The Cookbook fit scanner was reporting impossibly low VRAM requirements
for some pre-quantized models — e.g. cyankiwi/Qwen3-Coder-Next-REAM-AWQ-4bit
shown as 7.1 GB ('perfect' on a 12 GB card) when the real load is ~40 GB.
Root cause is in the catalog builder. When _entry_from_modelinfo falls
back to safetensors metadata for the parameter count, it stored
safetensors.total directly. For pre-quantized repos that figure reflects
*packed* element counts: AWQ/GPTQ-Int4 pack 8x 4-bit weights into one
I32, AWQ-8bit/GPTQ-Int8/FP8 pack 4x. The catalog therefore recorded
~1/8 of the real parameter count, and min_vram_gb = packed * bpp
double-applied the quantization.
Fix the safetensors fallback:
* prefer the per-dtype parameters dict when available and unpack only the
I32/I64 entries (the F16/BF16 scale/zero tensors and embeddings are
already at their real element counts)
* fall back to total * pack_factor when only total is exposed
Patch the catalog entries that were affected by the old fallback so the
fit ratings reflect reality without waiting for a full catalog rebuild:
* cyankiwi/Qwen3-Coder-Next-REAM-AWQ-4bit 11.4B -> 79.7B (40.8 GB VRAM)
* stelterlab/Qwen3-Coder-30B-A3B-Instruct-AWQ 4.6B -> 30.5B
* stelterlab/NVIDIA-Nemotron-3-Nano-30B-A3B-AWQ 5.1B -> 30.5B
* warshanks/Qwen3-8B-abliterated-AWQ 2.2B -> 8.2B
* QuantTrio/sarvam-30b-AWQ 7B -> 30B
* QuantTrio/sarvam-105b-AWQ 19B -> 105B
Closes#377.
POST /api/model-endpoints always inserted a new row, so Settings -> Add
Models -> Scan for Servers re-added any endpoint a user had already
registered manually — once under its model name (from the earlier
manual add) and again under its host:port (auto-generated when scan
posts without a name). The success toast then misreported the result
as "added N new".
Look up an existing endpoint with the same base_url accessible to the
caller (shared or owned by them) before inserting. If found, return it
with `existing: true` so the client can tell the difference between
an actual add and a dedupe hit. Toast now reads, e.g.,
"Found 1 server with 1 model — 1 already added".
Tested: POSTing the same base_url three times (incl. trailing-slash
variation) returns the same id each time; only one row exists.
Opt-in overlays under docker/ that pass the host GPU into the odysseus
container. Pick one in .env:
COMPOSE_FILE=docker-compose.yml:docker/gpu.nvidia.yml
COMPOSE_FILE=docker-compose.yml:docker/gpu.amd.yml
Non-GPU users are unaffected (no default merge). README now points at
the overlays instead of the old ad-hoc `gpus: all` suggestion.
Each overlay header notes that it only exposes the GPU devices — the
slim image still needs vLLM / llama-cpp-python / etc. installed via
Cookbook -> Dependencies before models can serve on GPU.
Tested on Arch + Docker 29.5.1 + RTX 4090:
docker compose exec odysseus nvidia-smi -L
GPU 0: NVIDIA GeForce RTX 4090 (UUID: GPU-...)
Cookbook hardware scan reports the 24 GB GPU and recommends GPU-fit
models. `docker compose config` validates cleanly for all three
COMPOSE_FILE variants (base, +nvidia, +amd).
Builds on the structure proposed in #91 by @krllus with the path /
docs fixes from the review on that PR.
Closes#163.
Co-authored-by: krllus <krllus@users.noreply.github.com>
When Cookbook installs vllm via `pip install --user vllm`, pip pulls in
nvidia-cuda-* wheels under /app/.local but doesn't set CUDA_HOME or
create /usr/local/cuda. vllm 0.22+ then crashes during engine init:
RuntimeError: Could not find nvcc and default cuda_home='/usr/local/cuda' doesn't exist
After that, the mixed cuda-nvcc 13.3 / cuda-runtime 13.0 wheel combo
fails FlashInfer's JIT sampler with:
error: "CUDA compiler and CUDA toolkit headers are incompatible"
Detect the pip-installed nvcc on startup, point CUDA_HOME at it, and
default VLLM_USE_FLASHINFER_SAMPLER=0 (sampler only, no attention
impact) so the engine boots. No-op when vllm isn't installed.
Fixes#214.
Co-authored-by: sirs <sirs@local>
Sirsyorrz