Upload app.py with huggingface_hub

app.py

@@ -264,13 +264,19 @@ OpenAI-style REST API for [PASITA](https://huggingface.co/OpceanAI/PASITA), an 8
 from-scratch model that turns plain text into faithful Markdown. No API key is required.
 Works with the OpenAI SDK by pointing `base_url` at `/v1`. The Gradio API and the MCP
 server stay available at `?view=api`.
+
+Honored request fields: `model`, `messages`, `stream`, `stream_options.include_usage`,
+`max_tokens` (upper bound), `candidates` (1-3), `structured_fallback`. Sampling
+temperature is fixed at 0.3 because that is the measured recipe; other OpenAI fields
+are accepted but ignored.
 """.strip()
 
 
-def _api_error(message: str, error_type: str = "invalid_request_error", status: int = 400, code: str | None = None) -> JSONResponse:
+def _api_error(message: str, error_type: str = "invalid_request_error", status: int = 400, code: str | None = None, headers: dict | None = None) -> JSONResponse:
     return JSONResponse(
         status_code=status,
         content={"error": {"message": message, "type": error_type, "param": None, "code": code}},
+        headers=headers,
     )
 
 
@@ -278,9 +284,9 @@ def _exception_to_error(exc: Exception) -> JSONResponse:
     message = f"{type(exc).__name__}: {exc}"
     text = message.lower()
     if "zerogpu" in text and any(k in text for k in ("quota", "exhausted", "insufficient", "limit")):
-        return _api_error(str(exc), "rate_limit_error", status=429, code="gpu_quota_exceeded")
+        return _api_error(str(exc), "rate_limit_error", status=429, code="gpu_quota_exceeded", headers={"Retry-After": "60"})
     if any(k in text for k in ("illegal duration", "task aborted", "queue is full", "queue full")):
-        return _api_error(str(exc), "rate_limit_error", status=429, code="gpu_busy")
+        return _api_error(str(exc), "rate_limit_error", status=429, code="gpu_busy", headers={"Retry-After": "30"})
     return _api_error(message, "server_error", status=500)
 
 
@@ -311,6 +317,7 @@ class ChatCompletionRequest(BaseModel):
     model: str | None = "pasita-v1"
     messages: list[ChatMessage]
     stream: bool = False
+    stream_options: dict | None = None
     max_tokens: int | None = None
     candidates: int = 3
     structured_fallback: bool = True
@@ -370,6 +377,18 @@ class _LegacyConvertBodyMiddleware:
 
 server.add_middleware(_LegacyConvertBodyMiddleware)
 
+try:
+    from fastapi.middleware.cors import CORSMiddleware
+
+    server.add_middleware(
+        CORSMiddleware,
+        allow_origins=["*"],
+        allow_methods=["GET", "POST", "OPTIONS"],
+        allow_headers=["*"],
+    )
+except Exception:
+    pass
+
 
 @server.exception_handler(fastapi.exceptions.RequestValidationError)
 async def validation_error_handler(request: fastapi.Request, exc: fastapi.exceptions.RequestValidationError):
@@ -486,6 +505,8 @@ def chat_completions(body: ChatCompletionRequest, request: fastapi.Request) -> d
     markdown = payload["markdown"]
 
     if body.stream:
+        include_usage = bool((body.stream_options or {}).get("include_usage"))
+
         def sse():
             def chunk(delta, finish=None, usage=None):
                 data = {
@@ -501,7 +522,7 @@ def chat_completions(body: ChatCompletionRequest, request: fastapi.Request) -> d
 
             yield chunk({"role": "assistant", "content": ""})
             yield chunk({"content": markdown})
-            yield chunk({}, finish="stop", usage=payload["usage"])
+            yield chunk({}, finish="stop", usage=payload["usage"] if include_usage else None)
             yield "data: [DONE]\n\n"
 
         return StreamingResponse(