Park Woorak

This seems to be the case with typical ARM devices. Also it fails to build from source. Please refer to #59.

Hi, is there anyone who made progress with the `webgpu.get_fmap` error?

I found a way to pass through `await tvm.asyncLoadWebGPUPipelines(vm.getInternalModule());` without `webgpu.get_fmap` error. I realized that models built from https://github.com/mlc-ai/mlc-llm project don't have that problem. There are quite many differences, but...

@Yumin-gd Hey, I could not help you for that issue, but let me just answer your questions. > _**Specific Implementation Details**_: Could you kindly provide more details about how and...

+1

I guess it is because of multimodal support of those models. only 1B model that doesn't have multimodal feature worked on my condition, while the multimodal-supporting models are failing.

> And also there is no `"gemma3_instruction"` in the list(https://github.com/mlc-ai/mlc-llm/blob/main/python/mlc_llm/interface/gen_config.py#L264) even though gemma 3 template is added(https://github.com/mlc-ai/mlc-llm/blob/main/python/mlc_llm/conversation_template/gemma.py#L23-L37). #3224 resolves this part. I wanted that this whole problem with gemma3 full...

The reason why gemma-3 does not work properly seems to be due to the lack of fp16 support in the gemma-3 model itself. Since gemma3 is made based on the...