The loader resolves every relative import to a FLAT file beside the module,
whatever the nesting. Three `from .fusion import ...` sites in alephlm.py and
one `from .data.curriculum import ...` in presets.py therefore made it demand
fusion.py and data.curriculum.py in this repo and refuse to load the model
(OSError / FileNotFoundError on transformers 5).
Both imports are inert for this craft: the config carries no `fusion` key, so
that branch is never taken, and the curriculum is a training-side concern. They
are now resolved by name at the point of use, so nothing is imported unless the
code path is actually run. No weights, config or math changed.
Proved before push, in a virtualenv with transformers 5.13 and no geolip
installed: AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)
loads 237.1M params with no missing or unexpected keys and runs a forward
pass over a byte prompt.
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.
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.
mirror alephllm 0.8.9 (same-session law): presets.py (v3 preset builder + registrations guarded for copies without the data stack; TrainConfig.phase_lr_scale/phase_seed_offset; Preset.data_scale/epoch_cap/rebalance_to), governor.py (raw_block: nested wrappers), alephlm.py (compile_hubs via raw_block). Gated: strict load of model.safetensors through the vendored path, logits parity 0.0 vs geolip.
mirror alephllm 0.8.9 (same-session law): presets.py (v3 preset builder + registrations guarded for copies without the data stack; TrainConfig.phase_lr_scale/phase_seed_offset; Preset.data_scale/epoch_cap/rebalance_to), governor.py (raw_block: nested wrappers), alephlm.py (compile_hubs via raw_block). Gated: strict load of model.safetensors through the vendored path, logits parity 0.0 vs geolip.
mirror alephllm 0.8.9 (same-session law): presets.py (v3 preset builder + registrations guarded for copies without the data stack; TrainConfig.phase_lr_scale/phase_seed_offset; Preset.data_scale/epoch_cap/rebalance_to), governor.py (raw_block: nested wrappers), alephlm.py (compile_hubs via raw_block). Gated: strict load of model.safetensors through the vendored path, logits parity 0.0 vs geolip.
