default NPI file added

QEfficient/transformers/models/gemma3/modeling_gemma3.py

@@ -648,6 +648,17 @@ def forward(self, input_ids, position_ids, pixel_values, image_idx, past_key_val
         image_idx = (indices1.max() + 1).unsqueeze(0).unsqueeze(0)
         return outputs.logits, pixel_values, image_idx, outputs.past_key_values
 
+    def get_npi_file(self, model_name: str, **compiler_options):
+        if model_name == "google/gemma-3-4b-it":
+            compiler_options["node_precision_info"] = constants.DEFAULT_GEMMA3_4B_NODE_PRECISION_INFO
+        elif model_name == "google/gemma-3-27b-it":
+            compiler_options["node_precision_info"] = constants.DEFAULT_GEMMA3_27B_NODE_PRECISION_INFO
+        else:
+            raise ValueError(
+                f"For Model {self.pretrained_model_name_or_path} default NPI file is not supported/added. Please use one of the following: google/gemma-3-4b-it, google/gemma-3-27b-it"
+            )
+        return compiler_options
+
     def get_specializations(
         self,
         batch_size: int,

QEfficient/transformers/models/modeling_auto.py

@@ -681,6 +681,9 @@ def compile(
             **compiler_options,
         )
 
+        if hasattr(self.model, "get_npi_file"):
+            compiler_options = self.model.get_npi_file(self.model.pretrained_model_name_or_path, **compiler_options)
+
         custom_io_vision = {}
         kv_cache_dtype = "mxint8" if mxint8_kv_cache else "float16"
         custom_io_vision["pixel_values"] = "float16"
@@ -1030,6 +1033,9 @@ def compile(
             **compiler_options,
         )
 
+        if hasattr(self.model, "get_npi_file"):
+            self.model.get_npi_file(self.pretrained_model_name_or_path)
+
         custom_io = {}
         kv_cache_dtype = "mxint8" if mxint8_kv_cache else "float16"
         # inputs

QEfficient/utils/constants.py

@@ -8,6 +8,8 @@
 import os
 from dataclasses import dataclass
 
+from QEfficient.transformers.models import gemma3 as gemma3
+
 UTILS_DIR = os.path.dirname(os.path.abspath(__file__))
 QEFF_DIR = os.path.dirname(UTILS_DIR)
 ROOT_DIR = os.path.dirname(QEFF_DIR)
@@ -100,7 +102,8 @@ def get_models_dir():
 LLAMA4_MAX_POSITION_EMBEDDINGS = 65536
 
 # Gemma3 Constant
-GEMMA3_MAX_POSITION_EMBEDDINGS = 32768
+DEFAULT_GEMMA3_4B_NODE_PRECISION_INFO = "QEfficient/transformers/models/gemma3/fp32_nodes_gemma3_4b_mm.yaml"
+DEFAULT_GEMMA3_27B_NODE_PRECISION_INFO = "QEfficient/transformers/models/gemma3/fp32_nodes_gemma3_27b_mm.yaml"
 
 
 class Constants:

examples/gemma3_example/gemma3_mm.py

@@ -22,11 +22,11 @@
 # pass HF_TOKEN if gated model
 # For running the model in single QPC approach use kv_offload=False. For Dual QPC approach use kv_offload=True ###
 qeff_model = QEFFAutoModelForImageTextToText.from_pretrained(
-    model_id, config=config, attn_implementation="eager", kv_offload=True
+    model_id, config=config, attn_implementation="eager", kv_offload=False
 )
 
 ### use skip_vision=Ture, if want to run only text, or false ###
-skip_vision = True
+skip_vision = False
 
 if skip_vision:
     ## Only Text ##
@@ -41,7 +41,6 @@
         aic_enable_depth_first=True,
         skip_vision=True,
         mos=1,
-        node_precision_info="examples/gemma3_example/fp32_mm.yaml",
     )
 
     messages = [
@@ -77,7 +76,6 @@
         mxint8_kv_cache=False,
         aic_enable_depth_first=True,
         mos=1,
-        node_precision_info="examples/gemma3_example/fp32_mm.yaml",
     )
 
     ### IMAGE + TEXT ###