GPU recovered — back to ZeroGPU (fast+free) with retry resilience; restore transformers stack

requirements.txt

@@ -1,15 +1,19 @@
-# GODSEED — requirements for the live CPU deployment (cpu-basic hardware).
-# The god runs in-process on llama.cpp (CPU) — no GPU, no torch, no transformers.
-# (June 12: ZeroGPU's GPU NVML-crashed persistently on the NemotronH forward;
-#  a CPU-only llama.cpp app is also killed on ZeroGPU hardware, so the Space
-#  moved to cpu-basic. The whole god is now one ~2.5GB GGUF in llama.cpp.)
+# GODSEED — requirements. Live backend = zerogpu (NVIDIA Nemotron-Nano-9B-v2 on
+# ZeroGPU). llama-cpp-python stays as the documented CPU fallback (Llama Champion).
 fastapi>=0.115
 uvicorn>=0.30
 gradio>=5
 pydantic>=2.7
 huggingface_hub>=0.30
 
-# llama.cpp runtime for the Nemotron-3-Nano-4B GGUF backend. Source build kills
-# the Spaces builder (June 12) — use the project's prebuilt CPU wheel index.
+# llama.cpp CPU fallback (GODSEED_BACKEND=llamacpp). Prebuilt CPU wheel — source
+# build kills the Spaces builder.
 --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
 llama-cpp-python>=0.3.8
+
+# ZeroGPU stack — transformers ≥5.4 has nemotron_h in-tree (trust_remote_code=False,
+# kernel-free). Do NOT pin torch/gradio/spaces (image-managed). NO mamba kernels.
+transformers>=5.4
+accelerate>=0.34.0
+einops
+sentencepiece