Commits · 4b223f7bf577b50139f5a8c8342d2d0015c8424a · nexlab / coderai · GitLab

Oct 24, 2026
- prefer the drafter that does not switch async scheduling off · eb6f2935
  Stefy Lanza (nextime / spora ) authored Oct 24, 2026
  eb6f2935
- tune how a model is served, per model, from the model page (v0.4.84) · 8f0b898a
  Stefy Lanza (nextime / spora ) authored Oct 24, 2026
  8f0b898a
Oct 22, 2026
- make the cache floor configurable, and widen the cluster page too (v0.4.79) · 21fccade
  Stefy Lanza (nextime / spora ) authored Oct 22, 2026
  21fccade
- never hand back the cache the window needs (v0.4.78) · bda5f045
  Stefy Lanza (nextime / spora ) authored Oct 22, 2026
  bda5f045
- show the limits that bind a pod, not the one that does not (v0.4.69) · 20bced3b
  Stefy Lanza (nextime / spora ) authored Oct 22, 2026
  20bced3b
- let vLLM split thinking from the answer: the tag is a special token (v0.4.65) · ccea5f43
  Stefy Lanza (nextime / spora ) authored Oct 22, 2026
  ccea5f43
- choose the tool-call style from the model, not from its name (v0.4.63) · 20112637
  Stefy Lanza (nextime / spora ) authored Oct 22, 2026
  20112637
Oct 21, 2026
- read the chat template to tell qwen3_xml tool calls from hermes (v0.4.61) · a9e7ed0e
  Stefy Lanza (nextime / spora ) authored Oct 21, 2026
  a9e7ed0e
Oct 19, 2026
- one vLLM per card: wait for a starting engine, keep a busy one (v0.4.54) · ed4e2c2a
  Stefy Lanza (nextime / spora ) authored Oct 19, 2026
  ed4e2c2a
Oct 17, 2026
- surya reports a page confidence, and whether anything measured it (v0.4.15) · 6053e8b1
  Stefy Lanza (nextime / spora ) authored Oct 17, 2026
  6053e8b1
- the accent repair had no input on a pod (v0.4.14) · 20594e97
  Stefy Lanza (nextime / spora ) authored Oct 17, 2026
  20594e97
Oct 16, 2026
- the adapter half of the same bug (v0.4.11) · dc40989f
  Stefy Lanza (nextime / spora ) authored Oct 16, 2026
  dc40989f
- one engine, every name — whatever section wrote it (v0.4.10) · 91098e7f
  Stefy Lanza (nextime / spora ) authored Oct 16, 2026
  91098e7f
Oct 15, 2026
- an adapter id named the container, not the content (v0.3.94) · 11569dd5
  Stefy Lanza (nextime / spora ) authored Oct 15, 2026
  11569dd5
- vLLM was never launched able to accept tool calls (v0.3.92) · 7601bdf9
  Stefy Lanza (nextime / spora ) authored Oct 15, 2026
  7601bdf9
- a retrain keeps the name and changes the bytes (v0.3.91) · 6ac424e9
  Stefy Lanza (nextime / spora ) authored Oct 15, 2026
  6ac424e9
- a pool pod keeps the LoRA door open (v0.3.89) · 5fb825ca
  Stefy Lanza (nextime / spora ) authored Oct 15, 2026
  5fb825ca
- the adapter was found by a name a seeded entry does not have (v0.3.88) · 36ed3da0
  Stefy Lanza (nextime / spora ) authored Oct 15, 2026
  36ed3da0
- a control plane that rented pods to deliver a teardown (v0.3.87) · 7d0f24aa
  Stefy Lanza (nextime / spora ) authored Oct 15, 2026
  7d0f24aa
- shared pools: a name is not a path (v0.3.86) · 48fbcf76
  Stefy Lanza (nextime / spora ) authored Oct 15, 2026
  48fbcf76
Oct 12, 2026
- lora: a trained text adapter can reach a pod, and be picked per request (v0.3.36) · 3a46f9a8
  Stefy Lanza (nextime / spora ) authored Oct 12, 2026
  3a46f9a8
Oct 05, 2026
- vllm: size the share down too, from the cache a successful boot reports (v0.2.28) · bb2389ff
  Stefy Lanza (nextime / spora ) authored Oct 05, 2026
  bb2389ff
- vllm: discover the share an engine needs instead of trusting a constant (v0.2.27) · aa085f85
  Stefy Lanza (nextime / spora ) authored Oct 05, 2026
  aa085f85
- ocr: size the VLM share to the measured footprint and cap its batch (v0.2.26) · f25592dc
  Stefy Lanza (nextime / spora ) authored Oct 05, 2026
  f25592dc
Oct 04, 2026
- ocr: size a VLM engine's vLLM against free VRAM, and never answer 503 (v0.2.25) · b1004ebd
  Stefy Lanza (nextime / spora ) authored Oct 04, 2026
  b1004ebd
Oct 02, 2026
- ocr: olmOCR-2 as an alternative engine, and stop the Surya-2 crash loop (v0.2.24) · 75200c0b
  Stefy Lanza (nextime / spora ) authored Oct 02, 2026
  75200c0b
Sep 19, 2026
- cluster: nodes as engines, one GGUF over llama.cpp RPC, vLLM/SGLang... · f5ca2d82
  Stefy Lanza (nextime / spora ) authored Sep 19, 2026
  f5ca2d82
Sep 18, 2026
- vllm / engines-kt pods: a C compiler for Triton, and a root-cause error tail · b3b4dc5c
  Stefy Lanza (nextime / spora ) authored Sep 18, 2026
  b3b4dc5c
Sep 17, 2026
- runpod: coderai-vllm and coderai-llama — the LLM servers as coderai pods · d11f14ae
  Stefy Lanza (nextime / spora ) authored Sep 17, 2026
  d11f14ae
Sep 14, 2026
- text: LoRA and QLoRA on text models, locally and on RunPod · 34d08890
  Stefy Lanza (nextime / spora ) authored Sep 14, 2026
  34d08890
Sep 13, 2026
- workers: let every isolated-venv worker point at a remote service · 4766c1df
  Stefy Lanza (nextime / spora ) authored Sep 13, 2026
  4766c1df
Aug 22, 2026
- ocr: make managed Surya-2 vLLM reclaimable by VRAM eviction · 2c419247
  Stefy Lanza (nextime / spora ) authored Aug 22, 2026
  2c419247
Aug 21, 2026
- vllm: cooperative VRAM co-tenancy + GPU selection (full engine parity) · 4223107e
  Stefy Lanza (nextime / spora ) authored Aug 21, 2026
  4223107e
- vllm: default VLLM_USE_FLASHINFER=0 (no runtime JIT toolchain) · 077e61bd
  Stefy Lanza (nextime / spora ) authored Aug 21, 2026
  077e61bd
- Add vLLM as a first-class managed engine backend (LLM + OCR Surya2) · a5d9fe82
  Stefy Lanza (nextime / spora ) authored Aug 21, 2026
  a5d9fe82