Upload folder using huggingface_hub

config.json

@@ -42,6 +42,7 @@
   "max_length": 448,
   "max_source_positions": 1500,
   "max_target_positions": 448,
+  "median_filter_width": 7,
   "model_type": "whisper",
   "num_hidden_layers": 4,
   "num_mel_bins": 80,
@@ -139,7 +140,7 @@
     50360,
     50361
   ],
-  "transformers_version": "4.30.2",
+  "transformers_version": "4.33.0.dev0",
   "use_cache": true,
   "use_weighted_layer_sum": false,
   "vocab_size": 51864

generation_config.json

@@ -1,5 +1,38 @@
 {
-  "alignment_heads": [[1, 0], [2, 0], [2, 5], [3, 0], [3, 1], [3, 2], [3, 3], [3, 4]],
+  "alignment_heads": [
+    [
+      1,
+      0
+    ],
+    [
+      2,
+      0
+    ],
+    [
+      2,
+      5
+    ],
+    [
+      3,
+      0
+    ],
+    [
+      3,
+      1
+    ],
+    [
+      3,
+      2
+    ],
+    [
+      3,
+      3
+    ],
+    [
+      3,
+      4
+    ]
+  ],
   "begin_suppress_tokens": [
     220,
     50256
@@ -111,5 +144,5 @@
     50360,
     50361
   ],
-  "transformers_version": "4.30.2"
+  "transformers_version": "4.33.0.dev0"
 }

onnx/decoder_model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f8bec5100bcac885428dc96253d46df583260bac7ec5c99a50897de8d6eb7f73
-size 198046458
+oid sha256:38f0b91131d1f1c8b2b5f1d1096f61e44ded18be49313c80abc5c11daf03bd28
+size 118405120

onnx/decoder_model_merged.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:441da7a85e2cac5508fad6795cf115333b8566cffc886326788e0a64366411a5
-size 198194466
+oid sha256:950978b1dbcbf250335358c1236053ba19a7f7849b33dc777f4421b72b7626fa
+size 118578063

onnx/decoder_model_merged_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:29e4ab70caaedbfd18a4cc4fab3d0708b9eb7c888f54f31f11949a6a040fcf36
-size 50583569
+oid sha256:dbb2e063b7fbc41d9803b9698f93ecb035c50cbb3fb87b56cb131e4a5eb99059
+size 30727382

onnx/decoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:699a1478cd7b1126a8f345281c1fb29fa72a18b8f45e4b31c6499653e3dee1d2
-size 50347713
+oid sha256:62b1761a6c23961cdabbf7caa25495525d78528a64baacf7e4283c83cb2941f1
+size 30459718

onnx/decoder_with_past_model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dca3143327bca6dacbb59f4858e7a4901d0ac719cf468ce48a8e21494f8f1cca
-size 193292243
+oid sha256:52b796cccc711df396afa098ae7cb6271d7fad42e4e7188564448cfa99aed830
+size 113653015

onnx/decoder_with_past_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1182513a4c3ccd16fbbabb2a760f86eac2e827d9f8ce4c78c33e3935af86526c
-size 49113801
+oid sha256:2c7390ddd38c567a51a73f175c6655afcbcd5cf1d5dc6be5badc89cba3f1d59c
+size 29228736

onnx/encoder_model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:530ff754d57158293aea418249399a20af813593005195f84e16ab54150e5589
-size 32900723
+oid sha256:a048dcf0cde98db805f46be32b75d778cf824aad20b51a02e5b9cff457426238
+size 32909539

onnx/encoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:66bb0bf174a3a3bcfdaccea161c811d0d90a25c927021ab9ffbf639b902225f1
-size 10113251
+oid sha256:8cc3c6f8563d1b3fbd2c5af9f64c2bed8b020bc593c402d1ef53b9f08fbf1b90
+size 10124913

quantize_config.json

@@ -0,0 +1,115 @@
+{
+    "per_channel": false,
+    "reduce_range": false,
+    "per_model_config": {
+        "decoder_model": {
+            "op_types": [
+                "ReduceMean",
+                "Concat",
+                "Sqrt",
+                "Less",
+                "Unsqueeze",
+                "Add",
+                "Cast",
+                "Div",
+                "Equal",
+                "Softmax",
+                "Pow",
+                "Squeeze",
+                "Where",
+                "Sub",
+                "ConstantOfShape",
+                "Erf",
+                "Transpose",
+                "Reshape",
+                "MatMul",
+                "Gather",
+                "Shape",
+                "Mul",
+                "Range",
+                "Constant",
+                "Slice",
+                "Expand"
+            ],
+            "weight_type": "QInt8"
+        },
+        "decoder_model_merged": {
+            "op_types": [
+                "ReduceMean",
+                "Concat",
+                "Sqrt",
+                "Less",
+                "If",
+                "Unsqueeze",
+                "Add",
+                "Cast",
+                "Div",
+                "Equal",
+                "Softmax",
+                "Pow",
+                "Squeeze",
+                "Where",
+                "Sub",
+                "ConstantOfShape",
+                "Erf",
+                "Transpose",
+                "Reshape",
+                "MatMul",
+                "Gather",
+                "Shape",
+                "Mul",
+                "Range",
+                "Constant",
+                "Slice",
+                "Expand"
+            ],
+            "weight_type": "QInt8"
+        },
+        "decoder_with_past_model": {
+            "op_types": [
+                "MatMul",
+                "Gather",
+                "ReduceMean",
+                "Sub",
+                "Div",
+                "Concat",
+                "Sqrt",
+                "Shape",
+                "Softmax",
+                "Pow",
+                "Erf",
+                "Transpose",
+                "Reshape",
+                "Mul",
+                "Constant",
+                "Unsqueeze",
+                "Add",
+                "Slice"
+            ],
+            "weight_type": "QInt8"
+        },
+        "encoder_model": {
+            "op_types": [
+                "MatMul",
+                "ReduceMean",
+                "Conv",
+                "Div",
+                "Sub",
+                "Sqrt",
+                "Gather",
+                "Shape",
+                "Concat",
+                "Softmax",
+                "Pow",
+                "Erf",
+                "Transpose",
+                "Reshape",
+                "Mul",
+                "Constant",
+                "Unsqueeze",
+                "Add"
+            ],
+            "weight_type": "QUInt8"
+        }
+    }
+}