Repository navigation
feat(kokoro): own opt-in CUDA worker admission and cleanup - #55
Conversation
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Merged as 8a1bae0. Verified that its tree exactly matches reviewed head 8609dcc and its sole parent is the captured previous main 5d07c37. Both PR workflows passed on attempt 1. Post-merge main CI/build are running; this does not activate the loader in existing servers or qualify CUDA speech readiness. The bounded real MCP trial accepted the launch and cleaned up the owned worker/reservation, but failed before authenticated CUDA readiness. No speed result exists. Hermes #445 retains readiness and host-integration work; no additional GPU trial is queued. |
|
Post-merge validation is complete: main CI run 42302851706 and package build run 42302851797 both passed on attempt 1 for 8a1bae0. Release attachment and PyPI publication were correctly skipped. No loader activation or further GPU trial occurred. |
Kokoro CUDA could only be launched directly by its caller, so vram-mcp could reserve capacity without owning the load. Add three opt-in, Windows-only tools for one operator-configured reviewed worker:
load_kokoro,kokoro_status, andstop_kokoro.The server reserves and serializes admission, checks fresh selected-device capacity, launches the actual verified base Python image, and retains custody until child exit and reservation/admission release are proven. Callers supply no command or paths. The existing pinned worker starts its Windows parent-handle watchdog before CUDA initialization. Device visibility is restricted to the admitted UUID. Unknown launch or cleanup blocks duplicate work; an unpublished admission failure explicitly requires closing the scoped server. There is no force, eviction, direct-launch fallback or default enablement.
A successful load means launch acceptance only:
gpu_readyremains false. The trusted client must independently validate the existing authenticated CUDA-ready/profile handshake. Operator capacity estimates are not hard memory caps or attributable peaks. This does not reload existing MCP servers or alter the CPU voice baseline.Validation:
One bounded integration trial used a separate real MCP server and the new load/status/stop tools. Launch was accepted and shared-ledger ownership was verified, but the authenticated CUDA-ready handshake was not reached, so no speech timings were measured. The owned worker stopped, its reservation was released, and the scoped server closed. Sampled pressure remained healthy; attributable peak memory and desktop usability were not measured. The probe retained only a generic startup failure, so its cause remains unproven. No retry or production activation is queued; the CPU voice baseline is unchanged. CUDA readiness and performance remain open under Hermes #445.
Refs Hermes optional-loader dependency. That record retains host-integration and hardware qualification ownership; this PR does not close it automatically.