DeepSeek-V2-Lite-Chat — CI mirror (no weights)

A weight-less mirror of deepseek-ai/DeepSeek-V2-Lite-Chat used by the LMCache CPU CI so that MoE + DeepSeek architecture code paths can be exercised on GitHub Actions runners without pulling ~30 GB of safetensors.

The tarball published as a GitHub Release asset only contains config / tokenizer / modelling / index metadata. Downstream tooling loads the model with --load-format dummy (vLLM) so weights are random-initialised at startup.

Snapshot

Files below are pinned to HF snapshot 85864749cd611b4353ce1decdb286193298f64c7:

File	Purpose
config.json	Model architecture
configuration_deepseek.py	Custom config class (needs trust_remote_code=True)
modeling_deepseek.py	Custom modelling class (kept for completeness)
tokenizer.json / tokenizer_config.json	Tokenizer
tokenization_deepseek_fast.py	Custom tokenizer class
generation_config.json	Default generation kwargs
model.safetensors.index.json	Weight shard index (kept for completeness)
LICENSE / .gitattributes	Upstream artefacts

*.safetensors are intentionally excluded.

How the CI consumes it

The LMCache CPU workflow calls .github/scripts/download_gh_release_model.sh with:

MODEL_ID=deepseek-ai/DeepSeek-V2-Lite-Chat
GH_REPO=opendataio/deepseek_v2_lite
GH_TAG=v1.0
TARBALL_NAME=deepseek_v2_lite.tar.gz
TARBALL_TOP_DIR=DeepSeek-V2-Lite-Chat
SNAPSHOT=85864749cd611b4353ce1decdb286193298f64c7

which downloads the release asset and lays it out under ~/.cache/huggingface/hub/models--deepseek-ai--DeepSeek-V2-Lite-Chat/snapshots/<sha>/.