Ju4tCode comments

Results 204 comments of


                                            Ju4tCode

[Usage]: How to bypass multimodal processor logic when inputs are already processed

Here is the processor main logic and the model: ```python class CustomMultiModalProcessor(BaseMultiModalProcessor[CustomProcessingInfo]): def _extract_image_placeholders( self, prompt_ids: list[int], image_data: dict[str, Any] ) -> list[PlaceholderFeaturesInfo]: # get feature token length from input...

[Usage]: How to bypass multimodal processor logic when inputs are already processed

> Did you enable chunked prefill? Yes. I enabled this.

[Usage]: How to bypass multimodal processor logic when inputs are already processed

disabling the chunked prefill seems not solve the issue. The model forward still received too many multimodal input when decode. ```python ValueError: Attempted to assign 660 + 660 = 1320...

[Usage]: How to bypass multimodal processor logic when inputs are already processed

The issue only occurs when one request is in decoding stage and the other is in prefill stage. Here is some detailed info about the tensor shapes: ```log # The...

[Usage]: How to bypass multimodal processor logic when inputs are already processed

I referred to this code. Should i parse the multimodal data in the order of the input placehold tokens? If the input tokens look like `[text, image, audio, text, image,...

[Usage]: How to bypass multimodal processor logic when inputs are already processed

I am a little confused. How this `get_multimodal_embedding` method affect the multimodal kwargs received during model forward? This issue is that when the input ids contain both decode and prefill...

[Usage]: How to bypass multimodal processor logic when inputs are already processed

After digging into the model input builder, i find more info about this issue. The builder gets the mm kwargs from the inter_data_list here: https://github.com/vllm-project/vllm/blob/ed6e9075d31e32c8548b480a47d1ffb77da1f54c/vllm/worker/model_runner.py#L974-L979 The inter_data is append when...

[Usage]: How to bypass multimodal processor logic when inputs are already processed

> In `get_multimodal_embeddings`, you should make sure `None` values aren't in the returned tuple. After changing the above error and enabling vllm v1 engine, everything works. 🥹 Thanks for your...

Plugin: nonebot_plugin_bilimusic

插件测试出错了，请修正插件

Adapter: nonebot-adapter-vocechat

适配器依赖不应该包含 nonebot2 驱动器。适配器的 bot_connect 应该在 self.on_ready 中进行，否则 hook 无法正确执行。此处 bot 为什么要修改 call api 逻辑？https://github.com/5656565566/nonebot-adapter-vocechat/blob/5621940c5077f7b18883ac41685089c95ef64c3c/nonebot/adapters/vocechat/bot.py#L35-L63 你应该在 adapter 中去做相应处理而不是在 bot 中，这样会影响 hook 的执行。其他调用的地方也应该对应修改。 event 类型的 name 和 type 不得为空，要么别提供默认值 https://github.com/5656565566/nonebot-adapter-vocechat/blob/5621940c5077f7b18883ac41685089c95ef64c3c/nonebot/adapters/vocechat/event.py#L13-L19 另外请阅读适配器规范...