Nevermind, dividing by 4 is actually correct...

This commit is contained in:
Gnome Ann 2022-03-19 00:55:04 -04:00
parent f16b61ec77
commit 6c20d0d657

View File

@ -885,7 +885,7 @@ def read_neox_checkpoint(state, path, config, checkpoint_shards=2):
original_shape = shards[0][key].shape
for checkpoint_shard in range(checkpoint_shards):
if key in ("attention.dense.bias", "mlp.dense_4h_to_h.bias"):
shards[checkpoint_shard][key] /= config["cores_per_replica"]
shards[checkpoint_shard][key] /= output_shards
if key != "word_embeddings.weight" and shards[checkpoint_shard][key].ndim == 2:
shards[checkpoint_shard][key] = shards[checkpoint_shard][key].T
tensor = shards[checkpoint_shard][key]