udpate setup

.github/workflows/test_onnxruntime.yml

@@ -42,7 +42,6 @@ jobs:
         run: |
           pip install --upgrade pip
           pip install --no-cache-dir torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-          pip install git+https://github.com/huggingface/transformers
           pip install .[tests,onnxruntime]
 
       - name: Test with pytest (in series)

optimum/bettertransformer/models/attention.py

@@ -207,9 +207,10 @@ def codegen_wrapped_scaled_dot_product(
         # causal_mask is always [True, ..., True] otherwise, so executing this
         # is unnecessary
         if query_length > 1:
-
             if not check_if_transformers_greater("4.44.99"):
-                causal_mask = self.causal_mask[:, :, key_length - query_length : key_length, :key_length].to(torch.bool)
+                causal_mask = self.causal_mask[:, :, key_length - query_length : key_length, :key_length].to(
+                    torch.bool
+                )
 
                 causal_mask = torch.where(causal_mask, 0, mask_value)
 
@@ -219,7 +220,6 @@ def codegen_wrapped_scaled_dot_product(
                 # we use torch.min to avoid having tensor(-inf)
                 attention_mask = torch.min(causal_mask, attention_mask)
             else:
-
                 attention_mask = attention_mask[:, :, :, : key.shape[-2]]
 
         sdpa_result = torch.nn.functional.scaled_dot_product_attention(
@@ -229,7 +229,6 @@ def codegen_wrapped_scaled_dot_product(
     return sdpa_result, None
 
 
-
 # Adapted from transformers.models.opt.modeling_opt.OPTAttention.forward
 def opt_forward(
     self,

optimum/onnxruntime/modeling_decoder.py

@@ -724,6 +724,7 @@ def _save_pretrained(self, save_directory: Union[str, Path]):
         super()._save_pretrained(save_directory)
         self.generation_config.save_pretrained(save_directory)
 
+
 class ORTGPTBigCodeForCausalLM(ORTModelForCausalLM):
     # Adapted from transformers.models.gpt_bigcode.modeling_gpt_bigcode.GPTBigCodeForCausalLM.prepare_inputs_for_generation
     def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):

setup.py

@@ -15,7 +15,8 @@
 REQUIRED_PKGS = [
     "coloredlogs",
     "sympy",
-    "transformers[sentencepiece]>=4.29,<4.46.0",
+    "transformers @ git+https://github.com/huggingface/transformers.git",
+    # "transformers[sentencepiece]>=4.29,<4.46.0",
     "torch>=1.11",
     "packaging",
     "numpy<2.0",  # transformers requires numpy<2.0 https://github.com/huggingface/transformers/pull/31569