fix(docker): add --no-build-isolation to pip download for mamba-ssm

Fix HF Spaces build timeout: limit mamba_ssm to sm_89 only

Cap CUDA-extension build parallelism (fix HF Spaces OOMKill, exit 137)

Add Gradio head-to-head baseline (V1/V2/V3) and bench Space entrypoint

Fix TF32 API

Fix

Fix selective scam

Fix nvcc check syntax in Dockerfile

Fix Dockerfile syntax: Remove invalid bash conditionals from RUN commands

Switch Dockerfile to CUDA 12.1 + cu121 (recommended), use venv, add nvcc check, verify extensions at build time

Final check: Ensure setup.sh is executable in Dockerfile

Docker: Add OpenCV libs, use PYTORCH_ALLOC_CONF only with max_split_size_mb=128, create postBuild for Python Spaces fallback

Fix: Remove unsupported max_split_size_mb=128, use PYTORCH_ALLOC_CONF instead of deprecated PYTORCH_CUDA_ALLOC_CONF

Fix: PyTorch 2.9.0 cu128, simplified setup.sh with fatal errors, TF32=tf32, cuDNN always on, TORCH_CUDA_ARCH_LIST=8.9

update Dockerfile for PyTorch index

Production build strategy: Docker support, pinned dependencies, OOM guardrails, optimized defaults