[MOE Quantization] Warn against "undercalibrated" modules

[dbogunowicz]

Note: this branch requires this PR: neuralmagic/compressed-tensors#46 to land in compressed-tensors.

Example Use:

from llmcompressor.transformers import oneshot
from llmcompressor.modifiers.quantization.gptq import GPTQModifier

# Sets parameters for the GPTQ algorithms - target Linear layer weights at 4 bits
recipe = GPTQModifier(scheme="FP8", targets="Linear", ignore=["lm_head"])

# Apply GPTQ algorithm using open_platypus dataset for calibration.
oneshot(
    model="Isotonic/TinyMixtral-4x248M-MoE",
    dataset="open_platypus",
    recipe=recipe,
    save_compressed=True,
    output_dir="llama-compressed-quickstart",
    overwrite_output_dir=True,
    max_seq_length=128,
    num_calibration_samples=2,
)

...
2024-07-10T21:33:18.273943+0200 | _build_quant_modifier | INFO - Building quantization modifier with args: {'targets': 'Linear', 'scheme': 'FP8', 'ignore': ['lm_head']}
2024-07-10T21:33:18.299325+0200 | _calibrate | INFO - Running QuantizationModifier calibration with 2 samples...
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:09<00:00,  4.53s/it]
2024-07-10T21:33:27.368876+0200 | _check_token_distribution | WARNING - The module_name: model.layers.0.block_sparse_moe.experts.1.w1 received less than 20% of calibration batch tokens (35/256 tokens). This could result may harm the quantization quality.
2024-07-10T21:33:27.369534+0200 | _check_token_distribution | WARNING - The module_name: model.layers.0.block_sparse_moe.experts.1.w2 received less than 20% of calibration batch tokens (35/256 tokens). This could result may harm the quantization quality.
2024-07-10T21:33:27.369577+0200 | _check_token_distribution | WARNING - The module_name: model.layers.0.block_sparse_moe.experts.1.w3 received less than 20% of calibration batch tokens (35/256 tokens). This could result may harm the quantization quality.
...