space: load Nemotron the normal way — transformers-native (no trust_remote_code), NO mamba_ssm/causal_conv1d. Those custom Triton CUDA kernels were the segfault (THCPModule_initExtension); native falls back to pure-torch Mamba on ZeroGPU.

space: adopt the org's proven NPCverse structure — gradio 6 Server + @app .api + app.launch() (installs ZeroGPU hooks), transformers 5 (compatible with gradio 6; trust_remote_code uses our repo's modeling). Replaces the custom engine.launch+route-surgery that broke the hooks and segfaulted.

space: finetuned Warden on ZeroGPU the documented way — bf16 + .to('cuda') module-level + @spaces.GPU(xlarge), no bitsandbytes/device_map (the actual fix). Direct run_in_threadpool call verified by the probe.

space: ZeroGPU diagnostic — measure CPU RAM/disk/VRAM + confirm a @spaces.GPU call works, before loading the model the documented .to('cuda') way

space: revert to Gradio SDK + CPU llama-cpp-python (keeps the prize; ZeroGPU was the problem, not the SDK)

space: serve via Blocks.launch (ZeroGPU handshake) + pin transformers<5 for the bnb4 checkpoint format

space: torch 2.10 — mamba wheels demand triton>=3.5, impossible under torch 2.8

space: pin torch 2.8 + prebuilt mamba-ssm/causal-conv1d wheels

space: ZeroGPU port — Gradio SDK runtime, on-Space Warden inference