Fix compatibility for latest transformers release (huggingface#570)

* fix compatibility for latest transformers release

* update setup

* update setup

* fix test input size

* fix prepare generation for llama models

Author: echarlaix

optimum/intel/ipex/modeling_base.py

@@ -46,7 +46,7 @@
 from optimum.utils import NormalizedConfigManager
 
 from ..generation.modeling import jit_trace, prepare_jit_inputs
-from ..utils.import_utils import is_torch_version
+from ..utils.import_utils import is_torch_version, is_transformers_version
 from ..utils.modeling_utils import MULTI_QUERY_ATTN_MODELS, patch_decoder_attention_mask
 
 
@@ -326,7 +326,8 @@ def __init__(
         # Perform the initial warmup at the end of __init__
         super().__init__(model, config, model_save_dir=model_save_dir, warmup=False)
 
-        self.normalized_config = NormalizedConfigManager.get_normalized_config_class(config.model_type)(config)
+        model_type = config.model_type.replace("_", "-")
+        self.normalized_config = NormalizedConfigManager.get_normalized_config_class(model_type)(config)
         self.model_dtype = kwargs.get("model_dtype", self.dtype)
         self.use_cache = "past_key_values" in self.input_names
 
@@ -339,6 +340,7 @@ def __init__(
             )
         config.is_decoder = True
         config.is_encoder_decoder = False
+
         self.generation_config = GenerationConfig.from_model_config(config)
         try:
             self.model_cls = get_class_from_dynamic_module(
@@ -347,7 +349,12 @@ def __init__(
         except AttributeError:
             self.model_cls = get_model_class(self.config, AutoModelForCausalLM._model_mapping)
         self._reorder_cache = self.model_cls._reorder_cache.__get__(self)
-        self.prepare_inputs_for_generation = self.model_cls.prepare_inputs_for_generation.__get__(self)
+
+        if is_transformers_version(">=", "4.38.0") and model_type in {"llama", "phi", "persimmon"}:
+            self.prepare_inputs_for_generation = _prepare_inputs_for_generation_for_llama
+        else:
+            self.prepare_inputs_for_generation = self.model_cls.prepare_inputs_for_generation.__get__(self)
+
         if hasattr(self.model_cls, "_convert_to_standard_cache"):
             self._convert_to_standard_cache = self.model_cls._convert_to_standard_cache
         if hasattr(self.model_cls, "_convert_to_bloom_cache"):
@@ -430,3 +437,62 @@ def forward(
             past_key_values = outputs["past_key_values"] if self.use_cache else None
 
         return CausalLMOutputWithPast(logits=logits, past_key_values=past_key_values)
+
+
+def _prepare_inputs_for_generation_for_llama(
+    input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+):
+    from transformers.cache_utils import Cache
+
+    if past_key_values is not None:
+        if isinstance(past_key_values, Cache):
+            cache_length = past_key_values.get_seq_length()
+            past_length = past_key_values.seen_tokens
+            max_cache_length = past_key_values.get_max_length()
+        else:
+            cache_length = past_length = past_key_values[0][0].shape[2]
+            max_cache_length = None
+
+        # Keep only the unprocessed tokens:
+        # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+        # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+        # input)
+        if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+            input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+        # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+        # input_ids based on the past_length.
+        elif past_length < input_ids.shape[1]:
+            input_ids = input_ids[:, past_length:]
+        # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+        # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+        if (
+            max_cache_length is not None
+            and attention_mask is not None
+            and cache_length + input_ids.shape[1] > max_cache_length
+        ):
+            attention_mask = attention_mask[:, -max_cache_length:]
+
+    position_ids = kwargs.get("position_ids", None)
+    if attention_mask is not None and position_ids is None:
+        # create position_ids on the fly for batch generation
+        position_ids = attention_mask.long().cumsum(-1) - 1
+        position_ids.masked_fill_(attention_mask == 0, 1)
+        if past_key_values:
+            position_ids = position_ids[:, -input_ids.shape[1] :]
+
+    # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+    if inputs_embeds is not None and past_key_values is None:
+        model_inputs = {"inputs_embeds": inputs_embeds}
+    else:
+        model_inputs = {"input_ids": input_ids}
+
+    model_inputs.update(
+        {
+            "position_ids": position_ids,
+            "past_key_values": past_key_values,
+            "use_cache": kwargs.get("use_cache"),
+            "attention_mask": attention_mask,
+        }
+    )
+    return model_inputs

optimum/intel/neural_compressor/trainer.py

@@ -941,3 +941,42 @@ def get_model_sparsity(self):
         if self._compression_manager is not None:
             sparsity = self._compression_manager.model.report_sparsity()[-1]
         return sparsity
+
+    def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch, ignore_keys_for_eval):
+        # TODO : can be removed once transformers >= v4.38.0
+        if self.control.should_log and self.state.global_step > self._globalstep_last_logged:
+            if is_torch_tpu_available():
+                xm.mark_step()
+
+            logs: Dict[str, float] = {}
+
+            # all_gather + mean() to get average loss over all processes
+            tr_loss_scalar = self._nested_gather(tr_loss).mean().item()
+
+            # reset tr_loss to zero
+            tr_loss -= tr_loss
+
+            logs["loss"] = round(tr_loss_scalar / (self.state.global_step - self._globalstep_last_logged), 4)
+            logs["learning_rate"] = self._get_learning_rate()
+
+            self._total_loss_scalar += tr_loss_scalar
+            self._globalstep_last_logged = self.state.global_step
+            self.store_flos()
+
+            self.log(logs)
+
+        metrics = None
+        if self.control.should_evaluate:
+            metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
+            self._report_to_hp_search(trial, self.state.global_step, metrics)
+
+            # Run delayed LR scheduler now that metrics are populated
+            if isinstance(self.lr_scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
+                metric_to_check = self.args.metric_for_best_model
+                if not metric_to_check.startswith("eval_"):
+                    metric_to_check = f"eval_{metric_to_check}"
+                self.lr_scheduler.step(metrics[metric_to_check])
+
+        if self.control.should_save:
+            self._save_checkpoint(model, trial, metrics=metrics)
+            self.control = self.callback_handler.on_save(self.args, self.state, self.control)

setup.py

@@ -13,8 +13,8 @@
 
 INSTALL_REQUIRE = [
     "torch>=1.11",
-    "optimum>=1.17.0",
-    "transformers>=4.29.0,<4.38.0",
+    "optimum~=1.17",
+    "transformers>=4.36.0,<4.39.0",
     "datasets>=1.4.0",
     "sentencepiece",
     "scipy",
@@ -45,14 +45,11 @@
         "neural-compressor>=2.2.0",
         "onnx",
         "onnxruntime<1.15.0",
-        "transformers>=4.34.0",
     ],
     "openvino": [
         "openvino>=2023.3",
         "onnx",
         "onnxruntime",
-        "transformers>=4.36.0",
-        "optimum>=1.16.1",
     ],
     "openvino-tokenizers": ["openvino-tokenizers[transformers]"],
     "nncf": ["nncf>=2.8.1"],

tests/ipex/test_inference.py

@@ -115,7 +115,7 @@ def test_text_generation_pipeline_inference(self, model_arch):
         model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, return_dict=False)
         model = model.eval()
         tokenizer = AutoTokenizer.from_pretrained(model_id)
-        inputs = "DeepSpeed is a machine learning framework for deep neural networks and deep reinforcement learning. It is written in C++ and is available for Linux, Mac OS X,"
+        inputs = "This is a simple input"
         text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
         with torch.inference_mode():
             output = text_generator(inputs)

tests/ipex/test_modeling.py

@@ -67,7 +67,6 @@
     "gptj": "hf-internal-testing/tiny-random-GPTJModel",
     "levit": "hf-internal-testing/tiny-random-LevitModel",
     "llama": "fxmarty/tiny-llama-fast-tokenizer",
-    "opt": "hf-internal-testing/tiny-random-OPTModel",
     "marian": "sshleifer/tiny-marian-en-de",
     "mbart": "hf-internal-testing/tiny-random-mbart",
     "mistral": "echarlaix/tiny-random-mistral",
@@ -76,6 +75,8 @@
     "mobilevit": "hf-internal-testing/tiny-random-mobilevit",
     "mpt": "hf-internal-testing/tiny-random-MptForCausalLM",
     "mt5": "stas/mt5-tiny-random",
+    "opt": "hf-internal-testing/tiny-random-OPTModel",
+    "phi": "hf-internal-testing/tiny-random-PhiForCausalLM",
     "resnet": "hf-internal-testing/tiny-random-resnet",
     "roberta": "hf-internal-testing/tiny-random-roberta",
     "roformer": "hf-internal-testing/tiny-random-roformer",
@@ -199,7 +200,7 @@ def test_pipeline(self, model_arch):
 class IPEXModelForCausalLMTest(unittest.TestCase):
     SUPPORTED_ARCHITECTURES = (
         "bart",
-        # "gpt_bigcode",
+        "gpt_bigcode",
         "blenderbot",
         "blenderbot-small",
         "bloom",
@@ -208,8 +209,9 @@ class IPEXModelForCausalLMTest(unittest.TestCase):
         "gpt_neo",
         "gpt_neox",
         "llama",
-        # "mistral",
-        # "mpt",
+        "mistral",
+        # "phi",
+        "mpt",
         "opt",
     )
     GENERATION_LENGTH = 100

tests/openvino/test_modeling.py

@@ -483,7 +483,7 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
         "gpt_neo",
         "gpt_neox",
         "llama",
-        "llama_gptq",
+        # "llama_gptq",
         "marian",
         "mistral",
         "mpt",
@@ -504,7 +504,7 @@ def test_compare_to_transformers(self, model_arch):
         ov_model = OVModelForCausalLM.from_pretrained(model_id, export=True, ov_config=F32_CONFIG)
         self.assertIsInstance(ov_model.config, PretrainedConfig)
         self.assertTrue(ov_model.use_cache)
-        self.assertEqual(ov_model.stateful, self.IS_SUPPORT_STATEFUL and model_arch != "gpt_bigcode")
+
         transformers_model = AutoModelForCausalLM.from_pretrained(model_id)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
         tokens = tokenizer(
@@ -520,10 +520,15 @@ def test_compare_to_transformers(self, model_arch):
         self.assertIsInstance(ov_outputs.logits, torch.Tensor)
         self.assertTrue("past_key_values" in ov_outputs)
         self.assertIsInstance(ov_outputs.past_key_values, tuple)
-        if self.IS_SUPPORT_STATEFUL and model_arch != "gpt_bigcode":
+
+        is_stateful = ov_model.config.model_type not in {"gpt_bigcode", "llama"} and self.IS_SUPPORT_STATEFUL
+        self.assertEqual(ov_model.stateful, is_stateful)
+        if is_stateful:
             self.assertTrue(len(ov_outputs.past_key_values) == 1 and len(ov_outputs.past_key_values[0]) == 0)
+
         with torch.no_grad():
             transformers_outputs = transformers_model(**tokens)
+
         # Compare tensor outputs
         self.assertTrue(torch.allclose(ov_outputs.logits, transformers_outputs.logits, atol=1e-4))
         del transformers_model
@@ -540,7 +545,7 @@ def test_pipeline(self, model_arch):
         model.half()
         model.compile()
         pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
-        outputs = pipe("This is a sample", max_length=10)
+        outputs = pipe("This is a sample", max_length=20)
         self.assertEqual(pipe.device, model.device)
         self.assertTrue(all("This is a sample" in item["generated_text"] for item in outputs))
         del pipe