Expand description
Checkpoint-level detection of MTP / next-token-prediction weights.
Atlas binds MTP through three unrelated loader paths — the Qwen-shaped
MtpWeights vec, DeepSeek-V4’s mtp.0.* module, and GLM-5.3’s
layers.{num_hidden_layers} block — but the “did the user get what they
asked for” check read only the first of them. On GLM-5.3 that produced
GLM-5.3 MTP draft module loaded (layers.45)
`--speculative` was requested but no MTP weights were loaded for this modelin the same startup log: the module HAD loaded, from
model.language_model.layers.45.*. This module is the layout-aware
predicate the check should have been asking, kept separate from any one
loader so a new architecture only has to be named here once.
Enums§
- MtpLayout
- The MTP weight layout a checkpoint ships, if any.
Functions§
- detect
- Detect the MTP layout from raw tensor names.
- detect_
in_ store detectover a loadedWeightStore.