Module host_memory 

How much memory the host actually has free, and whether a model fits.

Lives here rather than in a binary because both ferrox-cli and ferrox-server need the same answer, and two copies of a platform probe drift.

Enums

FitPlan

What to do about a model whose weights may not fit.

Constants

FIT_HEADROOM_BYTES

Bytes to leave for everything that is not weights: the KV cache, activations, and the pages of the checkpoint not yet touched.

Functions

available_bytes

Bytes a new allocation can reasonably expect to get.

derived_copy_budget

Bytes a process-wide cache of DERIVED copies of already-resident weights may retain – today, weight_matrix::repack_cache, which holds the interleaved (repacked) form of matrices the GEMV kernels read.

parse_vm_stat

Free plus inactive pages from vm_stat output, in bytes.

plan_for

Decide whether streaming is needed, given what the model weighs and what the host has.