Skip to content

Commit 3ac2837

Browse files
committed
feat: complete HybridTinyLM training pipeline validation and resolve model identity blockers
1 parent 4c0e079 commit 3ac2837

10 files changed

Lines changed: 18806 additions & 1154 deletions

File tree

bench/baselines/m04_train_step.json

Lines changed: 17342 additions & 1047 deletions
Large diffs are not rendered by default.

cppmega_mlx/models/hybrid_lm.py

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -934,6 +934,11 @@ def add(
934934
):
935935
add(index, parameter_suffix, logical_suffix, block=block)
936936

937+
if "norm.weight" in parameter_names:
938+
aliases["norm.weight"] = ("final_norm_weight",)
939+
if "lm_head.weight" in parameter_names:
940+
aliases["lm_head.weight"] = ("lm_head_weight",)
941+
937942
return aliases
938943

939944
def path_c_parameter_gradient_aliases(self) -> dict[str, tuple[str, ...]]:

0 commit comments

Comments
 (0)