Description
Describe the bug
- You can provide extra irrelevant arguments when specifying a modifier and no error is thrown
- This can mask issues for arguments that may no longer be supported or if they are provided incorrectly
E.g:
quant_stage:
quant_modifiers:
QuantizationModifier:
dampen: 0.1
ignore: ["lm_head"]
config_groups:
group_0:
weights:
num_bits: 8
type: "float"
symmetric: true
strategy: "tensor"
observer: "mse"
targets: ["Linear"]- The recipe below specifies
dampenfor the QuantizationModifier, however, this is not a valid argument and no error is thrown
Expected behavior
- Validate recipe arguments are valid for the modifiers being specified