Configuration

class EngineArgs:
    # Model Configuration
    model: str  # HuggingFace model name or local path
    tokenizer: Optional[str] = None  # Tokenizer name/path (defaults to model)
    tokenizer_mode: str = "auto"  # "auto", "slow", or "fast"
    revision: Optional[str] = None  # Model revision/branch
    code_revision: Optional[str] = None  # Code revision for remote code
    tokenizer_revision: Optional[str] = None  # Tokenizer revision
    trust_remote_code: bool = False  # Execute remote code
    download_dir: Optional[str] = None  # Model download directory
    load_format: str = "auto"  # Model loading format
    config_format: str = "auto"  # Config loading format

    # Model Execution
    dtype: str = "auto"  # Model precision ("auto", "half", "float16", "bfloat16", "float32")
    kv_cache_dtype: str = "auto"  # KV cache data type
    quantization_param_path: Optional[str] = None  # Quantization parameters
    quantization: Optional[str] = None  # Quantization method ("awq", "gptq", "squeezellm", "fp8")

    # Memory and Performance
    gpu_memory_utilization: float = 0.9  # GPU memory usage fraction
    swap_space: int = 4  # CPU swap space in GiB
    cpu_offload_gb: float = 0  # CPU offload memory in GB
    max_model_len: Optional[int] = None  # Maximum sequence length
    max_num_batched_tokens: Optional[int] = None  # Maximum batch size in tokens
    max_num_seqs: int = 256  # Maximum concurrent sequences
    max_logprobs: int = 20  # Maximum logprobs to return

    # Parallelism and Distribution
    tensor_parallel_size: int = 1  # Tensor parallelism degree
    pipeline_parallel_size: int = 1  # Pipeline parallelism degree
    max_parallel_loading_workers: Optional[int] = None  # Model loading workers
    ray_workers_use_nsight: bool = False  # Enable Nsight profiling
    block_size: int = 16  # Attention block size
    enable_prefix_caching: bool = False  # Enable prefix caching
    disable_custom_all_reduce: bool = False  # Disable custom all-reduce

    # Advanced Options
    preemption_mode: Optional[str] = None  # Preemption strategy
    num_lookahead_slots: int = 0  # Speculative decoding slots
    seed: int = 0  # Random seed
    num_gpu_blocks_override: Optional[int] = None  # Override GPU block count
    max_seq_len_to_capture: int = 8192  # Maximum sequence length for CUDA graphs
    disable_sliding_window: bool = False  # Disable sliding window attention

    # Multimodal Support
    image_input_type: Optional[str] = None  # Image input format
    image_token_id: Optional[int] = None  # Image token ID
    image_input_shape: Optional[str] = None  # Image input dimensions
    image_feature_size: Optional[int] = None  # Image feature size
    scheduler_delay_factor: float = 0.0  # Scheduler delay factor
    enable_chunked_prefill: Optional[bool] = None  # Chunked prefill optimization

class AsyncEngineArgs(EngineArgs):
    # Inherits all EngineArgs options plus:
    worker_use_ray: bool = False  # Use Ray for distributed workers
    engine_use_ray: bool = False  # Use Ray for engine management
    disable_log_requests: bool = False  # Disable request logging
    max_log_len: Optional[int] = None  # Maximum log length

# Model Data Types
class ModelDType(str, Enum):
    AUTO = "auto"
    HALF = "half"
    FLOAT16 = "float16"
    BFLOAT16 = "bfloat16"
    FLOAT32 = "float32"

# Quantization Methods
class QuantizationMethods(str, Enum):
    AWQ = "awq"
    GPTQ = "gptq"
    SQUEEZELLM = "squeezellm"
    FP8 = "fp8"
    BITSANDBYTES = "bitsandbytes"

# Load Formats
class LoadFormats(str, Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"

class Device(str, Enum):
    GPU = "gpu"
    CPU = "cpu"
    TPU = "tpu"
    XPU = "xpu"

class DeviceConfig:
    device: Device  # Target device type
    device_ids: Optional[List[int]] = None  # Specific device IDs
    placement_group: Optional[str] = None  # Ray placement group

from vllm import LLM, EngineArgs

# Simple GPU setup
args = EngineArgs(
    model="microsoft/DialoGPT-medium",
    tensor_parallel_size=1,
    gpu_memory_utilization=0.8,
    max_model_len=2048
)

llm = LLM(**args.to_dict())

from vllm import LLM, EngineArgs

# Multi-GPU configuration
args = EngineArgs(
    model="microsoft/DialoGPT-large",
    tensor_parallel_size=4,  # Use 4 GPUs
    pipeline_parallel_size=2,  # Pipeline across 2 stages
    gpu_memory_utilization=0.9,
    max_model_len=4096,
    trust_remote_code=True
)

llm = LLM(**args.to_dict())

from vllm import LLM, EngineArgs

# AWQ quantized model
args = EngineArgs(
    model="microsoft/DialoGPT-medium-awq",
    quantization="awq",
    dtype="half",
    gpu_memory_utilization=0.95,
    max_model_len=8192
)

llm = LLM(**args.to_dict())

from vllm import AsyncLLMEngine, AsyncEngineArgs

# Async engine with Ray
async_args = AsyncEngineArgs(
    model="microsoft/DialoGPT-medium",
    worker_use_ray=True,
    engine_use_ray=True,
    tensor_parallel_size=2,
    max_num_seqs=128,
    gpu_memory_utilization=0.9
)

engine = AsyncLLMEngine.from_engine_args(async_args)

from vllm import LLM, EngineArgs

# Optimize for memory efficiency
args = EngineArgs(
    model="microsoft/DialoGPT-small",
    gpu_memory_utilization=0.95,
    swap_space=8,  # 8GB CPU swap
    cpu_offload_gb=2,  # Offload 2GB to CPU
    max_num_batched_tokens=1024,  # Smaller batches
    enable_prefix_caching=True,  # Cache common prefixes
    block_size=8  # Smaller attention blocks
)

llm = LLM(**args.to_dict())

from vllm import LLM, EngineArgs

# Development setup with detailed logging
args = EngineArgs(
    model="microsoft/DialoGPT-medium",
    tensor_parallel_size=1,
    seed=42,  # Reproducible results
    disable_custom_all_reduce=True,  # Use standard operations
    max_logprobs=10,  # Detailed probability info
    trust_remote_code=True,  # For custom models
    revision="main"  # Specific model version
)

llm = LLM(**args.to_dict())

def validate_config(args: EngineArgs) -> None:
    """
    Validate engine configuration parameters.

    Raises:
    ValueError: If configuration is invalid
    RuntimeError: If hardware requirements are not met
    """

def get_default_config_for_device(device: Device) -> EngineArgs:
    """
    Get recommended default configuration for target device.

    Parameters:
    - device: Target deployment device

    Returns:
    EngineArgs with device-optimized defaults
    """

# Key environment variables
VLLM_WORKER_MULTIPROC_METHOD  # Worker process method
VLLM_USE_MODELSCOPE  # Use ModelScope for model downloads
VLLM_TARGET_DEVICE  # Override target device
VLLM_GPU_MEMORY_UTILIZATION  # Default GPU memory usage
VLLM_HOST  # Server host address
VLLM_PORT  # Server port
VLLM_USE_RAY_COMPILED_DAG  # Use compiled Ray DAGs