transformers
peft
accelerate
gradio
pillow
torch
torchvision
# Pure-Python + Triton fast path for Qwen3.5 Gated DeltaNet — no nvcc needed.
# (causal-conv1d and flash-attn require CUDA Toolkit nvcc to build, which the
# default HF Spaces docker image lacks. flash-linear-attention alone gives most
# of the speedup via its Triton-compiled kernels.)
flash-linear-attention