Flow-Drafter-Qwen3.5-27B

Flow-Drafter-Qwen3.5-27B

A speculative-decoding draft model for Qwen/Qwen3.5-27B, from the Chained-Flow project — the Qwen3.5 sibling of Flow-Drafter-27B (which targets the post-trained Qwen3.6-27B). Same joint-VAE tree-flow recipe.

Predicts all K future hidden states in a single flow pass, expands them into a draft tree, and the base model verifies the whole tree in one forward pass — accepting the longest agreed path (lossless).

Architecture — joint-VAE tree flow drafter

• A transformer-hidden VAE compresses the 5120-d backbone hidden into a 1024-d latent; the flow runs in that small latent (cheap, base-decoupled).
• A flow-matching drafter predicts K=8 future latents in 2 Euler steps, decoded back to hiddens.
• Token feedback at zero sequential cost: a low-rank Markov head plus path-conditioned residuals.
• The VAE is unfrozen and trained jointly with the flow for acceptance (not reconstruction), with a small reconstruction anchor.

Acceptance (RedHatAI/speculator_benchmarks + held-out gsm8k)

Live tree-accept = real greedy tokens accepted per verify pass vs the backbone; lossless.

Files

• model.safetensors — the drafter (flow experts, Markov head, path head, jointly-trained VAE weights)
• chained_flow_tree_config.json — architecture + loss config
• vae/ — the base VAE checkpoint (needed to construct the VAE at load time)

Trained on a diverse mix (gsm8k, nemotron-math, nemotron-stem, alpaca-code, dolly-chat), 2-GPU DDP.