spikerj/spikersoft-issues
1
0
Fork 0
Code Issues 196 Pull Requests Actions Packages Projects Releases Wiki Activity

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

New Issue
Closed
opened 2026-07-04 05:27:47 +00:00 by spikerj · 3 comments
spikerj commented 2026-07-04 05:27:47 +00:00
Owner
Copy Link
Copy Source

Sub-ticket of epic #346, Mac workstream. Depends on #356 (device seam). Target hardware: M5 Pro Max, 128 GB unified memory.

What: an end-to-end pipeline profile that runs every stage on Apple Silicon using only Tier-A backends (audit: portable with a device patch, no CUDA extensions). 128 GB unified memory removes the 24 GB VRAM ceiling that constrains the Linux host — models that queue behind GpuCoordinator leases on SERVER can run generously on the Mac.

Profile (default model per stage on macOS):

Stage Backend Audit status
concept SDXL-Lightning (fallback SDXL-Turbo) portable-as-is
modeling TripoSR (fallback Shap-E) needs-work: torchmcubes CPU build, rembg/onnxruntime
texturing SD-Turbo-Tex portable-as-is
rigging smart_rig (Blender addon, CPU) → RigNet CPU fallback (pyg CPU wheels, slow) CPU-fine
animation MDM device int 0 → mps mapping (#356)
export/enrichment headless Blender w/ Metal (#356) + OpenPose/MediaPipe (already CPU) fine

Scope:

  • Introduce a pipeline profile concept (linux-cuda vs macos-mps) selecting default + allowed models per stage — data-driven alongside the stage plans from #347; per-OS manifest overrides from #356
  • Validate each profile stage on the M5: run, record wall-time + peak memory + output quality on 3 student-realistic prompts (matrix mirrors spike #354 — post results there for comparison against the CUDA host)
  • FLUX.schnell as optional concept upgrade: bf16→fp16/fp32 on MPS, 22 GB weights — trivial fit in 128 GB unified memory; measure speed before enabling
  • MPS-specific fixes discovered during validation (op fallbacks via PYTORCH_ENABLE_MPS_FALLBACK, dtype issues) folded back into #356's device helper
  • hymotion reality-check: manifest says 4 GB, CLAUDE.md says 24 GB+ — measure on the Mac (unified memory can hold it either way); enable in the profile if throughput is acceptable
  • Document the profile + measured timings in art_pipe docs (upstream candidate)

Note: quality parity with the CUDA profile is NOT the goal here — a working, honest Mac chain is. TRELLIS-class modeling quality arrives via #355/M3.

Acceptance: on the M5 Mac, one command runs a character prompt through concept → modeling → texturing → rigging → animation → export producing a rigged, animated GLB with previews, all stages on MPS/CPU per the profile, with recorded timings per stage.

Sub-ticket of epic #346, Mac workstream. Depends on #356 (device seam). Target hardware: M5 Pro Max, 128 GB unified memory. **What:** an end-to-end pipeline profile that runs every stage on Apple Silicon using only Tier-A backends (audit: portable with a device patch, no CUDA extensions). 128 GB unified memory removes the 24 GB VRAM ceiling that constrains the Linux host — models that queue behind GpuCoordinator leases on SERVER can run generously on the Mac. **Profile (default model per stage on macOS):** | Stage | Backend | Audit status | |---|---|---| | concept | SDXL-Lightning (fallback SDXL-Turbo) | portable-as-is | | modeling | TripoSR (fallback Shap-E) | needs-work: torchmcubes CPU build, rembg/onnxruntime | | texturing | SD-Turbo-Tex | portable-as-is | | rigging | smart_rig (Blender addon, CPU) → RigNet CPU fallback (pyg CPU wheels, slow) | CPU-fine | | animation | MDM | device int `0` → mps mapping (#356) | | export/enrichment | headless Blender w/ Metal (#356) + OpenPose/MediaPipe (already CPU) | fine | **Scope:** - [ ] Introduce a **pipeline profile** concept (linux-cuda vs macos-mps) selecting default + allowed models per stage — data-driven alongside the stage plans from #347; per-OS manifest overrides from #356 - [ ] Validate each profile stage on the M5: run, record wall-time + peak memory + output quality on 3 student-realistic prompts (matrix mirrors spike #354 — post results there for comparison against the CUDA host) - [ ] FLUX.schnell as optional concept upgrade: bf16→fp16/fp32 on MPS, 22 GB weights — trivial fit in 128 GB unified memory; measure speed before enabling - [ ] MPS-specific fixes discovered during validation (op fallbacks via `PYTORCH_ENABLE_MPS_FALLBACK`, dtype issues) folded back into #356's device helper - [ ] hymotion reality-check: manifest says 4 GB, CLAUDE.md says 24 GB+ — measure on the Mac (unified memory can hold it either way); enable in the profile if throughput is acceptable - [ ] Document the profile + measured timings in art_pipe docs (upstream candidate) **Note:** quality parity with the CUDA profile is NOT the goal here — a working, honest Mac chain is. TRELLIS-class modeling quality arrives via #355/M3. **Acceptance:** on the M5 Mac, one command runs a character prompt through concept → modeling → texturing → rigging → animation → export producing a rigged, animated GLB with previews, all stages on MPS/CPU per the profile, with recorded timings per stage.
spikerj added the enhancement label 2026-07-04 05:27:47 +00:00
spikerj commented 2026-07-04 06:43:38 +00:00
Author
Owner
Copy Link
Copy Source

First real generations on the M5 Max — both Tier-A launch backends work on MPS through the canonical worker seam (artpipe serve → POST /jobs), zero PYTORCH_ENABLE_MPS_FALLBACK needed. Report committed at art_pipe/docs/mps-validation-m5.md (branch feature/device-abstraction-mps, commits ef32053/6463acb/5f0e04d).

Run Wall time Output
sdxl_lightning "robot mascot" (cold) 19.9 s clean 1024px concept PNG
sdxl_lightning "red dragon" 12.4 s prompt-faithful cartoon game art (eyeballed: excellent)
sdxl_lightning "treasure chest" 10.3 s clean
triposr dragon PNG → 3D 11.3 s 81k-vert OBJ + 3.2 MB GLB, clearly the dragon w/ vertex colors

~5–7 s of each SDXL run is per-job model reload (fresh subprocess per job) — warm-worker mode is the biggest Mac latency lever. Disk: 8.8 GB weights of the 25 GB budget.

Fixes committed (two are upstream candidates): torchmcubes has no MPS kernel → CPU marching-cubes wrap in triposr; transformers<5 pin (breaks TripoSR ckpt on CUDA too); TripoSR weights/torchmcubes provisioning via manifest setup_commands; psutil made optional in hardware.detect() (stdlib server venv crashed).

Remaining for this ticket: profile config concept, sdxl_turbo/shap_e/mdm validation, unified-memory admission budgeting (VRAM gate is nvidia-smi-only → everything admitted optimistically), Blender export stage (no Blender installed here yet). Cross-reference for the #354 matrix: Mac numbers above vs CUDA host TBD.

**First real generations on the M5 Max — both Tier-A launch backends work on MPS through the canonical worker seam** (`artpipe serve` → `POST /jobs`), zero `PYTORCH_ENABLE_MPS_FALLBACK` needed. Report committed at `art_pipe/docs/mps-validation-m5.md` (branch `feature/device-abstraction-mps`, commits `ef32053`/`6463acb`/`5f0e04d`). | Run | Wall time | Output | |---|---|---| | sdxl_lightning "robot mascot" (cold) | 19.9 s | clean 1024px concept PNG | | sdxl_lightning "red dragon" | 12.4 s | prompt-faithful cartoon game art (eyeballed: excellent) | | sdxl_lightning "treasure chest" | 10.3 s | clean | | triposr dragon PNG → 3D | 11.3 s | 81k-vert OBJ + 3.2 MB GLB, clearly the dragon w/ vertex colors | ~5–7 s of each SDXL run is per-job model reload (fresh subprocess per job) — **warm-worker mode is the biggest Mac latency lever**. Disk: 8.8 GB weights of the 25 GB budget. Fixes committed (two are upstream candidates): torchmcubes has no MPS kernel → CPU marching-cubes wrap in triposr; `transformers<5` pin (breaks TripoSR ckpt on CUDA too); TripoSR weights/torchmcubes provisioning via manifest `setup_commands`; psutil made optional in `hardware.detect()` (stdlib server venv crashed). Remaining for this ticket: profile config concept, sdxl_turbo/shap_e/mdm validation, unified-memory admission budgeting (VRAM gate is nvidia-smi-only → everything admitted optimistically), Blender export stage (no Blender installed here yet). Cross-reference for the #354 matrix: Mac numbers above vs CUDA host TBD.
spikerj commented 2026-07-04 07:12:13 +00:00
Author
Owner
Copy Link
Copy Source

Slice 2 done on feature/device-abstraction-mps (commits 82b422f…1d56900; pytest 596 passed — 22 new profile tests; independently re-verified).

  • Pipeline profiles implemented in config.py (PIPELINE_PROFILES: linux-cuda/windows-cuda/macos-mps), auto-detected (detect_profile_name(), ARTPIPE_PROFILE override). The existing PIPELINE_* constants now derive from the active profile, so preset resolution, submission normalization, node registry, and stage fallbacks were wired for free. Verified live: default submission on this Mac resolves triposr/image_to_3d; env override flips to linux-cuda.
  • Unified-memory VRAM admission implemented (not deferred): dispatcher free/total VRAM falls back to unified memory on Apple Silicon, vram_mb enforced with 1.4× overhead & 0.85 cap.
  • sdxl_turbo on MPS: zero fixes (6.2–9.3 s). shap_e on MPS: 2 fixes (undeclared ipywidgets dep — hits CUDA too; float64→float32 schedule extraction, upstream candidate) — 35 s/64-step and 116 s/128-step text→3D GLBs, prompt-faithful.
  • Disk now ~18 GB weights total. Deferred: flux (gated), mdm (SMPL licensing — flagged: it's the macos-mps motion default, needs the license step or a profile change), sf3d/instantmesh/riggers/Blender stages; shap_e fixed-output-path clobber noted.

Next: first end-to-end platform run on the Mac (#359 slice 1) using the validated concept+modeling chain.

Slice 2 done on `feature/device-abstraction-mps` (commits `82b422f`…`1d56900`; pytest **596 passed** — 22 new profile tests; independently re-verified). - **Pipeline profiles implemented** in `config.py` (`PIPELINE_PROFILES`: linux-cuda/windows-cuda/macos-mps), auto-detected (`detect_profile_name()`, `ARTPIPE_PROFILE` override). The existing `PIPELINE_*` constants now derive from the active profile, so preset resolution, submission normalization, node registry, and stage fallbacks were wired for free. Verified live: default submission on this Mac resolves triposr/image_to_3d; env override flips to linux-cuda. - **Unified-memory VRAM admission implemented** (not deferred): dispatcher free/total VRAM falls back to unified memory on Apple Silicon, `vram_mb` enforced with 1.4× overhead & 0.85 cap. - **sdxl_turbo on MPS: zero fixes** (6.2–9.3 s). **shap_e on MPS: 2 fixes** (undeclared ipywidgets dep — hits CUDA too; float64→float32 schedule extraction, upstream candidate) — 35 s/64-step and 116 s/128-step text→3D GLBs, prompt-faithful. - Disk now ~18 GB weights total. Deferred: flux (gated), mdm (SMPL licensing — flagged: it's the macos-mps motion default, needs the license step or a profile change), sf3d/instantmesh/riggers/Blender stages; shap_e fixed-output-path clobber noted. Next: first end-to-end platform run on the Mac (#359 slice 1) using the validated concept+modeling chain.
spikerj commented 2026-07-04 23:30:39 +00:00
Author
Owner
Copy Link
Copy Source

Resolved in spikersoft-artpipe PR #1 (merged to main). The macos-mps profile is shipped config (auto-detected, ARTPIPE_PROFILE override), unified-memory VRAM admission enforced, and the full Tier-A chain validated repeatedly on the M5 — culminating in complete 7-stage Character and Prop plans through the platform stack (45–72 s Prop smoke). Validated backends at close: sdxl_lightning, sdxl_turbo, triposr, shap_e, sd_turbo_tex, Blender stages, SafetyCheck, plus trellis_mac in allowed_models (#358). Remaining per-backend stragglers (flux gated, mdm SMPL licensing, sf3d/instantmesh) are documented in docs/mps-validation-m5.md and tracked via #358/#354 where relevant. Closing.

Resolved in spikersoft-artpipe PR #1 (merged to `main`). The `macos-mps` profile is shipped config (auto-detected, `ARTPIPE_PROFILE` override), unified-memory VRAM admission enforced, and the full Tier-A chain validated repeatedly on the M5 — culminating in complete 7-stage Character and Prop plans through the platform stack (45–72 s Prop smoke). Validated backends at close: sdxl_lightning, sdxl_turbo, triposr, shap_e, sd_turbo_tex, Blender stages, SafetyCheck, plus trellis_mac in allowed_models (#358). Remaining per-backend stragglers (flux gated, mdm SMPL licensing, sf3d/instantmesh) are documented in `docs/mps-validation-m5.md` and tracked via #358/#354 where relevant. Closing.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
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
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: spikerj/spikersoft-issues#357