# Publishing Instructions

Recommended Hub repository:

```text
kel-jp/llm-jp-4-8b-instruct-NVFP4-speculator.dflash
```

Publish this as a new model repository, not as a new revision of
`kel-jp/llm-jp-4-8b-instruct-NVFP4`. The NVFP4 repository is the verifier model;
this repository is a companion draft model used through vLLM speculative decoding.

## CLI Upload

```bash
cd /home/dgx/ege/kel-jp
hf auth login
HF_XET_HIGH_PERFORMANCE=1 hf upload \
  kel-jp/llm-jp-4-8b-instruct-NVFP4-speculator.dflash \
  release/llm-jp-4-8b-instruct-NVFP4-speculator.dflash \
  .
```

If the repo does not exist yet, create it first:

```bash
hf repos create kel-jp/llm-jp-4-8b-instruct-NVFP4-speculator.dflash --type model
```

To create it private instead:

```bash
hf repos create kel-jp/llm-jp-4-8b-instruct-NVFP4-speculator.dflash --type model --private
```

## Python Upload

```python
from huggingface_hub import HfApi, create_repo

repo_id = "kel-jp/llm-jp-4-8b-instruct-NVFP4-speculator.dflash"
local_dir = "/home/dgx/ege/kel-jp/release/llm-jp-4-8b-instruct-NVFP4-speculator.dflash"

create_repo(repo_id, repo_type="model", exist_ok=True)

HfApi().upload_folder(
    repo_id=repo_id,
    repo_type="model",
    folder_path=local_dir,
    commit_message="Release DFlash drafter for llm-jp-4-8b-instruct-NVFP4",
)
```

## Suggested First Release Tag

After upload, tag the initial checkpoint as:

```text
b4-l1-v28k-elyza100
```

Use later tags or branches for future drafters with different block sizes,
layer counts, vocab widths, verifier revisions, or benchmark suites.