Added default NPI file

QEfficient/transformers/models/gemma3/configs/__init__.py

@@ -0,0 +1,6 @@
+# -----------------------------------------------------------------------------
+#
+# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+# SPDX-License-Identifier: BSD-3-Clause
+#
+# -----------------------------------------------------------------------------

QEfficient/transformers/models/gemma3/modeling_gemma3.py

@@ -677,6 +677,14 @@ def forward(
         logits = logits.float()
         return logits, pixel_values, image_idx, outputs.past_key_values
 
+    def get_npi_file(self, model_name: str) -> str:
+        if constants.NPI_MAPPING[model_name] is not None:
+            return constants.NPI_MAPPING[model_name]
+        else:
+            raise ValueError(
+                f"For Model {self.pretrained_model_name_or_path} default NPI file is not supported/added for this particular model. Please use one of the following: google/gemma-3-4b-it, google/gemma-3-27b-it"
+            )
+
     def get_specializations(
         self,
         batch_size: int,

QEfficient/transformers/models/modeling_auto.py

@@ -1191,7 +1191,6 @@ def compile(
         compiler_options.pop("continuous_batching", None)
         compiler_options.pop("kv_cache_batch_size", None)
         compiler_options.pop("full_batch_size", None)
-
         if not skip_vision:
             self.vision_model._compile(
                 compile_dir=compile_dir,
@@ -1207,6 +1206,10 @@ def compile(
                 **compiler_options,
             )
 
+        # Custom NPI file options
+        if hasattr(self.model, "get_npi_file") and "node_precision_info" not in compiler_options:
+            compiler_options["node_precision_info"] = self.model.get_npi_file(self.model.name_or_path)
+
         if not skip_lang:
             custom_io_lang = {}
             # Inputs
@@ -1220,7 +1223,6 @@ def compile(
             for output_name in output_names["lang"]:
                 if output_name.endswith("_RetainedState"):
                     custom_io_lang[output_name] = "float16" if "vision_embeds" in output_name else kv_cache_dtype
-
             self.lang_model._compile(
                 compile_dir=compile_dir,
                 compile_only=True,
@@ -1817,6 +1819,9 @@ def compile(
             **compiler_options,
         )
 
+        if hasattr(self.model, "get_npi_file") and "node_precision_info" not in compiler_options:
+            compiler_options["node_precision_info"] = self.model.get_npi_file(self.model.name_or_path)
+
         custom_io = {}
         kv_cache_dtype = "mxint8" if mxint8_kv_cache else "float16"
         # inputs
@@ -1835,7 +1840,6 @@ def compile(
         compiler_options.pop("continuous_batching", None)
         compiler_options.pop("kv_cache_batch_size", None)
         compiler_options.pop("full_batch_size", None)
-
         self._compile(
             onnx_path=onnx_path,
             compile_dir=compile_dir,

QEfficient/utils/constants.py

@@ -24,6 +24,15 @@
 ONNX_EXPORT_IMAGE_DEPTH = 3
 ONNX_EXPORT_CTX_LEN = 1024
 
+NPI_MAPPING = {
+    "google/gemma-3-4b-it": os.path.join(
+        QEFF_DIR, "transformers", "models", "gemma3", "configs", "fp32_nodes_gemma3_4b.yaml"
+    ),
+    "google/gemma-3-27b-it": os.path.join(
+        QEFF_DIR, "transformers", "models", "gemma3", "configs", "fp32_nodes_gemma3_27b.yaml"
+    ),
+}
+
 # Compiler defaults
 DEFAULT_AIC_NUM_CORES = 16
 DEFAULT_AIC_MXPF6_MATMUL = False

examples/image_text_to_text/models/gemma_vision/gemma3_example.py

@@ -17,8 +17,8 @@
 config = AutoConfig.from_pretrained(model_id)
 
 # For Testing Purpose Only
-config.text_config.num_hidden_layers = 1
-config.vision_config.num_hidden_layers = 2
+# config.text_config.num_hidden_layers = 1
+# config.vision_config.num_hidden_layers = 2
 
 tokenizer = transformers.AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
 processor = AutoProcessor.from_pretrained(model_id)
@@ -44,7 +44,6 @@
         aic_enable_depth_first=True,
         skip_vision=True,
         mos=1,
-        node_precision_info="examples/gemma3_example/fp32_nodes_gemma3_4b.yaml",  # Change to fp32_nodes_gemma3_27b.yaml for 27B model
     )
 
     messages = [
@@ -80,7 +79,6 @@
         mxint8_kv_cache=False,
         aic_enable_depth_first=True,
         mos=1,
-        node_precision_info="examples/gemma3_example/fp32_nodes_gemma3_4b.yaml",  # Change to fp32_nodes_gemma3_27b.yaml for 27B model
     )
 
     ### IMAGE + TEXT ###

pyproject.toml

@@ -56,6 +56,10 @@ dependencies = [
 test = ["pytest","pytest-mock"]
 docs = ["Sphinx==7.1.2","sphinx-rtd-theme==2.0.0","myst-parser==3.0.1","sphinx-multiversion"]
 quality = ["black", "ruff", "hf_doc_builder@git+https://github.com/huggingface/doc-builder.git"]
+
+[tool.setuptools.package-data]
+"QEfficient.transformers.models.gemma3.configs" = ["*.yaml"]
+
 [build-system]
 requires = ["setuptools>=62.0.0"]
 build-backend = "setuptools.build_meta"