Merge branch 'main' into custom_dataset

Signed-off-by: Swati Allabadi <[email protected]>

Author: quic-swatia

.github/CODEOWNERS

@@ -1,12 +1,12 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # -----------------------------------------------------------------------------
 
 # Default owners
 # review when someone opens a pull request and assign appropriate reviewer
-* @quic-rishinr @ochougul @quic-hemagnih
+* @quic-rishinr @ochougul @quic-hemagnih @quic-amitraj
 pyproject.toml @carlstreeter-quic
 

LICENSE

@@ -1,4 +1,4 @@
-Copyright (c) 2024, Qualcomm Innovation Center, Inc. All rights reserved.
+Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 
 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions are met:
@@ -11,7 +11,7 @@ modification, are permitted provided that the following conditions are met:
       disclaimer in the documentation and/or other materials provided
       with the distribution.
 
-    * Neither the name of Qualcomm Innovation Center, Inc. nor the names of its
+    * Neither the name of Qualcomm Technologies, Inc. nor the names of its
       contributors may be used to endorse or promote products derived
       from this software without specific prior written permission.
 

QEfficient/__init__.py

@@ -1,29 +1,25 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2025 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # -----------------------------------------------------------------------------
 
 import os
+import warnings
+
+from QEfficient.utils import custom_format_warning
 
 # 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"
-
-from transformers import AutoConfig
-
-from QEfficient.transformers.modeling_utils import MODEL_TYPE_TO_CONFIG_CLS_AND_ARCH_CLS
+# Placeholder for all non-transformer models registered in QEfficient
+import QEfficient.utils.model_registery  # noqa: F401
 from QEfficient.utils.logging_utils import logger
 
-# loop over all the model types which are not present in transformers and register them
-for model_type, model_cls in MODEL_TYPE_TO_CONFIG_CLS_AND_ARCH_CLS.items():
-    # Register the model config class based on the model type. This will be first element in the tuple
-    AutoConfig.register(model_type, model_cls[0])
-
-    # Register the non transformer library Class and config class using AutoModelClass
-    model_cls[2].register(model_cls[0], model_cls[1])
+# custom warning for the better logging experience
+warnings.formatwarning = custom_format_warning
 
 
 def check_qaic_sdk():

QEfficient/base/__init__.py

@@ -1,6 +1,6 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # -----------------------------------------------------------------------------

QEfficient/base/common.py

@@ -1,6 +1,6 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # -----------------------------------------------------------------------------
@@ -41,15 +41,14 @@ def from_pretrained(cls, pretrained_model_name_or_path: str, *args, **kwargs) ->
         Downloads HuggingFace model if already doesn't exist locally, returns QEFFAutoModel object based on type of model.
         """
         config = AutoConfig.from_pretrained(pretrained_model_name_or_path)
-        architecture = config.architectures[0] if config.architectures else None
 
-        class_name = MODEL_CLASS_MAPPING.get(architecture)
+        class_name = MODEL_CLASS_MAPPING.get(config.__class__.__name__, None)
         if class_name:
             module = __import__("QEfficient.transformers.models.modeling_auto")
             model_class = getattr(module, class_name)
         else:
             raise NotImplementedError(
-                f"Unknown architecture={architecture}, either use specific auto model class for loading the model or raise an issue for support!"
+                f"Unknown architecture={config.__class__.__name__}, either use specific auto model class for loading the model or raise an issue for support!"
             )
 
         local_model_dir = kwargs.pop("local_model_dir", None)

QEfficient/base/modeling_qeff.py

@@ -1,6 +1,6 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # ----------------------------------------------------------------------------
@@ -241,10 +241,12 @@ def _compile(
             :mdp_ts_num_devices (int): Number of devices to partition to use Multi-Device Partitioning with tensor-slicing.
             :num_speculative_tokens (int, optional): Number of speculative tokens to take as input for Speculative Decoding Target Language Model.
             :enable_qnn (bool): Enables QNN Compilation. ``Defaults to False.``
-            :qnn_config (str): Path of QNN Config parameters file. ``Defaults to None.``
-            :compiler_options: Pass any compiler option as input. Any flag that is supported by `qaic-exec` can be passed. Params are converted to flags as below:
+            :qnn_config (str): Path of QNN Config parameters file. Any extra parameters for QNN compilation can be passed via this file. ``Defaults to None.``
+            :compiler_options: Pass any compiler option as input.
+                Any flag that is supported by `qaic-exec` can be passed. Params are converted to flags as below:
                 - aic_num_cores=16 -> -aic-num-cores=16
                 - convert_to_fp16=True -> -convert-to-fp16
+                For QNN Compilation path, when enable_qnn is set to True, any parameter passed in compiler_options will be ignored.
         """
         if onnx_path is None and self.onnx_path is None:
             self.export()
@@ -256,6 +258,11 @@ def _compile(
             raise FileNotFoundError(f"ONNX file not found at: {onnx_path}")
 
         if enable_qnn:
+            if compiler_options:
+                logger.warning(
+                    f"Extra arguments to QNN compilation are supported only via qnn_config file. Ignoring {compiler_options}"
+                )
+
             self.qpc_path = qnn_compile(
                 onnx_path=onnx_path,
                 qpc_base_path=compile_dir,

QEfficient/base/onnx_transforms.py

@@ -1,6 +1,6 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # ----------------------------------------------------------------------------

QEfficient/base/pytorch_transforms.py

@@ -1,6 +1,6 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # ----------------------------------------------------------------------------
@@ -9,6 +9,8 @@
 
 from torch import nn
 
+from QEfficient.utils.logging_utils import logger
+
 
 class PytorchTransform:
     """
@@ -110,3 +112,65 @@ def apply(cls, model: nn.Module) -> Tuple[nn.Module, bool]:
                     transformed = True
 
         return model, transformed
+
+
+class SplitGateUpWeightsTransform(PytorchTransform):
+    """
+    split fused Gate+Up weights and copy into the model
+
+    For every transformer layer inside `model`:
+      • expects   <PREFIX>.experts.gate_up_proj   in the *source* `sd`
+      • copies halves into
+            <PREFIX>.experts.gate_proj     <-- Gate   [E,H,I]
+            <PREFIX>.experts.up_proj       <-- Up     [E,H,I]
+    """
+
+    @classmethod
+    def apply(cls, model: nn.Module) -> Tuple[nn.Module, bool]:
+        transformed = False
+        model_class = model.__class__.__name__ if hasattr(model, "model") else model.__class__.__name__
+
+        if model_class not in VLM_SPLIT_GATE_UP_WEIGHTS:
+            return model, transformed
+
+        model_tmp = model.language_model if hasattr(model, "language_model") else model
+
+        num_layers = len(model_tmp.model.layers)
+        delete_fused_key = True
+        sd = model_tmp.state_dict()
+        for layer_idx in range(num_layers):
+            # ---- build the textual prefix once per layer ----------
+            prefix = f"model.layers.{layer_idx}.feed_forward.experts."
+
+            fused_key = prefix + "gate_up_proj"
+            gate_key = prefix + "gate_proj"
+            up_key = prefix + "up_proj"
+
+            # ---- split  [E,H,2I] → two  [E,H,I]  tensors ----------------------
+            fused = sd[fused_key]  # [E, H, 2I]  (no .weight here)
+            E, H, two_I = fused.shape
+            ffn_dim = two_I // 2
+            gate, up = fused.split(ffn_dim, dim=-1)  # views – no copy
+
+            experts = model_tmp.model.layers[layer_idx].feed_forward.experts
+            experts.gate_proj.data.copy_(gate)
+            experts.up_proj.data.copy_(up)
+
+            # ---- update the state-dict so load_state_dict sees the right keys
+            sd[gate_key] = gate
+            sd[up_key] = up
+
+            if delete_fused_key:
+                del sd[fused_key]
+
+            logger.info(f"[layer {layer_idx:02d}] loaded gate_proj & up_proj from fused tensor  (shape {fused.shape})")
+            transformed = True
+
+        if hasattr(model, "language_model"):
+            model.language_model = model_tmp
+        else:
+            model = model_tmp
+        return model, transformed
+
+
+VLM_SPLIT_GATE_UP_WEIGHTS = {"QEffLlama4ForConditionalGeneration"}

QEfficient/cloud/__init__.py

@@ -1,6 +1,6 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # -----------------------------------------------------------------------------

QEfficient/cloud/compile.py

@@ -1,6 +1,6 @@
 # -----------------------------------------------------------------------------
 #
-# Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
 # SPDX-License-Identifier: BSD-3-Clause
 #
 # -----------------------------------------------------------------------------
@@ -85,17 +85,29 @@
     parser.add_argument(
         "--enable_qnn",
         "--enable-qnn",
-        action="store_true",
+        nargs="?",
+        const=True,
+        type=str,
         default=False,
         help="Enables QNN. Optionally, a configuration file can be provided with [--enable_qnn CONFIG_FILE].\
              If not provided, the default configuration will be used.\
              Sample Config: QEfficient/compile/qnn_config.json",
     )
-    parser.add_argument(
-        "qnn_config",
-        nargs="?",
-        type=str,
-    )
-    # FIXME(ochougul): Allow extra compilation arguments
-    args = parser.parse_args()
-    QEfficient.compile(**vars(args))
+
+    args, compiler_options = parser.parse_known_args()
+
+    if isinstance(args.enable_qnn, str):
+        args.qnn_config = args.enable_qnn
+        args.enable_qnn = True
+
+    compiler_options_dict = {}
+    for i in range(0, len(compiler_options)):
+        if compiler_options[i].startswith("--"):
+            key = compiler_options[i].lstrip("-").replace("-", "_")
+            value = (
+                compiler_options[i + 1]
+                if i + 1 < len(compiler_options) and not compiler_options[i + 1].startswith("-")
+                else True
+            )
+            compiler_options_dict[key] = value
+    QEfficient.compile(**args.__dict__, **compiler_options_dict)