Mapper weights are supplied separately and referenced by a portable descriptor. No weight download occurs during import.
The existing head-local loader reads one layer_NNN.safetensors file for each
target layer, in increasing layer order. A complete artifact must contain every
target layer. The low-level layer loader is intentionally small. Use
cachebridge-verify path/to/mapper.json to validate the complete layer roster,
file hashes, tensor geometry and finite coefficients before loading. The
concrete serving adapter also checks model geometry, tokenizer equality and
loaded mapper hashes.
| Tensor | Shape / meaning |
|---|---|
selected |
[support_layers], int64 source-layer indices in feature order |
headwise |
Present for head-local artifacts; writer emits boolean true |
k_weight, v_weight |
[target_heads, support_layers * source_dim, target_dim] |
k_bias, v_bias |
[target_heads, target_dim] |
Weights and biases are stored as FP32. Source layer support is shared between
K and V for each target layer. Final V corrections are folded into v_weight
and v_bias; the online mapper does not need the calibration statistics.
offline.fitting.fit_layer returns this tensor dictionary for one target
layer, together with solve diagnostics. Save it using
safetensors.torch.save_file. Its output is an initializer. The full
sequential V-calibration driver must run before describing the artifact as
the complete method. The CPU example intentionally uses a one-layer fixture.
The trace API uses content-space K. The online page API uses RoPE-applied K:
the mapper removes source RoPE, transforms K/V, then applies target RoPE using
the page's true absolute position_start. Values have no RoPE transform.
Compact TP input also needs the correct source_head_offset. A preallocated
output must match the resulting shape, dtype and device.
The low-level reference mapper also retains its legacy dense-affine mode to preserve regression checks. That mode is not a strict-original CFM execution baseline and is not author-provided CFM code.
Before enabling a production serving adapter, bind the tensor files to source and target model revisions, tokenizer compatibility, RoPE configuration, fitting stage/recipe, precision, calibration-input fingerprints and file hashes. The descriptor records these identities and the fitting provenance. Foundation-model revisions are recorded separately from mapper file hashes; an architecture-compatible fine-tune is not automatically a compatible replacement for the model used during fitting.