[Bug] ReViSQL-BIRD-K2.6 generates gibberish output when deployed with SGLang (base Kimi-K2.6 works fine)
Summary
Hi. We deployed uiuc-kang-lab/ReViSQL-BIRD-K2.6 (64-shard safetensors, deployed via vLLM-compatible OpenAI API on an internal MaaS platform using SGLang) and the model produces completely incoherent output. The original base model moonshotai/Kimi-K2.6 deployed on the same infrastructure with the same serving stack works normally, so we suspect that this is specific to the released checkpoint.
What we checked
- Tokenizer/config files are byte-identical to the base model (
tiktoken.modelmd5 match;tokenizer_config.json,chat_template.jinja,config.json, all*.pyfiles identical). So it's not a tokenizer mismatch on our side. - Tried both Chat Completions and Responses APIs, various
max_tokens,temperature,top_p,stop/stop_token_ids[163585, 163586] — always gibberish. - Greedy decoding degenerating into an endless
maskedloop suggests the checkpoint's embedding / lm_head rows for special tokens are corrupted or misaligned — a symptom consistent with a broken LoRA merge (e.g.resize_token_embeddings/ row misalignment at merge time) or a corrupted/shifted shard upload.
Ask
Could you verify the released safetensors shards, and whether the checkpoint could be validated with SGLang after merge?
Hi, thank you for the detailed report and diagnostics. We’re investigating this merged checkpoint, including verifying the safetensors shards and testing the model with SGLang.
In the meantime, you can also try serving our original, unmerged LoRA adapter exported from Tinker: uiuc-kang-lab/ReViSQL-BIRD-K2.6-LoRA.
We’ll update this issue once we’ve completed validation of the merged checkpoint using SGLang.