# Runtime dependencies for the Gradio SDK / ZeroGPU Hugging Face Space.
# Read by HF Spaces' Gradio runner at build time.

gradio>=6.17.3
spaces>=0.30.0
# huggingface_hub: pinned <1.0 because transformers 4.55-4.57 (our pin below
# to match the MiniCPM bundled modeling code) requires huggingface_hub<1.0.
# That window also satisfies gradio's <2.0,>=0.33.5 requirement.
huggingface_hub>=0.34.0,<1.0
pydantic>=2.7
websockets>=14.0
fastapi
uvicorn

# Text LLM backend: HF transformers + PyTorch, MiniCPM4.1-8B is loaded
# CPU-side at import and moved to cuda inside the @spaces.GPU function.
#
# transformers pin: openbmb/MiniCPM4.1-8B's bundled modeling_minicpm.py
# imports CacheLayerMixin/DynamicLayer (added ~4.55) AND
# is_torch_fx_available from utils.import_utils (removed in 5.0). The 4.55-
# 4.57 window is the only range where both symbols coexist; outside it the
# KV-cache path is broken and we have to fall back to O(n^2) decoding.
torch>=2.4.0
transformers>=4.55.0,<5.0
accelerate>=1.0.0
sentencepiece