You can deploy this model and start earning money today!

dlab-cmu/
sf-map-Qwen__Qwen3.8-27B

Name: sf-map-Qwen__Qwen3.8-27B
Availability: InStock

Views

No views yet

SF map schema 3.0

An SF map is a lean, model-specific quality map. It contains only the values needed to choose dense, expert, and KV quantization actions and to recompute that plan when the available memory changes. Codec definitions and distortion come from the pinned catalog; physical bytes, backend legality, and speed come from the runtime and optional accelerator profile.

The map does not describe cache-window mechanics, eviction, routing-update timers, accelerator timings, or a frozen plan. Live routing observations may change which experts occupy a retained pool, but the map's measured retained count and codec prices remain fixed.

Abstract schema

jsonc

{
  "schema_version": "3.0",

  "model": {
    "id": "<model repository>",
    "revision": "<immutable model revision>"
  },

  "catalog_revision": "<fixed codec-catalog revision>",

  "reference": {
    "id": "sf-reference-v3",
    "ppl": "<measured reference perplexity>"
  },

  "dense": {
    "tensor_ids": ["<tensor ID>"],
    "numel": ["<element count>"],
    "direct_kl": ["<tensor × catalog-codec damage>"],
    "folded_kl": ["<tensor × catalog-codec damage>"],
    "fold_below_bpw": 2.0
  },

  "experts": {
    "layer_ids": ["<MoE layer ID>"],
    "n_experts": "<expert count>",
    "kept_counts": [
      256, 240, 230, 218, 205,
      192, 179, 166, 154, 141,
      128, 115, 102, 90, 77,
      64, 51, 38, 26, 16
    ],
    "measured_codecs": ["ib6p8", "ib9p4", "ib16p4"],
    "cells": ["<layer × retained-count × measured-codec log-damage>"],

    "interpolation": {
      "phi": ["<one value per catalog codec>"],
      "valid_between": ["ib6p8", "ib16p4"]
    },

    "routing_mass_prior": ["<layer × expert normalized routing mass>"]
  },

  "kv": {
    "layers": {
      "<layer ID>": {
        "mode": "separate_kv",
        "k": {
          "variance": "<measured value>",
          "beta": "<measured value>",
          "lowest_passing_codec": "<codec or null>"
        },
        "v": {
          "variance": "<measured value>",
          "beta": "<measured value>",
          "lowest_passing_codec": "<codec or null>"
        },
        "gamma": "<measured value>"
      },

      "<shared-KV layer ID>": {
        "mode": "shared_kv",
        "shared": {
          "variance": "<measured value>",
          "beta": "<measured value>",
          "lowest_passing_codec": "<codec or null>"
        }
      }
    }
  }
}

For a model without routed experts:

"experts": null

Reference

sf-reference-v3 is defined normatively as reference R: ib10p2 routed experts, FP16 dense weights, and FP16 KV. Components absent from an architecture are omitted. The map stores only the identifier and that model's measured reference perplexity.

Dense weights

Rows align directly with tensor_ids and numel.
Columns follow the catalog order identified by catalog_revision.
Use folded_kl below 2 bpw and direct_kl at or above 2 bpw.
Dense KL values are finite and nonnegative.

Experts

cells is the measured layer × retained-count × three-anchor surface.
Every background coordinate contains its literal damage; there is no zero sentinel.
Measured cells always override interpolation.
Expert cells are finite signed values. Negative damage is preserved through interpolation, comparison, Pareto pruning, and composition.
routing_mass_prior records the calibration identity ordering and seeds residency until live observations are sufficient. Routing adaptation changes identities only; retained-count and codec prices remain fixed.

Codec interpolation

text

range(t) = direct_kl[t, ib6p8] - direct_kl[t, ib16p4]

big_range = dense rows whose range is strictly greater than the median range
            across all rows in this map

phi(c) = median over big_range of
         (direct_kl[t,c] - direct_kl[t,ib16p4]) / range(t)

estimate(l,K,c) =
    cell(l,K,ib16p4)
    + (cell(l,K,ib6p8) - cell(l,K,ib16p4)) * phi(c)

phi follows the pinned catalog order.
phi(ib6p8) = 1.0 and phi(ib16p4) = 0.0 exactly.
ib9p4 always uses its literal measured cell.
Estimates are bracket-clamped against all measured anchors.
Only codecs inside valid_between have an interpolated expert quality action. Codecs outside it require another measured anchor.
The validator recomputes phi from dense.direct_kl and requires agreement within an absolute tolerance of 1e-12, then requires the two endpoints to be exactly 1 and 0.

KV cache

text

distortion(c) = (1 - kappa(c)) / kappa(c)

damage(l,s,c) = beta(l,s) * variance(l,s)
                * delta_distortion(c)

joint_damage(l) = gamma(l) * (damage_K(l) + damage_V(l))

Shared KV has one shared record and no gamma.
lowest_passing_codec is inclusive. A codec is legal exactly when its catalog distortion is no greater than the named codec's distortion.
null means that no measured cliff floor applies.

Planning and runtime updates

Choose one codec per dense tensor.
Choose a retained count and valid quality-priced codec per expert layer.
Choose a codec per KV side, or one codec for a shared-KV layer.
Obtain physical bytes and backend legality from the runtime.
Minimize predicted quality damage under the current memory budget.
When memory changes, recompute assignments from the same fixed map.
When routing changes, update retained identities without changing the map's retained-count or codec prices.

Total damage may be negative:

predicted_ppl = reference.ppl * exp(total_damage)

The predicted perplexity may therefore be below the reference perplexity.

Deliberately external

The SF map contains no codec menu, grid payloads, accelerator timings, speed ranks, backend legality, sink/window configuration, eviction policy, routing half-lives, speed-based exclusions, or frozen allocation plans.

dlab-cmu/sf-map-Qwen__Qwen3.8-27B