Added BF16 & FP16 models to PTQ tests (#2922)

### Changes

- Added BF16/FP16 models to PTQ tests scope.

### Reason for changes

- Extend e2e with different precisions.

### Related tickets

- 147481

### Tests

- manual post_training_quantization/562/ - passed

tests/post_training/data/ptq_reference_data.yaml

@@ -26,6 +26,14 @@ hf/hf-internal-testing/tiny-random-gpt2_backend_OV:
   metric_value: null
 hf/hf-internal-testing/tiny-random-gpt2_backend_TORCH:
   metric_value: null
+hf/bert-base-uncased_fp16_backend_FP32:
+  metric_value: null
+hf/bert-base-uncased_fp16_backend_OV:
+  metric_value: null
+hf/bert-base-uncased_bf16_backend_FP32:
+  metric_value: null
+hf/bert-base-uncased_bf16_backend_OV:
+  metric_value: null
 torchvision/resnet18_backend_FP32:
   metric_value: 0.6978
 torchvision/resnet18_backend_OV:

tests/post_training/model_scope.py

@@ -12,6 +12,8 @@
 import copy
 from typing import Dict, List
 
+import torch
+
 import nncf
 from nncf import ModelType
 from nncf import QuantizationPreset
@@ -81,6 +83,30 @@
         },
         "backends": [BackendType.TORCH, BackendType.OV, BackendType.OPTIMUM],
     },
+    {
+        "reported_name": "hf/bert-base-uncased_fp16",
+        "model_id": "bert-base-uncased",
+        "pipeline_cls": MaskedLanguageModelingHF,
+        "compression_params": {
+            "preset": QuantizationPreset.MIXED,
+            "model_type": ModelType.TRANSFORMER,
+            "subset_size": 2,
+        },
+        "backends": [BackendType.OV],
+        "params": {"base_precision": torch.float16},
+    },
+    {
+        "reported_name": "hf/bert-base-uncased_bf16",
+        "model_id": "bert-base-uncased",
+        "pipeline_cls": MaskedLanguageModelingHF,
+        "compression_params": {
+            "preset": QuantizationPreset.MIXED,
+            "model_type": ModelType.TRANSFORMER,
+            "subset_size": 2,
+        },
+        "backends": [BackendType.OV],
+        "params": {"base_precision": torch.bfloat16},
+    },
     # Torchvision models
     {
         "reported_name": "torchvision/resnet18",

tests/post_training/pipelines/masked_language_modeling.py

@@ -29,14 +29,26 @@ class MaskedLanguageModelingHF(PTQTestPipeline):
     """Pipeline for masked language models from Hugging Face repository"""
 
     def prepare_model(self) -> None:
+        torch_dtype = self.params.get("base_precision", torch.float32)
         if self.backend in PT_BACKENDS:
-            self.model_hf = transformers.AutoModelForSequenceClassification.from_pretrained(self.model_id)
+            self.model_hf = transformers.AutoModelForSequenceClassification.from_pretrained(
+                self.model_id, torch_dtype=torch_dtype
+            )
             self.model = self.model_hf
             self.model.config.torchscript = True  # Set to export by convert_model via torch.jit.trace
             self.dummy_tensor = self.model_hf.dummy_inputs["input_ids"]
         if self.backend in OV_BACKENDS + [BackendType.FP32]:
-            self.model_hf = OVModelForSequenceClassification.from_pretrained(self.model_id, export=True, compile=False)
-            self.model = self.model_hf.model
+            if torch_dtype != torch.float32:
+                # Since optimum-intel does not produce custom-type models, this workaround handles it.
+                self.model_hf = transformers.AutoModelForSequenceClassification.from_pretrained(
+                    self.model_id, torch_dtype=torch_dtype
+                )
+                self.model = ov.convert_model(self.model_hf, example_input=self.model_hf.dummy_inputs)
+            else:
+                self.model_hf = OVModelForSequenceClassification.from_pretrained(
+                    self.model_id, export=True, compile=False
+                )
+                self.model = self.model_hf.model
 
         if self.backend == BackendType.ONNX:
             self.model_hf = ORTModelForSequenceClassification.from_pretrained(self.model_id, export=True)
@@ -75,13 +87,10 @@ def transform_func(data):
                 return torch.tensor([data["input_ids"]]).type(dtype=torch.LongTensor).to(device)
 
         else:
+            input_names = [p.get_friendly_name() for p in self.model.get_parameters()]
 
             def transform_func(data):
-                return {
-                    "input_ids": np.expand_dims(data["input_ids"], axis=0),
-                    "token_type_ids": np.expand_dims(data["token_type_ids"], axis=0),
-                    "attention_mask": np.expand_dims(data["attention_mask"], axis=0),
-                }
+                return {n: np.expand_dims(data[n], axis=0) for n in input_names}
 
         return transform_func