Qwen3-8B-PPCM

PPCM is a ~9M three-layer encoder (CCEL → CTIL → CPRL) over Top-7 candidates. It ranks causally consistent paths in one pass, then the target model verifies the top path.

This repo is not a standalone chat model. Load Qwen3-8B as the target.

Code	github.com/Xinmu-Tantai/PPCM
Target	Qwen/Qwen3-8B
Speculative length	7 (block_size = 8), Top-7 candidates
Extra params	8.8M

On Qwen3-8B / GSM8K (T = 0, L = 7): τ = 5.97, 5.18× speedup.

Files

• model.safetensors — 5-layer causal draft + ppcm.ccel / ppcm.ctil / ppcm.cprl / ppcm.score
• config.json — PPCMDraftModel, num_speculative_tokens: 7
• ppcm.py — Hugging Face AutoModel class

Usage

Serve with the PPCM vLLM code:

TARGET_MODEL=Qwen/Qwen3-8B
DRAFT_MODEL=Xinmu7/PPCM
NUM_SPECULATIVE_TOKENS=7

License

Apache License 2.0.