Variable manifestConst

Type Declaration

• Readonlyid: "quantize"

• Readonlyjobs: readonly [{ gpu: "exclusive"; isolation: "process"; kind: "quantize" }]

• Readonlypanel: {
      entry: "@mlx-bun/module-quantize/panel";
      framed: false;
      path: "/quantize";
      tag: "mlx-quantize-panel";
      title: "Quantize";
  }

• Readonlyrequires: readonly ["jobs", "storage", "catalog"]

• Readonlyroutes: readonly [
      {
          id: "inspect";
          method: "POST";
          path: "/inspect";
          response: "json";
          summary: "Whether a model is quantizable and its on-disk size, from its configuration alone";
      },
      {
          id: "resolve-folder";
          method: "POST";
          path: "/resolve-folder";
          response: "json";
          summary: "Locate a folder picked in the browser on disk (hub snapshot, the app's models directory, or an indexed model)";
      },
      {
          id: "submit";
          method: "POST";
          path: "/submit";
          response: "json";
          summary: "Queue a quantize job into the app's models directory; answers with the job id and output directory";
      },
  ]

• Readonlystorage: readonly [
      {
          key: "models";
          kind: "directory";
          path: "models";
          purpose: "Quantized and converted models: plain model directories, indexed beside the hub cache";
      },
  ]

• Readonlysummary: "Quantize or convert a local MLX checkpoint: the managed quantize job, its web routes, and the convert verb."

• Readonlytitle: "Quantize"

• Readonlyverbs: readonly [
      {
          name: "convert";
          options: readonly [
              {
                  name: "hf-path";
                  summary: "Source model: local path, downloaded model, or HF repo id (fetched first); --model and the positional are aliases";
                  type: "string";
              },
              { name: "model"; summary: "Alias for --hf-path"; type: "string" },
              {
                  name: "mlx-path";
                  summary: "Output directory; must not already exist [default: ~/.mlx-bun/models/<model>-<bits>bit]";
                  type: "string";
              },
              {
                  name: "quantize";
                  short: "q";
                  summary: "Quantize the model (uniform affine); without -q or --target-bpw the model is rewritten with --dtype and/or -d only";
                  type: "boolean";
              },
              {
                  name: "q-bits";
                  summary: "Bits per weight: 4 or 8 [default: 4]";
                  type: "string";
              },
              {
                  name: "q-group-size";
                  summary: "Quantization group size: 32 or 64 [default: 64]";
                  type: "string";
              },
              {
                  missingValue: "--upload-repo expects a repo id (org/name)";
                  name: "upload-repo";
                  summary: "Push the converted model to this Hugging Face repo afterwards (write token checked first)";
                  type: "string";
              },
              {
                  name: "target-bpw";
                  summary: "Mixed precision target bits-per-weight, e.g. 4.5: OptiQ sensitivity sweep + per-layer knapsack; implies -q";
                  type: "string";
              },
              {
                  name: "candidate-bits";
                  summary: "Comma list the knapsack may pick from [default: 4,8]";
                  type: "string";
              },
              {
                  name: "calibration-mix";
                  summary: "\"optiq\" or a JSONL path [default: optiq]";
                  type: "string";
              },
              {
                  name: "n-calibration";
                  summary: "Calibration samples [default: 2]";
                  type: "string";
              },
              {
                  name: "rotate-weights";
                  summary: "Fold the model's offline TurboQuant rotation before quantization (auto-detects Llama/Qwen3.5/Qwen MTP)";
                  type: "boolean";
              },
              {
                  name: "rotation-seed";
                  summary: "Deterministic rotation seed [default: 42]";
                  type: "string";
              },
              {
                  name: "q-mode";
                  summary: "Quantization mode: affine, or trellis for packed Trellis (TCQ) MLP tensors of an HF-layout Qwen3.5-family model (rotation fold implied; implies -q) [default: affine]";
                  type: "string";
              },
              {
                  name: "trellis-bits";
                  summary: "--q-mode trellis: bits per coded weight, 1-8 [default: 3]";
                  type: "string";
              },
              {
                  name: "trellis-k-map";
                  summary: "--q-mode trellis: allocation JSON with per-tensor bits (budgets[--trellis-k-budget].kmap, optional affine_map)";
                  type: "string";
              },
              {
                  name: "trellis-k-budget";
                  summary: "--q-mode trellis: budget key inside --trellis-k-map [default: 3.00]";
                  type: "string";
              },
              {
                  name: "trellis-ldlq";
                  summary: "--q-mode trellis: directory of per-layer Hessian factors (layer-NNN-mlp|down.safetensors) enabling BlockLDLQ error feedback";
                  type: "string";
              },
              {
                  name: "trellis-reuse";
                  summary: "--q-mode trellis: comma list of packed artifacts to copy unchanged-geometry tensors from instead of re-encoding";
                  type: "string";
              },
              {
                  name: "trellis-down-axis";
                  summary: "--q-mode trellis: dim of down_proj the trellis runs along, out (rotated dim) or in [default: out]";
                  type: "string";
              },
              {
                  name: "trellis-interleave";
                  summary: "--q-mode trellis: reorder eligible 3-bit axis-0 codes into the two-block scatter layout";
                  type: "boolean";
              },
              {
                  name: "trellis-layers";
                  summary: "--q-mode trellis: trellis-code only the first N decoder layers (the rest ride the 3-bit affine base tier)";
                  type: "string";
              },
              {
                  name: "dtype";
                  summary: "Dtype of the non-quantized tensors and of the quantization scales/biases: float16 | bfloat16 | float32 [default: bf16 scales/biases, other tensors unchanged]";
                  type: "string";
              },
              {
                  name: "dequantize";
                  short: "d";
                  summary: "Dequantize a quantized model to dense weights (not with -q)";
                  type: "boolean";
              },
              {
                  name: "quant-predicate";
                  summary: "Not supported (mlx-lm's mixed_* recipes need 2/3/6-bit); use --target-bpw for mixed precision";
                  type: "string";
              },
          ];
          positional: readonly [
              {
                  name: "repo-or-path";
                  summary: "The source model: a local path, a downloaded model, or an HF repo id";
              },
          ];
          summary: "Quantize an HF model into a local MLX snapshot (mlx_lm.convert counterpart)";
      },
  ]