# sgl-project/SpecForge

Train speculative decoding models effortlessly and port them smoothly to SGLang serving.

Repository: https://github.com/sgl-project/SpecForge
Canonical: https://ross.abutalabs.com/products/specforge
Homepage: https://docs.sglang.ai/SpecForge/
Language: Python
License: MIT
License Family: permissive
Topics: eagle, eagle3, pytorch, sglang, fsdp, training, llm
Last push: 2026-09-03T01:24:11+00:00

## Health v2 (maintenance only)
Score: 64/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 100, release rhythm 35, longevity 32
- inputs: {"age_days": 450, "days_push": 0, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1145, forks 324 (observed 2026-09-03T02:15:17.802562+00:00)

## What it is
SpecForge is a Python framework from the SGLang team for training speculative decoding models such as EAGLE/EAGLE3 draft heads. Trained models port directly to SGLang serving to accelerate LLM inference.

## Use cases
- train an EAGLE3 speculative decoding head for a Llama model
- speed up LLM inference with speculative decoding in SGLang
- run distributed draft model training with FSDP
- train speculative decoding models on AMD ROCm or Ascend NPUs
- prepare datasets for speculative decoding training

## When to choose
- you serve LLMs with SGLang and want faster decoding via speculative decoding
- you need a maintained, out-of-the-box speculative decoding training pipeline
- you want disaggregated or colocated training with parallel topologies

## When to avoid
- you serve models with frameworks other than SGLang
- you need general-purpose LLM fine-tuning rather than speculative decoding draft models
- you have no multi-GPU hardware for training

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, machine-learning, gpu-computing
- domain: large-language-models, deep-learning, machine-learning, developer-tools
- platform: python
- tags: speculative-decoding, eagle3, sglang, pytorch, fsdp, distributed-training, llm-inference-acceleration, gpu, linux, docker

## Member repositories
- sgl-project/SpecForge (main) score 64

## Provenance
- Observed fields: from GitHub, fetched 2026-09-03T02:15:17.802562+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 2026-08-30T06:41:50.061484+00:00, confidence not recorded.
  - readme: https://github.com/sgl-project/SpecForge (fetched 2026-09-03T02:15:17.802562+00:00, sha 87bdeb1b016a)
  - homepage: https://docs.sglang.ai/SpecForge/ (fetched 2026-08-29T12:45:39.744376+00:00, sha 927ebb97ef22)
  - registry_pypi: https://pypi.org/pypi/specforge/json (fetched 2026-08-29T12:45:39.746593+00:00, sha 998e952cb298)
- Data as of 2026-08-30T08:39:29.467469+00:00.