Update for RTX 5090 (Blackwell sm_120) support

- Switch to PyTorch nightly with CUDA 12.8 (required for sm_120)
- Target TORCH_CUDA_ARCH_LIST="12.0" for Blackwell
- Remove nunchaku (incompatible with PyTorch nightly)
- Use latest SageAttention (has sm_120 kernel support)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Dockerfile

@@ -1,4 +1,4 @@
-# ComfyUI RunPod Serverless - CUDA 12.8.1, Python 3.12, PyTorch 2.8.0
+# ComfyUI RunPod Serverless - CUDA 12.8.1, Python 3.12, PyTorch nightly (RTX 5090/Blackwell)
 FROM nvidia/cuda:12.8.1-devel-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
@@ -44,34 +44,30 @@ RUN update-alternatives --install /usr/bin/python python /usr/bin/python3.12 1 &
     python3.12 -m ensurepip --upgrade && \
     python3.12 -m pip install --upgrade pip setuptools wheel
 
-# Install PyTorch 2.8.0+cu128 and triton 3.4.0 in single layer
+# Install PyTorch nightly with CUDA 12.8 (required for RTX 5090/Blackwell sm_120)
-RUN pip install \
+RUN pip install --pre \
-    torch==2.8.0+cu128 \
+    torch \
-    torchvision==0.23.0+cu128 \
+    torchvision \
-    torchaudio==2.8.0+cu128 \
+    torchaudio \
-    triton==3.4.0 \
+    --index-url https://download.pytorch.org/whl/nightly/cu128 && \
-    --index-url https://download.pytorch.org/whl/cu128 && \
     rm -rf /root/.cache/pip /tmp/*
 
-# Install nunchaku from GitHub wheel
+# Note: nunchaku disabled - requires stable PyTorch, not compatible with nightly
-RUN pip install https://github.com/nunchaku-tech/nunchaku/releases/download/v1.0.2/nunchaku-1.0.2+torch2.8-cp312-cp312-linux_x86_64.whl
+# RUN pip install https://github.com/nunchaku-tech/nunchaku/releases/download/v1.0.2/nunchaku-1.0.2+torch2.8-cp312-cp312-linux_x86_64.whl
 
 # Install key dependencies before SageAttention
 COPY requirements.txt /tmp/requirements.txt
 RUN pip install -r /tmp/requirements.txt && rm -rf /root/.cache/pip
 
-# Compile SageAttention 2.2 from source with no build isolation
+# Compile SageAttention from source for RTX 5090 (Blackwell sm_120)
 WORKDIR /tmp
 ENV EXT_PARALLEL=2
 ENV NVCC_APPEND_FLAGS="--threads 2"
 ENV MAX_JOBS=4
-# Target RunPod GPU architectures:
+# Target RTX 5090 (Blackwell) - sm_120 = compute capability 12.0
-# 8.0 = A100, 9.0 = H100/H200
+ENV TORCH_CUDA_ARCH_LIST="12.0"
-# Note: 8.6/8.9 excluded due to SageAttention sm90 kernel cross-compile issues
-ENV TORCH_CUDA_ARCH_LIST="8.0;9.0"
 RUN git clone https://github.com/thu-ml/SageAttention.git && \
     cd SageAttention && \
-    git checkout 2aecfa8 && \
     pip install --no-build-isolation . && \
     cd / && rm -rf /tmp/SageAttention /root/.cache/pip
 
@@ -93,7 +89,6 @@ RUN git clone https://github.com/ltdrdata/ComfyUI-Manager.git && \
     git clone https://github.com/willmiao/ComfyUI-Lora-Manager.git && \
     git clone https://github.com/city96/ComfyUI-GGUF.git && \
     git clone https://github.com/Fannovel16/ComfyUI-Frame-Interpolation.git && \
-    git clone https://github.com/nunchaku-tech/ComfyUI-nunchaku.git && \
     git clone https://github.com/evanspearman/ComfyMath.git && \
     git clone https://github.com/ssitu/ComfyUI_UltimateSDUpscale.git && \
     git clone https://github.com/pythongosssss/ComfyUI-Custom-Scripts.git && \
@@ -106,7 +101,6 @@ RUN (cd ComfyUI-KJNodes && pip install -r requirements.txt || true) && \
     (cd ComfyUI-VideoHelperSuite && pip install -r requirements.txt || true) && \
     (cd ComfyUI-GGUF && pip install -r requirements.txt || true) && \
     (cd ComfyUI-Frame-Interpolation && pip install -r requirements.txt || true) && \
-    (cd ComfyUI-nunchaku && pip install -r requirements.txt || true) && \
     (cd ComfyUI-Easy-Use && pip install -r requirements.txt || true) && \
     rm -rf /root/.cache/pip /tmp/*