Fix Gemma3 hidden-size handling and add built-in example NIfTI

README.md

@@ -25,6 +25,7 @@ UI updates:
 - EHR table uses the requested clinical field names (including cardiovascular comorbidity columns).
 - Embedded pipeline figure shown at top of interface.
 - MRI input supports both file upload and public URL.
+- MRI input supports file upload, public URL, and a built-in example NIfTI.
   - If Hugging Face upload fails with `AxiosError 403`, use URL mode.
 
 ## Solving the Large Checkpoint Push Issue
@@ -90,6 +91,7 @@ bash scripts/publish_assets_from_local.sh SalmaHassan/HyperClinical-assets ../av
 ## App Entry
 
 - Streamlit app: `src/streamlit_app.py`
+- Built-in sample MRI: `src/examples/example_case.nii.gz`
 
 ## Local Run
 

src/demo_backend/neurofusion/medgemma_encoder.py

@@ -41,6 +41,24 @@ def _normalize_loader_result(result):
     raise ValueError(f"Unexpected loader tuple length: {len(result)}")
 
 
+def _resolve_hidden_size(model) -> int:
+    """Infer LM embedding width across Gemma config variants."""
+    cfg = getattr(model, "config", None)
+    hidden_size = getattr(cfg, "hidden_size", None)
+
+    if hidden_size is None and cfg is not None and hasattr(cfg, "text_config"):
+        hidden_size = getattr(cfg.text_config, "hidden_size", None)
+
+    if hidden_size is None:
+        emb = model.get_input_embeddings()
+        hidden_size = getattr(emb, "embedding_dim", None)
+
+    if hidden_size is None:
+        raise AttributeError("Could not infer hidden_size from model config or input embeddings")
+
+    return int(hidden_size)
+
+
 def _try_load_medgemma(model_name: str, cache_dir: Optional[str] = None):
     """Load MedGemma via ``AutoModelForCausalLM`` (generative Gemma model).
 
@@ -70,7 +88,7 @@ def _try_load_medgemma(model_name: str, cache_dir: Optional[str] = None):
             model_name,
             **load_kwargs,
         )
-        hidden_size = model.config.hidden_size
+        hidden_size = _resolve_hidden_size(model)
         logger.info(f"Loaded MedGemma: {model_name}  (hidden_size={hidden_size})")
         return model, tokenizer, hidden_size, None
     except Exception as e:

src/examples/example_case.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b6c4f69e0bcce412b8aecbb47ade5b6e222723bdf5a98d90cef38f36b5341eb
+size 5301248

src/streamlit_app.py

@@ -29,6 +29,7 @@ from demo_backend.paths import (
 )
 from demo_backend.pipeline import run_full_inference
 
+EXAMPLE_NIFTI = PROJECT_ROOT / "src" / "examples" / "example_case.nii.gz"
 
 st.set_page_config(page_title="HyperClinical Demo", layout="wide")
 
@@ -164,13 +165,24 @@ require_true_hf_embeddings = (
 
 st.subheader("1) MRI Input")
 
-input_mode = st.radio("Select input mode", ["Upload file", "Public URL"], horizontal=True)
+input_mode = st.radio(
+    "Select input mode",
+    ["Upload file", "Public URL", "Built-in example"],
+    horizontal=True,
+)
 uploaded_file = None
 nifti_url = ""
+use_example_nifti = False
 if input_mode == "Upload file":
     uploaded_file = st.file_uploader("Upload a T1 MRI NIfTI (.nii or .nii.gz)", type=["nii", "gz"])
-else:
+elif input_mode == "Public URL":
     nifti_url = st.text_input("Public URL to .nii or .nii.gz")
+else:
+    use_example_nifti = True
+    if EXAMPLE_NIFTI.exists():
+        st.caption(f"Using built-in sample: `{EXAMPLE_NIFTI.name}`")
+    else:
+        st.warning("Built-in sample MRI is not available in this deployment.")
 
 st.subheader("2) Enter EHR Features")
 if "ehr_df" not in st.session_state:
@@ -195,6 +207,9 @@ if st.button("Run Full Inference", type="primary", use_container_width=True):
     if input_mode == "Public URL" and not nifti_url.strip():
         st.error("Please provide a public NIfTI URL.")
         st.stop()
+    if input_mode == "Built-in example" and not EXAMPLE_NIFTI.exists():
+        st.error("Built-in sample MRI is missing. Please use upload or URL mode.")
+        st.stop()
 
     if auto_download_assets:
         with st.spinner("Ensuring required assets are available..."):
@@ -234,13 +249,16 @@ if st.button("Run Full Inference", type="primary", use_container_width=True):
         nifti_path = run_dir / safe_name
         with nifti_path.open("wb") as f:
             f.write(uploaded_file.getbuffer())
-    else:
+    elif input_mode == "Public URL":
         with st.spinner("Downloading NIfTI from URL..."):
             try:
                 nifti_path = _download_nifti_from_url(nifti_url.strip(), run_dir)
             except Exception as exc:
                 st.error(f"Failed to download NIfTI from URL: {exc}")
                 st.stop()
+    else:
+        nifti_path = run_dir / EXAMPLE_NIFTI.name
+        shutil.copy2(EXAMPLE_NIFTI, nifti_path)
 
     ehr_row_internal = _ehr_ui_to_internal_row(ehr_df.iloc[0].to_dict())