Comfy-Org / Comfy-Org/Nvidia_RTX_Nodes_ComfyUI

Cant run inside wsl2 with 5060 ti 16gb is this only for w11?

Open
#20 6 comments 6 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
620
Forks
38
PR merge metrics
No merged PRs in 30d

Description

Here's a complete handoff summary:

---

**NVIDIA RTX Video Super Resolution (nvvfx) on WSL2 — Debug Session Summary**

**Goal:** Get the `comfyui_nvidia_rtx_nodes` ComfyUI custom node working on WSL2 Ubuntu with RTX 5060 Ti 16GB.

**Environment:**
- Windows 11 + WSL2 Ubuntu (kernel 6.6.87.2-microsoft-standard-WSL2)
- GPU: NVIDIA GeForce RTX 5060 Ti 16GB (+ RTX 3060 12GB secondary)
- Windows Driver: 595.79 (upgraded from 591.74 during session)
- WSL2 Linux Driver: 570.195.03
- Python 3.10.18 conda env named `comfyui`
- ComfyUI v0.16.4
- nvidia-vfx 0.1.0.1 (only version available)
- Launch script: `~/runcomfyui2.sh`

---

**The Node:**
`https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI`
- Imports `nvvfx` (not `nvidia.vfx` — the node is correct)
- Uses `nvvfx.VideoSuperRes` with `comfy_api.latest` async extension system
- Node loads fine at ComfyUI startup (0.0 seconds, no error)
- Crash happens at **execution time** when `sr.load()` is called

---

**What Was Ruled Out:**
- ✅ Package installs correctly under `nvvfx` namespace
- ✅ `import nvvfx` works
- ✅ `VideoSuperRes()` object creation works
- ✅ All `.so` dependencies resolved once `LD_LIBRARY_PATH` set to `~/anaconda3/envs/comfyui/lib/python3.10/site-packages/nvvfx/libs`
- ✅ TensorRT libs (`libnvinfer.so.10` etc.) are bundled inside the pip package itself
- ✅ NGX runtime (`libnvidia-ngx.so`) present in WSL2
- ✅ VSR feature DLLs (`nvsvsr.dll`, `nvvitvsr.dll`) present in WSL2 driver directory
- ✅ `/dev/dxg` opens successfully (WSL2 GPU path works)
- ✅ After driver update to 595.79, CUDA 13.2 available
- ❌ `sr.load()` silently calls `exit_group(1)` — no Python exception ever raised

---

**Root Cause (confirmed via `strace -f`):**

The failure is a **driver/pip package version mismatch**:

1. `sr.load()` spawns an NGX worker process
2. Worker opens: `/usr/lib/wsl/drivers/nvmdsi.inf_amd64_d39f1ab212fcacea/nvcubins.bin` (595.79 driver's NGX feature container, 73MB)
3. Worker seeks to offset 64251072, fails to parse it, calls `exit_group(1)`

The bundled `libnvngxruntime.so` was built from:
```
/dvs/p4/build/sw/rel/gpu_drv/r570/r573_76/drivers/ngx/
```
It **cannot parse the r595-format `nvcubins.bin`** from driver 595.79.

Attempting to swap in the r595 `libnvidia-ngx.so.1` from the WSL driver directory loads successfully but then `libnvVFXVideoSuperRes.so` fails with:
```
undefined symbol: _ZN19NvNGXRuntimeManager11GetInstanceEv
```
Meaning **both** `libnvngxruntime.so` and `libnvVFXVideoSuperRes.so` are r570-era and the entire VFX SDK in the pip package needs to be rebuilt against the r595 NGX API.

---

**WSL2-Specific Issues Found and Partially Worked Around:**

1. **`nvidia-modprobe` stub required** — the SDK spawns it with empty environment (0 vars), bypassing LD_PRELOAD. Real modprobe exits 1 on WSL2 (no kernel module). Fix: replace `/usr/bin/nvidia-modprobe` with a bash stub that exits 0. Note: gets overwritten on driver updates, must be recreated.
```bash
sudo tee /usr/bin/nvidia-modprobe << 'EOF'
#!/bin/bash
exit 0
EOF
sudo chmod +x /usr/bin/nvidia-modprobe
```

2. **`/usr/share/nvidia/ngx` must be created manually** — doesn't exist on WSL2:
```bash
sudo mkdir -p /usr/share/nvidia/ngx
sudo chmod 777 /usr/share/nvidia/ngx
```

3. **`/proc/driver/nvidia/capabilities/mig/monitor`** — doesn't exist on WSL2, appears non-fatal after modprobe stub is in place

4. **`LD_LIBRARY_PATH` must be set** for the bundled `.so` files — added to `runcomfyui2.sh`:
```bash
export LD_LIBRARY_PATH=~/anaconda3/envs/comfyui/lib/python3.10/site-packages/nvvfx/libs:$LD_LIBRARY_PATH
```

---

**Current Status:**
Blocked. The `nvidia-vfx 0.1.0.1` pip package ships r570-era binaries incompatible with the r595 driver. No newer pip package version exists. Bug reported to NVIDIA Developer Support, **case ID: 5986059**.

**Waiting for:** NVIDIA to release an updated `nvidia-vfx` package built against the r595 NGX SDK.

**Next steps when a new package version appears:**
1. `pip install --upgrade nvidia-vfx --index-url https://pypi.nvidia.com`
2. Recreate the `nvidia-modprobe` stub if it was overwritten
3. Verify `LD_LIBRARY_PATH` still set in `runcomfyui2.sh`
4. Test `sr.load()` directly before launching ComfyUI

Contributor guide

No contributing guide indexed for this repository

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.