Chat endpoint prematurely terminates when output contains "<im_end>" / ChatML special tokens (string-level stop matching)

[qoli]

Which version of LM Studio?
LM Studio 0.4.2+2 (0.4.2+2)

Which operating system?
macOS (Apple Silicon, M‑series)

What is the bug?
LM Studio’s built-in OpenAI-compatible chat server prematurely terminates generation when the output contains ChatML/Qwen special tokens such as:

"stop": [
  "<|im_end|>",
  "<im_end>",
  "</s>"
]

這些符號是什麼呢？

These tokens are treated as stop sequences even when they appear as normal text content. Generation is truncated and the UI shows:
"Stop reason: found EOS Token"

This makes it impossible to quote documentation, explain templates, or process text that contains these tokens.

Screenshots:
(attach screenshot showing “Stop reason: found EOS Token”)

Logs:
UI displays: Stop reason: found EOS Token
No crash or client disconnect. The server intentionally stops generation.

To Reproduce:

1. Start LM Studio
2. Load a Qwen / Qwen-Coder model (e.g. qwen3-coder-next)
3. Use the built-in chat
4. Send: Explain what <im_end> means in Qwen template.
5. Generation immediately stops or returns partial output

Expected behavior:
The model should treat "<im_end>" as plain text and continue generating normally.

Actual behavior:
Generation stops whenever the string appears.

Additional technical details:
LM Studio appears to automatically inject stop sequences like:
"stop": ["<|im_end|>", "<im_end>", ""]

This suggests string-level matching instead of tokenizer-level EOS detection, causing false positives.

[mktums]

Can confirm, still happens.
Debugging with Kilo's agent showed that <|im_end|> is the problem for me.

LM Studio 0.4.10 (Build 1)

@qoli Why did you closed this issue?

[qoli]

After looking at QwenLM/qwen-code#1936 and retesting this myself, I no longer think this is an LM Studio-specific bug.

I can reproduce the same EOS / special-token truncation behavior both in the official Qwen web interface and in LM Studio. Since the same symptom appears outside LM Studio as well, the issue seems more likely to be upstream — at the model / token-semantics / EOS-handling level — rather than something introduced solely by LM Studio.

In other words, LM Studio may expose the symptom, but based on cross-environment reproduction, it does not look like the root cause here.

So I think this issue is better understood as a broader EOS / self-referential termination problem, similar to what is described in QwenLM/qwen-code#1936, rather than a client-only bug in LM Studio.

QwenLM/qwen-code#1936

[mktums]

Well, it's not LM Studio's fault after all...
Further investigation showed that problem is in llama-cpp's parser, ggml-org/llama.cpp#20260 and numerous other issues...