CI failures addressed

Signed-off-by: Asmita Goswami <asmigosw@qti.qualcomm.com>

Author: asmigosw

QEfficient/transformers/models/modeling_auto.py

@@ -4221,7 +4221,7 @@ def cloud_ai_100_feature_generate(
             torch.nn.functional.pad(inputs["input_values"], (0, self.seq_len - input_ids_len), "constant", 0)
         )
         needed_dtype = getattr(self.model.config, "torch_dtype", torch.float32)
-        input_values = input_values.astype(CUSTOM_IO_DTYPE_MAP[needed_dtype])
+        input_values = input_values.astype(TORCH_TO_NUMPY_DTYPE_MAP[needed_dtype])
         inputs = dict(input_values=input_values)
         outputs = self.qpc_session.run(inputs)
 

examples/audio/wav2vec2_inference.py

@@ -7,6 +7,7 @@
 
 import argparse
 
+import torch
 from datasets import load_dataset
 from transformers import AutoProcessor
 
@@ -40,7 +41,7 @@ def main():
     processor = AutoProcessor.from_pretrained(args.model_name)
 
     ## STEP 2 -- Load the model
-    model = QEFFAutoModelForCTC.from_pretrained(args.model_name)
+    model = QEFFAutoModelForCTC.from_pretrained(args.model_name, torch_dtype=torch.float32)
 
     ## STEP 3 -- Compile the model
     model.compile(num_cores=args.num_cores)

tests/transformers/models/test_audio_embedding_models.py

@@ -31,7 +31,7 @@
     test_models = config_data["audio_embedding_models"]
 
 
-def load_ctc_model(model_config):
+def load_ctc_model(model_config, torch_dtype: Optional[torch.dtype] = torch.float32):
     """
     Function to load model from huggingface
     --------
@@ -48,6 +48,7 @@ def load_ctc_model(model_config):
         model_path,
         attn_implementation="eager",
         low_cpu_mem_usage=False,
+        torch_dtype=torch_dtype,
     )  # Run models for single layers only
     params = sum(p.numel() for p in model_hf.parameters())
     model_hf.eval()

tests/transformers/models/test_causal_lm_models.py

@@ -43,8 +43,11 @@
 test_models_qnn = [model["model_name"] for model in qnn_models]
 test_models_blockedKV = [model["model_name"] for model in blockedKV_models]
 
+all_models = causal_lm_models + causal_lm_fp16_models
+
 # Create a dictionary mapping model names to their configs
-model_config_dict = {model["model_name"]: model for model in causal_lm_models}
+model_config_dict = {model["model_name"]: model for model in all_models}
+breakpoint()
 
 
 def get_hf_config_from_custom_config(model_name):