engines card: canonical per-model memory info; cap GME vision token budget

The tooltip matched loaded_info entries (raw engine keys) against
canonicalized display names client-side and mostly missed — footprints
now canonicalize server-side with the same mapping as the model list,
keyed by canonical id, so every loaded model shows its VRAM (+RAM when
offloading).

llama-vl: bound mtmd image_max_tokens (config `image_max_tokens`,
default 1024) — large photos otherwise expand to thousands of vision
tokens whose transient compute buffer evicts co-resident models on a
small card.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014S8VtAvG499SsCbeESRK7V

codai/__init__.py

@@ -16,7 +16,7 @@
 
 # Canonical product version for CoderAI — single source of truth. Both the API
 # metadata and the admin web UI read from here.
-__version__ = "0.1.45"
+__version__ = "0.1.46"
 
 # Configure the CUDA caching allocator BEFORE torch is imported anywhere.
 # expandable_segments lets the allocator return freed pages to the driver even

codai/admin/templates/tasks.html

@@ -317,15 +317,9 @@ async function loadEngines(){
     const models = loaded.length;
     // Tooltip: one line per model with its memory footprint — VRAM, plus the
     // host-RAM slice separately when the model is CPU-offloading.
-    const info = e.loaded_info||[];
-    const infoByBase = {};
-    info.forEach(m => {
-      const b = (m.model||'').split('/').pop().replace(/\.gguf$/i,'');
-      infoByBase[b] = m; infoByBase[m.model] = m;
-    });
+    const info = e.loaded_info||{};   // canonical model id -> {vram_gb, ram_gb}
     const mtip = models ? esc(loaded.map(name => {
-      const b = (name||'').split('/').pop().replace(/\.gguf$/i,'');
-      const m = infoByBase[name] || infoByBase[b];
+      const m = info[name];
       if (!m || (!m.vram_gb && !m.ram_gb)) return name;
       let s = `${name} — VRAM ${m.vram_gb} GB`;
       if (m.ram_gb) s += ` · RAM ${m.ram_gb} GB (offload)`;

codai/api/embeddings.py

@@ -290,6 +290,16 @@ def _load_embedding_model(model_name: str, device: str, model_config: dict = Non
                 mp.use_gpu = (n_gpu_layers != 0)
                 mp.print_timings = False
                 mp.n_threads = 8
+                # Bound the vision token budget: qwen2-vl dynamic resolution
+                # turns a large photo into thousands of image tokens, and the
+                # vision tower's transient compute buffer then spikes VRAM —
+                # which evicts co-resident models mid-encode on a small card.
+                try:
+                    mp.image_max_tokens = int(
+                        cfg.get('image_max_tokens')
+                        or raw.get('image_max_tokens') or 1024)
+                except Exception:
+                    pass
                 mctx = _M.mtmd_init_from_file(
                     str(_mmproj).encode(), llm._model.model, mp)
                 if not mctx:

codai/frontproxy/app.py

@@ -1100,6 +1100,21 @@ class FrontProxy:
             seen.setdefault(canon, None)
         return list(seen.keys())
 
+    def _canonical_info(self, loaded_info) -> dict:
+        """loaded_info entries keyed by CANONICAL model id (same mapping as
+        _canonical_loaded), deduping the several key forms one resident model
+        reports under; keeps the largest footprint among duplicates."""
+        out: dict = {}
+        for m in (loaded_info or []):
+            if not isinstance(m, dict):
+                continue
+            k = str(m.get("model") or "")
+            canon = (self._model_info(k).get("model_id") or k)
+            prev = out.get(canon)
+            if prev is None or (m.get("vram_gb") or 0) > (prev.get("vram_gb") or 0):
+                out[canon] = {**m, "model": canon}
+        return out
+
     def engines_list(self) -> list:
         out = []
         for e in self.registry.all():
@@ -1116,7 +1131,8 @@ class FrontProxy:
                         "thermal_paused": bool(getattr(e, "therm_paused", False)),
                         "thermal_frozen": bool(getattr(e, "therm_sigstopped", False)),
                         "loaded_models": self._canonical_loaded(e.loaded_models),
-                        "loaded_info": list(getattr(e, "loaded_info", []) or []),
+                        "loaded_info": self._canonical_info(
+                            getattr(e, "loaded_info", None)),
                         "inflight": int(getattr(e, "inflight", 0) or 0),
                         "processing": (int(getattr(e, "inflight", 0) or 0) > 0
                                        or bool(getattr(e, "loading", None))),