Merge pull request huggingface#4 from microsoft/raviskolli/ort

Remove data based dependencies in T5 for ORT

examples/seq2seq/run_seq2seq.py

@@ -341,6 +341,7 @@ def main():
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         use_auth_token=True if model_args.use_auth_token else None,
+        ort=True if training_args.ort else None,
     )
     tokenizer = AutoTokenizer.from_pretrained(
         model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,

src/transformers/configuration_utils.py

@@ -175,6 +175,10 @@ class PretrainedConfig(object):
 
         - **use_bfloat16** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether or not the model should use
           BFloat16 scalars (only used by some TensorFlow models).
+
+    Onnxruntime specific parameters
+
+        - **ort** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether or not the model should use ORT.
     """
     model_type: str = ""
     is_composition: bool = False
@@ -186,6 +190,7 @@ def __init__(self, **kwargs):
         self.output_attentions = kwargs.pop("output_attentions", False)
         self.torchscript = kwargs.pop("torchscript", False)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop("use_bfloat16", False)
+        self.ort = kwargs.pop("ort", False)
         self.pruned_heads = kwargs.pop("pruned_heads", {})
         self.tie_word_embeddings = kwargs.pop(
             "tie_word_embeddings", True

src/transformers/models/t5/modeling_t5.py

@@ -591,6 +591,7 @@ class T5Block(nn.Module):
     def __init__(self, config, has_relative_attention_bias=False):
         super().__init__()
         self.is_decoder = config.is_decoder
+        self.ort = config.ort
         self.layer = nn.ModuleList()
         self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
         if self.is_decoder:
@@ -643,9 +644,16 @@ def forward(
         attention_outputs = self_attention_outputs[2:]  # Keep self-attention outputs and relative position weights
 
         # clamp inf values to enable fp16 training
-        if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
-            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
-            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+        if self.ort:
+            # Remove data-based control flow for static graph
+            if hidden_states.dtype == torch.float16:
+                clamp_value = torch.where(torch.isinf(hidden_states).any(), torch.finfo(hidden_states.dtype).max - 1000,
+                    torch.finfo(hidden_states.dtype).max)
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+        else:
+            if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
 
         do_cross_attention = self.is_decoder and encoder_hidden_states is not None
         if do_cross_attention:
@@ -670,9 +678,16 @@ def forward(
             hidden_states = cross_attention_outputs[0]
 
             # clamp inf values to enable fp16 training
-            if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
-                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
-                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+            if self.ort:
+                # Remove data-based control flow for static graph
+                if hidden_states.dtype == torch.float16:
+                    clamp_value = torch.where(torch.isinf(hidden_states).any(), torch.finfo(hidden_states.dtype).max - 1000,
+                        torch.finfo(hidden_states.dtype).max)
+                    hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+            else:
+                if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+                    clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+                    hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
 
             # Combine self attn and cross attn key value states
             if present_key_value_state is not None:
@@ -685,9 +700,16 @@ def forward(
         hidden_states = self.layer[-1](hidden_states)
 
         # clamp inf values to enable fp16 training
-        if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
-            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
-            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+        if self.ort:
+            # Remove data-based control flow for static graph
+            if hidden_states.dtype == torch.float16:
+                clamp_value = torch.where(torch.isinf(hidden_states).any(), torch.finfo(hidden_states.dtype).max - 1000,
+                    torch.finfo(hidden_states.dtype).max)
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+        else:
+            if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
 
         outputs = (hidden_states,)