[INC-302] Add Gemma 4 to inference

inference_models/examples/gemma4/count_backpacks.py

@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+"""End-to-end Gemma 4 example using ``inference_models.AutoModel`` (no CLI arguments).
+
+Downloads a public sample image, loads a hosted Gemma 4 checkpoint via Roboflow, and
+asks a focused counting question.
+
+Run from the ``inference_models`` package root::
+
+    uv run python examples/gemma4/run_gemma4_local.py
+
+Optional: set ``GEMMA4_MODEL_ID`` to override the default Roboflow registry id.
+"""
+
+from __future__ import annotations
+
+import io
+import os
+import sys
+
+import numpy as np
+import requests
+from PIL import Image
+
+from inference_models import AutoModel
+from inference_models.configuration import DEFAULT_DEVICE
+
+# Same image used in repo docs (e.g. workflows benchmarks).
+IMAGE_URL = "https://media.roboflow.com/inference/people-walking.jpg"
+
+# Roboflow registry id (must match a registered Gemma 4 package).
+DEFAULT_MODEL_ID = "gemma-4-e2b-it"
+
+SYSTEM_PROMPT = (
+    "You are a precise vision assistant. When asked about people or objects in a scene, "
+    "base your answer only on what is clearly visible. If you are uncertain, say so. "
+    "For counting questions, give a single best estimate and briefly note any ambiguity "
+    "(e.g. partially occluded figures or unclear backpacks)."
+)
+
+USER_PROMPT = (
+    "How many people in this image are clearly wearing a backpack? "
+    "Answer with a number first, then one short sentence explaining what you counted."
+)
+
+
+def _build_prompt(user: str, system: str) -> str:
+    return f"{user}<system_prompt>{system}"
+
+
+def _load_image_rgb(url: str) -> np.ndarray:
+    response = requests.get(url, timeout=60)
+    response.raise_for_status()
+    image = Image.open(io.BytesIO(response.content)).convert("RGB")
+    return np.array(image)
+
+
+def main() -> None:
+    load_kw = {
+        "device": DEFAULT_DEVICE,
+        "backend": "hugging-face",
+    }
+    print(f"Loading hosted model {DEFAULT_MODEL_ID!r} …")
+    model = AutoModel.from_pretrained(DEFAULT_MODEL_ID, **load_kw)
+
+    print(f"Fetching image {IMAGE_URL!r} …")
+    image_rgb = _load_image_rgb(IMAGE_URL)
+    prompt = _build_prompt(USER_PROMPT, SYSTEM_PROMPT)
+
+    print("Running inference …")
+    outputs = model.prompt(
+        images=image_rgb,
+        prompt=prompt,
+        input_color_format="rgb",
+        max_new_tokens=256,
+        do_sample=False,
+    )
+    print("---")
+    print(outputs[0] if outputs else outputs)
+
+
+if __name__ == "__main__":
+    main()

inference_models/examples/gemma4/model_config.example.json

@@ -0,0 +1,5 @@
+{
+  "model_architecture": "gemma-4-e2b-it",
+  "task_type": "vlm",
+  "backend_type": "hugging-face"
+}

inference_models/inference_models/configuration.py

@@ -234,6 +234,35 @@
     variable_name="INFERENCE_MODELS_QWEN25_VL_DEFAULT_SKIP_SPECIAL_TOKENS",
     default=True,
 )
+INFERENCE_MODELS_GEMMA4_DEFAULT_MAX_NEW_TOKENS = get_integer_from_env(
+    variable_name="INFERENCE_MODELS_GEMMA4_DEFAULT_MAX_NEW_TOKENS",
+    default=512,
+)
+INFERENCE_MODELS_GEMMA4_DEFAULT_DO_SAMPLE = get_boolean_from_env(
+    variable_name="INFERENCE_MODELS_GEMMA4_DEFAULT_DO_SAMPLE",
+    default=INFERENCE_MODELS_DEFAULT_DO_SAMPLE,
+)
+INFERENCE_MODELS_GEMMA4_DEFAULT_ENABLE_THINKING = get_boolean_from_env(
+    variable_name="INFERENCE_MODELS_GEMMA4_DEFAULT_ENABLE_THINKING",
+    default=False,
+)
+INFERENCE_MODELS_GEMMA4_DEFAULT_SKIP_SPECIAL_TOKENS = get_boolean_from_env(
+    variable_name="INFERENCE_MODELS_GEMMA4_DEFAULT_SKIP_SPECIAL_TOKENS",
+    default=True,
+)
+# Official Gemma 4 sampling recommendations when ``do_sample`` is True (HF model cards).
+INFERENCE_MODELS_GEMMA4_DEFAULT_TEMPERATURE = get_float_from_env(
+    variable_name="INFERENCE_MODELS_GEMMA4_DEFAULT_TEMPERATURE",
+    default=1.0,
+)
+INFERENCE_MODELS_GEMMA4_DEFAULT_TOP_P = get_float_from_env(
+    variable_name="INFERENCE_MODELS_GEMMA4_DEFAULT_TOP_P",
+    default=0.95,
+)
+INFERENCE_MODELS_GEMMA4_DEFAULT_TOP_K = get_integer_from_env(
+    variable_name="INFERENCE_MODELS_GEMMA4_DEFAULT_TOP_K",
+    default=64,
+)
 INFERENCE_MODELS_RESNET_DEFAULT_CONFIDENCE = get_float_from_env(
     variable_name="INFERENCE_MODELS_RESNET_DEFAULT_CONFIDENCE",
     default=INFERENCE_MODELS_DEFAULT_CONFIDENCE,

inference_models/inference_models/models/auto_loaders/models_registry.py

@@ -286,6 +286,22 @@ class RegistryEntry:
         module_name="inference_models.models.qwen3_5.qwen3_5_hf",
         class_name="Qwen35HF",
     ),
+    ("gemma-4-e2b-it", VLM_TASK, BackendType.HF): LazyClass(
+        module_name="inference_models.models.gemma4.gemma4_hf",
+        class_name="Gemma4HF",
+    ),
+    ("gemma-4-e4b-it", VLM_TASK, BackendType.HF): LazyClass(
+        module_name="inference_models.models.gemma4.gemma4_hf",
+        class_name="Gemma4HF",
+    ),
+    ("gemma-4-31b-it", VLM_TASK, BackendType.HF): LazyClass(
+        module_name="inference_models.models.gemma4.gemma4_hf",
+        class_name="Gemma4HF",
+    ),
+    ("gemma-4-26b-a4b-it", VLM_TASK, BackendType.HF): LazyClass(
+        module_name="inference_models.models.gemma4.gemma4_hf",
+        class_name="Gemma4HF",
+    ),
     ("florence-2", VLM_TASK, BackendType.HF): LazyClass(
         module_name="inference_models.models.florence2.florence2_hf",
         class_name="Florence2HF",

inference_models/inference_models/models/gemma4/__init__.py

@@ -0,0 +1 @@
+# Gemma 4 multimodal (Hugging Face) implementations