Upload 13 files

HF_V6_2_CHECKLIST.md

@@ -0,0 +1,25 @@
+# V6.2 deployment checklist
+
+1. Upload all files to the root of the Gradio Space. Do not keep them inside a subfolder.
+2. Upload `best_efficientnet.pth` separately (the ZIP intentionally does not contain the model weights).
+3. Use ZeroGPU hardware.
+4. Keep these secrets:
+   - `SUPABASE_URL`
+   - `SUPABASE_SERVICE_ROLE_KEY`
+5. Variables:
+   - `SUPABASE_STORAGE_BUCKET=xray-images`
+   - `SUPABASE_STORAGE_PUBLIC=true`
+   - `JANUS_ENABLED=true`
+   - `JANUS_BILINGUAL=true`
+   - `JANUS_MODEL_ID=ZrH42/Janus-Pro-CXR-Final`
+   - `MODEL_PATH=/home/oai/share/best_efficientnet.pth` only if that is the actual path; otherwise omit it because the default is the Space root.
+6. In Supabase, create/verify the `xray-images` bucket once. The server no longer tries to create it during inference.
+7. The server stores Janus/integrity/safety inside `medical_report_json`; it does not send non-existent columns to `analyses`.
+8. After build, open `/` and test `/health`, then `/gpu_health`, then Analyze.
+9. A successful AI test should show:
+   - `janus.enabled: true`
+   - non-empty `janus.raw_text`
+   - English report source from Janus
+   - Arabic verified or explicitly marked for manual review
+   - `database.saved: true`
+   - `database.storage_saved: true`

README.md

@@ -1,15 +1,13 @@
 ---
-title: FeatureX Clinical AI - ZeroGPU API
-author: Laith Nadeem Alabsi
+title: FeatureX Clinical AI
 emoji: 🩻
 colorFrom: blue
 colorTo: indigo
 sdk: gradio
-sdk_version: 6.27.0
-python_version: "3.12.12"
+sdk_version: 5.29.1
+python_version: "3.12"
 app_file: app.py
-preload_from_hub:
-  - ZrH42/Janus-Pro-CXR-Final
+suggested_hardware: cpu-basic
 ---
 
 # FeatureX Clinical AI — Hugging Face ZeroGPU

SHA256SUMS.txt

@@ -0,0 +1,13 @@
+2c7ee9e77fa13e08f3e62a64fb400d3d7eeb1cacfc409482b6940d94c386da71  DEPLOY_V6_GRADIO6.md
+5ff6aa4fba6c3b7474f32d9fcefdf060022cf03d7c025fbeb1fb08f39d3783ab  HF_V6_2_CHECKLIST.md
+e320901dec53dbb74dd0ffbbb631a98c036e9dab80f99e110c23e2da7f4c74cd  README.md
+d7632fa87691d1fbb05c64a58a31c2dcdddcbad7239f8b1136f4dc5e3b7c5bfb  app.py
+bb2a915c4e372a456d039dc7ec7350174b5032252ef610f17e8a602bc88976c6  arabic_translation.py
+82800f6795317e1d41a11c40100cc0e8aa2d4a73cf4ba4c99b73251111601882  database.py
+d5a0ea93b03a52eeb8fa12cc421be19626b0e1e5f3d5d56c67028ca56da0a344  env.example
+3b27bb81bf2b11b7d4b266b86a4426ec3469af0f3235df0eab1ae8859fd0f98d  flask_app_reference.py
+acb0d739947c5aeeff0d91c7add3d041f40188a5ea4a82994b978bfd9621f91e  inference.py
+2d76e40f526f982847334fbf71824a238ccabb704098933f6ddb32d1c90ee762  janus_engine.py
+fb3695147bc149917ab8264f4b0d6adc5115f1758801fc25db5b5a44f8f0e97e  requirements.txt
+b03844e23a2073511d00b05ff42045859e9d65fbe505b643dd7e31f148986410  supabase_002_bilingual_analyses.sql
+f7bec1683c145d65f07d4f30040ab58e29b7ff32fd16b8c4b95c4e0d790f0dfe  test_gradio_api.py

app.py

@@ -21,9 +21,30 @@ logging.basicConfig(
 )
 logger = logging.getLogger("featurex-zerogpu-server")
 
-SERVER_VERSION = "3.1.0-hf-zerogpu-gradio6"
+SERVER_VERSION = "3.2.0-hf-zerogpu-v6.2-janus49"
 MAX_UPLOAD = 15 * 1024 * 1024
 
+def _runtime_info():
+    info = {}
+    try:
+        import torch
+        info["torch"] = torch.__version__
+        info["cuda_available"] = bool(torch.cuda.is_available())
+        info["cuda_device"] = torch.cuda.get_device_name(0) if torch.cuda.is_available() else None
+    except Exception as exc:
+        info["torch_error"] = str(exc)
+    try:
+        import transformers
+        info["transformers"] = transformers.__version__
+    except Exception as exc:
+        info["transformers_error"] = str(exc)
+    info["janus_enabled"] = os.environ.get("JANUS_ENABLED", "true")
+    info["janus_model_id"] = os.environ.get("JANUS_MODEL_ID", "ZrH42/Janus-Pro-CXR-Final")
+    info["model_path"] = os.environ.get("MODEL_PATH", os.path.join(os.path.dirname(os.path.abspath(__file__)), "best_efficientnet.pth"))
+    info["model_file_exists"] = os.path.isfile(info["model_path"])
+    return info
+
+
 
 def _json(payload: Any) -> str:
     return json.dumps(payload, ensure_ascii=False, separators=(",", ":"))
@@ -302,6 +323,7 @@ def gpu_health() -> str:
             "server_version": SERVER_VERSION,
             "transport": "Gradio Server + ZeroGPU",
             "database": database.config_info(),
+            "runtime": _runtime_info(),
             "api": {"analyze": True, "health": True, "classes": True},
         }
     )
@@ -316,6 +338,7 @@ def health() -> str:
             "transport": "Gradio Server + ZeroGPU",
             "zero_gpu": True,
             "database": database.config_info(),
+            "runtime": _runtime_info(),
         }
     )
     return _json(info)
@@ -349,6 +372,7 @@ def _health_endpoint() -> str:
             "transport": "Gradio Blocks + ZeroGPU",
             "zero_gpu": True,
             "database": database.config_info(),
+            "runtime": _runtime_info(),
             "api": {"analyze": True, "health": True, "gpu_health": True, "classes": True},
         }
     )

database.py

@@ -15,7 +15,10 @@ logger = logging.getLogger("clinical-ai-db")
 
 SUPABASE_URL = os.environ.get("SUPABASE_URL", "").rstrip("/")
 SUPABASE_SERVICE_ROLE_KEY = os.environ.get("SUPABASE_SERVICE_ROLE_KEY", "")
-SUPABASE_BUCKET = os.environ.get("SUPABASE_STORAGE_BUCKET", "xray-images")
+_RAW_BUCKET = os.environ.get("SUPABASE_STORAGE_BUCKET", "xray-images").strip()
+# The deployed SQL and successful Kaggle server use the hyphenated bucket.
+# Accept the old xray_images spelling to avoid breaking an existing secret.
+SUPABASE_BUCKET = "xray-images" if _RAW_BUCKET == "xray_images" else (_RAW_BUCKET or "xray-images")
 PUBLIC_STORAGE = (
     os.environ.get("SUPABASE_STORAGE_PUBLIC", "true").strip().lower() == "true"
 )
@@ -41,35 +44,33 @@ def config_info() -> Dict[str, Any]:
 
 
 def _ensure_bucket() -> None:
-    """Ensure the configured Storage bucket exists when using the trusted server key."""
+    """Verify that the configured Storage bucket already exists.
+
+    The previous HF build tried to create the bucket on every request. On
+    Supabase projects with Storage RLS this can fail with:
+    "new row violates row-level security policy".
+    Bucket administration belongs in Supabase SQL/dashboard; the server only
+    verifies the bucket and uploads with the service-role key.
+    """
     if not enabled():
         raise RuntimeError("Supabase is not configured on the server.")
-    url = f"{SUPABASE_URL}/storage/v1/bucket"
+
+    url = f"{SUPABASE_URL}/storage/v1/bucket/{SUPABASE_BUCKET}"
     try:
-        response = requests.post(
+        response = requests.get(
             url,
             headers=_headers(),
-            json={"id": SUPABASE_BUCKET, "name": SUPABASE_BUCKET, "public": PUBLIC_STORAGE},
             timeout=REQUEST_TIMEOUT,
         )
-        if response.status_code in (200, 201, 409):
-            return
-        # Some projects already have the bucket but reject create; verify it exists.
-        if response.status_code >= 300:
-            check = requests.get(
-                f"{url}/{SUPABASE_BUCKET}",
-                headers=_headers(),
-                timeout=REQUEST_TIMEOUT,
-            )
-            if check.status_code == 200:
-                return
-            raise RuntimeError(
-                f"Storage bucket '{SUPABASE_BUCKET}' is unavailable: "
-                f"{response.status_code} {response.text[:800]}"
-            )
     except requests.RequestException as exc:
         raise RuntimeError(f"Storage bucket check failed: {exc}") from exc
 
+    if response.status_code != 200:
+        body = _safe_response_text(response, 800)
+        raise RuntimeError(
+            f"Storage bucket '{SUPABASE_BUCKET}' is unavailable: "
+            f"{response.status_code} {body}"
+        )
 
 def _headers(content_type: str = "application/json") -> Dict[str, str]:
     if not enabled():
@@ -422,17 +423,16 @@ def save_analysis(
         "threshold_url": visual_urls.get("threshold"),
         "comparison_url": visual_urls.get("comparison"),
         "medical_report_json": json.dumps(
-            selected_report,
+            {
+                **selected_report,
+                "report_integrity": result.get("report_integrity") or {},
+                "janus": result.get("janus") or {},
+                "safety": result.get("safety") or {},
+            },
             ensure_ascii=False,
         ),
         "report_ar": reports_ar,
         "report_en": reports_en,
-        "janus_raw_text": (result.get("janus") or {}).get("raw_text") or selected_report.get("janus_raw_text") or "",
-        "janus_model": (result.get("janus") or {}).get("model") or selected_report.get("report_model") or "",
-        "report_source": selected_report.get("report_source") or "",
-        "translation_methods": (result.get("report_integrity") or {}).get("translation_methods") or [],
-        "report_integrity": result.get("report_integrity") or {},
-        "safety": result.get("safety") or {},
         "patient_data": _localized_patient(patient),
         "predictions_data": predictions,
         "quality_metrics": quality_metrics,
@@ -440,7 +440,6 @@ def save_analysis(
         "model_info": model_info,
         "visualization_urls": visual_urls,
         "language_requested": requested_language,
-        "language": requested_language,
         "original_filename": original_filename,
         "image_sha256": image_sha256,
         "server_version": model_info.get("version", "unknown"),

janus_engine.py

@@ -10,13 +10,22 @@ from typing import Dict, List
 
 import torch
 
+def _log_runtime_versions():
+    try:
+        import transformers
+        logger.info("Transformers version: %s", transformers.__version__)
+    except Exception as exc:
+        logger.warning("Could not read Transformers version: %s", exc)
+    logger.info("Torch version: %s | CUDA available: %s", torch.__version__, torch.cuda.is_available())
+
+
 logger = logging.getLogger("featurex-janus")
 
 MODEL_ID = os.environ.get("JANUS_MODEL_ID", "ZrH42/Janus-Pro-CXR-Final")
 REPO_URL = "https://github.com/ZrH42/Janus-Pro-CXR.git"
 REPO_DIR = Path(os.environ.get("JANUS_REPO_DIR", "/tmp/Janus-Pro-CXR"))
 JANUS_ENABLED = os.environ.get("JANUS_ENABLED", "true").strip().lower() in {"1", "true", "yes", "on"}
-BILINGUAL = os.environ.get("JANUS_BILINGUAL", "true").strip().lower() in {"1", "true", "yes", "on"}
+BILINGUAL = os.environ.get("JANUS_BILINGUAL", "false").strip().lower() in {"1", "true", "yes", "on"}
 
 _PROCESSOR = None
 _MODEL = None
@@ -59,11 +68,12 @@ def _initialize():
     global _PROCESSOR, _MODEL, _INITIALIZED
     if _INITIALIZED:
         return
+    _log_runtime_versions()
     if not JANUS_ENABLED:
         _INITIALIZED = True
         return
     if not torch.cuda.is_available():
-        raise RuntimeError("Janus requires CUDA on the Kaggle GPU server.")
+        raise RuntimeError("Janus requires CUDA/ZeroGPU. The analyze function must execute on a ZeroGPU Space.")
 
     _clone_repo()
     _patch_compatibility()

requirements.txt

@@ -1,21 +1,19 @@
-# Hugging Face ZeroGPU provides the platform-managed Gradio, spaces and
-# huggingface-hub packages. Do not pin those packages here.
-#
-# The current HF Gradio 6 runtime requires huggingface-hub 1.x, so the old
-# Transformers 4.49.0 pin is intentionally removed. Transformers 5.4.0 is
-# used here as the compatibility bridge for the current Hub/Gradio stack.
-# Janus-Pro-CXR is still loaded through its custom remote-code implementation.
-
+# FeatureX Clinical AI — Hugging Face ZeroGPU
+# Gradio 5.29.1 is intentionally pinned because Janus-Pro-CXR was validated
+# with the Transformers 4.x stack; Gradio 6 + Hub 1.x forced an incompatible
+# Transformers 5.x bridge in the previous build.
+gradio==5.29.1
 torch==2.10.0
 torchvision==0.25.0
-transformers==5.4.0
-accelerate>=1.2.1
+transformers==4.49.0
+accelerate==1.2.1
+tokenizers==0.21.0
+huggingface-hub<1.0
 attrdict==2.0.1
 sentencepiece
 einops
 timm>=1.0.0
 safetensors>=0.5.0
-
 opencv-python-headless
 Pillow
 numpy==1.26.4