Upload folder using huggingface_hub

config.json

@@ -143,8 +143,7 @@
     50361,
     50362
   ],
-  "torch_dtype": "float32",
-  "transformers_version": "4.29.2",
+  "transformers_version": "4.30.2",
   "use_cache": true,
   "use_weighted_layer_sum": false,
   "vocab_size": 51865

generation_config.json

@@ -217,5 +217,5 @@
     "transcribe": 50359,
     "translate": 50358
   },
-  "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 sha256:289504e89417c90b51380c46cf839d192d9a9ac94ff50d64106322636963830a
+oid sha256:c09ce6b24db9c8a55eab37aba77fb217052da64fa7f30d5f0411008644bc0e15
 size 774594491

onnx/decoder_model_merged_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d207cac77f0112c57a0d45e43ea3e38996cea95dcc49da4f229e0c8dc578531e
-size 197291535
+oid sha256:df24179148e324c957c5740928e54c58f7894db9022949eb4ba4e5a5971aa4b9
+size 196432559

onnx/decoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f870b64ca059126c053fe199770e71f6f69e28a871ce212c266ae4fecf8aeeef
-size 196598144
+oid sha256:48bed0449e061d9ceef7a1b6b54800e9c730478e769404309e34dca27e4a89f4
+size 195739168

onnx/decoder_with_past_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bcf095f74de4ed450b92d871577848399f653d5d49cad07fec42acc333ec2c2d
-size 182185462
+oid sha256:f25f8c98570159bb7ed56df104a949ed523087a13b484f71bc9a351e50eec8b7
+size 181418766

onnx/encoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c6bab7874f6ec1e7ee1ae0c0fe0c08705e0810f2f68eb41b354f1fa0b160d9b7
-size 92704740
+oid sha256:736382d9bb09533d595c3c2d96e9f8716e9facfe10242b8ab4c8e802377665d3
+size 92289693

quant_config.json

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