Fix: download GGUF eagerly at container boot, outside the @spaces.GPU lease (was burning GPU-seconds on a 6-7min network download, causing the lease to be reclaimed mid-download with no error surfaced)

app.py

@@ -22,8 +22,15 @@ Real fix history (found + fixed 2026-07-23, in order):
      This stays entirely within Gradio's own request-handling pipeline, which is what ZeroGPU's
      detection actually requires.
 
-Cold-start note: first request after idle downloads the GGUF and allocates the GPU
-(~60-120 s). Subsequent requests within the same GPU lease are fast.
+Cold-start note: the 19.8GB GGUF is downloaded once at container boot (module import,
+below), NOT inside the @spaces.GPU-decorated function. Real bug found + fixed 2026-07-23:
+the download was originally triggered lazily from inside _generate_tokens(), which meant it
+ran *inside* the ZeroGPU lease window and burned GPU-seconds on a plain network transfer --
+confirmed via real Space logs showing a live download stuck at 81%/16.0GB after 5m24s
+wall-clock, well past the `@spaces.GPU(duration=180)` cap, with no further log lines ever
+written (the lease was silently reclaimed mid-download). Downloading eagerly at import time
+means the file is already local on the Space's persistent container disk by the time any
+real request arrives, so the GPU lease only has to cover model load + generation.
 """
 from __future__ import annotations
 
@@ -38,14 +45,9 @@ FILENAME = "chatpdb_32b_v1_q4km.gguf"
 N_CTX = 1536   # matches chatPDB's real training max_seq_length (config/train_config.yaml)
 N_GPU_LAYERS = -1  # offload all layers to GPU
 
-_model_path: str | None = None
-
-
-def _get_model_path() -> str:
-    global _model_path
-    if _model_path is None:
-        _model_path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)
-    return _model_path
+print(f"[chatPDB] downloading {FILENAME} from {REPO_ID} (one-time, outside GPU lease)...")
+_MODEL_PATH = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)
+print(f"[chatPDB] model ready at {_MODEL_PATH}")
 
 
 @spaces.GPU(duration=180)
@@ -97,7 +99,7 @@ def _generate_tokens(
     from llama_cpp import Llama
 
     llm = Llama(
-        model_path=_get_model_path(),
+        model_path=_MODEL_PATH,
         n_ctx=N_CTX,
         n_gpu_layers=N_GPU_LAYERS,
         verbose=False,
@@ -127,9 +129,10 @@ with gr.Blocks(title="chatPDB API") as demo:
     gr.Markdown(
         "## 🧬 chatPDB Inference API\n\n"
         "Internal endpoint for [chatpdb.mdeller.com](https://chatpdb.mdeller.com). "
-        "Consumed via Gradio's own API: `POST /call/generate` then `GET /call/generate/<event_id>` "
-        "(SSE).\n\n"
-        "**Cold start:** first request after idle takes ~60-120 s (GGUF download + GPU alloc)."
+        "Consumed via Gradio's own API: `POST /gradio_api/call/generate` then "
+        "`GET /gradio_api/call/generate/<event_id>` (SSE).\n\n"
+        "**Cold start:** model download happens once at container boot, outside the GPU "
+        "lease. First real request per GPU lease takes ~10-30 s to load the GGUF onto the GPU."
     )
 
     # Hidden inputs/outputs purely to register a real Gradio event with api_name="generate" --