Upload 133 files

app.py

@@ -2661,7 +2661,8 @@ def chat_respond(message: str, history: list[dict], system_prompt: str,
         text = "\n".join(f"{m['role']}: {m['content']}" for m in messages) + "\nassistant:"
 
     inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=context_length)
-    inputs = {k: v.to(model.device) for k, v in inputs.items()}
+    _model_device = next(model.parameters()).device
+    inputs = {k: v.to(_model_device) for k, v in inputs.items()}
 
     # Streaming generation — repetition_penalty (user-controllable, default 1.0)
     # can break degenerate refusal loops if increased.
@@ -3156,7 +3157,7 @@ def ab_chat_respond(message: str, history_left: list[dict], history_right: list[
     # --- Generate from abliterated model (streaming) ---
     stream_timeout = max(120, 120 + int(max_tokens * 0.1))
     streamer_abl = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=stream_timeout)
-    inputs_abl = {k: v.to(abliterated_model.device) for k, v in inputs.items()}
+    inputs_abl = {k: v.to(next(abliterated_model.parameters()).device) for k, v in inputs.items()}
     gen_kwargs_abl = {**inputs_abl, **gen_kwargs_base, "streamer": streamer_abl}
 
     gen_error_abl = [None]
@@ -3217,7 +3218,7 @@ def ab_chat_respond(message: str, history_left: list[dict], history_right: list[
         )
 
         streamer_orig = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=stream_timeout)
-        inputs_orig = {k: v.to(original_model.device) for k, v in inputs.items()}
+        inputs_orig = {k: v.to(next(original_model.parameters()).device) for k, v in inputs.items()}
         gen_kwargs_orig = {**inputs_orig, **gen_kwargs_base, "streamer": streamer_orig}
 
         gen_error_orig = [None]

obliteratus/abliterate.py

@@ -1052,7 +1052,8 @@ class AbliterationPipeline:
             else:
                 # Layer produced no activations (hook failure or skipped layer)
                 empty_layers.append(idx)
-                hidden = self._harmful_acts[0][0].shape[-1] if self._harmful_acts.get(0) else 768
+                _acts_0 = self._harmful_acts.get(0)
+                hidden = _acts_0[0].shape[-1] if _acts_0 and len(_acts_0) > 0 else 768
                 self._harmful_means[idx] = torch.zeros(1, hidden)
                 self._harmless_means[idx] = torch.zeros(1, hidden)
         if empty_layers:
@@ -1072,7 +1073,8 @@ class AbliterationPipeline:
                 if self._jailbreak_acts.get(idx):
                     self._jailbreak_means[idx] = torch.stack(self._jailbreak_acts[idx]).mean(dim=0)
                 else:
-                    hidden = self._harmful_acts[0][0].shape[-1] if self._harmful_acts.get(0) else 768
+                    _acts_0 = self._harmful_acts.get(0)
+                    hidden = _acts_0[0].shape[-1] if _acts_0 and len(_acts_0) > 0 else 768
                     self._jailbreak_means[idx] = torch.zeros(1, hidden)
             self.log("  Jailbreak activations collected for three-way contrastive analysis")
 
@@ -1138,15 +1140,19 @@ class AbliterationPipeline:
             pass  # Fall through to per-prompt with error handling
 
         wrapped = []
+        n_wrapped = 0
         for i, conv in enumerate(all_conversations):
             try:
                 text = tokenizer.apply_chat_template(
                     conv, tokenize=False, add_generation_prompt=True
                 )
                 wrapped.append(text)
+                n_wrapped += 1
             except Exception:
                 wrapped.append(prompts[i])  # fallback to raw if individual prompt fails
-        self.log(f"    chat template {n}/{n}")
+        self.log(f"    chat template {n_wrapped}/{n}")
+        if n_wrapped < n:
+            self.log(f"    WARNING: {n - n_wrapped} prompts fell back to raw text")
         return wrapped
 
     def _apply_spectral_cascade_weights(self):