fix(rope): YaRN attention_factor is (0.1*ln(factor)+1)*attn_factor, not the bare multiplier

Use chat_template.jinja as the single source: drop the {% include %} chat_template field from tokenizer_config.json

small self-contained fixes

Document SGLang support (sgl-project/sglang#24204)

Note FP8 KV cache needs vLLM 0.22.0; drop scrambled-output workaround (vllm#42650)

adjust expert scales for non-Hopper targets

update sampling parameters to match evals

Update model card for 256K context length

increase context length to 256k

Drop VLLM_USE_DEEP_GEMM=0 from vllm serve recipe (DeepGEMM is supported on Hopper and datacenter Blackwell)

Add base_model (#1)

Enable thinking by default in non-Hopper FP8-KV serve command

Update non-Hopper FP8-KV serve command and link to vLLM recipes page

Sync chat template with v5_1 (matches base/FP8/INT4)

Laguna XS.2 upload