From 87324564c637b793a0568c17d5c072b1f776d263 Mon Sep 17 00:00:00 2001
From: MickLesk <mickey.leskowitz@gmail.com>
Date: Sat, 7 Feb 2026 18:03:16 +0100
Subject: [PATCH] Make ComfyUI PyTorch versions dynamic from upstream README

Instead of hardcoding CUDA/ROCm/XPU versions, now dynamically extracts
the recommended PyTorch wheel URLs from ComfyUI's README.md at install time.

Benefits:
- Automatically stays up-to-date with upstream recommendations
- No manual updates needed when ComfyUI changes PyTorch versions
- Fallback to known-good versions (cu130, rocm6.4, xpu) if fetch fails

Implementation:
- Fetches https://raw.githubusercontent.com/comfyanonymous/ComfyUI/master/README.md
- Parses stable (non-nightly) pip install commands with regex
- Gracefully degrades to hardcoded fallbacks on network/parse errors
---
 install/comfyui-install.sh | 30 +++++++++++++++++++++++++++---
 1 file changed, 27 insertions(+), 3 deletions(-)

diff --git a/install/comfyui-install.sh b/install/comfyui-install.sh
index a1d9168dc..b1763f9d0 100644
--- a/install/comfyui-install.sh
+++ b/install/comfyui-install.sh
@@ -36,27 +36,51 @@ PYTHON_VERSION="3.12" setup_uv
 fetch_and_deploy_gh_release "ComfyUI" "comfyanonymous/ComfyUI" "tarball" "latest" "/opt/ComfyUI"
 
 msg_info "Python dependencies"
+
+# Extract PyTorch installation URLs from ComfyUI README for latest versions
+# Fallback to hardcoded values if extraction fails
+pytorch_nvidia_url="https://download.pytorch.org/whl/cu130"
+pytorch_amd_url="https://download.pytorch.org/whl/rocm6.4"
+pytorch_intel_url="https://download.pytorch.org/whl/xpu"
+
+if command -v curl >/dev/null 2>&1; then
+  readme_content=$(curl -fsSL "https://raw.githubusercontent.com/comfyanonymous/ComfyUI/master/README.md" 2>/dev/null || true)
+  if [[ -n "$readme_content" ]]; then
+    # Extract NVIDIA CUDA URL (looking for stable, not nightly)
+    nvidia_url=$(echo "$readme_content" | grep -oP 'pip install.*?--extra-index-url\s+\Khttps://download\.pytorch\.org/whl/cu\d+' | head -1 || true)
+    [[ -n "$nvidia_url" ]] && pytorch_nvidia_url="$nvidia_url"
+    
+    # Extract AMD ROCm URL (stable version, not nightly)
+    amd_url=$(echo "$readme_content" | grep -oP 'pip install.*?--index-url\s+\Khttps://download\.pytorch\.org/whl/rocm[\d.]+' | grep -v 'nightly' | head -1 || true)
+    [[ -n "$amd_url" ]] && pytorch_amd_url="$amd_url"
+    
+    # Extract Intel XPU URL
+    intel_url=$(echo "$readme_content" | grep -oP 'pip install.*?--index-url\s+\Khttps://download\.pytorch\.org/whl/xpu' | head -1 || true)
+    [[ -n "$intel_url" ]] && pytorch_intel_url="$intel_url"
+  fi
+fi
+
 $STD uv venv "/opt/ComfyUI/venv"
 if [[ "${comfyui_gpu_type,,}" == "nvidia" ]]; then
   $STD uv pip install \
     torch \
     torchvision \
     torchaudio \
-    --extra-index-url "https://download.pytorch.org/whl/cu130" \
+    --extra-index-url "$pytorch_nvidia_url" \
     --python="/opt/ComfyUI/venv/bin/python"
 elif [[ "${comfyui_gpu_type,,}" == "amd" ]]; then
   $STD uv pip install \
     torch \
     torchvision \
     torchaudio \
-    --index-url "https://download.pytorch.org/whl/rocm6.4" \
+    --index-url "$pytorch_amd_url" \
     --python="/opt/ComfyUI/venv/bin/python"
 elif [[ "${comfyui_gpu_type,,}" == "intel" ]]; then
   $STD uv pip install \
     torch \
     torchvision \
     torchaudio \
-    --index-url "https://download.pytorch.org/whl/xpu" \
+    --index-url "$pytorch_intel_url" \
     --python="/opt/ComfyUI/venv/bin/python"
 fi
 $STD uv pip install -r "/opt/ComfyUI/requirements.txt" --python="/opt/ComfyUI/venv/bin/python"