Upload folder using huggingface_hub

config.json

@@ -1,4 +1,5 @@
 {
+  "_attn_implementation_autoset": true,
   "_name_or_path": "distil-whisper/distil-large-v3",
   "activation_dropout": 0.0,
   "activation_function": "gelu",
@@ -39,7 +40,7 @@
   "num_mel_bins": 128,
   "pad_token_id": 50256,
   "scale_embedding": false,
-  "transformers_version": "4.45.2",
+  "transformers_version": "4.46.3",
   "use_cache": true,
   "use_weighted_layer_sum": false,
   "vocab_size": 51866

generation_config.json

@@ -303,5 +303,5 @@
     "transcribe": 50360,
     "translate": 50359
   },
-  "transformers_version": "4.45.2"
+  "transformers_version": "4.46.3"
 }

openvino_decoder_model.xml

@@ -10,7 +10,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="0" name="Parameter_26431" type="Parameter" version="opset1">
+		<layer id="0" name="Parameter_26704" type="Parameter" version="opset1">
 			<data shape="?,?,?" element_type="f32" />
 			<output>
 				<port id="0" precision="FP32" names="encoder_hidden_states">
@@ -47,7 +47,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="4" name="Constant_31865" type="Const" version="opset1">
+		<layer id="4" name="Constant_32205" type="Const" version="opset1">
 			<data element_type="i64" shape="2" offset="132776960" size="16" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -151,7 +151,7 @@
 				<port id="0" precision="I64" names="23" />
 			</output>
 		</layer>
-		<layer id="12" name="ShapeOf_31838" type="ShapeOf" version="opset3">
+		<layer id="12" name="ShapeOf_32121" type="ShapeOf" version="opset3">
 			<data output_type="i64" />
 			<input>
 				<port id="0" precision="I64">
@@ -165,7 +165,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="13" name="Constant_31839" type="Const" version="opset1">
+		<layer id="13" name="Constant_32122" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="133923868" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -174,7 +174,7 @@
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="14" name="Constant_31840" type="Const" version="opset1">
+		<layer id="14" name="Constant_32123" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="133923860" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -183,7 +183,7 @@
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="15" name="Gather_31841" type="Gather" version="opset8">
+		<layer id="15" name="Gather_32124" type="Gather" version="opset8">
 			<data batch_dims="0" />
 			<input>
 				<port id="0" precision="I64">
@@ -244,7 +244,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="20" name="Constant_26428" type="Const" version="opset1">
+		<layer id="20" name="__module.model.model.decoder.embed_positions/aten::index/Constant" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="132776976" size="4" />
 			<output>
 				<port id="0" precision="I32" />
@@ -321,7 +321,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="25" name="Constant_31606_compressed" type="Const" version="opset1">
+		<layer id="25" name="Constant_31887_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="133923884" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -331,7 +331,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="26" name="Constant_31606" type="Convert" version="opset1">
+		<layer id="26" name="Constant_31887" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -373,7 +373,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="28" name="Constant_31607_compressed" type="Const" version="opset1">
+		<layer id="28" name="Constant_31888_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="133926444" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -383,7 +383,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="29" name="Constant_31607" type="Convert" version="opset1">
+		<layer id="29" name="Constant_31888" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -425,7 +425,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="31" name="Constant_31509_compressed" type="Const" version="opset1">
+		<layer id="31" name="Constant_31790_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1280, 1280" offset="133929004" size="3276800" />
 			<output>
 				<port id="0" precision="FP16">
@@ -434,7 +434,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="32" name="Constant_31509" type="Convert" version="opset1">
+		<layer id="32" name="Constant_31790" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -452,7 +452,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="33" name="Multiply_31492" type="MatMul" version="opset1">
+		<layer id="33" name="Multiply_31773" type="MatMul" version="opset1">
 			<data transpose_a="false" transpose_b="true" />
 			<input>
 				<port id="0" precision="FP32">
@@ -473,7 +473,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="34" name="Constant_31608_compressed" type="Const" version="opset1">
+		<layer id="34" name="Constant_31889_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="137205804" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -483,7 +483,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="35" name="Constant_31608" type="Convert" version="opset1">
+		<layer id="35" name="Constant_31889" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -525,7 +525,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="37" name="Constant_31866" type="Const" version="opset1">
+		<layer id="37" name="Constant_32206" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -634,7 +634,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="44" name="Constant_31867" type="Const" version="opset1">
+		<layer id="44" name="Constant_32207" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -726,13 +726,13 @@
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="50" name="__module.model.model.decoder/aten::full/Convert" type="Const" version="opset1">
+		<layer id="50" name="17" type="Const" version="opset1">
 			<data element_type="f32" shape="" offset="140485212" size="4" />
 			<output>
-				<port id="0" precision="FP32" />
+				<port id="0" precision="FP32" names="17" />
 			</output>
 		</layer>
-		<layer id="51" name="Constant_31848" type="Const" version="opset1">
+		<layer id="51" name="Constant_32145" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923868" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -743,7 +743,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="52" name="Reshape_31849" type="Reshape" version="opset1">
+		<layer id="52" name="Reshape_32146" type="Reshape" version="opset1">
 			<data special_zero="false" />
 			<input>
 				<port id="0" precision="I64" />
@@ -776,7 +776,7 @@
 				<port id="2" precision="I64" names="62,64" />
 			</output>
 		</layer>
-		<layer id="55" name="Constant_24308" type="Const" version="opset1">
+		<layer id="55" name="Constant_24581" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="132776976" size="4" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -785,7 +785,7 @@
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="56" name="Unsqueeze_24309" type="Unsqueeze" version="opset1">
+		<layer id="56" name="Unsqueeze_24582" type="Unsqueeze" version="opset1">
 			<input>
 				<port id="0" precision="I64" />
 				<port id="1" precision="I32" />
@@ -827,7 +827,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="59" name="ShapeOf_31850" type="ShapeOf" version="opset3">
+		<layer id="59" name="ShapeOf_32152" type="ShapeOf" version="opset3">
 			<data output_type="i32" />
 			<input>
 				<port id="0" precision="FP32">
@@ -841,19 +841,19 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="60" name="Constant_31851" type="Const" version="opset1">
+		<layer id="60" name="Constant_32153" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="133923868" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="61" name="Constant_31852" type="Const" version="opset1">
+		<layer id="61" name="Constant_32154" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="133923860" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="62" name="Gather_31853" type="Gather" version="opset8">
+		<layer id="62" name="Gather_32155" type="Gather" version="opset8">
 			<data batch_dims="0" />
 			<input>
 				<port id="0" precision="I32">
@@ -905,7 +905,7 @@
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="67" name="Convert_31856" type="Convert" version="opset1">
+		<layer id="67" name="Convert_32158" type="Convert" version="opset1">
 			<data destination_type="i32" />
 			<input>
 				<port id="0" precision="I64" />
@@ -1059,7 +1059,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="80" name="Constant_24411" type="Const" version="opset1">
+		<layer id="80" name="Constant_24684" type="Const" version="opset1">
 			<data element_type="i64" shape="2" offset="140485218" size="16" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1179,7 +1179,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="88" name="Constant_31845" type="Const" version="opset1">
+		<layer id="88" name="Constant_32136" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923860" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1190,7 +1190,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="89" name="Constant_31846" type="Const" version="opset1">
+		<layer id="89" name="Constant_32137" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="133923860" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1199,7 +1199,7 @@
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="90" name="Gather_31847" type="Gather" version="opset8">
+		<layer id="90" name="Gather_32138" type="Gather" version="opset8">
 			<data batch_dims="0" />
 			<input>
 				<port id="0" precision="I64">
@@ -1216,7 +1216,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="91" name="Constant_30837" type="Const" version="opset1">
+		<layer id="91" name="Constant_31122" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923868" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1227,7 +1227,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="92" name="Constant_30839" type="Const" version="opset1">
+		<layer id="92" name="Constant_31124" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923868" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1238,7 +1238,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="93" name="Constant_30841" type="Const" version="opset1">
+		<layer id="93" name="Constant_31126" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923868" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1293,7 +1293,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="96" name="Constant_24845" type="Const" version="opset1">
+		<layer id="96" name="Constant_25118" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923860" size="8" />
 			<output>
 				<port id="0" precision="I64">
@@ -1301,7 +1301,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="97" name="ShapeOf_31861" type="ShapeOf" version="opset3">
+		<layer id="97" name="ShapeOf_32167" type="ShapeOf" version="opset3">
 			<data output_type="i64" />
 			<input>
 				<port id="0" precision="FP32">
@@ -1316,7 +1316,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="98" name="Constant_31862" type="Const" version="opset1">
+		<layer id="98" name="Constant_32168" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923868" size="8" />
 			<output>
 				<port id="0" precision="I64">
@@ -1324,13 +1324,13 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="99" name="Constant_31863" type="Const" version="opset1">
+		<layer id="99" name="Constant_32169" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="133923860" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="100" name="Gather_31864" type="Gather" version="opset8">
+		<layer id="100" name="Gather_32170" type="Gather" version="opset8">
 			<data batch_dims="0" />
 			<input>
 				<port id="0" precision="I64">
@@ -1347,7 +1347,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="101" name="Constant_24848" type="Const" version="opset1">
+		<layer id="101" name="Constant_25121" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="133923868" size="8" />
 			<output>
 				<port id="0" precision="I64">
@@ -1485,7 +1485,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="109" name="Constant_31609_compressed" type="Const" version="opset1">
+		<layer id="109" name="Constant_31890_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="143762042" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -1495,7 +1495,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="110" name="Constant_31609" type="Convert" version="opset1">
+		<layer id="110" name="Constant_31890" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -1537,7 +1537,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="112" name="Constant_31868" type="Const" version="opset1">
+		<layer id="112" name="Constant_32208" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1652,7 +1652,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="119" name="Constant_31869" type="Const" version="opset1">
+		<layer id="119" name="Constant_32209" type="Const" version="opset1">
 			<data element_type="i64" shape="3" offset="143764602" size="24" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -1732,7 +1732,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="124" name="Constant_31610_compressed" type="Const" version="opset1">
+		<layer id="124" name="Constant_31891_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="147041426" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -1742,7 +1742,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="125" name="Constant_31610" type="Convert" version="opset1">
+		<layer id="125" name="Constant_31891" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -1834,7 +1834,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="130" name="Constant_31611_compressed" type="Const" version="opset1">
+		<layer id="130" name="Constant_31892_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="147043986" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -1844,7 +1844,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="131" name="Constant_31611" type="Convert" version="opset1">
+		<layer id="131" name="Constant_31892" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -1886,7 +1886,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="133" name="Constant_31612_compressed" type="Const" version="opset1">
+		<layer id="133" name="Constant_31893_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="147046546" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -1896,7 +1896,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="134" name="Constant_31612" type="Convert" version="opset1">
+		<layer id="134" name="Constant_31893" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -1938,7 +1938,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="136" name="Constant_31512_compressed" type="Const" version="opset1">
+		<layer id="136" name="Constant_31793_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1280, 1280" offset="147049106" size="3276800" />
 			<output>
 				<port id="0" precision="FP16">
@@ -1947,7 +1947,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="137" name="Constant_31512" type="Convert" version="opset1">
+		<layer id="137" name="Constant_31793" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -1965,7 +1965,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="138" name="Multiply_31496" type="MatMul" version="opset1">
+		<layer id="138" name="Multiply_31777" type="MatMul" version="opset1">
 			<data transpose_a="false" transpose_b="true" />
 			<input>
 				<port id="0" precision="FP32">
@@ -1986,7 +1986,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="139" name="Constant_31613_compressed" type="Const" version="opset1">
+		<layer id="139" name="Constant_31894_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="150325906" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -1996,7 +1996,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="140" name="Constant_31613" type="Convert" version="opset1">
+		<layer id="140" name="Constant_31894" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -2038,7 +2038,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="142" name="Constant_31870" type="Const" version="opset1">
+		<layer id="142" name="Constant_32210" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -2147,7 +2147,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="149" name="Constant_25069" type="Const" version="opset1">
+		<layer id="149" name="Constant_24866" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="153605266" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -2158,7 +2158,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="150" name="Constant_25070" type="Const" version="opset1">
+		<layer id="150" name="Constant_24867" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="153605274" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -2169,7 +2169,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="151" name="Constant_25071" type="Const" version="opset1">
+		<layer id="151" name="Constant_24868" type="Const" version="opset1">
 			<data element_type="i64" shape="1" offset="153605282" size="8" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -2180,7 +2180,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="152" name="__module.model.model.decoder.layers.0.encoder_attn/prim::ListConstruct_1" type="Concat" version="opset1">
+		<layer id="152" name="__module.model.model.decoder.layers.0.self_attn/prim::ListConstruct_1" type="Concat" version="opset1">
 			<data axis="0" />
 			<input>
 				<port id="0" precision="I64">
@@ -2344,7 +2344,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="161" name="Constant_31614_compressed" type="Const" version="opset1">
+		<layer id="161" name="Constant_31895_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="156882090" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -2354,7 +2354,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="162" name="Constant_31614" type="Convert" version="opset1">
+		<layer id="162" name="Constant_31895" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -2500,7 +2500,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="170" name="Constant_31871" type="Const" version="opset1">
+		<layer id="170" name="Constant_32211" type="Const" version="opset1">
 			<data element_type="i64" shape="3" offset="156884650" size="24" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -2580,7 +2580,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="175" name="Constant_31615_compressed" type="Const" version="opset1">
+		<layer id="175" name="Constant_31896_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="160161474" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -2590,7 +2590,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="176" name="Constant_31615" type="Convert" version="opset1">
+		<layer id="176" name="Constant_31896" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -2682,7 +2682,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="181" name="Constant_31616_compressed" type="Const" version="opset1">
+		<layer id="181" name="Constant_31897_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="160164034" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -2692,7 +2692,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="182" name="Constant_31616" type="Convert" version="opset1">
+		<layer id="182" name="Constant_31897" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -2734,7 +2734,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="184" name="Constant_31617_compressed" type="Const" version="opset1">
+		<layer id="184" name="Constant_31898_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="160166594" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -2744,7 +2744,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="185" name="Constant_31617" type="Convert" version="opset1">
+		<layer id="185" name="Constant_31898" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -2834,7 +2834,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="190" name="Constant_31618_compressed" type="Const" version="opset1">
+		<layer id="190" name="Constant_31899_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 5120" offset="173276354" size="10240" />
 			<output>
 				<port id="0" precision="FP16">
@@ -2844,7 +2844,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="191" name="Constant_31618" type="Convert" version="opset1">
+		<layer id="191" name="Constant_31899" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -2951,7 +2951,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="197" name="Constant_31619_compressed" type="Const" version="opset1">
+		<layer id="197" name="Constant_31900_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="186393794" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -2961,7 +2961,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="198" name="Constant_31619" type="Convert" version="opset1">
+		<layer id="198" name="Constant_31900" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3053,7 +3053,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="203" name="Constant_31620_compressed" type="Const" version="opset1">
+		<layer id="203" name="Constant_31901_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="186396354" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3063,7 +3063,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="204" name="Constant_31620" type="Convert" version="opset1">
+		<layer id="204" name="Constant_31901" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3105,7 +3105,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="206" name="Constant_31621_compressed" type="Const" version="opset1">
+		<layer id="206" name="Constant_31902_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="186398914" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3115,7 +3115,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="207" name="Constant_31621" type="Convert" version="opset1">
+		<layer id="207" name="Constant_31902" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3157,7 +3157,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="209" name="Constant_31515_compressed" type="Const" version="opset1">
+		<layer id="209" name="Constant_31796_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1280, 1280" offset="186401474" size="3276800" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3166,7 +3166,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="210" name="Constant_31515" type="Convert" version="opset1">
+		<layer id="210" name="Constant_31796" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3184,7 +3184,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="211" name="Multiply_31500" type="MatMul" version="opset1">
+		<layer id="211" name="Multiply_31781" type="MatMul" version="opset1">
 			<data transpose_a="false" transpose_b="true" />
 			<input>
 				<port id="0" precision="FP32">
@@ -3205,7 +3205,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="212" name="Constant_31622_compressed" type="Const" version="opset1">
+		<layer id="212" name="Constant_31903_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="189678274" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3215,7 +3215,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="213" name="Constant_31622" type="Convert" version="opset1">
+		<layer id="213" name="Constant_31903" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3257,7 +3257,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="215" name="Constant_31872" type="Const" version="opset1">
+		<layer id="215" name="Constant_32212" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -3366,7 +3366,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="222" name="Constant_31873" type="Const" version="opset1">
+		<layer id="222" name="Constant_32213" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -3544,7 +3544,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="232" name="Constant_31623_compressed" type="Const" version="opset1">
+		<layer id="232" name="Constant_31904_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="196234434" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3554,7 +3554,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="233" name="Constant_31623" type="Convert" version="opset1">
+		<layer id="233" name="Constant_31904" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3596,7 +3596,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="235" name="Constant_31874" type="Const" version="opset1">
+		<layer id="235" name="Constant_32214" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -3711,7 +3711,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="242" name="Constant_31875" type="Const" version="opset1">
+		<layer id="242" name="Constant_32215" type="Const" version="opset1">
 			<data element_type="i64" shape="3" offset="143764602" size="24" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -3791,7 +3791,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="247" name="Constant_31624_compressed" type="Const" version="opset1">
+		<layer id="247" name="Constant_31905_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="199513794" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3801,7 +3801,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="248" name="Constant_31624" type="Convert" version="opset1">
+		<layer id="248" name="Constant_31905" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3893,7 +3893,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="253" name="Constant_31625_compressed" type="Const" version="opset1">
+		<layer id="253" name="Constant_31906_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="199516354" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3903,7 +3903,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="254" name="Constant_31625" type="Convert" version="opset1">
+		<layer id="254" name="Constant_31906" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3945,7 +3945,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="256" name="Constant_31626_compressed" type="Const" version="opset1">
+		<layer id="256" name="Constant_31907_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="199518914" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -3955,7 +3955,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="257" name="Constant_31626" type="Convert" version="opset1">
+		<layer id="257" name="Constant_31907" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -3997,7 +3997,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="259" name="Constant_31518_compressed" type="Const" version="opset1">
+		<layer id="259" name="Constant_31799_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1280, 1280" offset="199521474" size="3276800" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4006,7 +4006,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="260" name="Constant_31518" type="Convert" version="opset1">
+		<layer id="260" name="Constant_31799" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -4024,7 +4024,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="261" name="Multiply_31504" type="MatMul" version="opset1">
+		<layer id="261" name="Multiply_31785" type="MatMul" version="opset1">
 			<data transpose_a="false" transpose_b="true" />
 			<input>
 				<port id="0" precision="FP32">
@@ -4045,7 +4045,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="262" name="Constant_31627_compressed" type="Const" version="opset1">
+		<layer id="262" name="Constant_31908_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="202798274" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4055,7 +4055,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="263" name="Constant_31627" type="Convert" version="opset1">
+		<layer id="263" name="Constant_31908" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -4097,7 +4097,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="265" name="Constant_31876" type="Const" version="opset1">
+		<layer id="265" name="Constant_32216" type="Const" version="opset1">
 			<data element_type="i64" shape="4" offset="137208364" size="32" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -4348,7 +4348,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="280" name="Constant_31628_compressed" type="Const" version="opset1">
+		<layer id="280" name="Constant_31909_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="209354434" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4358,7 +4358,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="281" name="Constant_31628" type="Convert" version="opset1">
+		<layer id="281" name="Constant_31909" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -4504,7 +4504,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="289" name="Constant_31877" type="Const" version="opset1">
+		<layer id="289" name="Constant_32217" type="Const" version="opset1">
 			<data element_type="i64" shape="3" offset="156884650" size="24" />
 			<rt_info>
 				<attribute name="precise" version="0" />
@@ -4584,7 +4584,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="294" name="Constant_31629_compressed" type="Const" version="opset1">
+		<layer id="294" name="Constant_31910_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="212633794" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4594,7 +4594,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="295" name="Constant_31629" type="Convert" version="opset1">
+		<layer id="295" name="Constant_31910" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -4686,7 +4686,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="300" name="Constant_31630_compressed" type="Const" version="opset1">
+		<layer id="300" name="Constant_31911_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="212636354" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4696,7 +4696,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="301" name="Constant_31630" type="Convert" version="opset1">
+		<layer id="301" name="Constant_31911" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -4738,7 +4738,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="303" name="Constant_31631_compressed" type="Const" version="opset1">
+		<layer id="303" name="Constant_31912_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="212638914" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4748,7 +4748,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="304" name="Constant_31631" type="Convert" version="opset1">
+		<layer id="304" name="Constant_31912" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -4838,7 +4838,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="309" name="Constant_31632_compressed" type="Const" version="opset1">
+		<layer id="309" name="Constant_31913_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 5120" offset="225748674" size="10240" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4848,7 +4848,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="310" name="Constant_31632" type="Convert" version="opset1">
+		<layer id="310" name="Constant_31913" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -4955,7 +4955,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="316" name="Constant_31633_compressed" type="Const" version="opset1">
+		<layer id="316" name="Constant_31914_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="238866114" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -4965,7 +4965,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="317" name="Constant_31633" type="Convert" version="opset1">
+		<layer id="317" name="Constant_31914" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -5057,7 +5057,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="322" name="Constant_31634_compressed" type="Const" version="opset1">
+		<layer id="322" name="Constant_31915_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="238868674" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -5067,7 +5067,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="323" name="Constant_31634" type="Convert" version="opset1">
+		<layer id="323" name="Constant_31915" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -5109,7 +5109,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="325" name="Constant_31635_compressed" type="Const" version="opset1">
+		<layer id="325" name="Constant_31916_compressed" type="Const" version="opset1">
 			<data element_type="f16" shape="1, 1, 1280" offset="238871234" size="2560" />
 			<output>
 				<port id="0" precision="FP16">
@@ -5119,7 +5119,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="326" name="Constant_31635" type="Convert" version="opset1">
+		<layer id="326" name="Constant_31916" type="Convert" version="opset1">
 			<data destination_type="f32" />
 			<rt_info>
 				<attribute name="decompression" version="0" />
@@ -5182,7 +5182,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="329" name="Result_26430" type="Result" version="opset1">
+		<layer id="329" name="Result_26703" type="Result" version="opset1">
 			<input>
 				<port id="0" precision="FP32">
 					<dim>-1</dim>
@@ -5558,16 +5558,16 @@
 		<edge from-layer="328" from-port="2" to-layer="329" to-port="0" />
 	</edges>
 	<rt_info>
-		<Runtime_version value="2024.5.0-16901-32aaa2fbd96" />
+		<Runtime_version value="2024.5.0-17285-ea5c1dcfdf9-releases/2024/5" />
 		<conversion_parameters>
 			<framework value="pytorch" />
 			<is_python_object value="True" />
 		</conversion_parameters>
 		<optimum>
-			<optimum_intel_version value="1.20.0.dev0+2559620" />
-			<optimum_version value="1.23.1" />
-			<pytorch_version value="2.5.0" />
-			<transformers_version value="4.45.2" />
+			<optimum_intel_version value="1.21.0.dev0+d357376" />
+			<optimum_version value="1.23.3" />
+			<pytorch_version value="2.5.1" />
+			<transformers_version value="4.46.3" />
 		</optimum>
 	</rt_info>
 </net>

openvino_detokenizer.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b55ac04e90f19fede391281a9e5a90169fc646e1e302fe1208c782282b51ceaa
+oid sha256:1c95aa7a4b53ecb3768d2a6608aa9e419059aef1c75b9b99e6b1c56328281c18
 size 528306

openvino_detokenizer.xml

@@ -1,16 +1,16 @@
 <?xml version="1.0"?>
 <net name="detokenizer" version="11">
 	<layers>
-		<layer id="0" name="Parameter_33240" type="Parameter" version="opset1">
+		<layer id="0" name="Parameter_33580" type="Parameter" version="opset1">
 			<data shape="?,?" element_type="i64" />
 			<output>
-				<port id="0" precision="I64" names="Parameter_33240">
+				<port id="0" precision="I64" names="Parameter_33580">
 					<dim>-1</dim>
 					<dim>-1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1" name="Convert_33256" type="Convert" version="opset1">
+		<layer id="1" name="Convert_33596" type="Convert" version="opset1">
 			<data destination_type="i32" />
 			<input>
 				<port id="0" precision="I64">
@@ -25,7 +25,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="2" name="Constant_33201" type="Const" version="opset1">
+		<layer id="2" name="Constant_33541" type="Const" version="opset1">
 			<data element_type="u8" shape="528257" offset="0" size="528257" />
 			<output>
 				<port id="0" precision="U8">
@@ -33,7 +33,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="3" name="StringTensorUnpack_33202" type="StringTensorUnpack" version="extension">
+		<layer id="3" name="StringTensorUnpack_33542" type="StringTensorUnpack" version="extension">
 			<data mode="begins_ends" />
 			<input>
 				<port id="0" precision="U8">
@@ -52,7 +52,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="4" name="VocabDecoder_33241" type="VocabDecoder" version="extension">
+		<layer id="4" name="VocabDecoder_33581" type="VocabDecoder" version="extension">
 			<data skip_tokens="50257, 50258, 50259, 50260, 50261, 50262, 50263, 50264, 50265, 50266, 50267, 50268, 50269, 50270, 50271, 50272, 50273, 50274, 50275, 50276, 50277, 50278, 50279, 50280, 50281, 50282, 50283, 50284, 50285, 50286, 50287, 50288, 50289, 50290, 50291, 50292, 50293, 50294, 50295, 50296, 50297, 50298, 50299, 50300, 50301, 50302, 50303, 50304, 50305, 50306, 50307, 50308, 50309, 50310, 50311, 50312, 50313, 50314, 50315, 50316, 50317, 50318, 50319, 50320, 50321, 50322, 50323, 50324, 50325, 50326, 50327, 50328, 50329, 50330, 50331, 50332, 50333, 50334, 50335, 50336, 50337, 50338, 50339, 50340, 50341, 50342, 50343, 50344, 50345, 50346, 50347, 50348, 50349, 50350, 50351, 50352, 50353, 50354, 50355, 50356, 50357, 50358, 50359, 50360, 50361, 50362, 50363, 50364" />
 			<input>
 				<port id="0" precision="I32">
@@ -87,7 +87,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="5" name="FuzeRagged_33242" type="FuzeRagged" version="extension">
+		<layer id="5" name="FuzeRagged_33582" type="FuzeRagged" version="extension">
 			<input>
 				<port id="0" precision="I32">
 					<dim>-1</dim>
@@ -111,7 +111,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="6" name="Constant_33244" type="Const" version="opset1">
+		<layer id="6" name="Constant_33584" type="Const" version="opset1">
 			<data element_type="u8" shape="47" offset="528257" size="47" />
 			<output>
 				<port id="0" precision="U8">
@@ -119,7 +119,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="7" name="Constant_33246" type="Const" version="opset1">
+		<layer id="7" name="Constant_33586" type="Const" version="opset1">
 			<data element_type="u8" shape="2" offset="528304" size="2" />
 			<output>
 				<port id="0" precision="U8">
@@ -127,7 +127,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="8" name="RegexNormalization_33247" type="RegexNormalization" version="extension">
+		<layer id="8" name="RegexNormalization_33587" type="RegexNormalization" version="extension">
 			<data global_replace="true" />
 			<input>
 				<port id="0" precision="I32">
@@ -158,7 +158,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="9" name="StringTensorPack_33248" type="StringTensorPack" version="extension">
+		<layer id="9" name="StringTensorPack_33588" type="StringTensorPack" version="extension">
 			<data mode="begins_ends" />
 			<input>
 				<port id="0" precision="I32">
@@ -177,7 +177,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="10" name="Result_33249" type="Result" version="opset1">
+		<layer id="10" name="Result_33589" type="Result" version="opset1">
 			<input>
 				<port id="0" precision="STRING">
 					<dim>-1</dim>
@@ -207,9 +207,29 @@
 		<edge from-layer="9" from-port="3" to-layer="10" to-port="0" />
 	</edges>
 	<rt_info>
+		<add_attention_mask value="True" />
+		<add_prefix_space />
+		<add_special_tokens value="True" />
 		<bos_token_id value="50257" />
+		<clean_up_tokenization_spaces />
+		<detokenizer_input_type value="i64" />
 		<eos_token_id value="50257" />
+		<handle_special_tokens_with_re />
+		<number_of_inputs value="1" />
+		<openvino_tokenizers_version value="2024.5.0.0" />
+		<openvino_version value="2024.5.0" />
 		<original_tokenizer_class value="&lt;class 'transformers.models.whisper.tokenization_whisper_fast.WhisperTokenizerFast'>" />
 		<pad_token_id value="50257" />
+		<sentencepiece_version value="0.2.0" />
+		<skip_special_tokens value="True" />
+		<streaming_detokenizer value="False" />
+		<tiktoken_version value="0.8.0" />
+		<tokenizer_output_type value="i64" />
+		<tokenizers_version value="0.20.3" />
+		<transformers_version value="4.46.3" />
+		<use_max_padding value="False" />
+		<use_sentencepiece_backend value="False" />
+		<utf8_replace_mode />
+		<with_detokenizer value="True" />
 	</rt_info>
 </net>

openvino_tokenizer.xml

@@ -1,27 +1,27 @@
 <?xml version="1.0"?>
 <net name="tokenizer" version="11">
 	<layers>
-		<layer id="0" name="Parameter_33120" type="Parameter" version="opset1">
+		<layer id="0" name="Parameter_33460" type="Parameter" version="opset1">
 			<data shape="?" element_type="string" />
 			<output>
-				<port id="0" precision="STRING" names="Parameter_33120">
+				<port id="0" precision="STRING" names="Parameter_33460">
 					<dim>-1</dim>
 				</port>
 			</output>
 		</layer>
-		<layer id="1" name="Constant_33218" type="Const" version="opset1">
+		<layer id="1" name="Constant_33558" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="0" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="2" name="Constant_33219" type="Const" version="opset1">
+		<layer id="2" name="Constant_33559" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="4" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="3" name="Constant_33220" type="Const" version="opset1">
+		<layer id="3" name="Constant_33560" type="Const" version="opset1">
 			<data element_type="i32" shape="1" offset="8" size="4" />
 			<output>
 				<port id="0" precision="I32">
@@ -29,19 +29,19 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="4" name="Constant_33221" type="Const" version="opset1">
+		<layer id="4" name="Constant_33561" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="0" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="5" name="Constant_33222" type="Const" version="opset1">
+		<layer id="5" name="Constant_33562" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="4" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="6" name="Constant_33223" type="Const" version="opset1">
+		<layer id="6" name="Constant_33563" type="Const" version="opset1">
 			<data element_type="i32" shape="1" offset="12" size="4" />
 			<output>
 				<port id="0" precision="I32">
@@ -49,13 +49,13 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="7" name="Constant_33126" type="Const" version="opset1">
+		<layer id="7" name="Constant_33466" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="16" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="8" name="StringTensorUnpack_33121" type="StringTensorUnpack" version="extension">
+		<layer id="8" name="StringTensorUnpack_33461" type="StringTensorUnpack" version="extension">
 			<data mode="begins_ends" />
 			<input>
 				<port id="0" precision="STRING">
@@ -74,7 +74,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="9" name="ShapeOf_33122" type="ShapeOf" version="opset3">
+		<layer id="9" name="ShapeOf_33462" type="ShapeOf" version="opset3">
 			<data output_type="i64" />
 			<input>
 				<port id="0" precision="I32">
@@ -87,19 +87,19 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="10" name="Constant_33123" type="Const" version="opset1">
+		<layer id="10" name="Constant_33463" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="16" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="11" name="Constant_33124" type="Const" version="opset1">
+		<layer id="11" name="Constant_33464" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="16" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="12" name="Gather_33125" type="Gather" version="opset8">
+		<layer id="12" name="Gather_33465" type="Gather" version="opset8">
 			<data batch_dims="0" />
 			<input>
 				<port id="0" precision="I64">
@@ -112,13 +112,13 @@
 				<port id="3" precision="I64" />
 			</output>
 		</layer>
-		<layer id="13" name="Constant_33127" type="Const" version="opset1">
+		<layer id="13" name="Constant_33467" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="24" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="14" name="Range_33128" type="Range" version="opset4">
+		<layer id="14" name="Range_33468" type="Range" version="opset4">
 			<data output_type="i32" />
 			<input>
 				<port id="0" precision="I64" />
@@ -131,19 +131,19 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="15" name="Constant_33129" type="Const" version="opset1">
+		<layer id="15" name="Constant_33469" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="24" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="16" name="Constant_33130" type="Const" version="opset1">
+		<layer id="16" name="Constant_33470" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="24" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="17" name="Add_33131" type="Add" version="opset1">
+		<layer id="17" name="Add_33471" type="Add" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="I64" />
@@ -153,13 +153,13 @@
 				<port id="2" precision="I64" />
 			</output>
 		</layer>
-		<layer id="18" name="Constant_33132" type="Const" version="opset1">
+		<layer id="18" name="Constant_33472" type="Const" version="opset1">
 			<data element_type="i64" shape="" offset="24" size="8" />
 			<output>
 				<port id="0" precision="I64" />
 			</output>
 		</layer>
-		<layer id="19" name="Range_33133" type="Range" version="opset4">
+		<layer id="19" name="Range_33473" type="Range" version="opset4">
 			<data output_type="i32" />
 			<input>
 				<port id="0" precision="I64" />
@@ -172,7 +172,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="20" name="Constant_33195" type="Const" version="opset1">
+		<layer id="20" name="Constant_33535" type="Const" version="opset1">
 			<data element_type="u8" shape="26491" offset="32" size="26491" />
 			<output>
 				<port id="0" precision="U8">
@@ -180,7 +180,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="21" name="SpecialTokensSplit_33196" type="SpecialTokensSplit" version="extension">
+		<layer id="21" name="SpecialTokensSplit_33536" type="SpecialTokensSplit" version="extension">
 			<input>
 				<port id="0" precision="I32">
 					<dim>-1</dim>
@@ -222,7 +222,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="22" name="Constant_33198" type="Const" version="opset1">
+		<layer id="22" name="Constant_33538" type="Const" version="opset1">
 			<data element_type="u8" shape="64" offset="26523" size="64" />
 			<output>
 				<port id="0" precision="U8">
@@ -230,7 +230,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="23" name="RegexSplit_33199" type="RegexSplit" version="extension">
+		<layer id="23" name="RegexSplit_33539" type="RegexSplit" version="extension">
 			<data behaviour="isolate" invert="false" max_splits="-1" />
 			<input>
 				<port id="0" precision="I32">
@@ -276,7 +276,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="24" name="Constant_33201" type="Const" version="opset1">
+		<layer id="24" name="Constant_33541" type="Const" version="opset1">
 			<data element_type="u8" shape="528257" offset="26587" size="528257" />
 			<output>
 				<port id="0" precision="U8">
@@ -284,7 +284,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="25" name="StringTensorUnpack_33202" type="StringTensorUnpack" version="extension">
+		<layer id="25" name="StringTensorUnpack_33542" type="StringTensorUnpack" version="extension">
 			<data mode="begins_ends" />
 			<input>
 				<port id="0" precision="U8">
@@ -303,7 +303,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="26" name="Constant_33207" type="Const" version="opset1">
+		<layer id="26" name="Constant_33547" type="Const" version="opset1">
 			<data element_type="u8" shape="369958" offset="554844" size="369958" />
 			<output>
 				<port id="0" precision="U8">
@@ -311,7 +311,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="27" name="StringTensorUnpack_33208" type="StringTensorUnpack" version="extension">
+		<layer id="27" name="StringTensorUnpack_33548" type="StringTensorUnpack" version="extension">
 			<data mode="begins_ends" />
 			<input>
 				<port id="0" precision="U8">
@@ -330,7 +330,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="28" name="Constant_33210" type="Const" version="opset1">
+		<layer id="28" name="Constant_33550" type="Const" version="opset1">
 			<data element_type="u8" shape="336859" offset="924802" size="336859" />
 			<output>
 				<port id="0" precision="U8">
@@ -338,7 +338,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="29" name="StringTensorUnpack_33211" type="StringTensorUnpack" version="extension">
+		<layer id="29" name="StringTensorUnpack_33551" type="StringTensorUnpack" version="extension">
 			<data mode="begins_ends" />
 			<input>
 				<port id="0" precision="U8">
@@ -357,7 +357,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="30" name="Constant_33204" type="Const" version="opset1">
+		<layer id="30" name="Constant_33544" type="Const" version="opset1">
 			<data element_type="u8" shape="20172" offset="1261661" size="20172" />
 			<output>
 				<port id="0" precision="U8">
@@ -365,7 +365,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="31" name="StringTensorUnpack_33205" type="StringTensorUnpack" version="extension">
+		<layer id="31" name="StringTensorUnpack_33545" type="StringTensorUnpack" version="extension">
 			<data mode="begins_ends" />
 			<input>
 				<port id="0" precision="U8">
@@ -384,7 +384,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="32" name="Constant_33212" type="Const" version="opset1">
+		<layer id="32" name="Constant_33552" type="Const" version="opset1">
 			<data element_type="i32" shape="1609" offset="1281833" size="6436" />
 			<output>
 				<port id="0" precision="I32">
@@ -392,7 +392,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="33" name="BPETokenizer_33213" type="BPETokenizer" version="extension">
+		<layer id="33" name="BPETokenizer_33553" type="BPETokenizer" version="extension">
 			<data unk_token="" fuse_unk="false" suffix_indicator="" end_suffix="" byte_fallback="false" cache_capacity="20000" />
 			<input>
 				<port id="0" precision="I32">
@@ -462,7 +462,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="34" name="Subtract_33214" type="Subtract" version="opset1">
+		<layer id="34" name="Subtract_33554" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="I32">
@@ -478,13 +478,13 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="35" name="Constant_33215" type="Const" version="opset1">
+		<layer id="35" name="Constant_33555" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="1288269" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="36" name="Minimum_33216" type="Minimum" version="opset1">
+		<layer id="36" name="Minimum_33556" type="Minimum" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="I32">
@@ -498,7 +498,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="37" name="Add_33217" type="Add" version="opset1">
+		<layer id="37" name="Add_33557" type="Add" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="I32">
@@ -514,19 +514,19 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="38" name="Constant_33224" type="Const" version="opset1">
+		<layer id="38" name="Constant_33564" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="0" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="39" name="Constant_33225" type="Const" version="opset1">
+		<layer id="39" name="Constant_33565" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="4" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="40" name="Constant_33226" type="Const" version="opset1">
+		<layer id="40" name="Constant_33566" type="Const" version="opset1">
 			<data element_type="i32" shape="1" offset="1288273" size="4" />
 			<output>
 				<port id="0" precision="I32">
@@ -534,7 +534,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="41" name="Constant_33227" type="Const" version="opset1">
+		<layer id="41" name="Constant_33567" type="Const" version="opset1">
 			<data element_type="i32" shape="4" offset="1288277" size="16" />
 			<output>
 				<port id="0" precision="I32">
@@ -542,7 +542,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="42" name="CombineSegments_33228" type="CombineSegments" version="extension">
+		<layer id="42" name="CombineSegments_33568" type="CombineSegments" version="extension">
 			<input>
 				<port id="0" precision="I32" />
 				<port id="1" precision="I32" />
@@ -593,7 +593,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="43" name="Subtract_33229" type="Subtract" version="opset1">
+		<layer id="43" name="Subtract_33569" type="Subtract" version="opset1">
 			<data auto_broadcast="numpy" />
 			<input>
 				<port id="0" precision="I32">
@@ -609,13 +609,13 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="44" name="Constant_33230" type="Const" version="opset1">
+		<layer id="44" name="Constant_33570" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="0" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="45" name="ReduceMax_33231" type="ReduceMax" version="opset1">
+		<layer id="45" name="ReduceMax_33571" type="ReduceMax" version="opset1">
 			<data keep_dims="false" />
 			<input>
 				<port id="0" precision="I32">
@@ -627,13 +627,13 @@
 				<port id="2" precision="I32" />
 			</output>
 		</layer>
-		<layer id="46" name="Constant_33232" type="Const" version="opset1">
+		<layer id="46" name="Constant_33572" type="Const" version="opset1">
 			<data element_type="i32" shape="" offset="1288273" size="4" />
 			<output>
 				<port id="0" precision="I32" />
 			</output>
 		</layer>
-		<layer id="47" name="RaggedToDense_33233" type="RaggedToDense" version="extension">
+		<layer id="47" name="RaggedToDense_33573" type="RaggedToDense" version="extension">
 			<data pad_right="true" />
 			<input>
 				<port id="0" precision="I32">
@@ -659,7 +659,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="48" name="Convert_33234" type="Convert" version="opset1">
+		<layer id="48" name="Convert_33574" type="Convert" version="opset1">
 			<data destination_type="i32" />
 			<input>
 				<port id="0" precision="BOOL">
@@ -674,7 +674,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="49" name="Convert_33234" type="Convert" version="opset1">
+		<layer id="49" name="Convert_33574" type="Convert" version="opset1">
 			<data destination_type="i64" />
 			<input>
 				<port id="0" precision="I32">
@@ -689,7 +689,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="51" name="RaggedToDense_33233.0" type="Convert" version="opset1">
+		<layer id="51" name="RaggedToDense_33573.0" type="Convert" version="opset1">
 			<data destination_type="i64" />
 			<input>
 				<port id="0" precision="I32">
@@ -704,7 +704,7 @@
 				</port>
 			</output>
 		</layer>
-		<layer id="52" name="Result_33237" type="Result" version="opset1">
+		<layer id="52" name="Result_33577" type="Result" version="opset1">
 			<input>
 				<port id="0" precision="I64">
 					<dim>-1</dim>
@@ -712,7 +712,7 @@
 				</port>
 			</input>
 		</layer>
-		<layer id="50" name="Result_33239" type="Result" version="opset1">
+		<layer id="50" name="Result_33579" type="Result" version="opset1">
 			<input>
 				<port id="0" precision="I64">
 					<dim>-1</dim>
@@ -805,9 +805,29 @@
 		<edge from-layer="51" from-port="1" to-layer="52" to-port="0" />
 	</edges>
 	<rt_info>
+		<add_attention_mask value="True" />
+		<add_prefix_space />
+		<add_special_tokens value="True" />
 		<bos_token_id value="50257" />
+		<clean_up_tokenization_spaces />
+		<detokenizer_input_type value="i64" />
 		<eos_token_id value="50257" />
+		<handle_special_tokens_with_re />
+		<number_of_inputs value="1" />
+		<openvino_tokenizers_version value="2024.5.0.0" />
+		<openvino_version value="2024.5.0" />
 		<original_tokenizer_class value="&lt;class 'transformers.models.whisper.tokenization_whisper_fast.WhisperTokenizerFast'>" />
 		<pad_token_id value="50257" />
+		<sentencepiece_version value="0.2.0" />
+		<skip_special_tokens value="True" />
+		<streaming_detokenizer value="False" />
+		<tiktoken_version value="0.8.0" />
+		<tokenizer_output_type value="i64" />
+		<tokenizers_version value="0.20.3" />
+		<transformers_version value="4.46.3" />
+		<use_max_padding value="False" />
+		<use_sentencepiece_backend value="False" />
+		<utf8_replace_mode />
+		<with_detokenizer value="True" />
 	</rt_info>
 </net>