`scripts/convert_hf_checkpoint.py --verify true` fails

Nobody has claimed this yet.

```python
Traceback (most recent call last):
File "scripts/convert_hf_checkpoint.py", line 68, in
CLI(convert_hf_checkpoint)
File "/home/carmocca/.local/lib/python3.8/site-packages/jsonargparse/cli.py", line 82, in CLI
return _run_component(component, cfg_init)
File "/home/carmocca/.local/lib/python3.8/site-packages/jsonargparse/cli.py", line 138, in _run_component
return component(**cfg)
File "/usr/local/lib/python3.8/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "scripts/convert_hf_checkpoint.py", line 62, in convert_hf_checkpoint
assert torch.testing.assert_close(out, out_hf)
File "/usr/local/lib/python3.8/dist-packages/torch/testing/_comparison.py", line 1511, in assert_close
raise error_metas[0].to_error(msg)
AssertionError: Tensor-likes are not close!

Mismatched elements: 131071613 / 131072000 (100.0%)
Greatest absolute difference: 32.40781021118164 at index (0, 2739, 1) (up to 1e-05 allowed)
Greatest relative difference: 1.0 at index (0, 0, 0) (up to 1.3e-06 allowed)
```

Using this slightly modified version of the file

```python
import gc
import json
import shutil
import sys
from pathlib import Path

import torch

# support running without installing as a package
wd = Path(__file__).parent.parent.resolve()
sys.path.append(str(wd))

from lit_llama.model import LLaMA, LLaMAConfig
from lit_llama.utils import EmptyInitOnDevice

@torch.no_grad()
def convert_hf_checkpoint(
*,
output_dir: Path = Path("checkpoints/lit-llama"),
ckpt_dir: Path = Path("checkpoints/hf-llama/"),
model_size: str = "7B",
dtype: str = "float32",
verify: bool = False,
) -> None:
"""
Perform the reverse operation of: https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py
"""
output_dir = output_dir / model_size
ckpt_dir = ckpt_dir / model_size

dt = getattr(torch, dtype, None)
if not isinstance(dt, torch.dtype):
raise ValueError(f"{dtype} is not a valid dtype.")
dtype = dt

config = LLaMAConfig.from_name(model_size)

with EmptyInitOnDevice(device="cpu", dtype=dtype):
model = LLaMA(config)

if verify:
try:
from transformers import LlamaForCausalLM
except ImportError:
raise ImportError("verify=True requires transformers to be installed, please `pip install transformers`")
print("Verifying...")

token_sample = torch.randint(0, config.vocab_size, size=(1, config.block_size), dtype=torch.int64)
out = model(token_sample)
del model
gc.collect()

print("Loading original model for comparison")
model_hf = LlamaForCausalLM.from_pretrained(ckpt_dir)
out_hf = model_hf(token_sample)["logits"]

print("Comparing outputs")
assert out.device.type == out_hf.device.type
assert out.dtype == out_hf.dtype
assert torch.testing.assert_close(out, out_hf)

if __name__ == "__main__":
from jsonargparse import CLI

CLI(convert_hf_checkpoint)
```

Reproduce the failure with scripts/convert_hf_checkpoint.py --verify true and inspect the comparison at line 62, including the outputs from LLaMA and LlamaForCausalLM. Verify the conversion behavior against the traceback; done means the verification comparison completes without the tensor mismatch.

Written by the indexing model from the issue text.