Upload folder using huggingface_hub

config.json

@@ -139,8 +139,7 @@
     50360,
     50361
   ],
-  "torch_dtype": "float32",
-  "transformers_version": "4.29.2",
+  "transformers_version": "4.30.2",
   "use_cache": true,
   "use_weighted_layer_sum": false,
   "vocab_size": 51864

generation_config.json

@@ -110,5 +110,5 @@
     50360,
     50361
   ],
-  "transformers_version": "4.29.2"
+  "transformers_version": "4.30.2"
 }

onnx/decoder_model_merged.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9e85f9a29a0b0e1dfdb2211ac09ff1bda4faa5808b5a9855a843982367ac58c2
-size 198194454
+oid sha256:441da7a85e2cac5508fad6795cf115333b8566cffc886326788e0a64366411a5
+size 198194466

onnx/decoder_model_merged_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8b79745272590fbc4e841188d1ce522efbca6d593a7dbbc35a56b87908b211ee
-size 50942904
+oid sha256:29e4ab70caaedbfd18a4cc4fab3d0708b9eb7c888f54f31f11949a6a040fcf36
+size 50583569

onnx/decoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7e845d7c9dc4317bfe89fd01e6f2cd14fa0b045cd099f229ee067ad9f62a0ec7
-size 50707084
+oid sha256:699a1478cd7b1126a8f345281c1fb29fa72a18b8f45e4b31c6499653e3dee1d2
+size 50347713

onnx/decoder_with_past_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:11085368e8f75c52e9ed52d75f1bdf149aa9f0dcb29f9f2a3795975fb755f08a
-size 49457772
+oid sha256:1182513a4c3ccd16fbbabb2a760f86eac2e827d9f8ce4c78c33e3935af86526c
+size 49113801

onnx/encoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8149b739060eaa30b6211a5e107e33e422f5c4d914229b94357f31bac4ed68f0
-size 10182475
+oid sha256:66bb0bf174a3a3bcfdaccea161c811d0d90a25c927021ab9ffbf639b902225f1
+size 10113251

quant_config.json

@@ -0,0 +1,89 @@
+{
+    "per_channel": false,
+    "reduce_range": false,
+    "per_model_config": {
+        "encoder_model": {
+            "op_types": [
+                "Transpose",
+                "ReduceMean",
+                "Sub",
+                "Gather",
+                "Conv",
+                "Shape",
+                "Pow",
+                "MatMul",
+                "Div",
+                "Erf",
+                "Softmax",
+                "Reshape",
+                "Concat",
+                "Add",
+                "Unsqueeze",
+                "Constant",
+                "Mul",
+                "Sqrt"
+            ],
+            "weight_type": "QUInt8"
+        },
+        "decoder_model": {
+            "op_types": [
+                "ReduceMean",
+                "Sub",
+                "Pow",
+                "Cast",
+                "MatMul",
+                "Softmax",
+                "Where",
+                "Range",
+                "Equal",
+                "Less",
+                "Div",
+                "Erf",
+                "Expand",
+                "Concat",
+                "Mul",
+                "Sqrt",
+                "Transpose",
+                "Gather",
+                "Shape",
+                "ConstantOfShape",
+                "Slice",
+                "Squeeze",
+                "Reshape",
+                "Add",
+                "Unsqueeze",
+                "Constant"
+            ],
+            "weight_type": "QInt8"
+        },
+        "decoder_model_merged": {
+            "op_types": [
+                "If"
+            ],
+            "weight_type": "QInt8"
+        },
+        "decoder_with_past_model": {
+            "op_types": [
+                "ReduceMean",
+                "Gather",
+                "Sub",
+                "Transpose",
+                "Shape",
+                "Slice",
+                "Pow",
+                "MatMul",
+                "Div",
+                "Softmax",
+                "Erf",
+                "Reshape",
+                "Concat",
+                "Add",
+                "Unsqueeze",
+                "Constant",
+                "Mul",
+                "Sqrt"
+            ],
+            "weight_type": "QInt8"
+        }
+    }
+}