Merge pull request #103 from VE-FORBRYDERNE/neox

Divide GPT-NeoX replicated bias layers by 4 again instead of by 8

tpu_mtj_backend.py

@@ -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]