capability': str, # e.g., "7.5"
    'compatible': bool,       # True if supported
    'platform': str          # e.g., "kaggle", "local"
}

Example:

compat = llcuda.check_gpu_compatibility()
if compat['compatible']:
    print(f"✅ {compat['gpu_name']} is compatible!")
else:
    print(f"⚠️ {compat['gpu_name']} may not work")

Detailed Documentation¶

• InferenceEngine - Complete InferenceEngine documentation
• Models & GGUF - Model loading and GGUF format
• GPU & Device - GPU management and compatibility
• Examples - Code examples and use cases

See Also¶

• Quick Start Guide
• Tutorials
• Performance Benchmarks