Skip to content

Eval bug: llama3.2 answering a single JSON line #19155

@kpouget

Description

@kpouget

Name and Version

llama-cli, llama-server b7849

Operating systems

Mac

GGML backends

Metal

Hardware

M4 Pro

Models

llama3.2

Problem description & steps to reproduce

With the recent versions of llama.cpp, I hit this template issue:

> tell what is GGML API

{"name": "tell", "parameters": {"tt": "GGML API"}}

[ Prompt: 832.0 t/s | Generation: 82.2 t/s ]

First Bad Commit

Git bisection pointed to the merge of this PR: #18462

Relevant log output

Other model families don't hit this issue

ramalama   run  smollm:135m
🦭 > hello
Hello! How can I help you?
ramalama   run  ibm/granite:2b
🦭 > hello
Hello! It's a pleasure to meet you. How can I assist you today?

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions