Source: https://seerepo.com/compare/Dao-AILab/flash-attention/vs/vllm-project/vllm
Markdown source: https://seerepo.com/compare/Dao-AILab/flash-attention/vs/vllm-project/vllm.md
Description: flash-attention and vllm share 6 technologies: apache-tvm-ffi, NVIDIA, quack-kernels, torch, einops, and triton. Both are written in Python.

# flash-attention vs vllm

flash-attention and vllm share 6 technologies: apache-tvm-ffi, NVIDIA, quack-kernels, torch, einops, and triton. Both are written in Python.

- [Dao-AILab/flash-attention](https://seerepo.com/r/Dao-AILab/flash-attention): Fast and memory-efficient exact attention
- [vllm-project/vllm](https://seerepo.com/r/vllm-project/vllm): A high-throughput and memory-efficient inference and serving engine for LLMs

## What they share and what sets them apart

- In common: apache-tvm-ffi, NVIDIA, quack-kernels, torch, einops, triton
- Only in flash-attention: torch-c-dlpack-ext, triton-windows
- Only in vllm: FastAPI, axum, gitdb, pyzmq, Hugging Face, Tokio, starlette, Ray, AIOHTTP, tower, mkdocs-material, indicatif, AMD, auto-round-lib, conch-triton-kernels, cuda-bindings, cuda-python, cupy-cuda12x, cupy-cuda13x, flashinfer-python, humming-kernels, lm-format-enforcer, mooncake-transfer-engine, mooncake-transfer-engine-rocm, nvidia-cudnn-frontend, OpenAI, openai-harmony, oss-harmony, sentencepiece, vllm-bench, vllm-chat, vllm-engine-core-client, vllm-llm, vllm-managed-engine, vllm-parser, vllm-proto, vllm-server, vllm-text, vllm-tokenizer, transformers, xgrammar-structural-tag, llm-multimodal, compressed-tensors, safetensors, mistral-common, xgrammar, partial-json-parser, torchvision, lmcache, torchaudio, timm, vllm-xpu-kernels, runai-model-streamer, tokenizers, tokenspeed-mla, llguidance, MCP, outlines-core, tpu-inference, fastokens, fastsafetensors, hf-hub, instanttensor, nixl, riptoken, tekken, tensorizer, tiktoken, tiktoken-rs, markdown, markupsafe, mkdocs-material-extensions, parquet, properdocs, pymdown-extensions, selectolax, xlsxwriter, pillow, image, jinja2, minijinja, torchcodec, minijinja-contrib, annotated-types, Pydantic, msgspec, validator, opentelemetry-semantic-conventions-ai, python-json-logger, vllm-metrics, vllm-tracing, tracing, prometheus-client, psutil, tracing-futures, tracing-subscriber, model-hosting-container-standards, OpenTelemetry, prometheus-fastapi-instrumentator, tracing-log

## Side by side

| | flash-attention | vllm |
|---|---|---|
| Size | 139,521 lines written, 150 code files | 2,606,460 lines written, 3,630 code files |
| Who writes it | 1.7% of its commits signed by AI. With Claude Code, Cursor, and GitHub Copilot | 14.7% of its commits signed by AI. With Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, Devin, Mistral Vibe, CodeRabbit, Factory, and Qwen Code |
| Activity | Last push November 16, 2026, 1,613 commits, 25,304 stars | Last push November 17, 2026, 25,558 commits, 97,060 stars |
| Tests | 80 test files. 32% of its code files have a test that imports them | 2,586 test files. 60.7% of its code files have a test that imports them |

- Size: Lines written by people and agents, without generated ones; code files, without tests.
- Who writes it: A commit counts as AI's when it carries its signature: a Co-Authored-By line, its author name or the merge of its branch.
- Activity: From GitHub and its history, on the day SeeRepo read it.
- Tests: The files that test the code, and the share of code files a test imports.

## Frequently asked questions

### Which is bigger, flash-attention or vllm?

vllm, with 2,606,460 lines written; flash-attention has 139,521.

The proof: [vllm-project/vllm's code at commit 61c3895](https://github.com/vllm-project/vllm/tree/61c3895fc937b4ffc2a59abe188959de1e459c8a)

### Which has more GitHub stars, flash-attention or vllm?

vllm, with 97,060 stars; flash-attention has 25,304.

The proof: [vllm-project/vllm on GitHub](https://github.com/vllm-project/vllm)

### Which has more tests, flash-attention or vllm?

vllm, with 2,586 test files; flash-attention has 80.

The proof: [vllm-project/vllm's code at commit 61c3895](https://github.com/vllm-project/vllm/tree/61c3895fc937b4ffc2a59abe188959de1e459c8a)

### Which was updated more recently, flash-attention or vllm?

vllm: its last push was on November 17, 2026, and flash-attention's on November 16, 2026.

The proof: [vllm-project/vllm's commits on GitHub](https://github.com/vllm-project/vllm/commits)

### Which has more commits signed by AI, flash-attention or vllm?

vllm: an AI signed 14.7% of its commits, with Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, Devin, Mistral Vibe, CodeRabbit, Factory, and Qwen Code; in flash-attention, 1.7%.

The proof: [the commits each tool signed in vllm-project/vllm](https://seerepo.com/r/vllm-project/vllm/ai)

## Sources

- Dao-AILab/flash-attention: Read on November 16, 2026, at commit 328dd7b (https://github.com/Dao-AILab/flash-attention/tree/328dd7b5a6905c75500a9bce334ac012763409a9)
- vllm-project/vllm: Read on November 17, 2026, at commit 61c3895 (https://github.com/vllm-project/vllm/tree/61c3895fc937b4ffc2a59abe188959de1e459c8a)

## What is your repo built with?

Any public repo, with no account: https://seerepo.com/r/owner/repo/built, with the owner and name it has on GitHub.
