# tile-ai/TileRT

Tile-Based Runtime for Ultra-Low-Latency LLM Inference

Repository: https://github.com/tile-ai/TileRT
Canonical: https://ross.abutalabs.com/products/tilert
Homepage: https://www.tilert.ai
Language: Python
License: MIT
License Family: permissive
Last push: 2026-08-13T13:21:51+00:00

## Health v2 (maintenance only)
Score: 77/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 97, release rhythm 84, longevity 21
- inputs: {"age_days": 295, "days_push": 20, "days_rel": 27, "gap_med": 33, "n_releases_24m": 6}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1738, forks 121 (observed 2026-08-28T04:05:29.743149+00:00)

## What it is
TileRT is a tile-based runtime for ultra-low-latency LLM inference, achieving hundreds to over 1000 tokens/s decode speeds on frontier models like DeepSeek-V3.2 and GLM-5 using model-system co-design on multi-GPU nodes. It integrates with vLLM via prefill-decode disaggregation behind an OpenAI-compatible endpoint.

## Use cases
- run frontier LLMs with millisecond single-stream decode latency
- serve GLM-5 or DeepSeek-V3.2 at 500-1000+ tokens/s
- combine vLLM prefill with TileRT decode via PD disaggregation
- use multi-token prediction to boost decoding throughput
- sustain high tokens/s at very long context lengths up to 200K

## When to choose
- you need maximum single-stream decode speed for large MoE models on NVIDIA GPUs
- you want an OpenAI-compatible endpoint with vLLM prefill and TileRT decode
- you are building latency-sensitive agentic or real-time applications

## When to avoid
- you need broad model coverage beyond the supported GLM/DeepSeek/MiMo models
- you need a general-purpose serving stack with heavy multi-user batching
- you run on non-NVIDIA hardware or small consumer GPUs

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, gpu-computing, machine-learning
- domain: large-language-models, deep-learning, gpu-computing, performance
- platform: python
- tags: inference-runtime, low-latency, decode-optimization, multi-token-prediction, pd-disaggregation, vllm-integration, linux, gpu, docker

## Member repositories
- tile-ai/TileRT (main) score 77

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:29.743149+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:30:37.131662+00:00, confidence not recorded.
  - readme: https://github.com/tile-ai/TileRT (fetched 2026-08-28T04:05:29.743149+00:00, sha d657b2a9f0fd)
  - homepage: https://www.tilert.ai (fetched 2026-08-29T11:07:44.203317+00:00, sha 604707136d24)
  - registry_pypi: https://pypi.org/pypi/tilert/json (fetched 2026-08-29T11:07:44.205709+00:00, sha 6ea12521253c)
- Data as of 2026-08-30T08:39:29.467469+00:00.