[QEff. Finetune] : Bypass qeff import and add trust_remote_code flag …

…for samsum dataset (#206)

1. Bypass qeff import
2. Add trust_remote_code flag for samsum dataset

---------

Signed-off-by: Mamta Singh <[email protected]>

QEfficient/__init__.py

@@ -5,24 +5,47 @@
 #
 # -----------------------------------------------------------------------------
 
-from QEfficient.base import QEFFAutoModel, QEFFAutoModelForCausalLM, QEFFCommonLoader
-from QEfficient.compile.compile_helper import compile
-from QEfficient.exporter.export_hf_to_cloud_ai_100 import qualcomm_efficient_converter
-from QEfficient.generation.text_generation_inference import cloud_ai_100_exec_kv
-from QEfficient.peft import QEffAutoPeftModelForCausalLM
-from QEfficient.transformers.transform import transform
-
-# Users can use QEfficient.export for exporting models to ONNX
-export = qualcomm_efficient_converter
+
+def check_qaic_sdk():
+    """Check if QAIC SDK is installed"""
+    try:
+        import platform
+        import sys
+
+        sys.path.append(f"/opt/qti-aic/dev/lib/{platform.machine()}")
+        import qaicrt  # noqa: F401
+
+        return True
+    except ImportError:
+        return False
+
+
+QAIC_INSTALLED = check_qaic_sdk()
+
+# Conditionally import QAIC-related modules if the SDK is installed
 __version__ = "0.0.1.dev0"
+if QAIC_INSTALLED:
+    from QEfficient.base import QEFFAutoModel, QEFFAutoModelForCausalLM, QEFFCommonLoader
+    from QEfficient.compile.compile_helper import compile
+    from QEfficient.exporter.export_hf_to_cloud_ai_100 import qualcomm_efficient_converter
+    from QEfficient.generation.text_generation_inference import cloud_ai_100_exec_kv
+    from QEfficient.peft import QEffAutoPeftModelForCausalLM
+    from QEfficient.transformers.transform import transform
+
+    # Users can use QEfficient.export for exporting models to ONNX
+    export = qualcomm_efficient_converter
+
+    __all__ = [
+        "transform",
+        "export",
+        "compile",
+        "cloud_ai_100_exec_kv",
+        "QEFFAutoModel",
+        "QEFFAutoModelForCausalLM",
+        "QEffAutoPeftModelForCausalLM",
+        "QEFFCommonLoader",
+    ]
 
-__all__ = [
-    "transform",
-    "export",
-    "compile",
-    "cloud_ai_100_exec_kv",
-    "QEFFAutoModel",
-    "QEFFAutoModelForCausalLM",
-    "QEffAutoPeftModelForCausalLM",
-    "QEFFCommonLoader",
-]
+    print("QAIC SDK is installed.")
+else:
+    print("QAIC SDK is not installed. Proceeding without it.")

QEfficient/cloud/finetune.py

@@ -5,7 +5,6 @@
 #
 # -----------------------------------------------------------------------------
 
-import os
 import random
 import warnings
 
@@ -58,8 +57,6 @@ def main(**kwargs):
     update_config(train_config, **kwargs)
     device = train_config.device
 
-    os.environ["HF_DATASETS_TRUST_REMOTE_CODE"] = "True"
-
     # dist init
     if train_config.enable_ddp:
         # TODO: may have to init qccl backend, next try run with torchrun command

QEfficient/finetune/configs/peft_config.py

@@ -20,7 +20,7 @@ class lora_config:
     bias = "none"
     task_type: str = "CAUSAL_LM"
     lora_dropout: float = 0.05
-    inference_mode: bool = False
+    inference_mode: bool = False  # should be False for finetuning
 
 
 # CAUTION prefix tuning is currently not supported

QEfficient/finetune/configs/training.py

@@ -38,7 +38,7 @@ class train_config:
     save_metrics: bool = True  # saves training metrics to a json file for later plotting
     intermediate_step_save: int = 1000
     batching_strategy: str = "packing"
-    enable_sorting_for_ddp: bool = "True"
+    enable_sorting_for_ddp: bool = True
 
     # TODO: vbaddi: Uncomment post adding qaic to Pytorch Profiler
     # flop_counter: bool = False # Enable flop counter to measure model throughput, can not be used with pytorch profiler at the same time.

QEfficient/finetune/dataset/samsum_dataset.py

@@ -5,27 +5,11 @@
 #
 # -----------------------------------------------------------------------------
 
-from unittest.mock import patch
-
 import datasets
 
 
-@patch("builtins.input", return_value="N")
-def load_samsum(split, _):
-    try:
-        ds = datasets.load_dataset("Samsung/samsum", split=split)
-    except ValueError as e:
-        if "trust_remote_code" in str(e):
-            raise ValueError(
-                "Loading Samsung/samsum requires you to execute the dataset script in that repo on your local machine. Make sure you have read the code there to avoid malicious use, then set HF_DATASETS_TRUST_REMOTE_CODE env variable to True."
-            ) from e
-        else:
-            raise e
-    return ds
-
-
 def get_preprocessed_samsum(dataset_config, tokenizer, split, context_length=None):
-    dataset = load_samsum(split)
+    dataset = datasets.load_dataset("Samsung/samsum", split=split, trust_remote_code=True)
 
     prompt = "Summarize this dialog:\n{dialog}\n---\nSummary:\n"
 

scripts/finetune/run_ft_model.py

@@ -9,17 +9,18 @@
 import warnings
 
 import torch
-from configs.training import train_config as TRAIN_CONFIG
 from peft import AutoPeftModelForCausalLM
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
+from QEfficient.finetune.configs.training import train_config as TRAIN_CONFIG
+
 # Suppress all warnings
 warnings.filterwarnings("ignore")
 
 try:
     import torch_qaic  # noqa: F401
 
-    device = "qaic:1"
+    device = "qaic:0"
 except ImportError as e:
     print(f"Warning: {e}. Moving ahead without these qaic modules.")
     device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
@@ -29,7 +30,7 @@
     train_config.model_name,
     use_cache=False,
     attn_implementation="sdpa",
-    torch_dtype=torch.float16 if torch.cuda.is_available() or device == "qaic:1" else None,
+    torch_dtype=torch.float16 if torch.cuda.is_available() or device == "qaic:0" else None,
 )
 
 # Load the tokenizer and add special tokens