fix zerogpu quota: drop duration holds, cache results, cpu bulk scan, bound shap evals

app.py

@@ -281,17 +281,41 @@ def risk(conf, level=""):
     return "lo", "var(--lo)", "Low risk", "🟢"
 
 
+_CACHE, _CACHE_MAX = {}, 256
+
+
+def _cache_get(key):
+    return _CACHE.get(key)
+
+
+def _cache_put(key, value):
+    if len(_CACHE) >= _CACHE_MAX:
+        _CACHE.clear()
+    _CACHE[key] = value
+    return value
+
+
 def _run(message):
+    key = ("run", message)
+    hit = _cache_get(key)
+    if hit is not None:
+        return hit
     r = predictor.predict(message)
     r["_tokens"] = [w for w in str(message).lower().split() if w.isalpha() and len(w) > 3][:14]
-    return r
+    return _cache_put(key, r)
 
 
 def _explain(message):
+    key = ("ex", message)
+    hit = _cache_get(key)
+    if hit is not None:
+        return hit
     try:
-        return explainer.explain_text(message, num_features=12)
+        # 8 features keeps the SHAP kernel small — this is the expensive part.
+        out = explainer.explain_text(message, num_features=8)
     except Exception:
-        return {"top_spam_words": [], "top_ham_words": []}
+        out = {"top_spam_words": [], "top_ham_words": []}
+    return _cache_put(key, out)
 
 
 # ── renderers ────────────────────────────────────────────────────────────────
@@ -389,7 +413,7 @@ def h_warn(r):
 # ─────────────────────────────────────────────────────────────────────────────
 #  Handlers
 # ─────────────────────────────────────────────────────────────────────────────
-@spaces.GPU(duration=90)
+@spaces.GPU
 def scan(message):
     if not message or not message.strip():
         return ('<div class="card"><div class="row">Type or paste a message to analyse it.</div></div>',
@@ -403,6 +427,7 @@ def scan(message):
 
 @spaces.GPU
 def mob_detail(which, custom=""):
+    """Single seeded-message analysis. Kept on GPU for a responsive detail view."""
     i = SEED_LABELS.index(which) if which in SEED_LABELS else 0
     msg = custom.strip() or SEED_SMS[i]["b"]
     r = _run(msg)
@@ -411,8 +436,12 @@ def mob_detail(which, custom=""):
             + h_verdict(r, msg) + h_signals(r) + h_shap(r, ex) + h_warn(r))
 
 
-@spaces.GPU
 def mob_scan_all(progress=gr.Progress()):
+    """Bulk scan of the seeded inbox.
+
+    Deliberately NOT decorated with @spaces.GPU — it runs on CPU so that walking
+    16 messages does not burn the ZeroGPU quota. Single-message scans stay on GPU.
+    """
     rows, threats = [], 0
     for i, m in enumerate(SEED_SMS):
         progress(i / len(SEED_SMS), desc=f"Scanning {i+1}/{len(SEED_SMS)} · {m['n']}")
@@ -445,7 +474,7 @@ def mob_scan_all(progress=gr.Progress()):
 # ─────────────────────────────────────────────────────────────────────────────
 #  JSON API — used by the Android app and any external client
 # ─────────────────────────────────────────────────────────────────────────────
-@spaces.GPU(duration=60)
+@spaces.GPU
 def predict_api(message: str) -> dict:
     return _run(message)
 

smishing_detector/explainability/shap_explainer.py

@@ -50,15 +50,33 @@ class SmishingExplainer:
 
         return predict_fn
 
-    def explain_text(self, message: str, num_features: int = 10) -> dict:
+    def explain_text(self, message: str, num_features: int = 10,
+                     max_tokens: int = 24, max_evals: int = 96) -> dict:
+        """
+        Word-level SHAP attribution.
+
+        max_tokens / max_evals bound the compute: the Text masker perturbs every
+        token, so cost grows roughly as 2x token count. Capping them keeps a single
+        explanation to a handful of forward passes instead of an unbounded loop.
+        """
         predict_fn = self._get_prediction_function()
 
+        # Bound the text SHAP has to work on — long messages give diminishing
+        # returns well before the model even reaches MAX_LEN tokens.
+        words = str(message).split()
+        if len(words) > max_tokens:
+            message = " ".join(words[:max_tokens])
+
         tokenizer = self.predictor.tokenizer
         masker = shap.maskers.Text(tokenizer)
 
         explainer = shap.Explainer(predict_fn, masker)
 
-        shap_values = explainer([message])
+        try:
+            shap_values = explainer([message], max_evals=max_evals)
+        except Exception:
+            # Fall back to the uncapped path if the masker rejects max_evals.
+            shap_values = explainer([message])
 
         tokens = [str(t) for t in shap_values.data[0]]
         values = shap_values.values[0]