Install flashinfer jit-cache shim + sm120f provider (skip FP4 JIT)

requirements.txt

@@ -8,11 +8,15 @@ fastvideo @ git+https://github.com/hao-ai-lab/FastVideo.git@376e445f0e8d192c2d9a
 
 # NVFP4 FP4 GEMMs (packed linears + decoder) on Blackwell.
 flashinfer-python
-# Prebuilt flashinfer FP4 cubins for sm120 (RTX PRO 6000). flashinfer uses these
-# instead of JIT-compiling fp4_quantization_120f / fp4_gemm_cutlass_sm120 on the
-# first forward, cutting ~175 s off every cold boot (the FastVideo analog of the
-# official H3 Spaces' prebuilt-AOTI download).
+# Prebuilt flashinfer FP4 cubins for sm120 (RTX PRO 6000): skip the ~175 s
+# fp4_quantization_120f / fp4_gemm_cutlass_sm120 JIT compile on every cold boot
+# (the FastVideo analog of the official H3 Spaces' prebuilt-AOTI download).
+# Two packages: the shim (top-level `flashinfer_jit_cache` with __version__ and
+# provider discovery) + the sm120f provider (the actual cubins + entry point).
+# Installing the provider alone leaves a namespace package with no __version__.
+--find-links https://flashinfer.ai/whl/cu130/flashinfer-jit-cache/
 --find-links https://flashinfer.ai/whl/cu130/flashinfer-jit-cache-sm120f/
+flashinfer-jit-cache==0.7.2+cu130
 flashinfer-jit-cache-sm120f==0.7.2+cu130
 
 spaces