From d38de7de56cae9e0ba602d87878c6cb26081a721 Mon Sep 17 00:00:00 2001
From: TroyHernandez <troy.hernandez@pm.me>
Date: Fri, 16 Jan 2026 22:19:12 -0600
Subject: [PATCH] Add Dockerfile for Blackwell GPU support (RTX 50xx)

Blackwell architecture (sm_120) requires:
- CUDA 12.8+ (12.4 doesn't support sm_120)
- PyTorch 2.7.0+ (first version with Blackwell support)

Key changes from main Dockerfile:
- Base image: nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04
- PyTorch: 2.7.1+cu128
- CUDA arch: 12.0 (sm_120)

Tested on RTX 5060 Ti 16GB with:
- LTX-2 19B Distilled: ~3.5 min per video
- VACE 1.3B: ~10 min per video
- Profile 4 (low VRAM mode) works correctly

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
---
 Dockerfile.blackwell | 75 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 75 insertions(+)
 create mode 100644 Dockerfile.blackwell

diff --git a/Dockerfile.blackwell b/Dockerfile.blackwell
new file mode 100644
index 000000000..fbb40c0d1
--- /dev/null
+++ b/Dockerfile.blackwell
@@ -0,0 +1,75 @@
+FROM nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04
+
+# Blackwell (RTX 50xx) requires sm_120 architecture
+ARG CUDA_ARCHITECTURES="12.0"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Install system dependencies
+RUN apt update && \
+    apt install -y \
+    python3 python3-pip git wget curl cmake ninja-build \
+    libgl1 libglib2.0-0 ffmpeg && \
+    apt clean
+
+WORKDIR /workspace
+
+COPY requirements.txt .
+
+# Upgrade pip first
+RUN pip install --upgrade pip setuptools wheel
+
+# Install PyTorch 2.7.1 with CUDA 12.8 for Blackwell
+RUN pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128
+
+# Install requirements
+RUN pip install -r requirements.txt
+
+# Install SageAttention from git (patch GPU detection for Blackwell)
+ENV TORCH_CUDA_ARCH_LIST="${CUDA_ARCHITECTURES}"
+ENV FORCE_CUDA="1"
+ENV MAX_JOBS="1"
+
+COPY <<EOF /tmp/patch_setup.py
+import os
+with open('setup.py', 'r') as f:
+    content = f.read()
+
+# Get architectures from environment variable
+arch_list = os.environ.get('TORCH_CUDA_ARCH_LIST')
+arch_set = '{' + ', '.join([f'"{arch}"' for arch in arch_list.split(';')]) + '}'
+
+# Replace the GPU detection section
+old_section = '''compute_capabilities = set()
+device_count = torch.cuda.device_count()
+for i in range(device_count):
+    major, minor = torch.cuda.get_device_capability(i)
+    if major < 8:
+        warnings.warn(f"skipping GPU {i} with compute capability {major}.{minor}")
+        continue
+    compute_capabilities.add(f"{major}.{minor}")'''
+
+new_section = 'compute_capabilities = ' + arch_set + '''
+print(f"Manually set compute capabilities: {compute_capabilities}")'''
+
+content = content.replace(old_section, new_section)
+
+with open('setup.py', 'w') as f:
+    f.write(content)
+EOF
+
+RUN git clone https://github.com/thu-ml/SageAttention.git /tmp/sageattention && \
+    cd /tmp/sageattention && \
+    python3 /tmp/patch_setup.py && \
+    pip install --no-build-isolation .
+
+RUN useradd -u 1000 -ms /bin/bash user
+
+RUN chown -R user:user /workspace
+
+RUN mkdir /home/user/.cache && \
+    chown -R user:user /home/user/.cache
+
+COPY entrypoint.sh /workspace/entrypoint.sh
+
+ENTRYPOINT ["/workspace/entrypoint.sh"]