fix qparams decompression

[shanjiaz]

Updated the decompress_weight function to unpack zero_point/cast scale dtype during decompression. Replace the tensor in module with updated one.
Example script used:

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor.utils import dispatch_for_generation

#MODEL_ID = "nm-testing/TinyLlama-1.1B-Chat-v1.0-w4a16-asym-awq-e2e"
MODEL_ID = "nm-testing/TinyLlama-1.1B-Chat-v1.0-NVFP4"

model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

print("========== SAMPLE GENERATION ==============")
dispatch_for_generation(model)
input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=100)
print(tokenizer.decode(output[0]))''
print("==========================================\n\n")

Example script now generates coherent result:

(llm-compressor) [shanjiaz@nma-a100-solo-4-preserve llm-compressor]$ python zp_decompression.py 
`torch_dtype` is deprecated! Use `dtype` instead!
Compressing model: 154it [00:00, 747.12it/s]


========== SAMPLE GENERATION ==============
<s> Hello my name is John and I am a software engineer. I have been working in the tech industry for the past 10 years. I have worked on various projects and have gained a lot of experience. I am passionate about technology and have a keen interest in the latest technologies. I have a bachelor's degree in computer science and have completed several certifications in various technologies. I am currently working as a software engineer at a leading technology company. In my free time, I enjoy
==========================================

[dsikka]

Just an fyi: #509
This will also impact mxfp4.
I've turned off mxfp4 decompression in the meantime / lower priority anyway

[dsikka]

Would it be cleaner to add optional:
compress_scale / decompress_scale
and compress_zp / decompress_zp functions?

This would impact:

• PackedCompressor (packed zp)
• NVFP4PackedCompressor (fp8 scales)
• MXFP4PackedCompressor (uint8 scales)

[shanjiaz]

Would it be cleaner to add optional: compress_scale / decompress_scale and compress_zp / decompress_zp functions?

This would impact:

• PackedCompressor (packed zp)
• NVFP4PackedCompressor (fp8 scales)
• MXFP4PackedCompressor (uint8 scales)

Sure! I can do that.

[kylesayrs]

Woop

[dsikka]

Some comments, needs rebase

[kylesayrs]

Please make sure that compressed_data is only updated when you want it to be, and not accidentally updated as part of some other calculation

[kylesayrs]

Nice job