Update Deployment_UI_BE.py

Deployment_UI_BE.py

@@ -152,7 +152,6 @@ def _normalize_blob_url(u: str | None) -> str | None:
     u = str(u).strip()
     if u.startswith(("http://", "https://")):
         return u
-    # Treat '/x' or 'x' as local to this app (same origin as FE)
     if u.startswith("/"):
         return f"{_LOCAL_BASE}{u}"
     return f"{_LOCAL_BASE}/{u}"
@@ -170,37 +169,35 @@ def _fetch_url(u: str):
 def _fetch_blob_from_page():
     return _fetch_url(f"{_LOCAL_BASE}/modelblob.json")
 
-@router.post("/api/ingest/from_landing")
-def api_ingest_from_landing(blob_url: str | None = None):
-    """
-    Ingest the deployment blob for downstream use.
-    Mirrors FE behavior: resolve relative paths like '/modelblob.json'
-    against the app origin.
-    """
-    u = _normalize_blob_url(blob_url) or _normalize_blob_url("/modelblob.json")
-    parsed = _fetch_url(u)
-    if not parsed:
-        return JSONResponse({"error": "Blob not available"}, 404)
-    _ingest_blob(parsed, model_id_hint="", container_image_hint="")
-    return JSONResponse({"ok": True, "source": u})
-
-# (Optional compatibility: UI posting to /Deployment_UI; accepts blob_url via query)
-@router.post("/Deployment_UI")
-async def deployment_ui_ingest(request: Request,
-                               model_id: str = Form(""),
-                               container_image: str = Form(""),
-                               blob: str = Form("")):
-    """
-    Legacy entry used by the Deployment UI page.
-    Prefers blob_url from query string; falls back to the modelblob page JSON.
-    """
-    blob_url = request.query_params.get("blob_url")
-    u = _normalize_blob_url(blob_url) if blob_url else _normalize_blob_url("/modelblob.json")
-    parsed = _fetch_url(u)
-    if not parsed:
-        return HTMLResponse("<pre>Missing blob (no /modelblob.json and no blob_url)</pre>", 400)
-    _ingest_blob(parsed, model_id_hint=model_id, container_image_hint=container_image)
-    return RedirectResponse("/Deployment_UI", 303)
+# --- UPDATED: healthRoute + predictRoute caching ---
+def _ingest_blob(parsed: dict, model_id_hint: str = "", container_image_hint: str = ""):
+    if not isinstance(parsed, dict):
+        raise HTTPException(400, "Invalid blob (expected JSON object).")
+
+    _INST.update({
+        "blob": parsed,
+        "model_id": model_id_hint or "",
+        "container_image_hint": container_image_hint or "",
+    })
+
+    c = (((parsed.get("supportedActions") or {}).get("deploy") or {}).get("containerSpec")
+         or parsed.get("container") or {}) or {}
+
+    for k in ("predictRoute", "healthRoute", "readinessRoute", "livenessRoute"):
+        v = c.get(k)
+        if isinstance(v, str) and v.strip():
+            _INST[k] = v.strip()
+
+    image_uri = (c.get("imageUri") or "").strip().lower()
+    pr_hint, hr_hint = _infer_routes_from_image(image_uri)
+
+    # mirror Vertex defaults
+    if not _INST.get("predictRoute"):
+        _INST["predictRoute"] = pr_hint or "/predict"
+    if not _INST.get("healthRoute"):
+        _INST["healthRoute"] = hr_hint or "/health"
+
+    return True
 # ---------------------------------------------------------------------
 # Create instance
 # ---------------------------------------------------------------------
@@ -348,8 +345,8 @@ async def api_create_instance(req: Request):
     _INST["ip"] = _INST.get("ip") or ""
     _INST["port"] = _INST.get("port") or ""
     return JSONResponse(content, r.status_code)
-    # ---------------------------------------------------------------------
-# Poll / read instance status + explicit readiness fields
+# ---------------------------------------------------------------------
+# Poll / read instance status + explicit readiness fields (updated)
 # ---------------------------------------------------------------------
 @router.get("/api/compute/pods/{pod_id}")
 def api_get_instance(pod_id: str = None):
@@ -380,7 +377,6 @@ def api_get_instance(pod_id: str = None):
 
         if ip and isinstance(pm, dict) and pm:
             # choose mapped public port for the declared internal port; else first mapping
-            chosen = None
             if isinstance(declared, int) and str(declared) in pm:
                 chosen = str(pm[str(declared)])
             else:
@@ -398,45 +394,34 @@ def api_get_instance(pod_id: str = None):
             _log_status(f"PORT_MAPPING declared={declared} chosen={chosen} all={pm}")
             _log_status(f"RESOLVED_ENDPOINT base={base}")
 
-            # health + route discovery (no 8080 fallback, no proxy)
-            code, ms, snippet = _probe("GET", f"{base}/health")
-            _log_status(f"HEALTH code={code} ms={ms} body_snippet={snippet}")
-
-            # --- Vertex-mirrored route + final prompt URL log ---
-            pr_explicit = (c.get("predictRoute") or "").strip() if isinstance(c, dict) else ""
-            if pr_explicit:
-                _INST["predictRoute"] = pr_explicit
-                _log_status(f"PREDICT_ROUTE_SET {pr_explicit} (from blob)")
-            elif not _INST.get("predictRoute"):
-                _INST["predictRoute"] = "/predict"
-                _log_status("PREDICT_ROUTE_SET /predict (Vertex default)")
-
-            # If still unset for some reason, try lightweight discovery once
-            if not _INST.get("predictRoute"):
-                route = _discover_route(base)
-                if route:
-                    _INST["predictRoute"] = route
-                    _log_status(f"PREDICT_ROUTE_SET {route} (discovered)")
-                else:
-                    _log_status("PREDICT_ROUTE_SET none")
-
-            # Final prompt URL (prefer IP; else proxy host)
-            pr = _INST.get("predictRoute") or "/predict"
-            if _INST.get("ip") and _INST.get("port"):
-                prompt_url = f"http://{_INST['ip']}:{_INST['port']}{pr}"
+            # --- health-first readiness (mirror Vertex), fallback to predict existence ---
+            hr = (_INST.get("healthRoute") or "/health").strip()
+            pr = (_INST.get("predictRoute") or "/predict").strip()
+
+            code_h, ms_h, snippet_h = _probe("GET", f"{base}{hr}")
+            _log_status(f"HEALTH_PROBE path={hr} code={code_h} ms={ms_h} body_snippet={snippet_h[:120]}")
+
+            if code_h in (200, 204):
+                _INST["status"] = "READY"
             else:
-                proxy_base = f"https://{pid}-{declared}.proxy.runpod.net" if declared else ""
-                prompt_url = f"{proxy_base}{pr}" if proxy_base else ""
-            if prompt_url:
+                code_p, ms_p, _ = _probe("HEAD", f"{base}{pr}")
+                _log_status(f"PREDICT_PROBE path={pr} code={code_p} ms={ms_p}")
+                if code_p in (200, 204, 400, 405):
+                    _INST["status"] = "READY"
+
+            # final prompt URL log (unchanged)
+            if _INST.get("predictRoute"):
+                proute = _INST["predictRoute"]
+                prompt_url = f"http://{_INST['ip']}:{_INST['port']}{proute}"
                 _log_status(f"PROMPT_ENDPOINT {prompt_url}")
 
-    # Always include cached readiness data for the UI
     merged = {**last, "cachedState": {
         "podId": _INST.get("podId"),
         "status": _INST.get("status"),
         "ip": _INST.get("ip"),
         "port": _INST.get("port"),
         "predictRoute": _INST.get("predictRoute"),
+        "healthRoute": _INST.get("healthRoute"),
     }}
     return JSONResponse(merged)
 
@@ -492,7 +477,7 @@ async def api_end_all():
         return JSONResponse(status_code=500, content={"error": f"RunPod delete failed: {e}"})
 
 # ---------------------------------------------------------------------
-# Wait instance
+# Wait instance (updated: same health-first + predict fallback)
 # ---------------------------------------------------------------------
 @router.get("/api/compute/wait_instance")
 def api_wait_instance(pod_id: str = None):
@@ -519,43 +504,35 @@ def api_wait_instance(pod_id: str = None):
         pass
 
     if ip and pm:
-        try:
-            if isinstance(declared, int) and str(declared) in pm:
-                port = str(pm[str(declared)])
-        except Exception:
-            pass
-        if not port and "8080" in pm:
+        if isinstance(declared, int) and str(declared) in pm:
+            port = str(pm[str(declared)])
+        elif "8080" in pm:
             port = str(pm["8080"])
-        elif not port and pm:
+        elif pm:
             port = str(pm[next(iter(pm.keys()))])
 
     if ip and port:
         base = f"http://{ip}:{port}"
         _log_status(f"RESOLVED_IP {base}")
-        code, ms, snippet = _probe("GET", f"{base}/health")
-        _log_status(f"HEALTH code={code} ms={ms} body_snippet={snippet}")
-
-        # --- Vertex-mirrored route + final prompt URL log ---
-        pr_explicit = (c.get("predictRoute") or "").strip() if isinstance(c, dict) else ""
-        if pr_explicit:
-            _INST["predictRoute"] = pr_explicit
-            _log_status(f"PREDICT_ROUTE_SET {pr_explicit} (from blob)")
-        elif not _INST.get("predictRoute"):
-            _INST["predictRoute"] = "/predict"
-            _log_status("PREDICT_ROUTE_SET /predict (Vertex default)")
-
-        if not _INST.get("predictRoute"):
-            route = _discover_route(base)
-            if route:
-                _INST["predictRoute"] = route
-                _log_status(f"PREDICT_ROUTE_SET {route} (discovered)")
-            else:
-                _log_status("PREDICT_ROUTE_SET none")
 
-        pr = _INST.get("predictRoute") or "/predict"
-        prompt_url = f"{base}{pr}"
-        _log_status(f"PROMPT_ENDPOINT {prompt_url}")
+        hr = (_INST.get("healthRoute") or "/health").strip()
+        pr = (_INST.get("predictRoute") or "/predict").strip()
+
+        code_h, ms_h, snippet_h = _probe("GET", f"{base}{hr}")
+        _log_status(f"HEALTH_PROBE path={hr} code={code_h} ms={ms_h} body_snippet={snippet_h[:120]}")
+
+        if code_h in (200, 204):
+            _INST["status"] = "READY"
+        else:
+            code_p, ms_p, _ = _probe("HEAD", f"{base}{pr}")
+            _log_status(f"PREDICT_PROBE path={pr} code={code_p} ms={ms_p}")
+            if code_p in (200, 204, 400, 405):
+                _INST["status"] = "READY"
+
+        if _INST.get("predictRoute"):
+            _log_status(f"PROMPT_ENDPOINT {base}{_INST['predictRoute']}")
 
+    # proxy base (unchanged)
     try:
         cspec = _get_container_spec()
         internal, _ = _get_port_and_proto(cspec)
@@ -567,7 +544,7 @@ def api_wait_instance(pod_id: str = None):
     except Exception:
         pass
 
-    _INST.update({"ip": ip or "", "port": port or "", "status": last.get("desiredStatus", "")})
+    _INST.update({"ip": ip or "", "port": port or ""})
     _save_state()
 
     merged = {
@@ -579,6 +556,7 @@ def api_wait_instance(pod_id: str = None):
             "port": _INST.get("port"),
             "base": _INST.get("base"),
             "predictRoute": _INST.get("predictRoute"),
+            "healthRoute": _INST.get("healthRoute"),
         },
     }
     return JSONResponse(merged)