Add warning for non-divisible group quantization

[kylesayrs]

Purpose

• Test discrepancies between initialized parameters and values calculated by observers
• Reveal potential issue with how qparams are initialized [Bugfix] Fix shape calculation for group quantization neuralmagic/compressed-tensors#308
• Add warning for when users attempt to quantize groups that aren't perfectly divisible

Prerequisites

• [Logging] Support logging once #1431

Changes

• Added test_observers_update in tests/llmcompressor/modifiers/calibration/test_observers.py
• Add a warning for attempts to quantize indivisible groups

Attempting to quantize a module weight whose columns (3420) are not divisible by group_size (128). This scheme is not supported by vLLM, please consider adjusting the group_size for modules with this number of columns

Testing

• This test fails without CT changes, but succeeds with them

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[markurtz]

I think we should simplify this using the built-in filter pathways for the logger.add function from Loguru. This way, all log levels and standard functions are supported, and the root logger import will work without other special/additional imports.

Something along the lines of the following:

_logged_once = set()

def deduplicate_filter(record):
    message = record["message"]
    log_once = record["extra"].get("log_once", False)

    if log_once and message in _logged_once:
        return True
    
    if log_once:
        _logged_once.add(message)

    return False
    
logger.add(..., filter=deduplicate_filter)

logger.bind(log_once=True).info("This will only log once")
logger.bind(log_once=True).info("This will only log once")

Support log once #1431

[brian-dellabetta]

🔥

The base branch was changed.

[brian-dellabetta]

awesome!