Add a QuestionAnswering bert model for testing (#19)

* Add matching transformers version as current tt-metal

* Remove the type conversion passes and directly call transformers to use bfloat16 when initializing the model

* Convert torch.Tensor.to to ttnn.as_tensor

* Use ttnn.full correctly instead of aten.full for certain cases

* Convert bert model to unittest

* Move transformers installation to dev

* Refactor model input and output print statement for test_bert

requirements-dev.txt

@@ -2,3 +2,4 @@
 pytest==7.2.2
 pytest-timeout==2.2.0
 pre-commit==3.0.4
+transformers==4.38.0

tests/test_bert.py

@@ -0,0 +1,86 @@
+import torch
+import torch_ttnn
+import unittest
+from torch_ttnn import ttnn
+import collections
+
+# Load model directly
+from transformers import (
+    AutoTokenizer,
+    AutoModelForQuestionAnswering,
+)
+
+
+class TestBert(unittest.TestCase):
+    def setUp(self):
+        # Open device 0
+        self.device: ttnn.Device = ttnn.open_device(device_id=0)
+
+    def tearDown(self):
+        # Close the device
+        ttnn.close_device(self.device)
+
+    def test_bert(self):
+        # Download model from cloud
+        model_name = "phiyodr/bert-large-finetuned-squad2"
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_name, padding_side="left", torch_dtype=torch.bfloat16
+        )
+        m = AutoModelForQuestionAnswering.from_pretrained(
+            model_name, torch_dtype=torch.bfloat16
+        )
+        m.eval()
+
+        # Set up sample input
+        context = 'Johann Joachim Winckelmann was a German art historian and archaeologist. He was a pioneering Hellenist who first articulated the difference between Greek, Greco-Roman and Roman art. "The prophet and founding hero of modern archaeology", Winckelmann was one of the founders of scientific archaeology and first applied the categories of style on a large, systematic basis to the history of art. '
+        question = "What discipline did Winkelmann create?"
+
+        inputs = tokenizer.encode_plus(
+            question,
+            context,
+            add_special_tokens=True,
+            return_tensors="pt",
+            max_length=256,
+            padding="max_length",
+            truncation=True,
+        )
+
+        # Run inference with the original model
+        with torch.no_grad():
+            outputs_before = m(**inputs)
+
+        # Helper function to decode output to human-readable text
+        def decode_output(outputs):
+            response_start = torch.argmax(outputs.start_logits)
+            response_end = torch.argmax(outputs.end_logits) + 1
+            response_tokens = inputs.input_ids[0, response_start:response_end]
+            return tokenizer.decode(response_tokens)
+
+        answer_before = decode_output(outputs_before)
+
+        # Compile model with ttnn backend
+        option = torch_ttnn.TorchTtnnOption(device=self.device)
+        m = torch.compile(m, backend=torch_ttnn.backend, options=option)
+
+        # Run inference with the compiled model
+        with torch.no_grad():
+            outputs_after = m(**inputs)
+        option._out_fx_graphs[0].print_tabular()
+
+        answer_after = decode_output(outputs_after)
+
+        print(
+            f"""
+        model_name: {model_name}
+        input:
+            context: {context}
+            question: {question}
+        answer before: {answer_before}
+        answer after: {answer_after}
+        """
+        )
+
+        # TODO: Add more checks for the compiled graph
+
+        # Check inference result
+        self.assertEqual(answer_before, answer_after)

tests/test_more_ops.py

@@ -555,6 +555,29 @@ def input_shapes(self):
         return [(4, 4)]
 
 
+class ToCopyModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.to(torch.bfloat16)
+
+    def input_shapes(self):
+        return [(4, 4)]
+
+
+class ToCopyWithOpAfterModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        to = x.to(torch.bfloat16)
+        return torch.add(to, to)
+
+    def input_shapes(self):
+        return [(4, 4)]
+
+
 class TestModules(unittest.TestCase):
     def setUp(self):
         # Open device 0
@@ -894,18 +917,17 @@ def test_div_scalar_denom(self):
 
         # Check the graph has be rewritten and contain ttnn ops
         nodes = list(option._out_fx_graphs[0].nodes)
-        self.assertTrue(nodes[5].target == ttnn.reciprocal)
-        self.assertTrue(nodes[5].args[0].target == ttnn.to_device)
-        self.assertTrue(nodes[5].args[0].args[0].target == ttnn.to_layout)
-        self.assertTrue(nodes[5].args[0].args[0].args[0].target == ttnn.from_torch)
-        self.assertTrue(nodes[9].target == ttnn.mul)
-        self.assertTrue(nodes[9].args[0].target == ttnn.to_device)
-        self.assertTrue(nodes[9].args[0].args[0].target == ttnn.to_layout)
-        self.assertTrue(nodes[9].args[0].args[0].args[0].target == ttnn.from_torch)
-        self.assertTrue(nodes[9].args[1].target == ttnn.reciprocal)
-        self.assertTrue(nodes[10].target == ttnn.from_device)
-        self.assertTrue(nodes[11].target == ttnn.to_layout)
-        self.assertTrue(nodes[12].target == ttnn.to_torch)
+        self.assertTrue(nodes[1].target == ttnn.full)
+        self.assertTrue(nodes[3].target == ttnn.reciprocal)
+        self.assertTrue(nodes[3].args[0].target == ttnn.to_layout)
+        self.assertTrue(nodes[7].target == ttnn.mul)
+        self.assertTrue(nodes[7].args[0].target == ttnn.to_device)
+        self.assertTrue(nodes[7].args[0].args[0].target == ttnn.to_layout)
+        self.assertTrue(nodes[7].args[0].args[0].args[0].target == ttnn.from_torch)
+        self.assertTrue(nodes[7].args[1].target == ttnn.reciprocal)
+        self.assertTrue(nodes[8].target == ttnn.from_device)
+        self.assertTrue(nodes[9].target == ttnn.to_layout)
+        self.assertTrue(nodes[10].target == ttnn.to_torch)
         # Check inference result
         self.assertTrue(check_with_pcc(result_before, result_after))
 
@@ -972,14 +994,15 @@ def test_rsub_scalar(self):
 
         # Check the graph has be rewritten and contain ttnn ops
         nodes = list(option._out_fx_graphs[0].nodes)
-        # self.aseertTrue(nodes[1].target == ttnn.full)
-        self.assertTrue(nodes[8].target == ttnn.sub)
-        self.assertTrue(nodes[8].args[0].target == ttnn.to_device)
-        self.assertTrue(nodes[8].args[0].args[0].target == ttnn.to_layout)
-        self.assertTrue(nodes[8].args[0].args[0].args[0].target == ttnn.from_torch)
-        self.assertTrue(nodes[9].target == ttnn.from_device)
-        self.assertTrue(nodes[10].target == ttnn.to_layout)
-        self.assertTrue(nodes[11].target == ttnn.to_torch)
+        self.assertTrue(nodes[1].target == ttnn.full)
+        self.assertTrue(nodes[6].target == ttnn.sub)
+        self.assertTrue(nodes[6].args[0].target == ttnn.to_layout)
+        self.assertTrue(nodes[6].args[1].target == ttnn.to_device)
+        self.assertTrue(nodes[6].args[1].args[0].target == ttnn.to_layout)
+        self.assertTrue(nodes[6].args[1].args[0].args[0].target == ttnn.from_torch)
+        self.assertTrue(nodes[7].target == ttnn.from_device)
+        self.assertTrue(nodes[8].target == ttnn.to_layout)
+        self.assertTrue(nodes[9].target == ttnn.to_torch)
         # Check inference result
         self.assertTrue(check_with_pcc(result_before, result_after, 0.9998))
 
@@ -1816,6 +1839,51 @@ def test_sigmoid(self):
         # Check inference result
         self.assertTrue(torch.allclose(result_before, result_after, rtol=0.2))
 
+    def test_to_copy(self):
+        m = ToCopyModule()
+        input_shapes = m.input_shapes()
+        inputs = [torch.rand(shape) for shape in input_shapes]
+        result_before = m.forward(*inputs)
+        option = torch_ttnn.TorchTtnnOption(device=self.device)
+        option.gen_graphviz = True
+        # The compilation is lazy, so we need to run forward once to trigger the compilation
+        m = torch.compile(m, backend=torch_ttnn.backend, options=option)
+        result_after = m.forward(*inputs)
+        option._out_fx_graphs[0].print_tabular()
+
+        # Check the graph has be rewritten and contain ttnn ops
+        nodes = list(option._out_fx_graphs[0].nodes)
+        self.assertTrue(nodes[2].target == ttnn.as_tensor)
+        self.assertTrue(nodes[3].target == ttnn.from_device)
+        self.assertTrue(nodes[4].target == ttnn.to_layout)
+        self.assertTrue(nodes[5].target == ttnn.to_torch)
+        # Check inference result
+        self.assertTrue(torch.allclose(result_before, result_after, rtol=0.2))
+
+    def test_to_copy_with_op_after(self):
+        m = ToCopyWithOpAfterModule()
+        input_shapes = m.input_shapes()
+        inputs = [torch.rand(shape) for shape in input_shapes]
+        result_before = m.forward(*inputs)
+        option = torch_ttnn.TorchTtnnOption(device=self.device)
+        option.gen_graphviz = True
+        # The compilation is lazy, so we need to run forward once to trigger the compilation
+        m = torch.compile(m, backend=torch_ttnn.backend, options=option)
+        result_after = m.forward(*inputs)
+        option._out_fx_graphs[0].print_tabular()
+
+        # Check the graph has be rewritten and contain ttnn ops
+        nodes = list(option._out_fx_graphs[0].nodes)
+        self.assertTrue(nodes[2].target == ttnn.as_tensor)
+        self.assertTrue(nodes[3].target == ttnn.add)
+        self.assertTrue(nodes[3].args[0].target == ttnn.as_tensor)
+        self.assertTrue(nodes[3].args[1].target == ttnn.as_tensor)
+        self.assertTrue(nodes[4].target == ttnn.from_device)
+        self.assertTrue(nodes[5].target == ttnn.to_layout)
+        self.assertTrue(nodes[6].target == ttnn.to_torch)
+        # Check inference result
+        self.assertTrue(torch.allclose(result_before, result_after, rtol=0.2))
+
 
 if __name__ == "__main__":
     unittest.main()

tools/run_transformers.py

@@ -44,13 +44,17 @@ def run_model(
     ]
 
     if model.model_task in text_modules:
-        tokenizer = AutoTokenizer.from_pretrained(model.model_name, padding_side="left")
+        tokenizer = AutoTokenizer.from_pretrained(
+            model.model_name, padding_side="left", torch_dtype=torch.bfloat16
+        )
     elif model.model_task in vision_modules:
-        image_processor = AutoImageProcessor.from_pretrained(model.model_name)
+        image_processor = AutoImageProcessor.from_pretrained(
+            model.model_name, torch_dtype=torch.bfloat16
+        )
     else:
         raise ValueError(f"model task: {model.model_task} not supported.")
 
-    m = model.model_task.from_pretrained(model.model_name)
+    m = model.model_task.from_pretrained(model.model_name, torch_dtype=torch.bfloat16)
 
     if backward:
         try:

torch_ttnn/backend.py

@@ -32,12 +32,6 @@ def aten_backend(
 
     gm = remove_clones_for_input_aliasing(gm)
 
-    # Change float types in dtype kwargs to bfloat16
-    from .convert_type import convert_dtype_to_bfloat16, convert_float_to_bfloat16
-
-    gm = convert_float_to_bfloat16(gm)
-    gm = convert_dtype_to_bfloat16(gm)
-
     option: TorchTtnnOption = options["torch_ttnn_option"]
     torch.fx.graph._register_custom_builtin("ttnn_Specified_Device", "", option.device)
     torch.fx.graph._register_custom_builtin(