Repository navigation

fit: the four planner fixes together, and where they meet - #229

Merged

88plug merged 13 commits into

mainfrom

claude/fit-plans-fix

Oct 30, 2026

Merged

fit: the four planner fixes together, and where they meet#229
88plug merged 13 commits into
mainfrom
claude/fit-plans-fix

Commits on Oct 29, 2026
- fit: OVMS serves one GPU
  88plug
  committed
  9482b32
- fit: size sliding-window KV per engine
  88plug
  committed
  6f3ffa8
- fit: count llama-server's host caches
  88plug
  committed
  57cf38d
- fit: size llama.cpp's layer split card by card
  88plug
  committed
  edb80df
- fit: llama.cpp per-card bytes follow the bytes to serve; moe-offload picks its card count
  88plug
  committed
  d666d07
- fit: the per-card split counts llama.cpp's SWA cells and the context checkpoints
  88plug
  committed
  86d31c5
- fit: llama.cpp's output layer holds the copy of tied embeddings
  88plug
  committed
  5dbc4df
- fit: a GGUF's MTP blocks take llama.cpp layer slots on a split
  88plug
  committed
  04dfd9d
- fit: a GGUF conversion keeps the MTP blocks, so llama.cpp's split counts them
  88plug
  committed
  303d7d5
- llamacpp_test: the per-card fit's flags reach llama-server
  88plug
  committed
  9874581
- agent: K3D_GPU_AGENT_GPUS reaches the fit for llama.cpp as for vLLM
  88plug
  committed
  6e05e9e
- fit: cite llama.cpp v0.8.0, the llama-server the agent runs
  88plug
  committed
  46d55b3
Commits on Oct 30, 2026
- agent: K3D_GPU_AGENT_GPUS reaches the fit when the backend is unset and OVMS does not serve the model
  88plug
  committed
  e49d0f8