Distiller and onnx converter and pqmha

src/pquant/configs/config_ap.yaml

@@ -13,6 +13,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/config_autosparse.yaml

@@ -16,6 +16,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/config_cs.yaml

@@ -12,6 +12,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/config_dst.yaml

@@ -14,6 +14,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/config_fitcompress.yaml

@@ -11,6 +11,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/config_mdmm.yaml

@@ -24,6 +24,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/config_pdp.yaml

@@ -14,6 +14,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/config_wanda.yaml

@@ -16,6 +16,7 @@ quantization_parameters:
   default_data_keep_negatives: 0.
   default_data_integer_bits: 0.
   default_data_fractional_bits: 7.
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   quantize_input: true
   quantize_output: false

src/pquant/configs/finetuning.yaml

@@ -13,6 +13,7 @@ quantization_parameters:
       integer_bits: 4
       fractional_bits: 6
   use_high_granularity_quantization: false
+  dynamic_data_quantization: true
   granularity: "per_tensor"
   use_real_tanh: false
   use_symmetric_quantization: false

src/pquant/core/keras/activations.py

@@ -68,6 +68,7 @@ def __init__(
         self.hgq_gamma = config.quantization_parameters.hgq_gamma
         self.hgq_heterogeneous = config.quantization_parameters.hgq_heterogeneous
         self.use_fitcompress = config.fitcompress_parameters.enable_fitcompress
+        self.dynamic_data = config.quantization_parameters.dynamic_data_quantization
 
         self.post_fitcompress_calibration = False
         self.saved_inputs = []
@@ -89,6 +90,7 @@ def build(self, input_shape):
                 is_heterogeneous=self.use_hgq,
                 hgq_gamma=self.hgq_gamma,
                 place="datalane",
+                dynamic_data=self.dynamic_data,
             )
         if self.quantize_output:
             self.output_quantizer = Quantizer(
@@ -101,6 +103,7 @@ def build(self, input_shape):
                 is_heterogeneous=self.use_hgq,
                 hgq_gamma=self.hgq_gamma,
                 place="datalane",
+                dynamic_data=self.dynamic_data,
             )
 
         if self.use_multiplier: