--force-model now truly means it
The v0.11.19 VRAM safety check no longer applies to forced models at all: when --force-model is present, the engine loads exactly what you forced, regardless of what the card reports. That is the point of the flag — you take responsibility for the fit. (KERYX_LLAMA_VRAM_CHECK=0 remains as a diagnostic escape for auto mode.)
Relaxed safety margin for auto mode
The v0.11.19 pre-check used a 1200 MiB margin, which wrongly refused borderline-but-working fits — e.g. a 12B model on a 10 GB card (~200–400 MB of slack, proven to work on dedicated cards). The margin is now 128 MiB and guards only what can never fit (weight-level allocation failures, the reproduced crash class — a 9.3 GiB model on an 8 GB card is still refused). Context-size shortfalls continue to use llama's safe 4096→1024 retry as before.
If your multi-GPU forced setup failed on v0.11.19 with "could not SERVE the FORCED model", this release restores the v0.11.17-and-earlier loading behavior for it. Tip for tight fits: --low-ram --wait-ready stages cards one at a time, which gives the serving card the cleanest shot at loading.
Includes v0.11.19 (staging-crash fix), v0.11.18 (--wait-ready), v0.11.17 (chaseless solver).