removed platform sdk dependency (#609)

This PR updates QEff to support QPC generation on systems without the
Platform SDK by refactoring the module loading behavior. Users can now
compile models and generate QPCs using QEff with only the Apps SDK
installed.

Background: Previously, both Apps SDK and Platform SDK were required to
compile and generate QPCs using QEff. The goal is to allow QPC
generation with only the Apps SDK installed for systems without Ultra
cards.
Changes:
Refactored init.py and generation/cloud_infer.py to use lazy loading via
importlib for qaicrt and aicapi.
This ensures that Platform SDK-dependent modules are only loaded when
explicitly needed, avoiding import errors during initialization and QPC
generation.

Signed-off-by: Sharvari Medhe <[email protected]>
Co-authored-by: Hem Agnihotri <[email protected]>

Author: smedhe

QEfficient/__init__.py

@@ -9,19 +9,49 @@
 import warnings
 
 import QEfficient.utils.model_registery  # noqa: F401
+from QEfficient.base import (
+    QEFFAutoModel,
+    QEFFAutoModelForCausalLM,
+    QEFFAutoModelForCTC,
+    QEFFAutoModelForImageTextToText,
+    QEFFAutoModelForSpeechSeq2Seq,
+    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
 from QEfficient.utils import custom_format_warning
 from QEfficient.utils.logging_utils import logger
 
+# 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",
+    "QEFFAutoModelForCTC",
+    "QEffAutoPeftModelForCausalLM",
+    "QEFFAutoModelForImageTextToText",
+    "QEFFAutoModelForSpeechSeq2Seq",
+    "QEFFCommonLoader",
+]
 # For faster downloads via hf_transfer
 # This code is put above import statements as this needs to be executed before
 # hf_transfer is imported (will happen on line 15 via leading imports)
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 # Placeholder for all non-transformer models registered in QEfficient
 
-
 # custom warning for the better logging experience
 warnings.formatwarning = custom_format_warning
 
+# Conditionally import QAIC-related modules if the SDK is installed
+__version__ = "0.0.1.dev0"
+
 
 def check_qaic_sdk():
     """Check if QAIC SDK is installed"""
@@ -37,40 +67,5 @@ def check_qaic_sdk():
         return False
 
 
-# Conditionally import QAIC-related modules if the SDK is installed
-__version__ = "0.0.1.dev0"
-
-if check_qaic_sdk():
-    from QEfficient.base import (
-        QEFFAutoModel,
-        QEFFAutoModelForCausalLM,
-        QEFFAutoModelForCTC,
-        QEFFAutoModelForImageTextToText,
-        QEFFAutoModelForSpeechSeq2Seq,
-        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",
-        "QEFFAutoModelForCTC",
-        "QEffAutoPeftModelForCausalLM",
-        "QEFFAutoModelForImageTextToText",
-        "QEFFAutoModelForSpeechSeq2Seq",
-        "QEFFCommonLoader",
-    ]
-
-else:
+if not check_qaic_sdk():
     logger.warning("QAIC SDK is not installed, eager mode features won't be available!")

QEfficient/generation/cloud_infer.py

@@ -5,6 +5,8 @@
 #
 # -----------------------------------------------------------------------------
 
+import platform
+import sys
 from pathlib import Path
 from typing import Dict, List, Optional, Union
 from warnings import warn
@@ -13,32 +15,29 @@
 
 try:
     import qaicrt
+
+    is_qaicrt_imported = True
 except ImportError:
-    import platform
-    import sys
+    try:
+        sys.path.append(f"/opt/qti-aic/dev/lib/{platform.machine()}")
+        import qaicrt
 
-    sys.path.append(f"/opt/qti-aic/dev/lib/{platform.machine()}")
-    import qaicrt
+        is_qaicrt_imported = True
+    except ImportError:
+        is_qaicrt_imported = False
 
 try:
     import QAicApi_pb2 as aicapi
-except ImportError:
-    import sys
 
-    sys.path.append("/opt/qti-aic/dev/python")
-    import QAicApi_pb2 as aicapi
+    is_aicapi_imported = True
+except ImportError:
+    try:
+        sys.path.append("/opt/qti-aic/dev/python")
+        import QAicApi_pb2 as aicapi
 
-aic_to_np_dtype_mapping = {
-    aicapi.FLOAT_TYPE: np.dtype(np.float32),
-    aicapi.FLOAT_16_TYPE: np.dtype(np.float16),
-    aicapi.INT8_Q_TYPE: np.dtype(np.int8),
-    aicapi.UINT8_Q_TYPE: np.dtype(np.uint8),
-    aicapi.INT16_Q_TYPE: np.dtype(np.int16),
-    aicapi.INT32_Q_TYPE: np.dtype(np.int32),
-    aicapi.INT32_I_TYPE: np.dtype(np.int32),
-    aicapi.INT64_I_TYPE: np.dtype(np.int64),
-    aicapi.INT8_TYPE: np.dtype(np.int8),
-}
+        is_aicapi_imported = True
+    except ImportError:
+        is_qaicrt_imported = False
 
 
 class QAICInferenceSession:
@@ -58,6 +57,25 @@ def __init__(
         :activate: bool. If false, activation will be disabled. Default=True.
         :enable_debug_logs: bool. If True, It will enable debug logs. Default=False.
         """
+        if not (is_qaicrt_imported and is_aicapi_imported):
+            raise ImportError(
+                "Unable to import `qaicrt` and/or `QAicApi_pb2` libraries required for executing QPC files on the CLOUD AI platform.\n"
+                "Please ensure that the QAIC platform SDK and apps SDK are installed correctly."
+            )
+
+        # Build dtype mapping once (depends on aicapi constants)
+        self.aic_to_np_dtype_mapping = {
+            aicapi.FLOAT_TYPE: np.dtype(np.float32),
+            aicapi.FLOAT_16_TYPE: np.dtype(np.float16),
+            aicapi.INT8_Q_TYPE: np.dtype(np.int8),
+            aicapi.UINT8_Q_TYPE: np.dtype(np.uint8),
+            aicapi.INT16_Q_TYPE: np.dtype(np.int16),
+            aicapi.INT32_Q_TYPE: np.dtype(np.int32),
+            aicapi.INT32_I_TYPE: np.dtype(np.int32),
+            aicapi.INT64_I_TYPE: np.dtype(np.int64),
+            aicapi.INT8_TYPE: np.dtype(np.int8),
+        }
+
         # Load QPC
         if device_ids is not None:
             devices = qaicrt.QIDList(device_ids)
@@ -77,7 +95,7 @@ def __init__(
             raise RuntimeError("Failed to getIoDescriptor")
         iodesc.ParseFromString(bytes(iodesc_data))
         self.allowed_shapes = [
-            [(aic_to_np_dtype_mapping[x.type].itemsize, list(x.dims)) for x in allowed_shape.shapes]
+            [(self.aic_to_np_dtype_mapping[x.type].itemsize, list(x.dims)) for x in allowed_shape.shapes]
             for allowed_shape in iodesc.allowed_shapes
         ]
         self.bindings = iodesc.selected_set.bindings
@@ -97,7 +115,7 @@ def __init__(
         # Create input qbuffers and buf_dims
         self.qbuffers = [qaicrt.QBuffer(bytes(binding.size)) for binding in self.bindings]
         self.buf_dims = qaicrt.BufferDimensionsVecRef(
-            [(aic_to_np_dtype_mapping[binding.type].itemsize, list(binding.dims)) for binding in self.bindings]
+            [(self.aic_to_np_dtype_mapping[binding.type].itemsize, list(binding.dims)) for binding in self.bindings]
         )
 
     @property
@@ -205,6 +223,6 @@ def run(self, inputs: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
                 continue
             outputs[output_name] = np.frombuffer(
                 bytes(output_qbuffers[buffer_index]),
-                aic_to_np_dtype_mapping[self.bindings[buffer_index].type],
+                self.aic_to_np_dtype_mapping[self.bindings[buffer_index].type],
             ).reshape(self.buf_dims[buffer_index][1])
         return outputs