SpecSplit
Disaggregated speculative decoding (draft + verification) over gRPC.
Getting Started
- Project guide:
project_guide.md - Architecture overview:
architecture.md
Main Components
- Orchestrator (pipeline coordinator):
specsplit/workers/orchestrator/README.md - Draft Worker (generates speculative token trees):
specsplit/workers/draft/README.md - Target Worker (verifies/accepts token trees):
specsplit/workers/target/README.md
Protocol
spec_decoding.proto:proto/README.md
What to Read Next
- Core modules:
specsplit/core/README.md - Benchmarks & experiments:
benchmarks/README.md - Tests & CI:
tests/README.md - API reference:
api/index.md
GitHub Wiki
If you prefer a smaller “single page per topic” format, see the GitHub Wiki: SpecSplit.wiki.