Hardware profiles

A hardware profile is a named, versioned description of a machine: how much memory it has, whether that memory is unified, and the bandwidth/compute figures the throughput estimator needs. Pass one with --profile to score models against that machine instead of the detected one.

User walkthrough (bundled profiles + writing your own for unreleased hardware): docs/cli.md → Hardware profiles.

llmfit --profile ryzen-ai-max-plus-395 fit -n 10
llmfit --profile ./my-workstation.json recommend --json
llmfit hardware list
llmfit hardware show nvidia-rtx-4090

--profile replaces the --memory / --ram / --cpu-cores overrides (it conflicts with all three) because a profile describes a whole machine rather than one field.

Layout

hardware/
  schema.json          JSON Schema (draft-07) every profile must satisfy
  <name>.json          one profile; `name` must equal the file stem

Profiles in this directory are aggregated by llmfit-core/build.rs and embedded into the binary, so a merged profile ships in the next release and is available by name with no download. Users can add their own without rebuilding by dropping files into the directory printed by:

llmfit hardware path

A user profile whose name matches an embedded one takes precedence. LLMFIT_HARDWARE_PROFILES overrides that directory.

Fields

calibration records measured anchors with their provenance so they can be reviewed now and used by a later schema version; it changes no estimate today.

A discrete-GPU profile does not carry a VRAM figure, so VRAM stays as detected on the host. Unified profiles are fully described: VRAM tracks total_ram_gb.

Format

Each file conforms to schema.json. Example:

{
  "schema_version": 1,
  "name": "example-unified-256",
  "match": { "gpu_name_contains": "Radeon 8060S" },
  "hardware": {
    "total_ram_gb": 128.0,
    "unified_memory": true,
    "gpu_memory_bandwidth_gbps": 256.0,
    "ddr_bandwidth_gbps": 256.0,
    "gpu_compute_tflops_fp16": 29.7
  },
  "estimation": {
    "efficiency": 0.6,
    "run_mode_factors": { "cpu_only": 0.25 }
  },
  "calibration": [
    {
      "model": "openai/gpt-oss-120b",
      "quant": "MXFP4",
      "run_mode": "gpu",
      "measured_tps": 50.0,
      "source": "https://github.com/AlexsJones/llmfit/issues/969"
    }
  ]
}

Unknown keys are tolerated when loading (so a profile written for a newer llmfit still works) but rejected by validation, which is what catches a typo before it silently does nothing:

llmfit hardware validate ./my-workstation.json

Bundled profiles

Validation

cargo test -p llmfit-core validates every file here against schema.json and checks that each name matches its file stem, so a malformed contribution fails CI rather than shipping as a profile that cannot be selected.