Fix bug in tools/ckpts/convert_neox_to_hf.py for setting intermediate…

…_size (#1209)

In tools/ckpts/convert_neox_to_hf.py, for neox architecture the 'intermediate_size'
argument is not explicitly set, so it defaults to 24576 from:

https://github.com/huggingface/transformers/blob/9fe3f585bb4ea29f209dc705d269fbe292e1128f/src/transformers/models/gpt_neox/configuration_gpt_neox.py#L48

Proposed solution: set intermediate-size to 4 * hidden-size

tools/ckpts/convert_neox_to_hf.py

@@ -277,6 +277,11 @@ def __init__(self, neox_config):
                 ),
                 "use_parallel_residual": get_key(neox_config, "gpt-j-residual", False),
                 "layer_norm_eps": get_key(neox_config, "layernorm-epsilon", 1e-5),
+                "intermediate_size": get_key(
+                    neox_config,
+                    "intermediate-size",
+                    4 * get_key(neox_config, "hidden-size"),
+                ),
             }
         )
         hf_config = GPTNeoXConfig(**args)