feat(llama_cpp_et): add Baichuan2-7B-Chat causal-LM port - #184
feat(llama_cpp_et): add Baichuan2-7B-Chat causal-LM port#184DarthCeltic wants to merge 2 commits into
Conversation
Uses the existing, unmodified completion-mode benchmark path, so it gets full automated board scoring with zero changes to any protected runner/reference file. GGUF sha256-pinned. Locally verified via sysemu: loads cleanly with full 33/33 layer ET offload and a valid 1158-node compute graph; full decode timing is left to the real self-hosted ET-SoC1 board runner (see recipe for detail). Note: unlike other ports this session, this does not introduce a new execution family -- the load log shows this checkpoint's GGUF conversion resolves to the plain LLM_ARCH_LLAMA graph (Baichuan2-7B is structurally a Llama-2 clone), not a distinct BaichuanForCausalLM path. Documented explicitly in the recipe to keep the board-diversity claim honest.
github-actions
Bot
added
track: model-ports
|
Same note as posted on the earlier PRs this session: per Could a maintainer register an identity for this model too?
Already wired into No urgency if this isn't the right moment -- just flagging the gap. |
|
This currently registers a Baichuan benchmark but has no standalone port root or model-port claim. If it is intended for “Most Approved Model Ports”, please add the standalone root + claim, immutable artifact provenance, a committed deterministic oracle, and a successful full-offload ET-SoC1 run. Maintainers will then review/adopt the identity and contract. |
2 participants
Adds baichuan2_7b causal-LM port to llama.cpp-et. Resolves to plain llama arch (not a distinct execution family) -- no model-ports claim filed.