[ArtStudio M1] Device abstraction in art_pipe core — cuda/mps/cpu seam (upstream to ProArt) #356

New Issue

Closed

opened 2026-07-04 05:27:27 +00:00 by spikerj · 2 comments

spikerj commented 2026-07-04 05:27:27 +00:00

Owner

Copy Link

Copy Source

Sub-ticket of epic #346, Mac workstream. Foundation for #355 and the M2–M4 tickets. Based on the MPS portability audit (comment on #355).

What: art_pipe has no device abstraction — every backend independently does cuda-if-available-else-cpu and installs CUDA wheels unconditionally. Add a cuda | mps | cpu seam so macOS (Apple Silicon, PyTorch MPS) is a first-class target. This is generic, upstreamable work — plan to PR it to Eric's ProArt repo rather than fork.

Scope (exact touchpoints from the audit):

• src/artpipe/hardware.py — add torch.backends.mps.is_available() branch to detect() (:183-207); report unified memory as VRAM; no nvidia-smi path on Darwin (:104-166)
• src/artpipe/main.py::_check_gpu_compat (:22-70) — treat MPS as GPU-available; skip sm/TORCH_CUDA_ARCH_LIST checks on Darwin
• src/artpipe/torch_compat.py — Darwin branch: default PyPI torch (ships MPS), no cu-index mapping
• src/artpipe/venv_setup.py (:257-320, defaults :8-9) + bootstrap.py (:23,78) — on Darwin ignore torch_index_url, install stock torch; per-OS manifest override convention (e.g. torch_index_url_darwin: "")
• Shared device helper (e.g. artpipe.device.resolve() returning device str + preferred dtype) and adopt it in the 13 backends that hardcode "cuda": flux_schnell.py:67, sdxl_lightning.py:58, sdxl_turbo.py:41, sd_turbo_tex.py:60, sf3d.py:44, text3d.py:62, triposr.py:46, triposg.py:46, instantmesh.py:68, shap_e.py:123, mdm.py:87 (int 0 — needs mapping), hymotion.py:101, rignet_infer.py:325; fix instantmesh.py:87 unconditional fp16
• cv/checks.py:79-83 + cv/analysis.py:203-217 — mps branch for SigLIP/CLIP
• Blender: src/artpipe/blender/headless/_helpers.py:137-165 — set cprefs.compute_device_type = "METAL" on macOS before refresh_devices() (today Cycles silently falls back to CPU)
• Guard torch.cuda.empty_cache() call sites (pattern from trellis-mac patches/mps_compat.py)

Note: no Metal kernels in this ticket — CUDA-extension backends stay CUDA-only and must fail with a clear "requires CUDA" message on Mac (stub pattern from trellis-mac backends/stubs.py), not ImportError crashes.

Acceptance: on an Apple-Silicon Mac, artpipe setup installs stock torch, artpipe launch passes the GPU check reporting MPS + unified memory, a Tier-A backend (e.g. sdxl_lightning) runs on mps, a blocked backend (e.g. triposg) fails with the clear message, and Blender Cycles renders on Metal. CUDA behavior on Linux unchanged.

Sub-ticket of epic #346, Mac workstream. Foundation for #355 and the M2–M4 tickets. Based on the MPS portability audit (comment on #355). **What:** art_pipe has no device abstraction — every backend independently does `cuda-if-available-else-cpu` and installs CUDA wheels unconditionally. Add a `cuda | mps | cpu` seam so macOS (Apple Silicon, PyTorch MPS) is a first-class target. This is generic, upstreamable work — plan to PR it to Eric's ProArt repo rather than fork. **Scope (exact touchpoints from the audit):** - [ ] `src/artpipe/hardware.py` — add `torch.backends.mps.is_available()` branch to `detect()` (:183-207); report unified memory as VRAM; no `nvidia-smi` path on Darwin (:104-166) - [ ] `src/artpipe/main.py::_check_gpu_compat` (:22-70) — treat MPS as GPU-available; skip sm/`TORCH_CUDA_ARCH_LIST` checks on Darwin - [ ] `src/artpipe/torch_compat.py` — Darwin branch: default PyPI torch (ships MPS), no cu-index mapping - [ ] `src/artpipe/venv_setup.py` (:257-320, defaults :8-9) + `bootstrap.py` (:23,78) — on Darwin ignore `torch_index_url`, install stock torch; per-OS manifest override convention (e.g. `torch_index_url_darwin: ""`) - [ ] Shared device helper (e.g. `artpipe.device.resolve()` returning device str + preferred dtype) and adopt it in the 13 backends that hardcode `"cuda"`: `flux_schnell.py:67`, `sdxl_lightning.py:58`, `sdxl_turbo.py:41`, `sd_turbo_tex.py:60`, `sf3d.py:44`, `text3d.py:62`, `triposr.py:46`, `triposg.py:46`, `instantmesh.py:68`, `shap_e.py:123`, `mdm.py:87` (int `0` — needs mapping), `hymotion.py:101`, `rignet_infer.py:325`; fix `instantmesh.py:87` unconditional fp16 - [ ] `cv/checks.py:79-83` + `cv/analysis.py:203-217` — mps branch for SigLIP/CLIP - [ ] Blender: `src/artpipe/blender/headless/_helpers.py:137-165` — set `cprefs.compute_device_type = "METAL"` on macOS before `refresh_devices()` (today Cycles silently falls back to CPU) - [ ] Guard `torch.cuda.empty_cache()` call sites (pattern from trellis-mac `patches/mps_compat.py`) **Note:** no Metal kernels in this ticket — CUDA-extension backends stay CUDA-only and must fail with a clear "requires CUDA" message on Mac (stub pattern from trellis-mac `backends/stubs.py`), not ImportError crashes. **Acceptance:** on an Apple-Silicon Mac, `artpipe setup` installs stock torch, `artpipe launch` passes the GPU check reporting MPS + unified memory, a Tier-A backend (e.g. sdxl_lightning) runs on `mps`, a blocked backend (e.g. triposg) fails with the clear message, and Blender Cycles renders on Metal. CUDA behavior on Linux unchanged.

spikerj added the enhancement label 2026-07-04 05:27:27 +00:00

spikerj referenced this issue 2026-07-04 05:27:47 +00:00

[ArtStudio M2] Mac-native pipeline profile — full Tier-A stage chain on MPS (concept → model → texture → rig → animate → export) #357

spikerj referenced this issue 2026-07-04 05:28:07 +00:00

[ArtStudio M3] Metal extension reuse — TRELLIS via trellis-mac, then mtldiffrast/mtlgemm for InstantMesh, Hunyuan paint, UniRig #358

spikerj referenced this issue 2026-07-04 05:28:45 +00:00

[ArtStudio Spike] Apple Silicon dev path — can trellis-mac (MPS port) back the modeling stage for local development? #355

spikerj referenced this issue 2026-07-04 05:28:59 +00:00

[Epic] ProArt 3D Art Studio — student asset pipeline (concept → model → texture → rig → animate → game-ready GLB) #346

spikerj commented 2026-07-04 06:11:39 +00:00

Author

Owner

Copy Link

Copy Source

Implemented on local branch feature/device-abstraction-mps in art_pipe (commit a186626, +384/−82 across 25 files). All scope items done: artpipe/device.py (resolve() cuda>mps>cpu + guarded empty_cache(), lazy-torch so the stdlib-only server still imports), hardware detect/live_stats MPS branch (Apple GPU + unified-memory-as-VRAM), _check_gpu_compat MPS path, Darwin torch-index handling in torch_compat/venv_setup/bootstrap (+ optional torch_index_url_darwin manifest key), all 13 hardcoded-cuda backends moved to the shared helper (incl. mdm int-device fix and instantmesh unconditional-fp16 fix), clear requires-CUDA guards on trellis/unirig/triposg/hunyuan3d_paint, SigLIP/CLIP mps branch, Blender Cycles METAL enable on macOS.

Verified on the M5 Max (128 GB): pytest 574 passed / 10 skipped / 0 failed; resolve() → ('mps', float16); hardware.detect() → "Apple M5 Max GPU", 131072 MB, mps_available: true; python -m artpipe serve /hardware endpoint returns the Apple GPU. Audit line-drift corrections and a real pre-existing bug fixed en route (detect() force-set cuda_available=True for any GPU entry) are noted in the commit.

Blocked on a decision: art_pipe's origin is Eric's server (clam.local:3000/EricSpiker/art_pipe) — not pushing to someone else's Gitea without Joey's go-ahead. Options: (a) push branch + open PR on Eric's Gitea if Joey has an account/fork there, (b) mirror art_pipe into git.spikersoft.com and PR there, then upstream later. Runtime model validation on MPS is #357 scope.

Working-tree note: the checkout has repo-wide CRLF noise (Windows-origin) plus an uncommitted README author-handle edit — left untouched; the commit was kept EOL-clean for upstreamability.

Implemented on local branch `feature/device-abstraction-mps` in art_pipe (commit `a186626`, +384/−82 across 25 files). All scope items done: `artpipe/device.py` (`resolve()` cuda>mps>cpu + guarded `empty_cache()`, lazy-torch so the stdlib-only server still imports), hardware detect/live_stats MPS branch (Apple GPU + unified-memory-as-VRAM), `_check_gpu_compat` MPS path, Darwin torch-index handling in torch_compat/venv_setup/bootstrap (+ optional `torch_index_url_darwin` manifest key), all 13 hardcoded-cuda backends moved to the shared helper (incl. mdm int-device fix and instantmesh unconditional-fp16 fix), clear requires-CUDA guards on trellis/unirig/triposg/hunyuan3d_paint, SigLIP/CLIP mps branch, Blender Cycles METAL enable on macOS. **Verified on the M5 Max (128 GB):** pytest 574 passed / 10 skipped / 0 failed; `resolve()` → `('mps', float16)`; `hardware.detect()` → "Apple M5 Max GPU", 131072 MB, `mps_available: true`; `python -m artpipe serve` `/hardware` endpoint returns the Apple GPU. Audit line-drift corrections and a real pre-existing bug fixed en route (`detect()` force-set `cuda_available=True` for any GPU entry) are noted in the commit. **Blocked on a decision:** art_pipe's `origin` is Eric's server (`clam.local:3000/EricSpiker/art_pipe`) — not pushing to someone else's Gitea without Joey's go-ahead. Options: (a) push branch + open PR on Eric's Gitea if Joey has an account/fork there, (b) mirror art_pipe into git.spikersoft.com and PR there, then upstream later. Runtime model validation on MPS is #357 scope. Working-tree note: the checkout has repo-wide CRLF noise (Windows-origin) plus an uncommitted README author-handle edit — left untouched; the commit was kept EOL-clean for upstreamability.

spikerj referenced this issue 2026-07-04 16:22:25 +00:00

[ArtStudio M4] Entire stack on the MacBook — native ArtPipeProcessor (no GPU in Docker on macOS), compose.mac, dev-mode GpuCoordinator #359

spikerj commented 2026-07-04 23:30:34 +00:00

Author

Owner

Copy Link

Copy Source

Resolved in spikersoft-artpipe PR #1 (merged to main — ProArt now lives at spikerj/spikersoft-artpipe per the mirror decision). The cuda/mps/cpu device seam: artpipe.device.resolve() adopted across all 13 hardcoded-cuda backends, MPS-aware hardware detect/live_stats/_check_gpu_compat, Darwin torch installs (stock PyPI, manifest override key), clear requires-CUDA guards on trellis/unirig/triposg/hunyuan3d_paint, SigLIP/CLIP mps, Blender Cycles METAL enable. Validated across six doc slices on the M5; 682 pytest green at merge; CUDA behavior unchanged. Upstreaming to Eric's original repo is now optional/at-leisure since ProArt is consumed as SpikerSoft work. Closing.

Resolved in spikersoft-artpipe PR #1 (merged to `main` — ProArt now lives at `spikerj/spikersoft-artpipe` per the mirror decision). The cuda/mps/cpu device seam: `artpipe.device.resolve()` adopted across all 13 hardcoded-cuda backends, MPS-aware hardware detect/live_stats/_check_gpu_compat, Darwin torch installs (stock PyPI, manifest override key), clear requires-CUDA guards on trellis/unirig/triposg/hunyuan3d_paint, SigLIP/CLIP mps, Blender Cycles METAL enable. Validated across six doc slices on the M5; 682 pytest green at merge; CUDA behavior unchanged. Upstreaming to Eric's original repo is now optional/at-leisure since ProArt is consumed as SpikerSoft work. Closing.

spikerj closed this issue 2026-07-04 23:31:01 +00:00

Sign in to join this conversation.

No Branch/Tag Specified

Branches Tags

main

No results found.

Labels

Clear labels

agentic

Actionable surfaced by the agentic architecture & security review

bug

Something is not working

duplicate

This issue or pull request already exists

enhancement

New feature

help wanted

Need some help

invalid

Something is wrong

question

More information is needed

sonarqube

Code quality issue surfaced by SonarQube (bug / code smell)

wontfix

This won't be fixed

No labels enhancement

Milestone

No items

No Milestone

Projects

Clear projects

No projects

Assignees

Clear assignees

spikerj (Joseph Spiker)

No Assignees

1 Participants

Notifications

Subscribe

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: spikerj/spikersoft-issues#356