Spaces:
Running on Zero
Running on Zero
Add prebuilt flashinfer sm120 cubins (skip FP4 JIT on cold boot)
Browse files- requirements.txt +6 -0
requirements.txt
CHANGED
|
@@ -8,6 +8,12 @@ fastvideo @ git+https://github.com/hao-ai-lab/FastVideo.git@376e445f0e8d192c2d9a
|
|
| 8 |
|
| 9 |
# NVFP4 FP4 GEMMs (packed linears + decoder) on Blackwell.
|
| 10 |
flashinfer-python
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
spaces
|
| 13 |
pillow
|
|
|
|
| 8 |
|
| 9 |
# NVFP4 FP4 GEMMs (packed linears + decoder) on Blackwell.
|
| 10 |
flashinfer-python
|
| 11 |
+
# Prebuilt flashinfer FP4 cubins for sm120 (RTX PRO 6000). flashinfer uses these
|
| 12 |
+
# instead of JIT-compiling fp4_quantization_120f / fp4_gemm_cutlass_sm120 on the
|
| 13 |
+
# first forward, cutting ~175 s off every cold boot (the FastVideo analog of the
|
| 14 |
+
# official H3 Spaces' prebuilt-AOTI download).
|
| 15 |
+
--find-links https://flashinfer.ai/whl/cu130/flashinfer-jit-cache-sm120f/
|
| 16 |
+
flashinfer-jit-cache-sm120f==0.7.2+cu130
|
| 17 |
|
| 18 |
spaces
|
| 19 |
pillow
|