runpod: per-model boot/load timeouts (big models need longer)

The 300s boot cap was a hard-coded constant; large models need much more time for
the image pull + weight download. Both budgets are now per-model config:

- boot_timeout_s (default 300) — until the pod exposes its port; a stuck machine
  is still dropped fast and the next candidate tried.
- load_timeout_s (default 600) — until vLLM answers /v1/models; raise for big models.

Parsed in RunpodModelConfig (min 30s), used by _provision_one in place of the class
constants (which remain the fallback), whitelisted in api_model_configure, and
exposed on the model page with hints.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R6iugoNeEgt9PkajKrFyq4

codai/admin/routes.py

@@ -2789,7 +2789,7 @@ async def api_model_configure(request: Request, username: str = Depends(require_
         # ints
         for k in ("container_disk_gb", "volume_gb", "port", "ctx", "min_pods",
                   "max_pods", "scale_up_inflight_per_pod", "idle_timeout_s",
-                  "min_workers", "max_workers"):
+                  "boot_timeout_s", "load_timeout_s", "min_workers", "max_workers"):
             v = src.get(k)
             if v not in (None, ""):
                 try:

codai/admin/templates/models.html

@@ -743,6 +743,18 @@ window.__DEFAULT_WHISPER_SERVER_PATH__ = {{ default_whisper_server_path|tojson }
             <div class="form-row" style="margin:0"><label class="form-label">Max pods</label><input type="number" id="cfg-rp-max-pods" class="form-input" min="1" placeholder="1"></div>
             <div class="form-row" style="margin:0"><label class="form-label">Idle timeout (s)</label><input type="number" id="cfg-rp-idle" class="form-input" min="0" placeholder="300"></div>
           </div>
+          <div style="display:grid;grid-template-columns:1fr 1fr;gap:.5rem;margin-top:.5rem">
+            <div class="form-row" style="margin:0">
+              <label class="form-label">Boot timeout (s)</label>
+              <input type="number" id="cfg-rp-boot" class="form-input" min="30" placeholder="300">
+              <span class="form-hint" style="font-size:11px">Wait for the pod to expose its port; a stuck machine is dropped and the next one tried.</span>
+            </div>
+            <div class="form-row" style="margin:0">
+              <label class="form-label">Model load timeout (s)</label>
+              <input type="number" id="cfg-rp-load" class="form-input" min="30" placeholder="600">
+              <span class="form-hint" style="font-size:11px">Wait for vLLM to serve the model — raise this for big models (long image pull + weight download).</span>
+            </div>
+          </div>
           <div style="display:grid;grid-template-columns:2fr 1fr 1fr;gap:.5rem;margin-top:.5rem">
             <div class="form-row" style="margin:0"><label class="form-label">Pod image <span class="muted">(blank = default vLLM)</span></label><input type="text" id="cfg-rp-image" class="form-input" placeholder="vllm/vllm-openai:latest"></div>
             <div class="form-row" style="margin:0"><label class="form-label">Disk (GB)</label><input type="number" id="cfg-rp-disk" class="form-input" min="1" placeholder="40"></div>
@@ -3991,6 +4003,8 @@ function _populateRunpod(s){
   set('cfg-rp-min-pods', rp.min_pods != null ? rp.min_pods : '');
   set('cfg-rp-max-pods', rp.max_pods != null ? rp.max_pods : '');
   set('cfg-rp-idle', rp.idle_timeout_s != null ? rp.idle_timeout_s : '');
+  set('cfg-rp-boot', rp.boot_timeout_s != null ? rp.boot_timeout_s : '');
+  set('cfg-rp-load', rp.load_timeout_s != null ? rp.load_timeout_s : '');
   set('cfg-rp-image', rp.image || '');
   set('cfg-rp-disk', rp.container_disk_gb != null ? rp.container_disk_gb : '');
   set('cfg-rp-port', rp.port != null ? rp.port : '');
@@ -4030,6 +4044,8 @@ function _collectRunpod(){
     min_pods:           intOrU('cfg-rp-min-pods'),
     max_pods:           intOrU('cfg-rp-max-pods'),
     idle_timeout_s:     intOrU('cfg-rp-idle'),
+    boot_timeout_s:     intOrU('cfg-rp-boot'),
+    load_timeout_s:     intOrU('cfg-rp-load'),
     image:              (val('cfg-rp-image') || '').trim(),
     container_disk_gb:  intOrU('cfg-rp-disk'),
     port:               intOrU('cfg-rp-port'),

codai/api/runpod_worker.py

@@ -56,6 +56,9 @@ class RunpodModelConfig:
     max_pods: int = 1
     scale_up_inflight_per_pod: int = 4
     idle_timeout_s: int = 300                # destroy a pod this long after its last request
+    # Boot budgets — bigger models need longer (image pull + weight download).
+    boot_timeout_s: int = 300                # until the pod exposes its port
+    load_timeout_s: int = 600                # until vLLM answers /v1/models
     # --- serverless ---
     endpoint_id: str = ""                    # reference an existing serverless endpoint
     min_workers: int = 0
@@ -130,6 +133,8 @@ def parse_model_runpod(block: Optional[dict]) -> RunpodModelConfig:
     cfg.scale_up_inflight_per_pod = max(1, _as_int(b.get("scale_up_inflight_per_pod"),
                                                    cfg.scale_up_inflight_per_pod))
     cfg.idle_timeout_s = max(0, _as_int(b.get("idle_timeout_s"), cfg.idle_timeout_s))
+    cfg.boot_timeout_s = max(30, _as_int(b.get("boot_timeout_s"), cfg.boot_timeout_s))
+    cfg.load_timeout_s = max(30, _as_int(b.get("load_timeout_s"), cfg.load_timeout_s))
     cfg.endpoint_id = (b.get("endpoint_id") or "").strip()
     cfg.min_workers = max(0, _as_int(b.get("min_workers"), cfg.min_workers))
     cfg.max_workers = max(1, _as_int(b.get("max_workers"), cfg.max_workers))
@@ -429,17 +434,19 @@ class RunpodPodPool:
             print(f"[runpod] pod {pod_id} created for {self.model_key!r} "
                   f"({sel['display_name']} {sel['cloud_type']}) — booting; logs: {console}",
                   flush=True)
+            boot_to = float(self.mcfg.boot_timeout_s or self.PORT_TIMEOUT_S)
+            load_to = float(self.mcfg.load_timeout_s or self.HEALTH_TIMEOUT_S)
             try:
-                url = client.wait_ready(pod_id, port, ready_timeout=self.PORT_TIMEOUT_S)
+                url = client.wait_ready(pod_id, port, ready_timeout=boot_to)
                 # Wait for the OpenAI server inside the pod (image pull + model load).
-                deadline = time.time() + self.HEALTH_TIMEOUT_S
+                deadline = time.time() + load_to
                 while time.time() < deadline:
                     if _pod_health_ok(url):
                         break
                     time.sleep(5)
                 else:
                     raise RunpodError(f"pod {pod_id} OpenAI server (vLLM) did not answer "
-                                      f"/v1/models within {int(self.HEALTH_TIMEOUT_S)}s")
+                                      f"/v1/models within {int(load_to)}s")
             except Exception as exc:
                 # Self-diagnose: pull the vLLM/container log before tearing down, so
                 # the cause (OOM / bad args / model gate / stuck machine) is in the log.