Optimized LLM Serving - vLLM and TGI

# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# vLLM -- high-throughput LLM serving
# pip install vllm
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

from vllm import LLM, SamplingParams
from vllm.distributed import destroy_model_parallel

# THREE KEY INNOVATIONS OF vLLM:
vllm_innovations = {
    "PagedAttention": '''KV cache fragmentation is the #1 LLM memory issue.
HuggingFace: allocates max_length KV cache upfront per request (wasteful).
vLLM: manages KV cache like OS virtual memory -- allocate on demand in pages.
Result: 2-3x more concurrent requests fit in GPU memory.''',

    "Continuous Batching": '''Static batching: wait for a batch to fill, process together.
Problem: short sequences finish but long ones continue -- GPU sits idle.
Continuous batching: as soon as one sequence finishes, add a new one to the batch.
Result: GPU utilization 80-95% (vs 30-50% with static batching).''',

    "Speculative Decoding": '''Speed up generation by predicting multiple tokens with a small model,
then verify with the large model in one forward pass.
If speculation correct: multiple tokens per step.
Net effect: 2-3x speedup for outputs with predictable patterns (code, tables).''',
}

for innovation, explanation in vllm_innovations.items():
    print(f"{innovation}:")
    print(f"  {explanation[:100]}...")

# OFFLINE INFERENCE (batch jobs)
llm = LLM(
    model="meta-llama/Meta-Llama-3-8B-Instruct",
    tensor_parallel_size=2,     # split model across 2 GPUs
    max_model_len=8192,
    gpu_memory_utilization=0.9, # use 90% of GPU memory for KV cache
    dtype="bfloat16",
    quantization="awq",         # or "gptq"  -- optional quantization
)

sampling = SamplingParams(
    temperature=0.7, top_p=0.9, max_tokens=512,
    stop=["<|eot_id|>"],  # LLaMA 3 end-of-turn token
)

# vLLM handles batching automatically -- send 1000 requests, get 1000 responses
prompts = [f"Question {i}: What is AI?" for i in range(100)]
outputs = llm.generate(prompts, sampling)

for output in outputs[:3]:
    print(f"  Prompt: {output.prompt[:50]}...")
    print(f"  Response: {output.outputs[0].text[:100]}...")
    print()

# ONLINE SERVING (API server)
# vllm serve meta-llama/Meta-Llama-3-8B-Instruct --tensor-parallel-size 2 --port 8000
# This creates an OpenAI-compatible API:
from openai import OpenAI
vllm_client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")

response = vllm_client.chat.completions.create(
    model="meta-llama/Meta-Llama-3-8B-Instruct",
    messages=[{"role": "user", "content": "Explain transformer attention"}],
    temperature=0.7, max_tokens=300,
)

# Throughput comparison:
print("\nvLLM vs HuggingFace Throughput Comparison (A100, LLaMA-3-8B):")
serving_comparison = {
    "HuggingFace naive": "50-100 tokens/sec",
    "HuggingFace + AMP": "150-200 tokens/sec",
    "vLLM basic":        "800-1200 tokens/sec",
    "vLLM + AWQ INT4":   "2000-3000 tokens/sec",
}
for system, throughput in serving_comparison.items():
    print(f"  {system:30s}: {throughput}")