revert: restore working AMD gateway config with new Droplet IP 165.245.137.45

app.py

@@ -29,7 +29,6 @@ load_dotenv()
 
 import gradio as gr
 from openai import OpenAI
-import spaces  # ZeroGPU — provided by HF runtime
 
 from pipeline.memory       import build_system_prompt, get_context, save_session
 from pipeline.parser       import extract_facts_from_response, parse_structured_output
@@ -49,59 +48,10 @@ PATIENT_ID       = os.getenv("PATIENT_ID",   "demo_user_001")
 PATIENT_NAME     = os.getenv("PATIENT_NAME", "Margaret")
 STT_URL          = os.getenv("STT_URL",      "")
 TTS_URL          = os.getenv("TTS_URL",      "")
-# Optional: set LUMI_GATEWAY_URL to route to an external AMD backend instead of ZeroGPU
-LUMI_GATEWAY_URL = os.getenv("LUMI_GATEWAY_URL", "")
-HF_TOKEN         = os.getenv("HF_TOKEN", "")
+LUMI_GATEWAY_URL = os.getenv("LUMI_GATEWAY_URL", "http://165.245.137.45:8080")
 
 llm = OpenAI(base_url=VLLM_BASE_URL, api_key=os.getenv("OPENAI_API_KEY", "not-required"))
 
-# ---------------------------------------------------------------------------
-# ZeroGPU — lazy model loading
-# ---------------------------------------------------------------------------
-import torch
-
-_llm_model = None
-_llm_tokenizer = None
-
-def _ensure_llm_loaded():
-    global _llm_model, _llm_tokenizer
-    if _llm_model is not None:
-        return
-    from transformers import AutoTokenizer, AutoModelForCausalLM
-    print(f"[LLM] Loading {MODEL_NAME} ...")
-    _llm_tokenizer = AutoTokenizer.from_pretrained(
-        MODEL_NAME, token=HF_TOKEN or None
-    )
-    _llm_model = AutoModelForCausalLM.from_pretrained(
-        MODEL_NAME,
-        torch_dtype=torch.bfloat16,
-        device_map="cuda",
-        token=HF_TOKEN or None,
-    )
-    _llm_model.eval()
-    print("[LLM] Ready.")
-
-
-@spaces.GPU
-def _call_llm_gpu(messages: list[dict]) -> str:
-    """Run LLM inference on ZeroGPU. Returns the full decoded response."""
-    _ensure_llm_loaded()
-    text = _llm_tokenizer.apply_chat_template(
-        messages, tokenize=False, add_generation_prompt=True
-    )
-    inputs = _llm_tokenizer(text, return_tensors="pt").to(_llm_model.device)
-    with torch.no_grad():
-        output = _llm_model.generate(
-            **inputs,
-            max_new_tokens=512,
-            temperature=0.7,
-            do_sample=True,
-            repetition_penalty=1.2,
-            pad_token_id=_llm_tokenizer.eos_token_id,
-        )
-    new_tokens = output[0][inputs["input_ids"].shape[-1]:]
-    return _llm_tokenizer.decode(new_tokens, skip_special_tokens=True)
-
 PORTRAIT_DIR = Path(__file__).parent / "portraits"
 
 def _load_portrait(profile_id: str) -> Image.Image | None:
@@ -133,17 +83,12 @@ def _get_system_prompt() -> str:
 
 
 def _call_llm(messages: list[dict]) -> str:
-    # Priority 1: external AMD / custom gateway
     if LUMI_GATEWAY_URL:
         import httpx as _httpx
         r = _httpx.post(f"{LUMI_GATEWAY_URL}/chat",
                         json={"messages": messages}, timeout=60.0)
         r.raise_for_status()
         return r.json()["response"]
-    # Priority 2: ZeroGPU local inference
-    return _call_llm_gpu(messages)
-    # Priority 3 (unreachable on ZeroGPU, kept for local dev reference):
-    # vLLM OpenAI-compatible fallback
     resp = llm.chat.completions.create(
         model=MODEL_NAME,
         messages=messages,
@@ -155,16 +100,17 @@ def _call_llm(messages: list[dict]) -> str:
 
 
 def _call_llm_stream(messages: list[dict]):
-    """
-    Simulated streaming: run full generation then yield word-by-word.
-    ZeroGPU generator functions are unreliable for long-running GPU calls,
-    so we generate all tokens first and stream the result from CPU.
-    """
-    full = _call_llm(messages)
-    words = full.split(" ")
-    for i, word in enumerate(words):
-        yield ("" if i == 0 else " ") + word
-        time.sleep(0.025)
+    for chunk in llm.chat.completions.create(
+        model=MODEL_NAME,
+        messages=messages,
+        temperature=0.7,
+        max_tokens=512,
+        stream=True,
+        extra_body={"repetition_penalty": 1.2},
+    ):
+        delta = chunk.choices[0].delta.content
+        if delta:
+            yield delta
 
 
 def _build_messages(

requirements.txt

@@ -4,6 +4,7 @@ chromadb>=0.5.0
 sentence-transformers>=3.0.0
 numpy>=1.26.0
 python-dotenv
+audioop-lts
 Pillow>=10.0.0
 soundfile>=0.12.0
 uvicorn>=0.29.0
@@ -11,6 +12,3 @@ fastapi>=0.110.0
 gradio>=5.0.0
 edge-tts
 kokoro-onnx>=0.4.0
-# ZeroGPU (spaces injected by HF build)
-transformers>=4.45.0
-accelerate>=0.33.0