restrict transformers version for now...

huggingface · Dec 14, 2023 · daefcf3 · daefcf3
1 parent 5950d0d
commit daefcf3
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 15 deletions.
diff --git a/optimum/intel/openvino/trainer.py b/optimum/intel/openvino/trainer.py
@@ -637,13 +637,10 @@ def _inner_training_loop(
                     if args.max_grad_norm is not None and args.max_grad_norm > 0:
                         # deepspeed does its own clipping
 
-<<<<<<< HEAD
-=======
                         if getattr(self, "do_grad_scaling", False):
                             # AMP: gradients need unscaling
                             self.scaler.unscale_(self.optimizer)
 
->>>>>>> fix tests
                         if is_sagemaker_mp_enabled() and args.fp16:
                             self.optimizer.clip_master_grads(args.max_grad_norm)
                         elif self.use_apex:
@@ -659,14 +656,6 @@ def _inner_training_loop(
                             )
 
                     # Optimizer step
-<<<<<<< HEAD
-                    self.optimizer.step()
-                    optimizer_was_run = not self.accelerator.optimizer_step_was_skipped
-                    if optimizer_was_run:
-                        # Delay optimizer scheduling until metrics are generated
-                        if not isinstance(self.lr_scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
-                            self.lr_scheduler.step()
-=======
                     optimizer_was_run = True
                     if self.deepspeed:
                         pass  # called outside the loop
@@ -681,7 +670,6 @@ def _inner_training_loop(
 
                     if optimizer_was_run and not self.deepspeed:
                         self.lr_scheduler.step()
->>>>>>> fix tests
 
                     model.zero_grad()
                     self.state.global_step += 1

diff --git a/tests/openvino/test_modeling.py b/tests/openvino/test_modeling.py
@@ -494,7 +494,7 @@ def test_compare_to_transformers(self, model_arch):
         set_seed(SEED)
         ov_model = OVModelForCausalLM.from_pretrained(model_id, export=True)
         self.assertIsInstance(ov_model.config, PretrainedConfig)
-        transformers_model = AutoModelForCausalLM.from_pretrained(model_id)
+        transformers_model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
         tokens = tokenizer(
             "This is a sample", return_tensors="pt", return_token_type_ids=False if model_arch == "llama" else None
@@ -510,7 +510,8 @@ def test_compare_to_transformers(self, model_arch):
         with torch.no_grad():
             transformers_outputs = transformers_model(**tokens)
         # Compare tensor outputs
-        self.assertTrue(torch.allclose(ov_outputs.logits, transformers_outputs.logits, atol=1e-2))
+        self.assertTrue(torch.allclose(ov_outputs.logits, transformers_outputs.logits, atol=1e-4),
+                         f"Max diff {torch.abs(ov_outputs.logits - transformers_outputs.logits).max()}")
         del transformers_model
         del ov_model
         gc.collect()
@@ -1244,7 +1245,6 @@ def test_compare_to_transformers(self, model_arch):
 
             ov_outputs = ov_model(**features, **decoder_inputs)
             self.assertIn("logits", ov_outputs)
-            self.assertIsInstance(ov_outputs.logits, TENSOR_ALIAS_TO_TYPE[input_type])
             # Compare tensor outputs
             self.assertTrue(torch.allclose(torch.Tensor(ov_outputs.logits), transformers_outputs.logits, atol=1e-3))