Update app.py

app.py

@@ -69,11 +69,10 @@ RATE_LIMIT_RPM   = int(os.getenv("RATE_LIMIT_RPM",   "60"))
 RATE_BURST       = int(os.getenv("RATE_LIMIT_BURST",  "10"))
 RATE_VIP_IPS     = set(os.getenv("RATE_VIP_IPS", "127.0.0.1").split(","))
 # RAM thresholds — auto-calculated as % of total RAM if not set explicitly
-_total_ram_gb = __import__('psutil').virtual_memory().total / (1024**3)
-# Thresholds: 90/95/98% of total RAM — safe on any machine size
-RAM_WARN_GB   = float(os.getenv("RAM_WARN_GB",   str(round(_total_ram_gb * 0.90, 1))))
-RAM_REJECT_GB = float(os.getenv("RAM_REJECT_GB", str(round(_total_ram_gb * 0.95, 1))))
-RAM_FLUSH_GB  = float(os.getenv("RAM_FLUSH_GB",  str(round(_total_ram_gb * 0.98, 1))))
+# RAM thresholds — computed dynamically in watchdog_task() based on actual machine RAM
+RAM_WARN_GB   = float(os.getenv("RAM_WARN_GB",   "0"))  # 0 = auto 82% of total RAM
+RAM_REJECT_GB = float(os.getenv("RAM_REJECT_GB", "0"))  # 0 = auto 90% of total RAM
+RAM_FLUSH_GB  = float(os.getenv("RAM_FLUSH_GB",  "0"))  # 0 = auto 95% of total RAM
 REQUEST_TIMEOUT  = float(os.getenv("REQUEST_TIMEOUT","90.0"))
 QUEUE_TIMEOUT    = float(os.getenv("QUEUE_TIMEOUT",  "30.0"))
 ENRICH_TIMEOUT   = float(os.getenv("ENRICH_TIMEOUT", "12.0"))
@@ -523,7 +522,12 @@ async def _flush_kv() -> bool:
     return False
 
 async def watchdog_task():
-    logger.info(f"[wd] warn={RAM_WARN_GB}G reject={RAM_REJECT_GB}G flush={RAM_FLUSH_GB}G")
+    # Compute dynamic thresholds based on actual machine RAM
+    _total = psutil.virtual_memory().total / (1024**3)
+    _warn   = RAM_WARN_GB   if RAM_WARN_GB   > 0 else round(_total * 0.82, 1)
+    _reject = RAM_REJECT_GB if RAM_REJECT_GB > 0 else round(_total * 0.90, 1)
+    _flush  = RAM_FLUSH_GB  if RAM_FLUSH_GB  > 0 else round(_total * 0.95, 1)
+    logger.info(f"[wd] total={_total:.1f}G warn={_warn}G reject={_reject}G flush={_flush}G")
     while True:
         try:
             await asyncio.sleep(2.0)
@@ -531,9 +535,9 @@ async def watchdog_task():
             used = mem.used / (1024**3)
             wd.ram_gb = used
 
-            if used >= RAM_FLUSH_GB:
+            if used >= _flush:
                 if wd.level != "flush":
-                    logger.warning(f"[wd] 🚨 FLUSH {used:.1f}GB")
+                    logger.warning(f"[wd] 🚨 FLUSH {used:.1f}/{_total:.0f}GB")
                 wd.level = "flush"; wd.flushes += 1
                 gc.collect(2)
                 exact_cache.cleanup()
@@ -543,13 +547,13 @@ async def watchdog_task():
                 await _flush_kv()
                 after = psutil.virtual_memory().used / (1024**3)
                 logger.info(f"[wd] post-flush {after:.1f}GB (freed {used-after:.1f}GB)")
-                if after < RAM_REJECT_GB: wd.level = "ok"
-            elif used >= RAM_REJECT_GB:
+                if after < _reject: wd.level = "ok"
+            elif used >= _reject:
                 if wd.level not in ("reject", "flush"):
-                    logger.warning(f"[wd] ⚠ REJECT {used:.1f}GB")
+                    logger.warning(f"[wd] ⚠ REJECT {used:.1f}/{_total:.0f}GB")
                 wd.level = "reject"; wd.rejects += 1; gc.collect(1)
-            elif used >= RAM_WARN_GB:
-                if wd.level == "ok": logger.info(f"[wd] ⚡ WARN {used:.1f}GB")
+            elif used >= _warn:
+                if wd.level == "ok": logger.info(f"[wd] ⚡ WARN {used:.1f}/{_total:.0f}GB")
                 wd.level = "warn"
             else:
                 if wd.level != "ok": logger.info(f"[wd] ✅ OK {used:.1f}GB")
@@ -1099,6 +1103,10 @@ async def chat(request: Request) -> Response | StreamingResponse:
                             headers={"X-Request-ID":rid,"X-Cache":"SEMANTIC-HIT"})
 
     # Build payload
+    # Thinking mode OFF by default — Qwen3 thinks for EVERY request otherwise
+    # User must explicitly pass "thinking": true to enable reasoning
+    thinking_requested = body.get("thinking", False)
+
     payload: dict = {
         "messages":     messages,
         "cache_prompt": True,
@@ -1109,12 +1117,11 @@ async def chat(request: Request) -> Response | StreamingResponse:
         "top_k":        int(body.get("top_k", 20)),
         "min_p":        float(body.get("min_p", 0.0)),
         "stream":       is_stream,
+        # Always set thinking explicitly to avoid Qwen3 auto-enabling it
+        "chat_template_kwargs": {"enable_thinking": bool(thinking_requested)},
     }
-    if body.get("thinking") is True:
-        payload["chat_template_kwargs"] = {"enable_thinking": True}
+    if thinking_requested:
         payload["reasoning_budget"] = int(body.get("thinking_budget", 4096))
-    elif body.get("thinking") is False:
-        payload["chat_template_kwargs"] = {"enable_thinking": False}
     if body.get("tools"):
         payload["tools"]               = body["tools"]
         payload["tool_choice"]         = body.get("tool_choice", "auto")
@@ -1592,4 +1599,4 @@ if __name__ == "__main__":
     import uvicorn
     uvicorn.run("app:app", host="0.0.0.0",
                 port=int(os.getenv("API_PORT","7860")),
-                loop="uvloop", http="h11", workers=1)
+                loop="uvloop", http="h11", workers=1)