GPU final: model resident on cuda (no per-call transfer) + 240s budget + one-shot — fixes empty-output timeout

mind/backends.py

@@ -327,7 +327,7 @@ class ZeroGPUBackend:
     # kernels, which cannot import on the ZeroGPU image (no compatible CUDA libs
     # at startup — June 12). 4B GGUF remains the llama.cpp local mode.
     DEFAULT_MODEL = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
-    GPU_DURATION_S = 120
+    GPU_DURATION_S = 240
 
     def __init__(self, model_id: str | None = None, max_input_tokens: int = 4096):
         self.model_id = model_id or os.environ.get(
@@ -367,19 +367,21 @@ class ZeroGPUBackend:
                 low_cpu_mem_usage=True,
             )
             self._model.eval()
-            # Model stays on CPU at load. The .to("cuda") happens INSIDE the
-            # @spaces.GPU function on EVERY call — this is exactly what the
-            # working bench Space does. ZeroGPU detaches the GPU between calls,
-            # so the model must be (re)homed to cuda within each GPU context;
-            # caching it (eager load or a one-time _on_cuda flag) leaves stale
-            # cuda tensors that NVML-crash on the next call (June 12). With
-            # one-shot generation there's one such call per wish.
+            # Move to cuda ONCE at load, guarded by `import spaces` (the lib
+            # virtualizes it and keeps the model resident across @spaces.GPU
+            # calls). With ONE-SHOT generation a wish is a single GPU call, so
+            # the multi-call NVML crash can't happen — and keeping the model
+            # resident avoids the ~67s CPU→GPU transfer that, done per-call, ate
+            # the whole GPU time budget and left no time to generate (→ empty
+            # output → fallback, June 12). Resident + one call = fast real output.
+            try:
+                import spaces  # noqa: F401
+                self._model.to("cuda")
+            except ImportError:
+                pass  # local/CPU dev
 
             def _generate(prompt: str, max_new_tokens: int, temperature: float) -> str:
-                import torch as _torch
-
-                if _torch.cuda.is_available():
-                    self._model.to("cuda")
+                import torch as _torch  # noqa: F401
 
                 # Nemotron-Nano-9B-v2 is a reasoning model that THINKS by
                 # default; "/no_think" in the system slot disables it (June 12

requirements.txt

@@ -1,5 +1,6 @@
-# GODSEED — CPU deployment. The god runs in-process on llama.cpp (4B Nemotron
-# GGUF). No GPU, no torch — reliable, real model, grammar-valid output.
+# GODSEED — live backend = zerogpu (NVIDIA Nemotron-Nano-9B-v2 on ZeroGPU,
+# one-shot generation = one @spaces.GPU call per wish). llama-cpp stays as the
+# documented CPU fallback (Llama Champion badge).
 fastapi>=0.115
 uvicorn>=0.30
 gradio>=5
@@ -8,3 +9,8 @@ huggingface_hub>=0.30
 
 --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
 llama-cpp-python>=0.3.8
+
+transformers>=5.4
+accelerate>=0.34.0
+einops
+sentencepiece