Fix ZeroGPU Blackwell AttributeError and CUDA kernel error

- Add missing use_fp8_gemm/quant_type/include_patterns attrs to the
ArgsNamespace shim passed to initialize_infer_state (upstream's
InferState dataclass reads them unconditionally).
- Pin torch/torchvision/torchaudio to a compatible triplet (2.10.0 /
0.25.0 / 2.10.0); the previous torchvision==0.21.0 pin forced pip to
downgrade torch to 2.6.0, which predates Blackwell (sm_120) support
and caused 'CUDA error: no kernel image is available'.
- Install flash-attn from the prebuilt Blackwell (sm_120) wheel instead
of the plain PyPI wheel, which lacks sm_120 kernels.

app.py

@@ -1,5 +1,10 @@
 import subprocess
-subprocess.run('pip install flash-attn==2.7.4.post1 --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+# Prebuilt flash-attn wheel matching torch==2.10.0 / cu128 / cp310, built with Blackwell (sm_120) kernels.
+# The plain PyPI wheel for flash-attn has no sm_120 kernels and fails at runtime with
+# "CUDA error: no kernel image is available for execution on the device" on ZeroGPU's
+# RTX PRO 6000 Blackwell hardware.
+FLASH_ATTN_WHEEL = "https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels/resolve/main/wheels/pt210-cu128-cp310/flash_attn-2.7.4.post1-cp310-cp310-linux_x86_64.whl"
+subprocess.run(f'pip install {FLASH_ATTN_WHEEL}', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
 import os
 import sys
@@ -232,7 +237,10 @@ class ArgsNamespace:
         self.cache_end_step = 45
         self.total_steps = 50
         self.cache_step_interval = 4
-        
+        self.use_fp8_gemm = False
+        self.quant_type = "fp8-per-token-sgl"
+        self.include_patterns = "double_blocks"
+
 
 initialize_infer_state(ArgsNamespace())
 

requirements.txt

@@ -11,10 +11,11 @@ pillow==11.3.0
 imageio==2.37.0
 imageio-ffmpeg==0.6.0
 omegaconf>=2.3.0
-torchaudio==2.6.0
+torch==2.10.0
+torchaudio==2.10.0
 diffusers==0.35.0
 safetensors==0.4.5
-torchvision==0.21.0
+torchvision==0.25.0
 qwen-vl-utils==0.0.8
 huggingface-hub==0.34.0
 huggingface_hub[cli]