Clean up observer defaulting logic, better error message

[kylesayrs]

Purpose

• Better error message when attempting to load an unregistered class instance
• Avoid annoying users with configs which have observer=memoryless for dynamic quantization

Changes

• Clean up observer defaulting logic
• Return the type name in the case that it is not found in the list of aliases
• Condition warning on if observer != memoryless

Testing

Registry Error

from llmcompressor.observers import Observer
Observer.load_from_registry("asdf", quantization_args=None)

- KeyError: "Unable to find None registered under type <class 'llmcompressor.observers.base.Observer'>.\nRegistered values for <class 'llmcompressor.observers.base.Observer'>: ['minmax', 'mse']\nRegistered aliases for <class 'llmcompressor.observers.base.Observer'>: []"
+ KeyError: "Unable to find asdf registered under type <class 'llmcompressor.observers.base.Observer'>.\nRegistered values for <class 'llmcompressor.observers.base.Observer'>: ['minmax', 'mse']\nRegistered aliases for <class 'llmcompressor.observers.base.Observer'>: []" 

Observer Warning

from compressed_tensors.quantization import QuantizationArgs
args = QuantizationArgs(dynamic=True, observer="minmax-whatever")
>>> /home/ksayers/compressed-tensors/src/compressed_tensors/quantization/quant_args.py:213: UserWarning: No observer is used for dynamic quantization, setting to None
  warnings.warn(

from compressed_tensors.quantization import QuantizationArgs
args = QuantizationArgs(dynamic=True, observer="memoryless")
>>> 

[dsikka]

this LGTM. could you verify this works as expected trying our e2e test or example case with dynamic per token quant?

[dsikka]

LGTM pending we've ran through dynamic examples for a new config (with "memoryless") and new config

[kylesayrs]

Ran examples/quantization_w8a8_fp8/llama3_example.py without issue

<|begin_of_text|>Hello my name is Mirella, I am a 25-year-old Brazilian who has been living in the UK for

grep -B8 -A4 'observer' llm-compressor/output/config.json 
    "config_groups": {
      "group_0": {
        "input_activations": {
          "actorder": null,
          "block_structure": null,
          "dynamic": true,
          "group_size": null,
          "num_bits": 8,
          "observer": null,
          "observer_kwargs": {},
          "strategy": "token",
          "symmetric": true,
          "type": "float"
        },
--
          "Linear"
        ],
        "weights": {
          "actorder": null,
          "block_structure": null,
          "dynamic": false,
          "group_size": null,
          "num_bits": 8,
          "observer": "minmax",
          "observer_kwargs": {},
          "strategy": "channel",
          "symmetric": true,
          "type": "float"
        }

[kylesayrs]

Test in PR description shows that warning does not trigger for observer=memoryless

[rahul-tuli]

LGTM!