requirements.txt

# HF Space (Gradio SDK + ZeroGPU), structured like the org's working spaces:
# gradio.Server + @app.api + app.launch(). Load the model the NORMAL way —
# transformers-native NemotronH, NO mamba_ssm / causal_conv1d. Those custom
# Triton CUDA kernels are what segfaulted on ZeroGPU (THCPModule_initExtension);
# working Nemotron ZeroGPU spaces never install them and let transformers fall
# back to pure PyTorch for the Mamba ops. bf16, no bitsandbytes.
spaces
torch==2.10.0
transformers>=5
accelerate
huggingface_hub>=1.2
sentencepiece
textual>=1.0
rich>=13.0
pyyaml>=6.0
httpx>=0.27