15 days ago
-
pipeline_mesh: correct-by-construction multi-host (synchronize-after-readback + world-size-gated MPI)
22 days ago
-
Fix device-memory OOM: per-resolution L1_SMALL/program-cache reclaim + deallocate() (studio-0.6-loras)
23 days ago
-
API: prompt-tagging (PromptTagBody) + stricter validation on /generate and OpenAI endpoint
-
Device VAE on the 1xN mesh path (reclaims ~5s; 2-card 1024 E2E 1.68x->1.75x)
-
Card: 2/4-card reset step
-
Fix: mesh-aware text-encoder to_torch (2-card serve); card: mesh reset step
-
Card: explicit 1/2/4-card run instructions
-
Docs: release update
-
Docs: release update
-
Release: device VAE default (all-res), multichip MOSS_MESH, UI params + /v1/parameters, API hardening
-
Card: host VAE default; device VAE 256-only (known high-res bug)
-
Fix: host VAE default (device VAE wrong >256); keep group_norm grid fix
-
Spec: save/delete endpoints
-
Card: device VAE now default
-
Device VAE default + save/delete API
-
Add HTTP API spec
-
Add full model card
-
Upload README.md with huggingface_hub
-
MOSS-Image v6 thin tt-dit-server bundle (ttnn 0.80, Blackhole)
-
initial commit
