Hugging Face's logo

poolside
/
Laguna-XS.2-NVFP4

Text Generation
Safetensors
vllm
laguna
laguna-xs.2
conversational
custom_code
8-bit precision
compressed-tensors
Model card Files
xet
 Community
2
Laguna-XS.2-NVFP4

Commit History

  1. Make compatible with Transformers 5.17 (#2)⁠

    poolside-eng +1: Yu-Peng-Poolside
    973ad42

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

    joerowell
    Verified signature
    d3cca23

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

    joerowell
    Verified signature
    ad2b974

  1. small self-contained fixes

    joerowell
    Verified signature
    578e5d9

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

    joerowell
    Verified signature
    1a001ef

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

    joerowell
    571346d

  2. adjust expert scales for non-Hopper targets

    joerowell
    Verified signature
    3ed234d

  1. update sampling parameters to match evals

    joerowell
    d9c13a0

  1. Update model card for 256K context length

    varunrandery
    Verified signature
    12a5f20

  2. increase context length to 256k

    joerowell
    d2a5f9c

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

    joerowell
    Verified signature
    514daf4

  1. Add base_model (#1)⁠

    joerowell +1: mgoin
    2f694d9

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

    joerowell
    Verified signature
    62a5860

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

    joerowell
    Verified signature
    92f8b44

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

    joerowell
    Verified signature
    c6a0e4c

  1. Laguna XS.2 upload

    joerowell
    98ebde9