Fix-3

Signed-off-by: amitraj <[email protected]>

Author: quic-amitraj

QEfficient/transformers/models/modeling_auto.py

@@ -66,21 +66,6 @@ def model_name(self) -> str:
             mname = mname[4:]
         return mname
 
-    @property
-    def model_hash(self) -> str:
-        # NOTE: model_config.to_diff_dict() has "_name_or_path" attribute which is the model card name or path.
-        # Using same card name will result in same hash. But, using a relative path for one run and
-        # absolute path for another run will result in different hash.
-        # The added complexity to resolve different paths to same location is not worth pursuing.
-        # Instead, advise the user to always provide same relative paths or absolute paths for local models.
-
-        # Compute the hash with: model_config, transforms
-        mhash = hashlib.sha256()
-        mhash.update(to_hashable(self.model.config.to_diff_dict()))
-        mhash.update(to_hashable(self._transform_names()))
-        mhash = mhash.hexdigest()[:16]
-        return mhash
-
 
 class QEFFAutoModelForCausalLM(QEFFTransformersBase):
     """
@@ -107,6 +92,10 @@ class QEFFAutoModelForCausalLM(QEFFTransformersBase):
     _onnx_transforms = [FP16ClipTransform, SplitTensorsTransform]
 
     def __init__(self, model: nn.Module, continuous_batching: bool = False, **kwargs):
+        model_class_name = model.__class__.__name__
+        if not (model_class_name.endswith("ForCausalLM") or model_class_name.endswith("LMHeadModel")):
+            raise TypeError(f"Required pytorch module for CausalLM or LMHeadModel, got {model_class_name}")
+
         if kwargs.pop("full_batch_size", None):
             continuous_batching = True
             warnings.warn(
@@ -382,6 +371,21 @@ def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
 
         return self
 
+    @property
+    def model_hash(self) -> str:
+        # NOTE: model_config.to_diff_dict() has "_name_or_path" attribute which is the model card name or path.
+        # Using same card name will result in same hash. But, using a relative path for one run and
+        # absolute path for another run will result in different hash.
+        # The added complexity to resolve different paths to same location is not worth pursuing.
+        # Instead, advise the user to always provide same relative paths or absolute paths for local models.
+
+        # Compute the hash with: model_config, transforms
+        mhash = hashlib.sha256()
+        mhash.update(to_hashable(self.model.config.to_diff_dict()))
+        mhash.update(to_hashable(self._transform_names()))
+        mhash = mhash.hexdigest()[:16]
+        return mhash
+
     def export(self, export_dir: Optional[str] = None) -> str:
         """
         Exports the model to ``ONNX`` format using ``torch.onnx.export``.

QEfficient/utils/constants.py

@@ -46,7 +46,7 @@ def get_models_dir():
 ONNX_EXPORT_EXAMPLE_BATCH_SIZE = 1
 ONNX_EXPORT_EXAMPLE_SEQ_LEN = 32
 ONNX_EXPORT_EXAMPLE_FBS = 4
-ONNX_EXPORT_OPSET = 13
+ONNX_EXPORT_OPSET = 14
 
 COMPILER = ["/opt/qti-aic/exec/qaic-exec", "-aic-hw", "-aic-hw-version=2.0"]
 

tests/transformers/models/test_causal_lm_models.py

@@ -183,11 +183,21 @@ def check_embed_pytorch_vs_ort_vs_ai100(
     # Try to initialize with add_pooling_layer parameter
     try:
         qeff_model = QEffAutoModel.from_pretrained(
-            pretrained_model_name_or_path=model_path, add_pooling_layer=False, num_hidden_layers=n_layer
+            pretrained_model_name_or_path=model_path,
+            add_pooling_layer=False,
+            num_hidden_layers=n_layer,
+            attn_implementation="eager",
+            trust_remote_code=True,
         )
     except TypeError:
         # If it fails, initialize without the parameter
-        qeff_model = QEffAutoModel.from_pretrained(pretrained_model_name_or_path=model_path, num_hidden_layers=n_layer)
+        qeff_model = QEffAutoModel.from_pretrained(
+            pretrained_model_name_or_path=model_path,
+            num_hidden_layers=n_layer,
+            attn_implementation="eager",
+            trust_remote_code=True,
+        )
+
     text = "My name is"
     tokenizer = AutoTokenizer.from_pretrained(model_name)
     inputs = tokenizer(text, return_tensors="pt", padding="max_length", max_length=seq_len)
@@ -206,7 +216,7 @@ def check_embed_pytorch_vs_ort_vs_ai100(
     onnx_embeddings = onnx_outputs[0]
     mad = np.mean(np.abs(pt_embeddings - onnx_embeddings))
     print("Mad for onnx and pytorch is ", mad)
-    assert mad <= 10**-6, f"MAD is too high for onnx and Pytorch: {mad}"
+    assert mad <= 10**-3, f"MAD is too high for onnx and Pytorch: {mad}"
 
     qeff_model.compile(
         num_cores=14,
@@ -277,17 +287,20 @@ def test_causal_lm_pytorch_vs_kv_vs_ort_vs_ai100_pl1():
 
 
 embed_test_models = [
-    "intfloat/e5-mistral-7b-instruct",  # MistralModel
+    # model_name, architecture
+    "nomic-ai/nomic-embed-text-v1.5",  # NomicBertModel
     "sentence-transformers/multi-qa-mpnet-base-cos-v1",  # MPNetForMaskedLM
     "BAAI/bge-reranker-v2-m3",  # XLMRobertaForSequenceClassification
     "BAAI/bge-small-en-v1.5",  # BertModel
+    # "intfloat/e5-mistral-7b-instruct",  # MistralModel
+    # "dunzhang/stella_en_1.5B_v5", # Qwen2ForCausalLM
 ]
 
 
 @pytest.mark.on_qaic
 @pytest.mark.parametrize("model_name", embed_test_models)
 def test_embed_model_pytorch_vs_onnx_vs_ai100(model_name):
     """
-    Test function to validate the Pytorch model, ONNX model and
+    Test function to validate output of the Pytorch, ONNX and AI 100 runtime model output.
     """
     check_embed_pytorch_vs_ort_vs_ai100(model_name=model_name, seq_len=32, n_layer=1)