Ling-3.0-Flash-Agentic-Chat-Template-Jinja

Ling 3.0 flash chat template

A drop-in Jinja chat template for Ling 3.0 flash, rendering the model's native conversation format. It carries the same agentic hardening as the Qwen fork (https://huggingface.co/sanjxz/Qwen-Fixed-Chat-Templates/tree/main), re-expressed in Ling's <system> / <user> / <assistant> token scheme.

• Engines: llama.cpp / LM Studio / MLX / vLLM and any engine with a HuggingFace-Jinja (minja) runtime.

---

Files

---

Feature set

The agentic internals are (almost, WIP) identical to the Qwen fork — see the "What the fork adds over the original" section of README.md for the full write-up. In brief, this template adds over chat_template_laguna-orig.jinja:

• Escalating loop-guard — consecutive-failure escalation (nudge → warn → hard halt), plus tool suspension: on hard halt a pre-scan replaces the <available_tools> block with a suspension notice so the model cannot emit a tool call.
• Repeat / ping-pong detection — fingerprints each turn by tool name + arguments over a two-turn window (catches A-B-A-B alternation); mutating tools exempted from result-repeat, caught instead by the ping-pong nudge.
• Four-tier failure heuristics — hard / harness / generic / permission errors, kept in sync between the pre-scan and the main loop, with $ (shell) and took (timing) false-positive escapes. Read tools are exempt from error-text matching; results are attributed via tool_call_id → name.
• Security hardening — only system/developer messages may toggle thinking; user content has the tokens stripped but never applied. sanitize_tool_tokens strips literal <|think_off|> / <|think_on|> from untrusted tool output.
• Harness integration — structural skill-tool triggering, vanished-tool warnings, strftime_now date injection, OpenAI {"type":"function",…} wrapper unwrapping, middle-out tool-response truncation, <__media__> marker stripping, unclosed-reasoning recovery.

---

Tool call formats

History tool calls and the format reminder are rendered in one of two formats, selected with tool_call_format.

XML (default) — arguments as <arg_key>/<arg_value> pairs:

<tool_call>get_weather<arg_key>location</arg_key><arg_value>NYC</arg_value></tool_call>

---

Running with llama.cpp

Point --chat-template-file at the template and enable --jinja. Example for Laguna S 2.1 APEX:

import os
import subprocess
import sys

env = os.environ.copy()

cmd = [
    r".\llama-server.exe",
    "-m", r"D:\Ling-3.0-flash-AD-IQ3_S-00001-of-00002.gguf",
    "--reasoning-preserve",
    "--reasoning-budget", "16000",
    "--reasoning-budget-message", "I've thought enough. Answering now with what I have.",
    "--fit", "on",
    "--n-cpu-moe", "39",
    "-lm", "mlock", #recent llama.cpp only ; --mlock & --no-mmap alternative
    "-c", "120000",
    "--cache-type-k", "q8_0",
    "--cache-type-v", "q8_0",
    "-np", "1",
    "-fa", "on",
    "-t", "8",
    "-tb", "8",
    "-b", "2048",
    "-ub", "2048",
    "--jinja",
    "-kvu",
    "--temp", "0.6",
    "--top-p", "0.95",
    "--top-k", "20",
    "--samplers", "top_k;top_p;temperature",
    "--alias", "laguna-s-2.1",
    "--cache-reuse", "256",
    "--cache-ram", "1024",
    "--host", "127.0.0.1",
    "--port", "8080",
    "--verbosity", "4",
    "--chat-template-file", r"G:\xlam3\chat_template_ling.jinja",
]

if __name__ == "__main__":
    try:
        print("Starting llama-server...")
        subprocess.run(cmd, env=env, check=True)
    except KeyboardInterrupt:
        print("\nServer stopped by user.")
    except Exception as e:
        print(f"\nError running server: {e}")

---

Authorship & license

MIT.