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

New Issue

2026-07-04T05:28:28Z

spikerj commented

2026-07-04 05:28:28 +00:00

Sub-ticket of epic #346, Mac workstream. Depends on #347/#348 (worker exists) + #357 (Mac profile). Goal: the full SpikerSoft stack — infra, backend, GPU pipeline, Angular — running on one M5 Pro Max (128 GB).

What's already solved: spikersoft-backend/docker-compose.mac.yml runs the platform on Apple Silicon (full sharded Mongo, RabbitMQ, Seq, Jaeger, single-node Redis Stack; ClamAV intentionally skipped). .NET 10 and Angular/Nx run natively on arm64. The gap is the GPU worker.

The architectural constraint: Docker on macOS has no GPU/MPS passthrough — the Linux deployment runs ArtPipeProcessor in a CUDA container (#348), but on a Mac it must run as a native host process talking to the containerized RabbitMQ/Mongo. #348's worker must therefore keep model-dispatch config-driven and host-agnostic (no container-path assumptions).

Scope:

ArtPipeProcessor native-run story on macOS: dotnet run (arm64) + art_pipe venvs on the host, config profile pointing at compose.mac endpoints (localhost RabbitMQ/Mongo/Seq/Jaeger); document in the worker README; note in #348 that paths/venv locations must be configurable, not baked into the Dockerfile
GpuCoordinator dev mode: either bypass leases on macOS (config flag; single-user machine) or run the coordinator natively with TotalGpuVramMB set from unified memory — pick one, document it (bypass recommended: the lease broker exists to arbitrate a shared card, which doesn't apply to a dev laptop)
Keycloak locally: add to the mac compose story (quay.io Keycloak images are multi-arch) or document pointing at ids.spikersoft.com with the existing spikersoft realm — decide + document (e2e realm import script if local)
One-command bring-up: script/Make target that starts compose.mac services, the API, the worker, and pnpm start for Angular; smoke-check endpoints (/healthz, hub negotiate)
End-to-end proof: submit from the local Angular Art Studio (#351) → native worker runs the macos-mps profile (#357) → GLB lands in local GridFS → SignalR progress renders — the entire student flow with zero non-Mac dependencies
Fold the instructions into the repo docs/skills so both Cursor and Claude pick them up (mirror convention)

Note: this also becomes the demo rig — the whole platform plus 3D generation on one laptop is exactly the story to show schools (no NVIDIA server required for pilots; a Mac Studio could even be a school's on-prem option later).

Acceptance: on a clean-ish M5 Mac: one documented command sequence brings up the full stack; a prompt submitted in the local browser produces a rigged animated GLB end-to-end with no connection to SERVER; teardown is clean.

Sub-ticket of epic #346, Mac workstream. Depends on #347/#348 (worker exists) + #357 (Mac profile). Goal: the full SpikerSoft stack — infra, backend, GPU pipeline, Angular — running on one M5 Pro Max (128 GB). **What's already solved:** `spikersoft-backend/docker-compose.mac.yml` runs the platform on Apple Silicon (full sharded Mongo, RabbitMQ, Seq, Jaeger, single-node Redis Stack; ClamAV intentionally skipped). .NET 10 and Angular/Nx run natively on arm64. The gap is the GPU worker. **The architectural constraint:** Docker on macOS has **no GPU/MPS passthrough** — the Linux deployment runs `ArtPipeProcessor` in a CUDA container (#348), but on a Mac it must run as a **native host process** talking to the containerized RabbitMQ/Mongo. #348's worker must therefore keep model-dispatch config-driven and host-agnostic (no container-path assumptions). **Scope:** - [ ] `ArtPipeProcessor` native-run story on macOS: `dotnet run` (arm64) + art_pipe venvs on the host, config profile pointing at compose.mac endpoints (localhost RabbitMQ/Mongo/Seq/Jaeger); document in the worker README; note in #348 that paths/venv locations must be configurable, not baked into the Dockerfile - [ ] GpuCoordinator dev mode: either bypass leases on macOS (config flag; single-user machine) or run the coordinator natively with `TotalGpuVramMB` set from unified memory — pick one, document it (bypass recommended: the lease broker exists to arbitrate a shared card, which doesn't apply to a dev laptop) - [ ] Keycloak locally: add to the mac compose story (quay.io Keycloak images are multi-arch) or document pointing at ids.spikersoft.com with the existing spikersoft realm — decide + document (e2e realm import script if local) - [ ] One-command bring-up: script/Make target that starts compose.mac services, the API, the worker, and `pnpm start` for Angular; smoke-check endpoints (`/healthz`, hub negotiate) - [ ] End-to-end proof: submit from the local Angular Art Studio (#351) → native worker runs the macos-mps profile (#357) → GLB lands in local GridFS → SignalR progress renders — the entire student flow with zero non-Mac dependencies - [ ] Fold the instructions into the repo docs/skills so both Cursor and Claude pick them up (mirror convention) **Note:** this also becomes the demo rig — the whole platform plus 3D generation on one laptop is exactly the story to show schools (no NVIDIA server required for pilots; a Mac Studio could even be a school's on-prem option later). **Acceptance:** on a clean-ish M5 Mac: one documented command sequence brings up the full stack; a prompt submitted in the local browser produces a rigged animated GLB end-to-end with no connection to SERVER; teardown is clean.

spikerj added the enhancement label 2026-07-04 05:28:28 +00:00

Stage	Model	Result	Time	Artifact
concept	sdxl_lightning	✅	10.4 s	1024² PNG (visually verified red dragon figurine)
modeling	triposr	✅	9.1 s	glTF v2 GLB, 41,420 verts (trimesh-verified)
texturing	sd_turbo_tex	✅	46.0 s (incl. first-use weight download)	texture PNG
export	—	designed clean failure (`No StageSettings for 'export'`), zero partial artifacts, failure events published	—	—