Feat: add medgemma

Author: Chang

src/emd/models/engines.py

@@ -90,6 +90,18 @@ class KtransformersEngine(OpenAICompitableEngine):
 )
 
 
+vllm_medgemma082 = VllmEngine(**{
+            "engine_type":EngineType.VLLM,
+            "engine_dockerfile_config": {"VERSION":"v0.8.2"},
+            "engine_cls":"vllm.vllm_backend.VLLMBackend",
+            "base_image_host":"public.ecr.aws",
+            "use_public_ecr":True,
+            "docker_login_region":"us-east-1",
+            "default_cli_args": " --max_num_seq 10 --disable-log-stats"
+}
+)
+
+
 vllm_mistral_small_engine082 = VllmEngine(
     **{
     **vllm_engine064.model_dump(),

src/emd/models/llms/__init__.py

@@ -6,5 +6,6 @@
     deepseek,
     baichuan,
     jina,
-    txgemma
+    txgemma,
+    medgemma
 )

src/emd/models/llms/medgemma.py

@@ -0,0 +1,93 @@
+from ..engines import vllm_medgemma082
+from .. import Model
+from ..frameworks import fastapi_framework
+from ..services import (
+    sagemaker_service,
+    sagemaker_async_service,
+    ecs_service,
+    local_service
+)
+from emd.models.utils.constants import ModelType
+from ..model_series import MEDGEMMA_SERIES
+from ..instances import (
+    g5d2xlarge_instance,
+    g5d4xlarge_instance,
+    g5d8xlarge_instance,
+    g5d12xlarge_instance,
+    g5d16xlarge_instance,
+    g5d24xlarge_instance,
+    g5d48xlarge_instance,
+    g6e2xlarge_instance,
+    local_instance
+)
+from ..utils.constants import ModelFilesDownloadSource
+
+
+Model.register(
+    dict(
+        model_id = "medgemma-27b-text-it",
+        supported_engines=[vllm_medgemma082],
+        supported_instances=[
+            g5d12xlarge_instance,
+            g5d24xlarge_instance,
+            g5d48xlarge_instance,
+            g5d2xlarge_instance,
+            g5d4xlarge_instance,
+            g5d8xlarge_instance,
+            g5d16xlarge_instance,
+            local_instance
+        ],
+        disable_hf_transfer=True,
+        supported_services=[
+            sagemaker_service,
+            sagemaker_async_service,
+            ecs_service,
+            local_service
+        ],
+        supported_frameworks=[
+            fastapi_framework
+        ],
+        allow_china_region=True,
+        huggingface_model_id="google/medgemma_27b_text_it",
+        modelscope_model_id="google/medgemma_27b_text_it",
+        model_files_download_source=ModelFilesDownloadSource.MODELSCOPE,
+        # require_huggingface_token=True,
+        application_scenario="llm for medical text and image comprehension",
+        description="The latest series of medgemma",
+        model_type=ModelType.LLM,
+        model_series=MEDGEMMA_SERIES,
+    )
+)
+
+
+Model.register(
+    dict(
+        model_id = "medgemma-4b-text-it",
+        supported_engines=[vllm_medgemma082],
+        supported_instances=[
+            g5d12xlarge_instance,
+            g5d24xlarge_instance,
+            g5d48xlarge_instance,
+            local_instance
+        ],
+        disable_hf_transfer=True,
+        supported_services=[
+            sagemaker_service,
+            sagemaker_async_service,
+            ecs_service,
+            local_service
+        ],
+        supported_frameworks=[
+            fastapi_framework
+        ],
+        allow_china_region=True,
+        huggingface_model_id="google/medgemma-4b-text-it",
+        modelscope_model_id="google/medgemma-4b-text-it",
+        model_files_download_source=ModelFilesDownloadSource.MODELSCOPE,
+        # require_huggingface_token=True,
+        application_scenario="llm for medical text and image comprehension",
+        description="The latest series of medgemma",
+        model_type=ModelType.LLM,
+        model_series=MEDGEMMA_SERIES,
+    )
+)

src/emd/models/model_series.py

@@ -117,6 +117,11 @@
     reference_link="https://huggingface.co/collections/google/txgemma-release-67dd92e931c857d15e4d1e87"
 )
 
+MEDGEMMA_SERIES = ModelSeries(
+    model_series_name=ModelSeriesType.MEDGEMMA,
+    description="MedGemma is a collection of Gemma 3 variants that are trained for performance on medical text and image comprehension.",
+    reference_link="https://huggingface.co/collections/google/medgemma-release-680aade845f90bec6a3f60c4"
+)
 
 MISTRAL_SERIES = ModelSeries(
     model_series_name=ModelSeriesType.MISTRAL,

src/emd/models/utils/constants.py

@@ -215,6 +215,7 @@ def get_service_quota_code(cls, instance_type: str):
 class ModelSeriesType(ConstantBase):
     GEMMA3 = "gemma3"
     TXGEMMA = "txgemma"
+    MEDGEMMA = "medgemma"
     MISTRAL = "mistral"
     QWEN2D5 = "qwen2.5"
     QWEN3 = "qwen3"