#6480 [Engine] qwen36 baked chat template breaks prompt reuse

open · @Mark-1982 · 0 commentaires · Sur GitHub

Multi-GPUNVIDIA / CUDALinux

Description

### Area

[Engine] (speed, kernels, models, the server and API)

### Desktop app, Terminal or source?

Source

### What happened

With thinking off, the chat template baked into the Qwen3.6 pack writes an empty think block on the first answer, then leaves that block out of history. The next prompt is no longer the same prefix, so reuse is 0.

A follow-up at about 70k tokens waited 200–213 seconds before the answer started. The engine itself is fine: a hand-fed prefix reused 14 of 15 tokens. After rewriting the one broken condition, the same two short turns went from 0 reused to 36 of 52.

Estimate only, the 70k case was not re-run: the mismatch is a fixed header, so a long follow-up should start in a few seconds instead of about 200. That is roughly 20x to 40x less waiting. The answer's typing speed does not change.

Same report as #5758, which was closed for a missing title prefix, not because it was fixed.

### Reproduction status

Reproduced locally

### Strata version

control 0.1.63, pack qwen36-ud-q4_k_xl

### Graphics card, memory and system

2x RTX 3060 12 GB, 64 GB RAM, Linux

Sur le site

Liens install, modèles, releases.