[Bug]: jinja2: raise_exception is undefined when a model's chat template tries to reject unsupported input

[KLL535]

Hello.
I encountered this error when running a model that does not support video in video processing mode; it seems the chat template is trying to notify me of this by triggering something like raise_exception("Video not supported") or similar. However, instead of a clear error message from the template, the entire inference process crashes with the following message, which does not reveal the cause of the crash:

Details

  File "I:\ComfyUI\python_embeded\Lib\site-packages\llama_cpp\llama.py", line 4497, in create_chat_completion
    return handler(
        llama=self,
    ...<52 lines>...
        reasoning_start_max_tokens=reasoning_start_max_tokens,
    )
  File "I:\ComfyUI\python_embeded\Lib\site-packages\llama_cpp\llama_multimodal.py", line 4354, in __call__
    return super().__call__(**kwargs)
           ~~~~~~~~~~~~~~~~^^^^^^^^^^
  File "I:\ComfyUI\python_embeded\Lib\site-packages\llama_cpp\llama_multimodal.py", line 1740, in __call__
    full_prompt_ids, chunk_token_spans, chunks, bitmap_cleanup = self._process_mtmd_prompt(
                                                                 ~~~~~~~~~~~~~~~~~~~~~~~~~^
        llama=llama,
        ^^^^^^^^^^^^
    ...<5 lines>...
        add_generation_prompt=add_generation_prompt,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "I:\ComfyUI\python_embeded\Lib\site-packages\llama_cpp\llama_multimodal.py", line 1518, in _process_mtmd_prompt
    text = self._render_and_replace_media(
        messages=messages,
    ...<5 lines>...
        add_generation_prompt=add_generation_prompt,
    )
  File "I:\ComfyUI\python_embeded\Lib\site-packages\llama_cpp\llama_multimodal.py", line 1329, in _render_and_replace_media
    text = self._render_mtmd_prompt(
        messages=messages,
    ...<4 lines>...
        add_generation_prompt=add_generation_prompt,
    )
  File "I:\ComfyUI\python_embeded\Lib\site-packages\llama_cpp\llama_multimodal.py", line 1249, in _render_mtmd_prompt
    return self.chat_template.render(
           ~~~~~~~~~~~~~~~~~~~~~~~~~^
        messages=messages,
        ^^^^^^^^^^^^^^^^^^
    ...<7 lines>...
        **getattr(self, "extra_template_arguments", {}),
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "I:\ComfyUI\python_embeded\Lib\site-packages\jinja2\environment.py", line 1295, in render
    self.environment.handle_exception()
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
  File "I:\ComfyUI\python_embeded\Lib\site-packages\jinja2\environment.py", line 942, in handle_exception
    raise rewrite_traceback_stack(source=source)
  File "<template>", line 36, in top-level template code
  File "I:\ComfyUI\python_embeded\Lib\site-packages\jinja2\sandbox.py", line 401, in call
    return __context.call(__obj, *args, **kwargs)
           ~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^
  File "I:\ComfyUI\python_embeded\Lib\site-packages\jinja2\runtime.py", line 784, in _invoke
    rv = self._func(*arguments)
  File "<template>", line 1, in template
  File "I:\ComfyUI\python_embeded\Lib\site-packages\jinja2\sandbox.py", line 399, in call
    if not __self.is_safe_callable(__obj):
           ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^
  File "I:\ComfyUI\python_embeded\Lib\site-packages\jinja2\sandbox.py", line 265, in is_safe_callable
    getattr(obj, "unsafe_callable", False) or getattr(obj, "alters_data", False)
    ~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
jinja2.exceptions.UndefinedError: 'raise_exception' is undefined

Thank you for supporting this library.

[JamePeng]

Could you provide the specific model and the code used to call it? It’s not clear to me how the reproduction was achieved.

[JamePeng]

@KLL535 You can try pulling the latest code updates to replace the Python file. and try again.

[KLL535]

Here are the verbose logs for various video content scenarios.

Old library and model: gemma-4-12B-it-qat-uncensored-heretic-UDmerge-Q4_K_XXL.gguf + mmproj-gemma-4-12B-it-QAT-BF16.gguf

log1.txt

⚠️ Could not capture the full log due to spam messages like llama_graph_n_input_tensors
✅ Response received.

---

Old library and model: Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf

log2.txt

❌ Inference failed:
'raise_exception' is undefined

---

New library and model: Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf

log3.txt

❌ Inference failed:
llama.cpp does not currently support video.

I now receive a correct message stating that llama.cpp does not support video, even though the model is claimed to support it and video works with Gemma 4.

[JamePeng]

@KLL535 I think I understand what you mean. The main issue is that you are still using the old designated handler to process multimodal inputs; since llama.cpp didn't support video when the original Qwen-3.5 template was created, I implemented an exception to block video processing. I suggest you consider switching to the GenericMTMDChatHandler approach. Of course, I’ll also look into updating the Qwen-3.5 template accordingly，as I upgraded the templates when testing video processing for Gemma-4.

[KLL535]

Well, I figured that since there’s a specialized handler for these models, it should be the first choice, with the generic handler reserved for non-standard or new models.

Speaking of the generic handler - I tried it with the Muse Glimmer 30B model, but it didn't work; it threw an error regarding a media placeholder mismatch. Sure enough, its built-in model template uses <|patch|> as the image placeholder. It would be great to have the option to specify custom placeholders for images, video, audio when creating the chat handler, alongside the other parameters - but I couldn't find such a setting.

[JamePeng]

@KLL535 I've submitted the new code; please give it another try.

[KLL535]

Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf + qwen3.5 chat handler → works

---

Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf + generic chat handler → wrong template

Setup

                handler_kwargs = {
                    "verbose": verbose,
                }

                video_ffmpeg_bin_dir = config.get("video_ffmpeg_bin_dir", None)
                if video_ffmpeg_bin_dir:
                    handler_kwargs["video_ffmpeg_bin_dir"] = video_ffmpeg_bin_dir
                    handler_kwargs["video_fps_target"] = config.get("video_fps_target", 1.0)
                    handler_kwargs["video_timestamp_interval_ms"] = config.get("video_timestamp_interval_ms", 5000)
                    handler_kwargs["batch_max_tokens"] = config.get("mmproj_batch_max_tokens", 1024)

                if image_min_tokens is not None:
                    handler_kwargs["image_min_tokens"] = image_min_tokens

                if image_max_tokens is not None:
                    handler_kwargs["image_max_tokens"] = image_max_tokens

                handler_class, handler_error = _resolve_handler_class(chat_handler_type)
                if handler_class is None:
                    return {"status": "error", "message": handler_error}, None

                extra_handler_kwargs = _rename_template_args(_handler_options(chat_handler_type, config))

                if chat_handler_type == "generic":

                    if extra_handler_kwargs:
                        handler_kwargs["extra_template_arguments"] = extra_handler_kwargs

                    chat_handler = handler_class(
                        mmproj_path=mmproj_path,
                        chat_format = None, 
                        **handler_kwargs,
                    )

The handler does not pick up tokenizer.chat_template from the GGUF. It renders the built-in fallback instead:

USER: <__media__>Please watch this video...
ASSISTANT:

The model produces output, but the prompt shape is wrong and extra_template_arguments are ignored, because the fallback does not read them.

Workaround that works - patch between Llama(...) and the first create_chat_completion(...):

Patch

                    elif chat_handler_type == "generic":

                        gguf_template_str = None
                        try:
                            gguf_template_str = current_cache["llm"].metadata.get("tokenizer.chat_template")
                        except Exception:
                            gguf_template_str = None

                        if isinstance(gguf_template_str, str) and gguf_template_str:

                            from jinja2 import Template

                            chat_handler.chat_format = gguf_template_str # (1)
                            chat_handler.chat_template = Template(gguf_template_str) # (2)

(1) alone does not help — the renderer still uses the old fallback chat_template.
(2) alone does not help — _chat_format_parser_tags is rebuilt from the fallback chat_format, so <|video_pad|> / <|image_pad|> never get normalized to <media>, and the run fails with media marker mismatch.
Both (1) and (2) together are required.

After the patch, both images and video work, enable_thinking is passed through, and thinking can be turned on/off.

log without patch.txt

log with patch.txt

---

Muse-Glimmer-30B-Heretic-IQ3_XXS.gguf + generic chat handler

With the same patch, video works and template_arguments_reasoning_strength is passed through (its effect is visible in Rendered prompt: Reasoning strength: low.). For images, one more adjustment is needed on the caller side: the Muse-Glimmer template checks part['type'] == 'image', so the OpenAI-style {"type": "image_url", "image_url": {...}} produces no media marker. Sending {"type": "image", "image": url} works.

[JamePeng]

another fix patch online, pls try again!

[KLL535]

Both of these models now work without a patch. Thank you.