1.5 KiB
GPU Resource Accounting
M10 treats NVML as physical truth while retaining ModelForge attribution as a separate logical
ledger. A resident model is already present in observed_used_vram_bytes and must therefore never
be subtracted a second time.
physical_used = max(observed_nvml_used, attributed_managed_resident)
unmanaged_external = max(0, observed_nvml_used - attributed_managed_resident)
future_lease = max(0, reserved_bytes - materialized_bytes)
schedulable = max(0, total - physical_used - future_leases - dynamic_reserve)
Every term is non-negative. The exposed invariant delta is
observed - managed_resident - unmanaged_external; it diagnoses sampling or attribution drift but
is never converted into capacity. KNOWN, ESTIMATED, and UNKNOWN describe attribution quality.
Unknown or stale telemetry fails closed with zero schedulable capacity; it never means zero use.
The versioned m10-v1 reserve is the maximum of 1 GiB, 5% of total VRAM, and a 256 MiB runtime
margin. A deployment requires its measured peak plus the maximum of 128 MiB or 10% of that peak.
Only unmaterialized leases reduce future headroom. Accelerator-row locking makes the database the
coordination boundary for concurrent admissions.
Resource envelopes remain bound to the exact artifact, runtime profile, accelerator and environment fingerprint. A stale or mismatched heavy envelope cannot authorize placement. Persistent history is bounded placement evidence, not a one-second telemetry archive.