vLLM 내부 구조: 고처리량 LLM 추론 시스템 분석 (2025)

In this post, I'll gradually introduce all of the core system components and advanced features that make up a modern high-throughput LLM inference system. In particular I'll be doing a breakdown of how vLLM [1] works. This post is the first in a series. It starts broad and then layers in detail (following an inverse-pyramid approach) so you can form an accurate high-level mental model of the complete system without drowning in minutiae. Later posts will dive into specific subsystems. This post is structured into five parts: LLM engine & engine core : fundamentals of vLLM (scheduling, paged attention, continuous batching, etc.) Advanced features : chunked prefill, prefix caching, guided & speculative decoding, disaggregated P/D Scaling up : from single-GPU to multi-GPU execution Serving layer : distributed / concurrent web scaffolding Benchmarks and auto-tuning : measuring latency and throughput 📝 Notes Analysis is based on commit 42172ad (August 9th, 2025). Target audience: anyone curious about how state-of-the-art LLM engines work, as well as those interested in contributing to vLLM, SGLang, etc. I'll focus on the V1 engine . I also explored V0 ( now deprecated ), which was valuable for understanding how the project evolved, and many concepts still carry over. The first section on LLM Engine / Engine Core might be a bit overwhelming/dry - but the rest of the blog has plenty examples and visuals. :) LLM Engine & Engine Core The LLM engine is the fundamental building block of vLLM. On its own, it already enables high-throughput inference - but only in an offline setting. You can't serve it to customers over the web yet. We'll use the following offline inference snippet as our running example (adapted from basic.py ). from vllm import LLM , SamplingParams prompts = [ "Hello, my name is" , "The president of the United States is" , ] sampling_params = SamplingParams ( temperature = 0.8 , top_p = 0.95 ) def main ( ) : llm = LLM ( model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0" ) outputs = llm . generate ( prompts , sampling_params ) if __name__ == "__main__" : main ( ) 📝 Environment vars: VLLM_USE_V1="1" # we're using engine V1 VLLM_ENABLE_V1_MULTIPROCESSING="0" # we're running in a single process This configuration is: offline (no web/distributed system scaffolding) synchronous (all execution happens in a single blocking process) single-GPU (no data/model/pipeline/expert parallelism; DP/TP/PP/EP = 1) using standard transformer [2] (supporting hybrid models like Jamba requires a more complex hybrid KV-cache memory allocator) From here, we'll gradually build up to an online, async, multi-GPU, multi-node inference system - but still serving a standard transformer. In this example we do two things, we: Instantiate an engine Call generate on it to sample from the given prompts Let's start analyzing the constructor. LLM Engine constructor The main components of the engine are: vLLM config (contains all of the knobs for configuring model, cache, parallelism, etc.) processor (turns raw inputs → EngineCoreRequests via validation, tokenization, and processing) engine core client (in our running example we're using InprocClient which is basically == EngineCore ; we'll gradually build up to DPLBAsyncMPClient which allows serving at scale) output processor (converts raw EngineCoreOutputs → RequestOutput that the user sees) 📝 Note: With the V0 engine being deprecated, class names and details may shift. I'll emphasize the core ideas rather than exact signatures. I'll abstract away some but not all of those details. Engine core itself is made up of several sub components: Model Executor (drives forward passes on the model, we're currently dealing with UniProcExecutor which has a single Worker process on a single GPU). We'll gradually build up to MultiProcExecutor which supports multiple GPUs Structured Output Manager (used for guided decoding - we'll cover this later) Scheduler (decides which requests go into the next engine step) - it further contains: policy setting - it can be either FCFS (first come first served) or priority (higher priority requests are served first) waiting and running queues KV cache manager - the heart of paged attention [3] The KV-cache manager maintains a free_block_queue - a pool of available KV-cache blocks (often on the order of hundreds of thousands, depending on VRAM size and block size). During paged attention, the blocks serve as the indexing structure that map tokens to their computed KV cache blocks. Core components described in this section and their relationships Block size for a standard transformer layer (non-MLA [4] ) is computed as follows: 2 (key/value) * block_size (default=16) * num_kv_heads * head_size * dtype_num_bytes (e.g. 2 for bf16) During model executor construction, a Worker object is created, and three key procedures are executed. (Later, with MultiProcExecutor , these same procedures run independently on each worker process across different GPUs.) Init device: Assign a CUDA device (e.g. "cuda:0") to the worker and check that the model dtype is supported (e.g. bf16) Verify enough VRAM is available, given the requested gpu_memory_utilization (e.g. 0.8 → 80% of total VRAM) Set up distributed settings (DP / TP / PP / EP, etc.) Instantiate a model_runner (holds the sampler, KV cache, and forward-pass buffers such as input_ids , positions , etc.) Instantiate an InputBatch object (holds CPU-side forward-pass buffers, block tables for KV-cache indexing, sampling metadata, etc.) Load model: Instantiate the model architecture Load the model weights Call model.eval() (PyTorch's inference mode) Optional: call torch.compile() on the model Initialize KV cache Get per-layer KV-cache spec. Historically this was always FullAttentionSpec (homogeneous transformer), but with hybrid models (sliding window, Transformer/SSM like Jamba) it became more complex (see Jenga [5] ) Run a dummy/profiling forward pass and take a GPU memory snapshot to compute how many KV cache blocks fit in available VRAM Allocate, reshape and bind KV cache tensors to attention layers Prepare attention metadata (e.g. set the backend to FlashAttention) later consumed by kernels during the fwd pass Unless --enforce-eager is provided, for each of warmup batch sizes do a dummy run and capture CUDA graphs. CUDA graphs record the whole sequence of GPU work into a DAG. Later during fwd pass we launch/replay pre-baked graphs and cut on kernel launch overhead and thus improve latency. I've abstracted away many low-level details here — but these are the core pieces I'll introduce now, since I'll reference them repeatedly in the following sections. Now that we have the engine initialized let's proceed to the generate function. Generate function The first step is to validate and feed requests into the engine. For each prompt we: Create a unique request ID and capture its arrival time Call an input preprocessor that tokenizes the prompt and returns a dictionary containing prompt , prompt_token_ids , and a type (text, tokens, embeds, etc.) Pack this info into an EngineCoreRequest , adding priority, sampling params, and other metadata Pass the request into the engine core, which wraps it in a Request object and sets its status to WAITING . This request is then added to the scheduler's waiting queue (append if FCFS, or heap-push if priority) At this point the engine has been fed and execution can begin. In the synchronous engine example, these initial prompts are the only ones we'll process — there's no mechanism to inject new requests mid-run. In contrast, the asynchronous engine supports this (aka continuous batching [6] ): after each step, both new and old requests are considered. Because the forward pass flattens the batch into a single sequence and custom kernels handle it efficiently, continuous batching is fundamentally supported even in the synchronous