# Split downstream outputs

Several downstream profiling caches are larger than Hugging Face Hub's 50 GB
single-file limit. Those files are uploaded as ordered 10 GiB chunks with names
ending in `.pt.part-000`, `.pt.part-001`, and so on.

The whitening-only `.pt` files are uploaded directly because they are below the
Hub hard limit.

## Reassembly

Download the relevant directory and concatenate the parts in lexical order.
Shell glob ordering is sufficient because the suffixes are zero-padded.

```bash
cat artifacts/downstream/outputs/llama2_r0.2_baseline/meta-llama_Llama-2-7b-chat-hf_profiling___calib_mix_jsonl_128_0.pt.part-* \
  > artifacts/downstream/outputs/llama2_r0.2_baseline/meta-llama_Llama-2-7b-chat-hf_profiling___calib_mix_jsonl_128_0.pt

cat artifacts/downstream/outputs/llama2_r0.2_decodeshare_a2/meta-llama_Llama-2-7b-chat-hf_profiling___calib_mix_jsonl_128_0.pt.part-* \
  > artifacts/downstream/outputs/llama2_r0.2_decodeshare_a2/meta-llama_Llama-2-7b-chat-hf_profiling___calib_mix_jsonl_128_0.pt

cat artifacts/downstream/outputs/svdllm_whiten_r0.2/meta_llama_Llama_2_7b_chat_hf_profiling_wikitext2_128_0.pt.part-* \
  > artifacts/downstream/outputs/svdllm_whiten_r0.2/meta_llama_Llama_2_7b_chat_hf_profiling_wikitext2_128_0.pt
```

## Uploaded downstream files

- `llama2_r0.2_baseline/`: baseline whitening output and split profiling cache.
- `llama2_r0.2_decodeshare_a2/`: DecodeShare A2 whitening output and split
  profiling cache.
- `svdllm_whiten_r0.2/`: SVD-LLM whitening output and split profiling cache.