seems promising , this is the way , can someone benchmark this
I'm getting 6.55t/s using the Qwen3.5-397B-A17B-4bit model with the command:
./infer --prompt "Explain quantum computing" --tokens 100
MacBook Pro M5 Pro (64GB RAM)
Appreciate the data point. M5 Max would also be interesting to see once available in desktop form.
can you post the final result (or as far as you got before you killed it) to show us how cohesive and good it is? I'd like to see an example of the output of this.
Since the output is quite long, here is a link:
https://pastebin.com/k76wiVGP
Why does this G character appear to prefix most of the output? ("Ġlike")
It is a tokenizer artifact most likely (https://github.com/huggingface/transformers/issues/4786). So the output is not properly decoded in this case, it should just be a space.
The original tokens have Ġ instead of space. I had this issue too when writing an inference engine for Qwen. You have to "normalize" those special characters.