Upload Multilingual-e5-large-Q8

config.json

@@ -80,9 +80,11 @@
   "use_cache": true,
   "classifier_dropout": null,
   "revis_quantization": {
-    "format": "revis-xlm-roberta-e5-q8",
+    "format": "revis-xlm-roberta-e5-mlx-native-q8",
     "bits": 8,
-    "type": "symmetric-per-row",
+    "group_size": 64,
+    "mode": "affine",
+    "type": "mlx-native-affine",
     "manifest": "quantization.json"
   }
 }

quantization.json

@@ -1,15 +1,19 @@
 {
-  "schemaVersion": 1,
-  "format": "revis-xlm-roberta-e5-q8",
+  "schemaVersion": 2,
+  "format": "revis-xlm-roberta-e5-mlx-native-q8",
   "source": "intfloat/multilingual-e5-large",
   "baseWeights": "weights.00.safetensors",
   "quantizedWeights": "weights.00.safetensors",
   "quantization": {
-    "type": "symmetric-per-row",
+    "type": "mlx-native-affine",
     "bits": 8,
+    "group_size": 64,
+    "mode": "affine",
     "quantizedTensorSuffix": ".qweight",
-    "scaleTensorSuffix": ".scale",
-    "dequantization": "weight = qweight.astype(float32) * scale[:, None]"
+    "scaleTensorSuffix": ".scales",
+    "biasTensorSuffix": ".biases",
+    "linear": "mx.quantized_matmul(x, qweight, scales=scales, biases=biases, transpose=True, group_size=group_size, bits=bits, mode=mode)",
+    "embedding": "mx.dequantize(qweight[ids], scales=scales[ids], biases=biases[ids], group_size=group_size, bits=bits, mode=mode)"
   },
   "quantized": [
     {
@@ -20,7 +24,8 @@
       ],
       "dtype": "float16",
       "qweight": "embeddings.position_embeddings.weight.qweight",
-      "scale": "embeddings.position_embeddings.weight.scale"
+      "scales": "embeddings.position_embeddings.weight.scales",
+      "biases": "embeddings.position_embeddings.weight.biases"
     },
     {
       "name": "embeddings.token_type_embeddings.weight",
@@ -30,7 +35,8 @@
       ],
       "dtype": "float16",
       "qweight": "embeddings.token_type_embeddings.weight.qweight",
-      "scale": "embeddings.token_type_embeddings.weight.scale"
+      "scales": "embeddings.token_type_embeddings.weight.scales",
+      "biases": "embeddings.token_type_embeddings.weight.biases"
     },
     {
       "name": "embeddings.word_embeddings.weight",
@@ -40,7 +46,8 @@
       ],
       "dtype": "float16",
       "qweight": "embeddings.word_embeddings.weight.qweight",
-      "scale": "embeddings.word_embeddings.weight.scale"
+      "scales": "embeddings.word_embeddings.weight.scales",
+      "biases": "embeddings.word_embeddings.weight.biases"
     },
     {
       "name": "encoder.layer.0.attention.output.dense.weight",
@@ -50,7 +57,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.0.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.0.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.0.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.0.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.0.attention.self.key.weight",
@@ -60,7 +68,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.0.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.0.attention.self.key.weight.scale"
+      "scales": "encoder.layer.0.attention.self.key.weight.scales",
+      "biases": "encoder.layer.0.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.0.attention.self.query.weight",
@@ -70,7 +79,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.0.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.0.attention.self.query.weight.scale"
+      "scales": "encoder.layer.0.attention.self.query.weight.scales",
+      "biases": "encoder.layer.0.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.0.attention.self.value.weight",
@@ -80,7 +90,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.0.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.0.attention.self.value.weight.scale"
+      "scales": "encoder.layer.0.attention.self.value.weight.scales",
+      "biases": "encoder.layer.0.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.0.intermediate.dense.weight",
@@ -90,7 +101,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.0.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.0.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.0.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.0.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.0.output.dense.weight",
@@ -100,7 +112,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.0.output.dense.weight.qweight",
-      "scale": "encoder.layer.0.output.dense.weight.scale"
+      "scales": "encoder.layer.0.output.dense.weight.scales",
+      "biases": "encoder.layer.0.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.1.attention.output.dense.weight",
@@ -110,7 +123,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.1.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.1.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.1.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.1.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.1.attention.self.key.weight",
@@ -120,7 +134,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.1.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.1.attention.self.key.weight.scale"
+      "scales": "encoder.layer.1.attention.self.key.weight.scales",
+      "biases": "encoder.layer.1.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.1.attention.self.query.weight",
@@ -130,7 +145,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.1.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.1.attention.self.query.weight.scale"
+      "scales": "encoder.layer.1.attention.self.query.weight.scales",
+      "biases": "encoder.layer.1.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.1.attention.self.value.weight",
@@ -140,7 +156,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.1.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.1.attention.self.value.weight.scale"
+      "scales": "encoder.layer.1.attention.self.value.weight.scales",
+      "biases": "encoder.layer.1.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.1.intermediate.dense.weight",
@@ -150,7 +167,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.1.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.1.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.1.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.1.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.1.output.dense.weight",
@@ -160,7 +178,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.1.output.dense.weight.qweight",
-      "scale": "encoder.layer.1.output.dense.weight.scale"
+      "scales": "encoder.layer.1.output.dense.weight.scales",
+      "biases": "encoder.layer.1.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.10.attention.output.dense.weight",
@@ -170,7 +189,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.10.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.10.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.10.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.10.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.10.attention.self.key.weight",
@@ -180,7 +200,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.10.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.10.attention.self.key.weight.scale"
+      "scales": "encoder.layer.10.attention.self.key.weight.scales",
+      "biases": "encoder.layer.10.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.10.attention.self.query.weight",
@@ -190,7 +211,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.10.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.10.attention.self.query.weight.scale"
+      "scales": "encoder.layer.10.attention.self.query.weight.scales",
+      "biases": "encoder.layer.10.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.10.attention.self.value.weight",
@@ -200,7 +222,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.10.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.10.attention.self.value.weight.scale"
+      "scales": "encoder.layer.10.attention.self.value.weight.scales",
+      "biases": "encoder.layer.10.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.10.intermediate.dense.weight",
@@ -210,7 +233,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.10.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.10.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.10.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.10.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.10.output.dense.weight",
@@ -220,7 +244,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.10.output.dense.weight.qweight",
-      "scale": "encoder.layer.10.output.dense.weight.scale"
+      "scales": "encoder.layer.10.output.dense.weight.scales",
+      "biases": "encoder.layer.10.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.11.attention.output.dense.weight",
@@ -230,7 +255,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.11.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.11.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.11.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.11.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.11.attention.self.key.weight",
@@ -240,7 +266,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.11.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.11.attention.self.key.weight.scale"
+      "scales": "encoder.layer.11.attention.self.key.weight.scales",
+      "biases": "encoder.layer.11.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.11.attention.self.query.weight",
@@ -250,7 +277,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.11.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.11.attention.self.query.weight.scale"
+      "scales": "encoder.layer.11.attention.self.query.weight.scales",
+      "biases": "encoder.layer.11.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.11.attention.self.value.weight",
@@ -260,7 +288,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.11.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.11.attention.self.value.weight.scale"
+      "scales": "encoder.layer.11.attention.self.value.weight.scales",
+      "biases": "encoder.layer.11.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.11.intermediate.dense.weight",
@@ -270,7 +299,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.11.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.11.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.11.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.11.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.11.output.dense.weight",
@@ -280,7 +310,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.11.output.dense.weight.qweight",
-      "scale": "encoder.layer.11.output.dense.weight.scale"
+      "scales": "encoder.layer.11.output.dense.weight.scales",
+      "biases": "encoder.layer.11.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.12.attention.output.dense.weight",
@@ -290,7 +321,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.12.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.12.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.12.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.12.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.12.attention.self.key.weight",
@@ -300,7 +332,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.12.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.12.attention.self.key.weight.scale"
+      "scales": "encoder.layer.12.attention.self.key.weight.scales",
+      "biases": "encoder.layer.12.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.12.attention.self.query.weight",
@@ -310,7 +343,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.12.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.12.attention.self.query.weight.scale"
+      "scales": "encoder.layer.12.attention.self.query.weight.scales",
+      "biases": "encoder.layer.12.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.12.attention.self.value.weight",
@@ -320,7 +354,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.12.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.12.attention.self.value.weight.scale"
+      "scales": "encoder.layer.12.attention.self.value.weight.scales",
+      "biases": "encoder.layer.12.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.12.intermediate.dense.weight",
@@ -330,7 +365,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.12.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.12.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.12.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.12.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.12.output.dense.weight",
@@ -340,7 +376,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.12.output.dense.weight.qweight",
-      "scale": "encoder.layer.12.output.dense.weight.scale"
+      "scales": "encoder.layer.12.output.dense.weight.scales",
+      "biases": "encoder.layer.12.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.13.attention.output.dense.weight",
@@ -350,7 +387,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.13.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.13.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.13.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.13.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.13.attention.self.key.weight",
@@ -360,7 +398,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.13.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.13.attention.self.key.weight.scale"
+      "scales": "encoder.layer.13.attention.self.key.weight.scales",
+      "biases": "encoder.layer.13.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.13.attention.self.query.weight",
@@ -370,7 +409,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.13.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.13.attention.self.query.weight.scale"
+      "scales": "encoder.layer.13.attention.self.query.weight.scales",
+      "biases": "encoder.layer.13.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.13.attention.self.value.weight",
@@ -380,7 +420,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.13.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.13.attention.self.value.weight.scale"
+      "scales": "encoder.layer.13.attention.self.value.weight.scales",
+      "biases": "encoder.layer.13.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.13.intermediate.dense.weight",
@@ -390,7 +431,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.13.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.13.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.13.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.13.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.13.output.dense.weight",
@@ -400,7 +442,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.13.output.dense.weight.qweight",
-      "scale": "encoder.layer.13.output.dense.weight.scale"
+      "scales": "encoder.layer.13.output.dense.weight.scales",
+      "biases": "encoder.layer.13.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.14.attention.output.dense.weight",
@@ -410,7 +453,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.14.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.14.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.14.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.14.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.14.attention.self.key.weight",
@@ -420,7 +464,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.14.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.14.attention.self.key.weight.scale"
+      "scales": "encoder.layer.14.attention.self.key.weight.scales",
+      "biases": "encoder.layer.14.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.14.attention.self.query.weight",
@@ -430,7 +475,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.14.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.14.attention.self.query.weight.scale"
+      "scales": "encoder.layer.14.attention.self.query.weight.scales",
+      "biases": "encoder.layer.14.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.14.attention.self.value.weight",
@@ -440,7 +486,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.14.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.14.attention.self.value.weight.scale"
+      "scales": "encoder.layer.14.attention.self.value.weight.scales",
+      "biases": "encoder.layer.14.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.14.intermediate.dense.weight",
@@ -450,7 +497,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.14.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.14.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.14.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.14.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.14.output.dense.weight",
@@ -460,7 +508,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.14.output.dense.weight.qweight",
-      "scale": "encoder.layer.14.output.dense.weight.scale"
+      "scales": "encoder.layer.14.output.dense.weight.scales",
+      "biases": "encoder.layer.14.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.15.attention.output.dense.weight",
@@ -470,7 +519,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.15.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.15.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.15.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.15.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.15.attention.self.key.weight",
@@ -480,7 +530,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.15.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.15.attention.self.key.weight.scale"
+      "scales": "encoder.layer.15.attention.self.key.weight.scales",
+      "biases": "encoder.layer.15.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.15.attention.self.query.weight",
@@ -490,7 +541,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.15.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.15.attention.self.query.weight.scale"
+      "scales": "encoder.layer.15.attention.self.query.weight.scales",
+      "biases": "encoder.layer.15.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.15.attention.self.value.weight",
@@ -500,7 +552,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.15.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.15.attention.self.value.weight.scale"
+      "scales": "encoder.layer.15.attention.self.value.weight.scales",
+      "biases": "encoder.layer.15.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.15.intermediate.dense.weight",
@@ -510,7 +563,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.15.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.15.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.15.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.15.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.15.output.dense.weight",
@@ -520,7 +574,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.15.output.dense.weight.qweight",
-      "scale": "encoder.layer.15.output.dense.weight.scale"
+      "scales": "encoder.layer.15.output.dense.weight.scales",
+      "biases": "encoder.layer.15.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.16.attention.output.dense.weight",
@@ -530,7 +585,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.16.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.16.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.16.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.16.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.16.attention.self.key.weight",
@@ -540,7 +596,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.16.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.16.attention.self.key.weight.scale"
+      "scales": "encoder.layer.16.attention.self.key.weight.scales",
+      "biases": "encoder.layer.16.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.16.attention.self.query.weight",
@@ -550,7 +607,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.16.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.16.attention.self.query.weight.scale"
+      "scales": "encoder.layer.16.attention.self.query.weight.scales",
+      "biases": "encoder.layer.16.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.16.attention.self.value.weight",
@@ -560,7 +618,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.16.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.16.attention.self.value.weight.scale"
+      "scales": "encoder.layer.16.attention.self.value.weight.scales",
+      "biases": "encoder.layer.16.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.16.intermediate.dense.weight",
@@ -570,7 +629,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.16.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.16.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.16.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.16.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.16.output.dense.weight",
@@ -580,7 +640,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.16.output.dense.weight.qweight",
-      "scale": "encoder.layer.16.output.dense.weight.scale"
+      "scales": "encoder.layer.16.output.dense.weight.scales",
+      "biases": "encoder.layer.16.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.17.attention.output.dense.weight",
@@ -590,7 +651,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.17.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.17.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.17.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.17.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.17.attention.self.key.weight",
@@ -600,7 +662,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.17.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.17.attention.self.key.weight.scale"
+      "scales": "encoder.layer.17.attention.self.key.weight.scales",
+      "biases": "encoder.layer.17.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.17.attention.self.query.weight",
@@ -610,7 +673,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.17.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.17.attention.self.query.weight.scale"
+      "scales": "encoder.layer.17.attention.self.query.weight.scales",
+      "biases": "encoder.layer.17.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.17.attention.self.value.weight",
@@ -620,7 +684,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.17.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.17.attention.self.value.weight.scale"
+      "scales": "encoder.layer.17.attention.self.value.weight.scales",
+      "biases": "encoder.layer.17.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.17.intermediate.dense.weight",
@@ -630,7 +695,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.17.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.17.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.17.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.17.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.17.output.dense.weight",
@@ -640,7 +706,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.17.output.dense.weight.qweight",
-      "scale": "encoder.layer.17.output.dense.weight.scale"
+      "scales": "encoder.layer.17.output.dense.weight.scales",
+      "biases": "encoder.layer.17.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.18.attention.output.dense.weight",
@@ -650,7 +717,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.18.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.18.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.18.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.18.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.18.attention.self.key.weight",
@@ -660,7 +728,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.18.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.18.attention.self.key.weight.scale"
+      "scales": "encoder.layer.18.attention.self.key.weight.scales",
+      "biases": "encoder.layer.18.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.18.attention.self.query.weight",
@@ -670,7 +739,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.18.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.18.attention.self.query.weight.scale"
+      "scales": "encoder.layer.18.attention.self.query.weight.scales",
+      "biases": "encoder.layer.18.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.18.attention.self.value.weight",
@@ -680,7 +750,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.18.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.18.attention.self.value.weight.scale"
+      "scales": "encoder.layer.18.attention.self.value.weight.scales",
+      "biases": "encoder.layer.18.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.18.intermediate.dense.weight",
@@ -690,7 +761,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.18.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.18.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.18.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.18.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.18.output.dense.weight",
@@ -700,7 +772,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.18.output.dense.weight.qweight",
-      "scale": "encoder.layer.18.output.dense.weight.scale"
+      "scales": "encoder.layer.18.output.dense.weight.scales",
+      "biases": "encoder.layer.18.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.19.attention.output.dense.weight",
@@ -710,7 +783,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.19.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.19.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.19.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.19.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.19.attention.self.key.weight",
@@ -720,7 +794,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.19.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.19.attention.self.key.weight.scale"
+      "scales": "encoder.layer.19.attention.self.key.weight.scales",
+      "biases": "encoder.layer.19.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.19.attention.self.query.weight",
@@ -730,7 +805,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.19.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.19.attention.self.query.weight.scale"
+      "scales": "encoder.layer.19.attention.self.query.weight.scales",
+      "biases": "encoder.layer.19.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.19.attention.self.value.weight",
@@ -740,7 +816,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.19.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.19.attention.self.value.weight.scale"
+      "scales": "encoder.layer.19.attention.self.value.weight.scales",
+      "biases": "encoder.layer.19.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.19.intermediate.dense.weight",
@@ -750,7 +827,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.19.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.19.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.19.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.19.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.19.output.dense.weight",
@@ -760,7 +838,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.19.output.dense.weight.qweight",
-      "scale": "encoder.layer.19.output.dense.weight.scale"
+      "scales": "encoder.layer.19.output.dense.weight.scales",
+      "biases": "encoder.layer.19.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.2.attention.output.dense.weight",
@@ -770,7 +849,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.2.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.2.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.2.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.2.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.2.attention.self.key.weight",
@@ -780,7 +860,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.2.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.2.attention.self.key.weight.scale"
+      "scales": "encoder.layer.2.attention.self.key.weight.scales",
+      "biases": "encoder.layer.2.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.2.attention.self.query.weight",
@@ -790,7 +871,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.2.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.2.attention.self.query.weight.scale"
+      "scales": "encoder.layer.2.attention.self.query.weight.scales",
+      "biases": "encoder.layer.2.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.2.attention.self.value.weight",
@@ -800,7 +882,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.2.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.2.attention.self.value.weight.scale"
+      "scales": "encoder.layer.2.attention.self.value.weight.scales",
+      "biases": "encoder.layer.2.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.2.intermediate.dense.weight",
@@ -810,7 +893,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.2.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.2.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.2.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.2.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.2.output.dense.weight",
@@ -820,7 +904,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.2.output.dense.weight.qweight",
-      "scale": "encoder.layer.2.output.dense.weight.scale"
+      "scales": "encoder.layer.2.output.dense.weight.scales",
+      "biases": "encoder.layer.2.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.20.attention.output.dense.weight",
@@ -830,7 +915,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.20.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.20.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.20.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.20.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.20.attention.self.key.weight",
@@ -840,7 +926,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.20.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.20.attention.self.key.weight.scale"
+      "scales": "encoder.layer.20.attention.self.key.weight.scales",
+      "biases": "encoder.layer.20.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.20.attention.self.query.weight",
@@ -850,7 +937,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.20.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.20.attention.self.query.weight.scale"
+      "scales": "encoder.layer.20.attention.self.query.weight.scales",
+      "biases": "encoder.layer.20.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.20.attention.self.value.weight",
@@ -860,7 +948,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.20.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.20.attention.self.value.weight.scale"
+      "scales": "encoder.layer.20.attention.self.value.weight.scales",
+      "biases": "encoder.layer.20.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.20.intermediate.dense.weight",
@@ -870,7 +959,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.20.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.20.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.20.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.20.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.20.output.dense.weight",
@@ -880,7 +970,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.20.output.dense.weight.qweight",
-      "scale": "encoder.layer.20.output.dense.weight.scale"
+      "scales": "encoder.layer.20.output.dense.weight.scales",
+      "biases": "encoder.layer.20.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.21.attention.output.dense.weight",
@@ -890,7 +981,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.21.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.21.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.21.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.21.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.21.attention.self.key.weight",
@@ -900,7 +992,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.21.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.21.attention.self.key.weight.scale"
+      "scales": "encoder.layer.21.attention.self.key.weight.scales",
+      "biases": "encoder.layer.21.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.21.attention.self.query.weight",
@@ -910,7 +1003,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.21.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.21.attention.self.query.weight.scale"
+      "scales": "encoder.layer.21.attention.self.query.weight.scales",
+      "biases": "encoder.layer.21.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.21.attention.self.value.weight",
@@ -920,7 +1014,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.21.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.21.attention.self.value.weight.scale"
+      "scales": "encoder.layer.21.attention.self.value.weight.scales",
+      "biases": "encoder.layer.21.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.21.intermediate.dense.weight",
@@ -930,7 +1025,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.21.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.21.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.21.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.21.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.21.output.dense.weight",
@@ -940,7 +1036,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.21.output.dense.weight.qweight",
-      "scale": "encoder.layer.21.output.dense.weight.scale"
+      "scales": "encoder.layer.21.output.dense.weight.scales",
+      "biases": "encoder.layer.21.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.22.attention.output.dense.weight",
@@ -950,7 +1047,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.22.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.22.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.22.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.22.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.22.attention.self.key.weight",
@@ -960,7 +1058,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.22.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.22.attention.self.key.weight.scale"
+      "scales": "encoder.layer.22.attention.self.key.weight.scales",
+      "biases": "encoder.layer.22.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.22.attention.self.query.weight",
@@ -970,7 +1069,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.22.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.22.attention.self.query.weight.scale"
+      "scales": "encoder.layer.22.attention.self.query.weight.scales",
+      "biases": "encoder.layer.22.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.22.attention.self.value.weight",
@@ -980,7 +1080,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.22.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.22.attention.self.value.weight.scale"
+      "scales": "encoder.layer.22.attention.self.value.weight.scales",
+      "biases": "encoder.layer.22.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.22.intermediate.dense.weight",
@@ -990,7 +1091,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.22.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.22.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.22.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.22.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.22.output.dense.weight",
@@ -1000,7 +1102,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.22.output.dense.weight.qweight",
-      "scale": "encoder.layer.22.output.dense.weight.scale"
+      "scales": "encoder.layer.22.output.dense.weight.scales",
+      "biases": "encoder.layer.22.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.23.attention.output.dense.weight",
@@ -1010,7 +1113,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.23.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.23.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.23.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.23.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.23.attention.self.key.weight",
@@ -1020,7 +1124,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.23.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.23.attention.self.key.weight.scale"
+      "scales": "encoder.layer.23.attention.self.key.weight.scales",
+      "biases": "encoder.layer.23.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.23.attention.self.query.weight",
@@ -1030,7 +1135,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.23.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.23.attention.self.query.weight.scale"
+      "scales": "encoder.layer.23.attention.self.query.weight.scales",
+      "biases": "encoder.layer.23.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.23.attention.self.value.weight",
@@ -1040,7 +1146,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.23.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.23.attention.self.value.weight.scale"
+      "scales": "encoder.layer.23.attention.self.value.weight.scales",
+      "biases": "encoder.layer.23.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.23.intermediate.dense.weight",
@@ -1050,7 +1157,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.23.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.23.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.23.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.23.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.23.output.dense.weight",
@@ -1060,7 +1168,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.23.output.dense.weight.qweight",
-      "scale": "encoder.layer.23.output.dense.weight.scale"
+      "scales": "encoder.layer.23.output.dense.weight.scales",
+      "biases": "encoder.layer.23.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.3.attention.output.dense.weight",
@@ -1070,7 +1179,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.3.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.3.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.3.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.3.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.3.attention.self.key.weight",
@@ -1080,7 +1190,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.3.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.3.attention.self.key.weight.scale"
+      "scales": "encoder.layer.3.attention.self.key.weight.scales",
+      "biases": "encoder.layer.3.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.3.attention.self.query.weight",
@@ -1090,7 +1201,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.3.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.3.attention.self.query.weight.scale"
+      "scales": "encoder.layer.3.attention.self.query.weight.scales",
+      "biases": "encoder.layer.3.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.3.attention.self.value.weight",
@@ -1100,7 +1212,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.3.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.3.attention.self.value.weight.scale"
+      "scales": "encoder.layer.3.attention.self.value.weight.scales",
+      "biases": "encoder.layer.3.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.3.intermediate.dense.weight",
@@ -1110,7 +1223,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.3.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.3.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.3.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.3.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.3.output.dense.weight",
@@ -1120,7 +1234,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.3.output.dense.weight.qweight",
-      "scale": "encoder.layer.3.output.dense.weight.scale"
+      "scales": "encoder.layer.3.output.dense.weight.scales",
+      "biases": "encoder.layer.3.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.4.attention.output.dense.weight",
@@ -1130,7 +1245,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.4.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.4.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.4.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.4.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.4.attention.self.key.weight",
@@ -1140,7 +1256,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.4.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.4.attention.self.key.weight.scale"
+      "scales": "encoder.layer.4.attention.self.key.weight.scales",
+      "biases": "encoder.layer.4.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.4.attention.self.query.weight",
@@ -1150,7 +1267,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.4.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.4.attention.self.query.weight.scale"
+      "scales": "encoder.layer.4.attention.self.query.weight.scales",
+      "biases": "encoder.layer.4.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.4.attention.self.value.weight",
@@ -1160,7 +1278,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.4.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.4.attention.self.value.weight.scale"
+      "scales": "encoder.layer.4.attention.self.value.weight.scales",
+      "biases": "encoder.layer.4.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.4.intermediate.dense.weight",
@@ -1170,7 +1289,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.4.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.4.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.4.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.4.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.4.output.dense.weight",
@@ -1180,7 +1300,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.4.output.dense.weight.qweight",
-      "scale": "encoder.layer.4.output.dense.weight.scale"
+      "scales": "encoder.layer.4.output.dense.weight.scales",
+      "biases": "encoder.layer.4.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.5.attention.output.dense.weight",
@@ -1190,7 +1311,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.5.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.5.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.5.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.5.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.5.attention.self.key.weight",
@@ -1200,7 +1322,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.5.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.5.attention.self.key.weight.scale"
+      "scales": "encoder.layer.5.attention.self.key.weight.scales",
+      "biases": "encoder.layer.5.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.5.attention.self.query.weight",
@@ -1210,7 +1333,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.5.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.5.attention.self.query.weight.scale"
+      "scales": "encoder.layer.5.attention.self.query.weight.scales",
+      "biases": "encoder.layer.5.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.5.attention.self.value.weight",
@@ -1220,7 +1344,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.5.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.5.attention.self.value.weight.scale"
+      "scales": "encoder.layer.5.attention.self.value.weight.scales",
+      "biases": "encoder.layer.5.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.5.intermediate.dense.weight",
@@ -1230,7 +1355,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.5.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.5.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.5.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.5.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.5.output.dense.weight",
@@ -1240,7 +1366,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.5.output.dense.weight.qweight",
-      "scale": "encoder.layer.5.output.dense.weight.scale"
+      "scales": "encoder.layer.5.output.dense.weight.scales",
+      "biases": "encoder.layer.5.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.6.attention.output.dense.weight",
@@ -1250,7 +1377,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.6.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.6.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.6.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.6.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.6.attention.self.key.weight",
@@ -1260,7 +1388,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.6.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.6.attention.self.key.weight.scale"
+      "scales": "encoder.layer.6.attention.self.key.weight.scales",
+      "biases": "encoder.layer.6.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.6.attention.self.query.weight",
@@ -1270,7 +1399,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.6.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.6.attention.self.query.weight.scale"
+      "scales": "encoder.layer.6.attention.self.query.weight.scales",
+      "biases": "encoder.layer.6.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.6.attention.self.value.weight",
@@ -1280,7 +1410,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.6.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.6.attention.self.value.weight.scale"
+      "scales": "encoder.layer.6.attention.self.value.weight.scales",
+      "biases": "encoder.layer.6.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.6.intermediate.dense.weight",
@@ -1290,7 +1421,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.6.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.6.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.6.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.6.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.6.output.dense.weight",
@@ -1300,7 +1432,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.6.output.dense.weight.qweight",
-      "scale": "encoder.layer.6.output.dense.weight.scale"
+      "scales": "encoder.layer.6.output.dense.weight.scales",
+      "biases": "encoder.layer.6.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.7.attention.output.dense.weight",
@@ -1310,7 +1443,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.7.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.7.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.7.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.7.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.7.attention.self.key.weight",
@@ -1320,7 +1454,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.7.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.7.attention.self.key.weight.scale"
+      "scales": "encoder.layer.7.attention.self.key.weight.scales",
+      "biases": "encoder.layer.7.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.7.attention.self.query.weight",
@@ -1330,7 +1465,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.7.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.7.attention.self.query.weight.scale"
+      "scales": "encoder.layer.7.attention.self.query.weight.scales",
+      "biases": "encoder.layer.7.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.7.attention.self.value.weight",
@@ -1340,7 +1476,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.7.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.7.attention.self.value.weight.scale"
+      "scales": "encoder.layer.7.attention.self.value.weight.scales",
+      "biases": "encoder.layer.7.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.7.intermediate.dense.weight",
@@ -1350,7 +1487,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.7.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.7.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.7.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.7.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.7.output.dense.weight",
@@ -1360,7 +1498,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.7.output.dense.weight.qweight",
-      "scale": "encoder.layer.7.output.dense.weight.scale"
+      "scales": "encoder.layer.7.output.dense.weight.scales",
+      "biases": "encoder.layer.7.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.8.attention.output.dense.weight",
@@ -1370,7 +1509,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.8.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.8.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.8.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.8.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.8.attention.self.key.weight",
@@ -1380,7 +1520,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.8.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.8.attention.self.key.weight.scale"
+      "scales": "encoder.layer.8.attention.self.key.weight.scales",
+      "biases": "encoder.layer.8.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.8.attention.self.query.weight",
@@ -1390,7 +1531,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.8.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.8.attention.self.query.weight.scale"
+      "scales": "encoder.layer.8.attention.self.query.weight.scales",
+      "biases": "encoder.layer.8.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.8.attention.self.value.weight",
@@ -1400,7 +1542,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.8.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.8.attention.self.value.weight.scale"
+      "scales": "encoder.layer.8.attention.self.value.weight.scales",
+      "biases": "encoder.layer.8.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.8.intermediate.dense.weight",
@@ -1410,7 +1553,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.8.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.8.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.8.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.8.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.8.output.dense.weight",
@@ -1420,7 +1564,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.8.output.dense.weight.qweight",
-      "scale": "encoder.layer.8.output.dense.weight.scale"
+      "scales": "encoder.layer.8.output.dense.weight.scales",
+      "biases": "encoder.layer.8.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.9.attention.output.dense.weight",
@@ -1430,7 +1575,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.9.attention.output.dense.weight.qweight",
-      "scale": "encoder.layer.9.attention.output.dense.weight.scale"
+      "scales": "encoder.layer.9.attention.output.dense.weight.scales",
+      "biases": "encoder.layer.9.attention.output.dense.weight.biases"
     },
     {
       "name": "encoder.layer.9.attention.self.key.weight",
@@ -1440,7 +1586,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.9.attention.self.key.weight.qweight",
-      "scale": "encoder.layer.9.attention.self.key.weight.scale"
+      "scales": "encoder.layer.9.attention.self.key.weight.scales",
+      "biases": "encoder.layer.9.attention.self.key.weight.biases"
     },
     {
       "name": "encoder.layer.9.attention.self.query.weight",
@@ -1450,7 +1597,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.9.attention.self.query.weight.qweight",
-      "scale": "encoder.layer.9.attention.self.query.weight.scale"
+      "scales": "encoder.layer.9.attention.self.query.weight.scales",
+      "biases": "encoder.layer.9.attention.self.query.weight.biases"
     },
     {
       "name": "encoder.layer.9.attention.self.value.weight",
@@ -1460,7 +1608,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.9.attention.self.value.weight.qweight",
-      "scale": "encoder.layer.9.attention.self.value.weight.scale"
+      "scales": "encoder.layer.9.attention.self.value.weight.scales",
+      "biases": "encoder.layer.9.attention.self.value.weight.biases"
     },
     {
       "name": "encoder.layer.9.intermediate.dense.weight",
@@ -1470,7 +1619,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.9.intermediate.dense.weight.qweight",
-      "scale": "encoder.layer.9.intermediate.dense.weight.scale"
+      "scales": "encoder.layer.9.intermediate.dense.weight.scales",
+      "biases": "encoder.layer.9.intermediate.dense.weight.biases"
     },
     {
       "name": "encoder.layer.9.output.dense.weight",
@@ -1480,7 +1630,8 @@
       ],
       "dtype": "float16",
       "qweight": "encoder.layer.9.output.dense.weight.qweight",
-      "scale": "encoder.layer.9.output.dense.weight.scale"
+      "scales": "encoder.layer.9.output.dense.weight.scales",
+      "biases": "encoder.layer.9.output.dense.weight.biases"
     },
     {
       "name": "pooler.dense.weight",
@@ -1490,7 +1641,8 @@
       ],
       "dtype": "float16",
       "qweight": "pooler.dense.weight.qweight",
-      "scale": "pooler.dense.weight.scale"
+      "scales": "pooler.dense.weight.scales",
+      "biases": "pooler.dense.weight.biases"
     }
   ],
   "kept": [

weights.00.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:300cb491160d5d6341397f850e3f1b8bdfb3b7ecf21b1f77eeb5f75c731b99ee
-size 561220870
+oid sha256:97aaf1cdb777e1506db17616885f1d9b05a90d3c981022c7a4ee31dc9f1fddc6
+size 595268674