feat(model): support FunctionGemma-270m-it for local inference (#1279)

FunctionGemma is a specialized variant of Gemma 3 270M optimized for
function calling.
However, it introduces a unique formatting convention for tool
definitions and calls that is incompatible with the existing
`GemmaHandler`.

This commit adds local support for FunctionGemma-270m-it by:
- Added `FunctionGemmaHandler` for local inference.
- Implemented specific prompt formatting that matches the official
documentation.

References:
-
https://ai.google.dev/gemma/docs/functiongemma/formatting-and-best-practices#base-prompt-structure
-
https://ai.google.dev/gemma/docs/functiongemma/model_card#benchmark-results

Co-authored-by: Huanzhi Mao <huanzhimao@gmail.com>

Author: FilipFan

berkeley-function-call-leaderboard/SUPPORTED_MODELS.md

@@ -50,6 +50,7 @@ For model names containing `{...}`, multiple versions are available. For example
 | Gemini-3-Pro-Preview                   | Function Calling | Google         | gemini-3-pro-preview-FC                                     |
 | Gemini-3-Pro-Preview                   | Prompt           | Google         | gemini-3-pro-preview                                        |
 | Gemma-3-{1b,4b,12b,27b}-it             | Prompt           | Self-hosted 💻 | google/gemma-3-{1b,4b,12b,27b}-it                           |
+| FunctionGemma-270m-it                  | Function Calling | Self-hosted 💻 | google/functiongemma-270m-it-FC                             |
 | GLM-4-9b-Chat                          | Function Calling | Self-hosted 💻 | THUDM/glm-4-9b-chat                                         |
 | GLM-4.5                                | Function Calling | Zhipu AI       | glm-4.5-FC                                                  |
 | GLM-4.5-Air                            | Function Calling | Zhipu AI       | glm-4.5-air-FC                                              |

berkeley-function-call-leaderboard/bfcl_eval/constants/model_config.py

@@ -39,6 +39,7 @@
 )
 from bfcl_eval.model_handler.local_inference.falcon_fc import Falcon3FCHandler
 from bfcl_eval.model_handler.local_inference.gemma import GemmaHandler
+from bfcl_eval.model_handler.local_inference.functiongemma import FunctionGemmaHandler
 from bfcl_eval.model_handler.local_inference.glm import GLMHandler
 from bfcl_eval.model_handler.local_inference.granite import (
     GraniteFunctionCallingHandler,
@@ -1247,6 +1248,18 @@ class ModelConfig:
         is_fc_model=False,
         underscore_to_dot=False,
     ),
+    "google/functiongemma-270m-it-FC": ModelConfig(
+        model_name="google/functiongemma-270m-it",
+        display_name="FunctionGemma-270m-it (FC)",
+        url="https://ai.google.dev/gemma/docs/functiongemma",
+        org="Google",
+        license="gemma-terms-of-use",
+        model_handler=FunctionGemmaHandler,
+        input_price=None,
+        output_price=None,
+        is_fc_model=True,
+        underscore_to_dot=False,
+    ),
     "meta-llama/Llama-3.1-8B-Instruct-FC": ModelConfig(
         model_name="meta-llama/Llama-3.1-8B-Instruct",
         display_name="Llama-3.1-8B-Instruct (FC)",

berkeley-function-call-leaderboard/bfcl_eval/constants/supported_models.py

@@ -107,6 +107,7 @@
     "google/gemma-3-4b-it",
     "google/gemma-3-12b-it",
     "google/gemma-3-27b-it",
+    "google/functiongemma-270m-it-FC",
     "meta-llama/Llama-3.1-8B-Instruct-FC",
     "meta-llama/Llama-3.1-8B-Instruct",
     "meta-llama/Llama-3.1-70B-Instruct-FC",