zero-gpu: log the parameter device and VRAM so we can prove the weights are on the GPU

BUILD.json

@@ -1,6 +1,6 @@
 {
-  "built_at": "2026-10-29T15:52:38",
-  "commit": "e9f5ce0",
+  "built_at": "2026-10-29T16:33:11",
+  "commit": "8f3c1ee",
   "files": 53,
   "with_store": false,
   "with_card": true

traceforge/space/local_llm.py

@@ -353,6 +353,9 @@ def generate(messages: list[dict], *, max_tokens: int = 512,
         stats["gpu"] = torch.cuda.get_device_name(0)
         stats["vram_gb"] = round(
             torch.cuda.get_device_properties(0).total_memory / 1e9)
+        stats["param_device"] = str(next(model.parameters()).device)
+        stats["vram_used_gb"] = round(torch.cuda.memory_allocated() / 1e9, 1)
+        stats["attn"] = getattr(model.config, "_attn_implementation", "")
     except Exception:
         pass
     LAST_GEN.update(stats)
@@ -442,6 +445,9 @@ class _Handler(BaseHTTPRequestHandler):
               f"{stats.get('seconds')}s {stats.get('tok_s')} tok/s "
               f"capped={stats.get('capped')} think={stats.get('think')} "
               f"skipped={stats.get('think_skipped')} "
+              f"dev={stats.get('param_device')} "
+              f"vram={stats.get('vram_used_gb')}GB "
+              f"attn={stats.get('attn')} "
               f"head={stats.get('raw_head')!r}", flush=True)
         self._send(200, {
             "id": "local-0",