TVM Quantizer¶
Category: quantizer.tvm
Version: 1.0.0
UUID: 01816306398d81ae2efc44e7fb48e34444bac083551dfb5ef68f24bacbf09dfe
Execution context: leip_forge
Subcomponent Parameters¶
| Name | Synonyms | Allowable Categories |
|---|---|---|
| Calibrator | calibrator |
calibrator |
Value Parameters¶
| Name | Synonyms | Type | Values | Help |
|---|---|---|---|---|
| Activation Data Type | quantizer.activation_dtype |
choice | int8uint |
|
| Bias Data Type | quantizer.bias_dtype |
choice | int32None |
Data type for biases in nn.bias_add operators. Defaults to activation_dtype if None. |
| Calibration Method | quantizer.calib_method |
choice | averageminmaxpercentile |
|
| Kernel Data Type | quantizer.kernel_dtype |
choice | int8uintNone |
|
| Per-channel Quantization | quantizer.per_channel |
scalar | bool | |
| Quantization Type | quantizer.quant_type |
choice | staticdynamicany |
|
| Enable Tensor Splitting | quantizer.split_tensors |
scalar | bool |
This component fits into¶
| Name | UUID | Synonyms |
|---|---|---|
| TVM Compiler | 23680b... |
quantizer |