Sep 22 09:24 Make compatible with Transformers 5.17 Apply the same fix as https://huggingface.co/poolside/Laguna-XS.2/discussions/12 Jul 14 08:34 fix(rope): YaRN attention_factor is (0.1*ln(factor)+1)*attn_factor, not the bare multiplier Jul 11 09:59 docs: note INT4 needs vLLM #47154 Jul 1 08:57 Use chat_template.jinja as the single source: drop the {% include %} chat_template field from tokenizer_config.json Jun 18 18:34 small self-contained fixes Jun 5 11:04 Note FP8 KV cache needs vLLM 0.22.0; drop scrambled-output workaround (vllm#42650) May 27 09:00 update sampling parameters to match evals May 26 15:24 Update model card for 256K context length 14:59 increase context length to 256k May 20 20:12 Drop VLLM_USE_DEEP_GEMM=0 from vllm serve recipe (DeepGEMM is supported on Hopper and datacenter Blackwell) May 10 15:00 Add base_model (#1 ) joerowell +1: mgoin - Add base_model (79f71ad63a3501acbee1714baa515c6b79a2c5bb)
Co-authored-by: Michael Goin <mgoin@users.noreply.huggingface.co> Apr 29 22:33 Enable thinking by default in non-Hopper FP8-KV serve command Adds --default-chat-template-kwargs '{"enable_thinking": true}' to the workaround command so it matches the main card's recommended invocation. 22:29 Update non-Hopper FP8-KV serve command and link to vLLM recipes page Brings the non-Hopper FP8-KV workaround command in line with the new vllm serve flags from the main model card and adds --moe_backend marlin. Also links to https://recipes.vllm.ai/poolside/Laguna-XS.2. Apr 28 14:18 Laguna XS.2 upload 
