zero-gpu: report tokens/s so the rewrite budget can be sized to the slice

BUILD.json

@@ -1,5 +1,5 @@
 {
-  "built_at": "2026-10-29T14:00:39",
+  "built_at": "2026-10-29T14:15:09",
   "commit": "c778551",
   "files": 53,
   "with_store": false,

traceforge/space/local_llm.py

@@ -66,6 +66,9 @@ LAST_RESPONSE: Optional[str] = None
 # needed.  A request that arrives before the checkpoint is on disk must fail
 # fast: downloading 48 GB inside a 120 s GPU slice would burn the whole slice.
 PREFETCH: dict = {"state": "idle", "error": "", "source": ""}
+# Throughput of the last generation: the GPU slice is 120 s, so tokens/s
+# decides whether a rewrite can finish inside one call.
+LAST_GEN: dict = {}
 
 # A Space can carry HTTP(S)_PROXY for outbound traffic; urllib would then
 # route the loopback call through it and the local server would never see
@@ -303,6 +306,7 @@ def generate(messages: list[dict], *, max_tokens: int = 512,
     enc = tok(prompt, return_tensors="pt", truncation=True,
               max_length=LOCAL_MAX_INPUT_TOKENS).to(LOCAL_DEVICE)
     n = max(1, min(int(max_tokens or 512), LOCAL_MAX_NEW_TOKENS))
+    t0 = time.time()
     with torch.inference_mode():
         out = model.generate(
             **enc,
@@ -315,6 +319,15 @@ def generate(messages: list[dict], *, max_tokens: int = 512,
             repetition_penalty=1.05,
             pad_token_id=tok.eos_token_id,
         )
+    elapsed = time.time() - t0
+    new_tokens = int(out.shape[1] - enc["input_ids"].shape[1])
+    LAST_GEN.update(
+        prompt_tokens=int(enc["input_ids"].shape[1]),
+        new_tokens=new_tokens,
+        seconds=round(elapsed, 1),
+        tok_s=round(new_tokens / elapsed, 1) if elapsed > 0 else 0.0,
+        capped=new_tokens >= n,
+    )
     text = tok.decode(out[0][enc["input_ids"].shape[1]:],
                       skip_special_tokens=True)
     return _strip_reasoning(text)
@@ -583,6 +596,7 @@ def health() -> dict:
         "gpu_duration_s": GPU_DURATION_S,
         "gpu_size": GPU_SIZE,
         "prefetch": dict(PREFETCH),
+        "last_gen": dict(LAST_GEN),
         "volume_read": _volume_read_mbps(),
         "server": _server_url or "",
         "ping": _ping(),