Summary

MemorylessMSEObserver.__init__ validates expand >= 1.0, but the default "mse" observer, MovingAverageMSEObserver, is missing the same check and silently accepts expand < 1.0.

expand scales the initial MSE search range (min_val = amin * expand, max_val = amax * expand in _grid_search_mse), and the grid search can only shrink it from there (p = 1 - i / grid <= 1). So an expand < 1.0 starts the search below the observed min/max and can never grow to cover the data — every candidate range clips the calibration values, producing a degenerate result. That's exactly why MemorylessMSEObserver rejects it; MovingAverageMSEObserver should too.

Reproduce

from compressed_tensors.quantization import QuantizationArgs
from llmcompressor.observers.mse import MemorylessMSEObserver, MovingAverageMSEObserver

args = QuantizationArgs(num_bits=8, symmetric=True)

MemorylessMSEObserver(base_name="weight", args=args, expand=0.5)
# ValueError: expand value must be at least 1.0, got 0.5   ✅

MovingAverageMSEObserver(base_name="weight", args=args, expand=0.5)
# (no error) ❌ — silently accepted, degenerate search range

Changes

• Add the expand >= 1.0 validation to MovingAverageMSEObserver.__init__, matching MemorylessMSEObserver.
• Add a parametrized regression test asserting both MSE observers reject expand < 1.0 and accept expand >= 1.0.

Testing

tests/llmcompressor/observers/test_mse.py passes (the new test_mse_observer_rejects_expand_below_one covers both observers). ruff check / ruff format --check clean on the changed files. (Note: the pre-existing test_mse_observer_torch_compile fails locally on macOS due to a torch.compile backend issue unrelated to this change — it fails the same way on main.)