AbstractPhil commited on
Commit
f0f1565
·
verified ·
1 Parent(s): 0d63228

mirror alephllm 0.9.0 (same-session law): presets.py (AlephLMConfig.fusion field, default None; the v3 preset builder + registrations guarded for copies without the data stack), governor.py (raw_block), alephlm.py (the optional fusion stage imported lazily — unfused crafts load without model/fusion.py; the trunk loop factored into _trunk). Gated: strict load of model.safetensors through the vendored path, logits parity 0.0 vs geolip.

Browse files
Files changed (1) hide show
  1. presets.py +6 -0
presets.py CHANGED
@@ -60,6 +60,12 @@ class AlephLMConfig:
60
  # x 32 layers) the retained scan tensors alone exceed a 95GB card —
61
  # measured OOM, Blackwell preflight 2026-08-26. ~2x hub recompute cost.
62
  hub_ckpt: int = 0
 
 
 
 
 
 
63
 
64
  def to_dict(self):
65
  d = asdict(self)
 
60
  # x 32 layers) the retained scan tensors alone exceed a 95GB card —
61
  # measured OOM, Blackwell preflight 2026-08-26. ~2x hub recompute cost.
62
  hub_ckpt: int = 0
63
+ # v3 (2026-09-19): weak-token fusion at the input plane. None = the
64
+ # byte-resolution trunk verbatim. A dict selects the hourglass form:
65
+ # {"rule": "entropy" | "spacelike", "theta": bits, "witness_floor": n,
66
+ # "table": "<npz path>", "k_lo": front blocks, "k_hi": back blocks} —
67
+ # see model/fusion.py. Old manifests load via the default.
68
+ fusion: Optional[dict] = None
69
 
70
  def to_dict(self):
71
  d = asdict(self)