vllm-offline

name	vllm-offline
description	Offline inference, batch processing, script integration
triggers	["When user wants to run offline inference without starting a server","When user needs batch processing for datasets","When user wants to integrate vLLM into Python scripts","When user needs synchronous or asynchronous inference"]

vllm-offline

Overview

vLLM's LLM class provides a simple Python API for offline inference. This skill covers the LLM class usage, batch processing, and integration into Python scripts.

Prerequisites

• vLLM installed (pip install vllm)
• Sufficient GPU memory for the target model
• Input data prepared for processing

Main Workflow

Step 1: Basic Synchronous Inference

from vllm import LLM, SamplingParams

# Initialize model
llm = LLM(model="meta-llama/Llama-2-7b-chat-hf")

# Define sampling parameters
sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.95,
    max_tokens=100
)

# Single prompt
prompt = "What is machine learning?"
output = llm.generate(prompt, sampling_params)
print(output[0].outputs[0].text)

Step 2: Batch Processing

prompts = [
    "What is AI?",
    "Explain Python programming",
    "Define cloud computing",
    "What is deep learning?"
]

# Process all prompts in batch
outputs = llm.generate(prompts, sampling_params)

for prompt, output in zip(prompts, outputs):
    print(f"Prompt: {prompt}")
    print(f"Response: {output.outputs[0].text}")
    print("---")

Step 3: Async Inference

import asyncio
from vllm import AsyncLLMEngine, AsyncEngineArgs, SamplingParams

# Initialize async engine
engine_args = AsyncEngineArgs(
    model="meta-llama/Llama-2-7b-chat-hf"
)
engine = AsyncLLMEngine.from_engine_args(engine_args)

async def generate_async(prompt):
    sampling_params = SamplingParams(temperature=0.7, max_tokens=100)
    async for output in engine.generate(prompt, sampling_params, request_id="1"):
        final_output = output
    return final_output

# Run async generation
result = asyncio.run(generate_async("What is AI?"))
print(result.outputs[0].text)

Step 4: Save Results

import json

outputs = llm.generate(prompts, sampling_params)

results = []
for prompt, output in zip(prompts, outputs):
    results.append({
        "prompt": prompt,
        "response": output.outputs[0].text,
        "tokens": len(output.outputs[0].token_ids)
    })

# Save to JSON
with open("outputs.json", "w") as f:
    json.dump(results, f, indent=2)

Common Patterns

Pattern 1: Process Large Dataset

import pandas as pd
from vllm import LLM, SamplingParams

# Load dataset
df = pd.read_csv("input.csv")

# Initialize model once
llm = LLM(model="meta-llama/Llama-2-7b-chat-hf")
sampling_params = SamplingParams(temperature=0.7, max_tokens=150)

# Process in batches
batch_size = 32
results = []

for idx in range(0, len(df), batch_size):
    batch = df.iloc[idx:idx+batch_size]
    prompts = batch["input_text"].tolist()

    outputs = llm.generate(prompts, sampling_params)
    responses = [o.outputs[0].text for o in outputs]

    results.extend(responses)
    print(f"Processed {min(idx+batch_size, len(df))}/{len(df)}")

df["output"] = results
df.to_csv("output.csv", index=False)

Pattern 2: Chat Template Application

from vllm import LLM, SamplingParams

llm = LLM(model="meta-llama/Llama-2-7b-chat-hf")

# Apply chat template
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is Python?"}
]

# Tokenizer applies chat template automatically
prompt = llm.get_tokenizer().apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

output = llm.generate(prompt, SamplingParams(max_tokens=100))
print(output[0].outputs[0].text)

Pattern 3: Streaming Output

from vllm import LLM, SamplingParams

llm = LLM(model="meta-llama/Llama-2-7b-chat-hf")

# For streaming, use AsyncLLMEngine
# See Step 3 for async example

# Non-streaming with progress tracking
prompts = [f"Question {idx}" for idx in range(100)]
outputs = llm.generate(prompts, SamplingParams(max_tokens=50))

for idx, output in enumerate(outputs):
    print(f"[{idx+1}/{len(prompts)}] Done")

Pattern 4: Multi-GPU Inference

from vllm import LLM

# Tensor parallelism across multiple GPUs
llm = LLM(
    model="meta-llama/Llama-2-70b-chat-hf",
    tensor_parallel_size=4  # Use 4 GPUs
)

output = llm.generate("Hello, how are you?")
print(output[0].outputs[0].text)

Troubleshooting

Problem: CUDA Out of Memory

Solution:

# Reduce GPU memory utilization
llm = LLM(
    model="meta-llama/Llama-2-7b-chat-hf",
    gpu_memory_utilization=0.8,  # Default is 0.9
    max_model_len=4096  # Reduce max sequence length
)

# Or use quantization
llm = LLM(
    model="meta-llama/Llama-2-7b-chat-hf",
    quantization="awq"  # or "gptq", "fp8"
)

Problem: Slow first inference

Solution:

First inference includes model loading time. Warm up:

# Warm up with a dummy prompt
llm = LLM(model="meta-llama/Llama-2-7b-chat-hf")
_ = llm.generate("Hello", SamplingParams(max_tokens=10))

# Now actual inference is faster
outputs = llm.generate(actual_prompts)

Problem: AsyncEngineDeadError

Solution:

# Ensure proper cleanup
import asyncio
from vllm import AsyncLLMEngine

async def main():
    engine = AsyncLLMEngine.from_engine_args(engine_args)
    try:
        result = await generate_async(prompt)
        return result
    finally:
        # Clean up
        await engine.abort_all_requests()

asyncio.run(main())

References

• LLM Class API - Detailed LLM class documentation
• Batch Processing Best Practices - Optimizing batch inference