feat: Docker/Portainer deployment for the text-to-motion demo (CPU + optional Vulkan GPU) - #6
Draft
lgcavalheiro wants to merge 1 commit into
Draft
feat: Docker/Portainer deployment for the text-to-motion demo (CPU + optional Vulkan GPU)#6lgcavalheiro wants to merge 1 commit into
lgcavalheiro wants to merge 1 commit into
Conversation
- Multi-stage Dockerfile (Go demo + kmd-generate, ~290 MB) with the GGML Vulkan backend and automatic CPU fallback; GGML is cloned at the pinned submodule commit so the build context needs no submodule init - docker/entrypoint.sh downloads and SHA-256-verifies the GGUF weights into /data on first start, then serves the demo on 0.0.0.0:8094 - docker-compose.yml for Portainer stacks plus docs/docker_portainer.md covering GPU passthrough (CDI), VRAM sizing and day-2 ops - demo: honour a pre-set KIMODO_BACKEND for the generator subprocess instead of hardcoding vulkan; the vulkan default still falls back to CPU when no device is present, and containers can now force cpu Verified end to end on a 24-core/16 GB host: CPU generates a 60-frame 10-step clip in 45 s; with an optional GTX 1050 Ti (Vulkan, text encoder chunked at KIMODO_TEXT_LAYER_CHUNK=4) in 36 s at ~2.8 GB VRAM.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Runs the kimodo.cpp text-to-motion demo as a container (CPU by default, optional Vulkan GPU), packaged for Docker Compose / Portainer:
Dockerfile— multi-stage (~290 MB final image): buildskmd-generate/kmd-inspect(CMake/Ninja, C++23,KIMODO_ENABLE_VULKAN=ON) and the Go demo, and checks out GGML itself at the pinned submodule commit, so the build context does not requiregit submodule update --init(plain git clones and Portainer builds work).docker/entrypoint.sh— on first start, downloads the GGUF weights (~1.1 GB motion + ~15 GB shared LLM2Vec text encoder, F32) into/datawith SHA-256 verification against the published manifests (resumable, skipped when complete), then execs the demo on0.0.0.0:${KIMODO_PORT:-8094}. An explicitly emptyKIMODO_MODELSskips downloads entirely.docker-compose.yml+docs/docker_portainer.md— Portainer-ready stack and a full deployment guide: build on the Docker host, deploy via the web editor, GPU passthrough via CDI (devices: [nvidia.com/gpu=all]), VRAM sizing (KIMODO_TEXT_LAYER_CHUNK), adding models, retrieving animated GLB / skinned GLB / BVH, troubleshooting.demo/main.go— one behavioural change: the generator subprocess inherits a pre-setKIMODO_BACKENDinstead of the hardcodedvulkan. The default is unchanged (vulkan auto-falls back to CPU when no device exists); this just lets a container forcecpu— which is also the escape hatch for hosts where Vulkan enumeration aborts instead of returning zero devices.Runtime dependencies worth noting
The image carries a few non-obvious runtime packages, each found the hard way on real hardware:
spirv-headersin the builder — GGML's CMake needsSPIRV-HeadersConfig.cmakefor the Vulkan backend.libx11-6/libxext6— the NVIDIA ICD library injected by nvidia-container-toolkit (libGLX_nvidia.so.0) links them.libegl1/libglx0/libglvnd0(GLVND) — the NVIDIA ICDdlopenslibEGL.so.1during Vulkan init even with no display; without it the ICD fails inside the container withCould not get 'vkCreateInstance'while the host works fine.libgomp1— OpenMP runtime for the GGML backends ondebian:trixie-slim.Verification
Deployed end to end on a 24-core / 16 GB homelab (Ubuntu, Docker 29, Portainer CE):
KIMODO_TEXT_LAYER_CHUNK=4(the default 8 OOMs a 4 GB card during text-encoder loading; documented in the guide)./api/models, weight download persistence across restarts, and theKIMODO_BACKEND=cpufallback all verified.Draft while I tidy the doc wording / check CI expectations — happy to adjust to whatever packaging conventions the project prefers (e.g. a different base, or splitting the demo patch into its own PR).