feat: detect AMD/ROCm GPUs and install ROCm PyTorch for extensions

docs/rocm-amd-setup.md

@@ -0,0 +1,66 @@
+# Running Modly on AMD GPUs (ROCm)
+
+Modly's extension installer only auto-detects **NVIDIA** GPUs (it probes
+`nvidia-smi`). On an AMD system it falls through to the CPU/CUDA path, so an
+extension's `torch` install either fails outright or installs a CPU build — no
+GPU acceleration.
+
+With a small app-side fix plus two environment prerequisites, you can run
+Modly's **pure-PyTorch** extensions on an AMD GPU via **ROCm**.
+
+**Tested on:** AMD Radeon RX 7600 (`gfx1102`) · Ubuntu 26.04 · ROCm 6.3 ·
+Hunyuan3D 2 Mini (image → 3D mesh, end-to-end).
+
+## Prerequisites
+
+### 1. Python 3.10–3.13 (NOT 3.14)
+Ubuntu 26.04 ships **Python 3.14** by default, and PyTorch publishes **no
+wheels for 3.14**. Every extension install fails with
+`No matching distribution found for torch`. Install 3.12 (no sudo, via
+[uv](https://docs.astral.sh/uv/)):
+
+```bash
+curl -LsSf https://astral.sh/uv/install.sh | sh
+uv python install 3.12
+ln -sf "$(uv python find 3.12)" ~/.local/bin/python3.12   # put python3.12 on PATH
+```
+
+Modly's `findSystemPython()` tries `python3.12` first, so on the next launch it
+rebuilds its backend venv with 3.12.
+
+### 2. ROCm
+Install ROCm. [TheRock](https://github.com/ROCm/TheRock) (AMD's modern ROCm
+distribution) **officially supports `gfx1102`** (RX 7600). Verify your card is
+seen:
+
+```bash
+rocminfo | grep -i gfx        # expect a gfx110x agent
+rocm-smi --showproductname    # GFX Version: gfx1102
+```
+
+### 3. The app-side fix (this change)
+`electron/main/ipc-handlers.ts` now calls `detectRocmGpu()`: when **no NVIDIA
+GPU** is present but ROCm is installed and `rocminfo` reports an AMD GPU
+(`gfx*`), the app passes `torch_flavor: 'rocm'` to each extension's `setup.py`.
+Extension setup scripts that accept `torch_flavor` (e.g. Hunyuan3D 2 Mini) then
+install PyTorch from the ROCm wheel index instead of CUDA/CPU.
+
+> Behavior is unchanged for NVIDIA and Apple Silicon users — the ROCm branch
+> only triggers on Linux with no NVIDIA GPU + ROCm present.
+
+## Install and run
+1. Launch Modly (it recreates its backend venv with Python 3.12 the first time).
+2. **Extensions → Install from GitHub**, paste e.g.
+   `https://github.com/lightningpixel/modly-hunyuan3d-mini-extension`.
+3. The setup log should show `[setup] -> PyTorch + ROCm 7.2` (not `cu118`/CPU).
+4. Download the model weights, then **Generate**.
+
+## What works on AMD — and what doesn't
+- ✅ **Pure-PyTorch extensions:** Hunyuan3D 2 Mini / Turbo / Fast, Trellis2 GGUF.
+- ❌ **TripoSG:** depends on `diso`, a CUDA-only C++ extension. It cannot compile
+  or run on AMD (no `nvcc`, and the prebuilt wheel is NVIDIA-only). Use the
+  Hunyuan3D variants instead.
+- ℹ️ PyTorch-ROCm reports the device as **`cuda`** (e.g. `Loaded on cuda`).
+  That's the AMD GPU via HIP — expected, not an error.
+- ℹ️ The RX 7600 has **8 GB VRAM**; prefer the "Mini" variants and keep an eye
+  on memory.

electron/main/ipc-handlers.ts

@@ -19,7 +19,7 @@ import { checkSetupNeeded, markSetupDone, runFullSetup, getVenvPythonExe, ensure
 import { logger } from './logger'
 import { getProcessRunner, getPythonProcessRunner, getExtPythonExe, terminateProcessRunner, terminateAllProcessRunners } from './process-runner'
 import { getBuiltinExtensionsDir } from './builtin-sync'
-import { spawn, execFile } from 'child_process'
+import { spawn, execFile, execFileSync } from 'child_process'
 import {
   EXT_BACKUP_PREFIX,
   EXT_INCOMPLETE_MARKER,
@@ -87,6 +87,25 @@ function detectGpuInfo(): Promise<GpuInfo> {
   })
 }
 
+// ─── ROCm detect (AMD GPU) ────────────────────────────────────────────────────
+// True when ROCm userspace is installed and rocminfo reports at least one AMD
+// GPU compute agent (gfx*). Used to select the ROCm PyTorch wheel flavor so
+// AMD GPUs install the rocm build instead of falling through to CUDA/CPU.
+function detectRocmGpu(): boolean {
+  if (process.platform !== 'linux') return false
+  if (!existsSync('/opt/rocm') && !existsSync('/usr/bin/rocminfo')) return false
+  try {
+    const out = execFileSync('rocminfo', [], {
+      encoding: 'utf8',
+      timeout: 5000,
+      stdio: ['ignore', 'pipe', 'ignore'],
+    })
+    return /gfx\d{3}/.test(out)
+  } catch {
+    return false
+  }
+}
+
 // ─── Run an extension's setup.py directly (no FastAPI needed) ─────────────────
 
 function runExtensionSetup(
@@ -108,12 +127,16 @@ function runExtensionSetup(
     try { mkdirSync(pipCacheDir, { recursive: true }) } catch { /* pip creates it too */ }
 
     const accelerator = process.platform === 'darwin' && process.arch === 'arm64' ? 'mps' : gpuSm > 0 ? 'cuda' : 'cpu'
+    // AMD GPU under ROCm: use the ROCm PyTorch wheel flavor. Falls back to CUDA
+    // when an NVIDIA GPU is present (gpuSm > 0) or ROCm is unavailable.
+    const torchFlavor = process.platform === 'linux' && gpuSm === 0 && detectRocmGpu() ? 'rocm' : 'cuda'
     const args = JSON.stringify({
       python_exe: pythonExe,
       ext_dir: extDir,
       gpu_sm: gpuSm,
       cuda_version: cudaVersion,
       accelerator,
+      torch_flavor: torchFlavor,
       platform: process.platform,
       arch: process.arch,
     })