fixed contrastive search error

app.py

@@ -18,7 +18,7 @@ from transformers import (
     CLIPModel, CLIPProcessor, BitsAndBytesConfig
 )
 
-app = FastAPI(title="XAI Auditor: Balanced 4-Caption Ensemble")
+app = FastAPI(title="XAI Auditor: Symmetrical Deterministic Ensemble")
 
 app.add_middleware(
     CORSMiddleware,
@@ -38,7 +38,7 @@ async def startup_event():
     token = os.getenv("HF_Token")
     if token: login(token=token)
     
-    print("Pinning 8-bit quantized models to memory...")
+    print("Pinning 8-bit quantized models to memory space...")
     local_dir = snapshot_download(repo_id="SaniaE/Image_Captioning_Ensemble", token=token, local_dir="weights")
 
     quantization_config = BitsAndBytesConfig(
@@ -56,7 +56,7 @@ async def startup_event():
         "processor": BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-large")
     }
 
-    # 2. Load Compressed ViT Track (Your retrained variant target)
+    # 2. Load Compressed ViT Track (Fixed Shape Mappings)
     MODELS["vit"] = {
         "model": AutoModelForCausalLM.from_pretrained(
             os.path.join(local_dir, "vit"),
@@ -77,19 +77,19 @@ async def startup_event():
         "processor": CLIPProcessor.from_pretrained(os.path.join(local_dir, "clip/clip_processor"))
     }
     
-    print("Symmetrical ensemble pipeline initialized and pinned. Ready.")
+    print("All system weights safely pinned. Matrix crash protections active.")
 
-# --- High-Speed 4-Caption Generation Engine ---
+# --- High-Speed Fixed Matrix Execution Engine ---
 
-def _generate_balanced_4_track(image, temp, max_len=15):
+def _generate_balanced_4_track(image, max_len=15):
     """
-    Symmetrical 4-Caption Engine optimized with Contrastive Search
-    to eliminate padding-token calculation latency.
+    Generates exactly 4 captions (2 BLIP, 2 ViT) using safe, high-speed 
+    deterministic decoding loops that eliminate padding math overhead.
     """
     captions = []
     
     with torch.inference_mode():
-        # Track A: Contrastive BLIP Pass (2 Parallel Variations)
+        # Track A: Batched BLIP Pass (2 Parallel Paths)
         b_data = MODELS["blip"]
         b_inputs = b_data["processor"](images=image, return_tensors="pt")
         b_pixels = b_inputs.pixel_values.to(DEVICE)
@@ -98,16 +98,15 @@ def _generate_balanced_4_track(image, temp, max_len=15):
         b_ids = b_data["model"].generate(
             pixel_values=batched_b_pixels,
             max_new_tokens=max_len,
-            do_sample=False,            # Disable random sampling overhead
-            penalty_alpha=0.6,          # Contrastive penalty anchor
-            top_k=4,                    # Narrow contrastive candidate pool
-            early_stopping=True,
-            use_cache=True
+            do_sample=False,          # Purely deterministic execution
+            num_beams=2,              # 2 separate beams for internal diversity split
+            early_stopping=True,      # Break out the millisecond end token drops
+            use_cache=True            # Leverage cached keys to prevent recomputation
         )
         b_caps = b_data["processor"].batch_decode(b_ids, skip_special_tokens=True)
         captions.extend([cap.strip() for cap in b_caps])
 
-        # Track B: Contrastive ViT Pass (2 Parallel Variations)
+        # Track B: Batched ViT Pass (2 Parallel Paths - Safe Matrix Sizing)
         v_data = MODELS["vit"]
         i_proc, t_proc = v_data["processor"]
         v_inputs = i_proc(images=image, return_tensors="pt")
@@ -123,9 +122,8 @@ def _generate_balanced_4_track(image, temp, max_len=15):
             pixel_values=batched_v_pixels,
             attention_mask=batched_mask,
             max_new_tokens=max_len,
-            do_sample=False,            # Disable random sampling overhead
-            penalty_alpha=0.6,          # Contrastive penalty anchor
-            top_k=4,                    # Narrow contrastive candidate pool
+            do_sample=False,          # Swapped off sample/contrastive to prevent shape mismatch
+            num_beams=2,              # Safe, standardized beam expansion tracking
             early_stopping=True,
             use_cache=True
         )
@@ -133,16 +131,17 @@ def _generate_balanced_4_track(image, temp, max_len=15):
         captions.extend([cap.strip() for cap in v_caps])
             
     return captions
+
 # --- Endpoints ---
 
 @app.post("/generate")
-async def generate_captions(file: UploadFile = File(...), temp: float = Query(0.7)):
+async def generate_captions(file: UploadFile = File(...)):
     """Generates 4 diverse captions split evenly across architectures for UI balance."""
     start_time = time.perf_counter()
     image = Image.open(file.file).convert("RGB")
     
-    # Run the balanced 2+2 pipeline
-    captions = await asyncio.to_thread(_generate_balanced_4_track, image, temp, 15)
+    # Fire off the optimized, crash-safe execution matrix
+    captions = await asyncio.to_thread(_generate_balanced_4_track, image, 15)
     
     elapsed_time = time.perf_counter() - start_time
     print(f"[BENCHMARK] /generate 4-caption turnaround: {elapsed_time:.4f}s")
@@ -198,37 +197,4 @@ async def internal_debate_audit(file: UploadFile = File(...), user_prompt: str =
     image_bytes = await file.read()
     image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
     
-    # Extract baseline descriptive anchor using a single-pass track
-    blip_caption = (await asyncio.to_thread(_generate_balanced_4_track, image, 1.0, 15))[0]
-    
-    clip_m = MODELS["clip"]["model"]
-    clip_p = MODELS["clip"]["processor"]
-    clip_dtype = torch.float16 if DEVICE == "cuda" else torch.float32
-    
-    image_inputs = clip_p(images=image, return_tensors="pt")
-    text_inputs = clip_p(text=[user_prompt, blip_caption], return_tensors="pt", padding=True)
-    
-    with torch.inference_mode():
-        img_pixels = image_inputs.pixel_values.to(device=DEVICE, dtype=clip_dtype)
-        txt_ids = text_inputs.input_ids.to(DEVICE)
-        txt_mask = text_inputs.attention_mask.to(DEVICE)
-        
-        image_features = clip_m.get_image_features(pixel_values=img_pixels)
-        image_features = image_features / image_features.norm(dim=-1, keepdim=True)
-        
-        text_features = clip_m.get_text_features(input_ids=txt_ids, attention_mask=txt_mask)
-        text_features = text_features / text_features.norm(dim=-1, keepdim=True)
-        
-        logits_per_image = (image_features @ text_features.T) * clip_m.logit_scale.exp()
-        probs = F.softmax(logits_per_image, dim=-1).cpu().to(torch.float32).numpy()[0]
-    
-    u_score, m_score = float(probs[0]), float(probs[1])
-    verdict = "Model Bias Detected." if abs(u_score - m_score) >= 0.15 else "Consensus: High Alignment."
-    if u_score < 0.35: verdict = "Perspective Divergence: Intent not grounded in image."
-
-    return {
-        "perspectives": {"user": user_prompt, "ai": blip_caption},
-        "audit_scores": {"intent_grounding": round(u_score, 4), "ai_grounding": round(m_score, 4)},
-        "verdict": verdict,
-        "metadata": {"processing_time_sec": round(time.perf_counter() - start_time, 4)}
-    }
+    # Extract