SmoothQuant-for-Qwen3/smoothquant_qwen.py

[LiDing666]

I have some trouble for this document.
fake_quant.activate_bit = args.activation
fake_quant.weight_bit = args.weight
model_quantized = fake_quant.quantize_Qwen3_like(model_fp16)
the code snippet you provided (fake_quant.quantize_Qwen3_like(model_fp16)) ​​does lack the SmoothQuant smoothing step​​, making it a naive W8A8 quantization rather than true SmoothQuant. Is this complete code?

[cjm715]

I noticed the same thing as you @yuni-star . @Yuye541 @fly-XarF is the smoothquant_qwen.py the most up to date version? few things I noticed:

• I had to modify quantize_llama_like to quantize_Qwen3_like.
• smooth_lm is imported but never applied. Also this process needs activation scales . As is, it looks like its applying naive int8 quantization without smoothing.
• in the line: evaluate_.eval_ours(model_quantized, tokenizer, model_name=f"{args.model_name}_w{fake_quant.weight_bit}a{fake_quant.activate_bit}") . I get an unexpected keyword error for model_name. The eval_ours method does not have that.