quantize per channel

config.json

@@ -1,5 +1,5 @@
 {
-  "_name_or_path": "pszemraj_nougat-base-onnx",
+  "_name_or_path": "pszemraj_nougat-small-onnx",
   "architectures": [
     "VisionEncoderDecoderModel"
   ],
@@ -21,7 +21,7 @@
     "decoder_attention_heads": 16,
     "decoder_ffn_dim": 4096,
     "decoder_layerdrop": 0.0,
-    "decoder_layers": 10,
+    "decoder_layers": 4,
     "decoder_start_token_id": null,
     "diversity_penalty": 0.0,
     "do_sample": false,
@@ -50,7 +50,7 @@
     },
     "length_penalty": 1.0,
     "max_length": 20,
-    "max_position_embeddings": 4096,
+    "max_position_embeddings": 3584,
     "min_length": 0,
     "model_type": "mbart",
     "no_repeat_ngram_size": 0,
@@ -185,5 +185,5 @@
   "model_type": "vision-encoder-decoder",
   "tie_word_embeddings": false,
   "torch_dtype": "float32",
-  "transformers_version": "4.35.2"
+  "transformers_version": "4.38.1"
 }

decoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2fef9b60876d3101101fbbd076ddd51cb7d34ebf46756ed407f37cb5f19ed66d
-size 275952336
+oid sha256:e9c8f1bb7456441320245595f93fa052612b9f477b71f4822a46c070faf48e8d
+size 174444105

decoder_with_past_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:828769fec141abee78ecd80c40fdc74876c409a179b080bca2dc52d62bd40ed5
-size 254784072
+oid sha256:5a68ddb99f8e476a6ba4e356145440de0fa70a5d2cff89c136feba9691f27c98
+size 165923718

encoder_model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6e5a8b5be2bafd166b95f2bd0f387a457c8ba90a36d1c7bee422ea197d4fbae7
-size 81514091
+oid sha256:6bec906871770afe111f0f3474bd5cb143cfd7f659c242431ee5431be580684d
+size 82052587

ort_config.json

@@ -2,7 +2,7 @@
   "one_external_file": true,
   "opset": null,
   "optimization": {},
-  "optimum_version": "1.14.1",
+  "optimum_version": "1.17.1",
   "quantization": {
     "activations_dtype": "QUInt8",
     "activations_symmetric": false,
@@ -20,7 +20,7 @@
       "Transpose",
       "EmbedLayerNormalization"
     ],
-    "per_channel": false,
+    "per_channel": true,
     "qdq_add_pair_to_weight": false,
     "qdq_dedicated_pair": false,
     "qdq_op_type_per_channel_support_to_axis": {
@@ -30,6 +30,6 @@
     "weights_dtype": "QUInt8",
     "weights_symmetric": true
   },
-  "transformers_version": "4.35.2",
+  "transformers_version": "4.38.1",
   "use_external_data_format": false
 }

tokenizer.json

@@ -2,13 +2,13 @@
   "version": "1.0",
   "truncation": {
     "direction": "Right",
-    "max_length": 4096,
+    "max_length": 3584,
     "strategy": "LongestFirst",
     "stride": 0
   },
   "padding": {
     "strategy": {
-      "Fixed": 4096
+      "Fixed": 3584
     },
     "direction": "Right",
     "pad_to_multiple_of": null,

tokenizer_config.json

@@ -189,7 +189,7 @@
   "bos_token": "<s>",
   "clean_up_tokenization_spaces": false,
   "eos_token": "</s>",
-  "max_length": 4096,
+  "max_length": 3584,
   "model_max_length": 3584,
   "pad_to_multiple_of": null,
   "pad_token": "<pad>",