no stop added

handler.py

@@ -1,8 +1,9 @@
 # -*- coding: utf-8 -*-
 """
-PULSE ECG Handler - Demo-like (sampling) LLaVA endpoint
-- Demo davranışı: do_sample=True, temperature/top_p payload'dan alınır
-- max_new_tokens: payload/slider değeri; bağlam limitine göre güvenli kırpma
+PULSE ECG Handler - Demo-like (sampling) + no_stop bayrağı
+- Demo davranışı: do_sample=True, temperature/top_p payload'dan
+- max_new_tokens: payload/slider değeri (KIRPMA YOK, direkt kullanılır)
+- İsteğe bağlı: no_stop=True ile stopping_criteria devre dışı
 - Tek görsel işleme; IM_START/END otomatik; 3D/4D/5D tensör uyumlu
 - Çıktıya post-format/deduplicate UYGULANMAZ (demo ile bire bir)
 """
@@ -99,7 +100,8 @@ def _safe_upload(path: str):
 def _conv_log_path():
     t = datetime.datetime.now()
     p = os.path.join(LOGDIR, f"{t.year:04d}-{t.month:02d}-{t.day:02d}-user_conv.json")
-    os.makedirs(os.path.dirname(p), exist_ok=True)
+    os.makedirs(os.path.dirname(p), exist_ok=True
+    )
     return p
 
 def load_image_any(image_input):
@@ -171,16 +173,6 @@ def _stopping(chatbot, input_ids):
     stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
     return KeywordsStoppingCriteria([stop_str], chatbot.tokenizer, input_ids)
 
-def _safe_max_new_tokens(requested: int, input_len: int, ctx_limit: int) -> int:
-    """
-    Demo'da slider değeri doğrudan kullanılıyor; burada ek güvenlik:
-    toplam (input + new + rezerv) <= ctx_limit olacak şekilde kırp.
-    """
-    requested = max(1, min(int(requested), 8192))
-    reserve = 16
-    available = max(32, ctx_limit - input_len - reserve)
-    return max(1, min(requested, available))
-
 def generate_response(
     message_text: str,
     image_input,
@@ -191,6 +183,8 @@ def generate_response(
     repetition_penalty: float = 1.0,
     conv_mode_override: str | None = None,
     det_seed: int | None = None,
+    no_stop: bool = False,
+    min_new_tokens: int | None = None,   # opsiyonel, uzunluğu zorlamak istersen
 ):
     if not LLAVA_AVAILABLE:
         return {"error": "LLaVA modules not available"}
@@ -248,43 +242,62 @@ def generate_response(
         else:
             return {"error": "Image processing returned empty"}
 
-        # Demo'da çoğunlukla half + to(device) kullanılıyor
-        image_tensor = image_tensor.to(device=device, dtype=getattr(torch, "float16", torch.float16))
+        # Demo tarafında half + to(device) kalıbı yaygın
+        image_tensor = image_tensor.to(device=device, dtype=dtype)
     except Exception as e:
         return {"error": f"Image processing failed: {e}"}
 
     # Prompt & tokenizasyon
     prompt, input_ids = _build_prompt_and_ids(chatbot, message_text, device)
-    stopping = _stopping(chatbot, input_ids)
-
-    # max_new_tokens'ı güvenle kırp (demo slider + bağlam tavanı)
-    ctx_limit = context_len or getattr(chatbot.model.config, "max_position_embeddings", 8192)
-    max_new_tokens = _safe_max_new_tokens(max_new_tokens, input_ids.shape[1], ctx_limit)
+    stopping = None if no_stop else _stopping(chatbot, input_ids)
 
-    # Demo: sampling açık; istenirse deterministik sample için seed verilebilir
+    # (opsiyonel) deterministik sampling
     if det_seed is not None:
-        torch.manual_seed(det_seed)
-        if torch.cuda.is_available():
-            torch.cuda.manual_seed(det_seed)
-            torch.cuda.manual_seed_all(det_seed)
+        try:
+            det_seed = int(det_seed)
+            torch.manual_seed(det_seed)
+            if torch.cuda.is_available():
+                torch.cuda.manual_seed(det_seed)
+                torch.cuda.manual_seed_all(det_seed)
+        except Exception:
+            pass
+
+    # EOS/PAD güvenli al
+    eos_id = chatbot.tokenizer.eos_token_id
+    if eos_id is None:
+        try:
+            eos_id = chatbot.tokenizer.convert_tokens_to_ids("</s>")
+        except Exception:
+            eos_id = 0
+
+    # generate kwargs (demo-like)
+    gen_kwargs = dict(
+        inputs=input_ids,
+        images=image_tensor,
+        do_sample=True,
+        temperature=float(temperature),
+        top_p=float(top_p),
+        repetition_penalty=float(repetition_penalty),
+        max_new_tokens=int(max_new_tokens),  # KIRPMA YOK
+        use_cache=False,
+        pad_token_id=eos_id,
+        eos_token_id=eos_id,
+        length_penalty=1.0,
+        early_stopping=False,
+        stopping_criteria=None if no_stop else [stopping],
+    )
+    if min_new_tokens is not None:
+        try:
+            mn = int(min_new_tokens)
+            if mn > 0 and mn <= int(max_new_tokens):
+                gen_kwargs["min_new_tokens"] = mn
+        except Exception:
+            pass
 
+    # Üretim
     try:
         with torch.no_grad():
-            outputs = chatbot.model.generate(
-                inputs=input_ids,
-                images=image_tensor,
-                do_sample=True,
-                temperature=float(temperature),
-                top_p=float(top_p),
-                repetition_penalty=float(repetition_penalty),
-                max_new_tokens=int(max_new_tokens),
-                use_cache=False,
-                pad_token_id=chatbot.tokenizer.eos_token_id,
-                eos_token_id=chatbot.tokenizer.eos_token_id,
-                length_penalty=1.0,
-                early_stopping=False,
-                stopping_criteria=[stopping],
-            )
+            outputs = chatbot.model.generate(**gen_kwargs)
         gen = outputs[0][input_ids.shape[1]:]
         text = chatbot.tokenizer.decode(gen, skip_special_tokens=True)
 
@@ -303,7 +316,7 @@ def generate_response(
             "image_hash": img_hash,
             "image_path": img_path or "",
         }
-        with open(_conv_log_path(), "a") as f:
+        with open(_conv_log_path(), "a", encoding="utf-8") as f:
             f.write(json.dumps(row, ensure_ascii=False) + "\n")
         _safe_upload(_conv_log_path())
         if img_path:
@@ -340,7 +353,7 @@ def query(payload: dict):
         repetition_penalty  = float(payload.get("repetition_penalty", 1.0))
         conv_mode_override  = payload.get("conv_mode", None)
 
-        # (Opsiyonel) deterministik sample için seed (demo defaultu: None)
+        # (Opsiyonel) deterministik sample için seed
         det_seed = payload.get("det_seed", None)
         if det_seed is not None:
             try:
@@ -348,6 +361,17 @@ def query(payload: dict):
             except Exception:
                 det_seed = None
 
+        # (Yeni) stopping_criteria kapatma bayrağı
+        no_stop = bool(payload.get("no_stop", False))
+
+        # (Opsiyonel) min_new_tokens
+        mnt = payload.get("min_new_tokens", None)
+        if mnt is not None:
+            try:
+                mnt = int(mnt)
+            except Exception:
+                mnt = None
+
         return generate_response(
             message_text=message,
             image_input=image,
@@ -357,6 +381,8 @@ def query(payload: dict):
             repetition_penalty=repetition_penalty,
             conv_mode_override=conv_mode_override,
             det_seed=det_seed,
+            no_stop=no_stop,
+            min_new_tokens=mnt,
         )
     except Exception as e:
         return {"error": f"Query failed: {e}"}
@@ -471,4 +497,3 @@ class EndpointHandler:
 
 if __name__ == "__main__":
     print("Handler ready. Use `EndpointHandler` or `query` for HF Inference Endpoints.")
-