Add more supported base_configs to QATConfig

[andrewor14]

Today we have:

Int8DynamicActivationInt4WeightConfig
Int4WeightOnlyConfig
Float8DynamicActivationFloat8WeightConfig
Float8DynamicActivationInt4WeightConfig
NVFP4InferenceConfig

We should add:

(for lowering to executorch)

• Int8DynamicActivationIntxWeightConfig
• IntxWeightOnlyConfig

(requested by Axolotl, see axolotl-ai-cloud/axolotl#3107)

• Int8DynamicActivationInt8WeightConfig
• Int4DynamicActivationInt4WeightConfig
• Int8WeightOnlyConfig

Registration happens here today:

ao/torchao/quantization/qat/fake_quantize_config.py

Line 342 in c4d4799

def _infer_fake_quantize_configs(

Example usage:

# Not supported yet today
base_config = Int8DynamicActivationIntxWeightConfig(group_size=32)
quantize_(model, QATConfig(base_config, step="prepare"))
train(model)
quantize_(model, QATConfig(base_config, step="convert"))

[andrewor14]

Int8DynamicActivationIntxWeightConfig and IntxWeightOnlyConfig are added in #3001

[jerryzh168]

@andrewor14 maybe make this a tracker?