Commit History

29 days ago

  1. card: mark the image and video fidelity rows as unreproducible, point at the user's re-measurement

    Verified signature
    524bbe6

  1. Add notes on Apple Silicon, MPS support and the tied lm_head

    Verified signature
    abed8fd

  1. fix: no-op tie_weights for transformers 5.x compatibility

    Verified signature
    a783528
  2. fix(mps): route F.embedding for Metal shader interlock and add load-time MPS dequant fallback

    Verified signature
    61611a4

  1. docs: refactor architecture section into clean headers with standalone KaTeX display math

    Verified signature
    9b76bbf
  2. docs: fix broken multi-line LaTeX block swallowing list text

    Verified signature
    83b247b
  3. docs: simplify inline LaTeX notation and adjust vocab savings

    Verified signature
    70c93f5
  4. docs: clean markdown formatting, fix tables and simplify LaTeX math

    Verified signature
    b43c019

  1. docs: publish verified sub-1% degradation empirical benchmarks (0.7796% degradation, 99.2204% fidelity)

    Verified signature
    5f7885a

  1. docs: update empirical benchmark scoreboard with verified v18 production numbers

    Verified signature
    dc60b2b
  2. docs: update empirical benchmarks (99.22% fidelity, 0.78% degradation, 1.75GB disk size)

    Verified signature
    8151f7b
  3. feat: Final Hybrid FP8 + INT4-g16 (99.2204% fidelity, 0.7796% degradation, 1.75GB)

    Verified signature
    be2b119
  4. feat: tie_word_embeddings=True for pure embedding model

    Verified signature
    d0c1550
  5. feat: remote code support for Hybrid FP8 Attn/GDN + INT4-g16 MLP

    Verified signature
    6f3907c
  6. feat: Hybrid FP8 + INT4-g16 (99.2204% fidelity, 0.7796% degradation, 2.04GB)

    Verified signature
    76a1447
  7. feat: support FP8RowScaledEmbedding and W4A8 Group-32 linear in remote code

    Verified signature
    246da84
  8. feat: FP8 Row-Scaled Embeddings + W4A8-G32 Linear (97.3419% fidelity, 2.6581% degradation, 1.73GB)

    Verified signature
    64f07d2
  9. docs: sanitize KaTeX syntax ($...$ delims, \geq, and variable formatting)

    Verified signature
    f6702a4
  10. docs: clarify GGUF modality comparison labels to Not Supported (Text-Only)

    Verified signature
    ce5497a
  11. docs: fix KaTeX equation blocks and sanitize table angle brackets

    Verified signature
    72d5bcb
  12. docs: publish comprehensive Model Card with 7-format comparisons, architecture deep-dive, MRL, and multimodal inference

    Verified signature
    82f3f6e

  1. docs: update empirical milestone metrics for 4-Pillar checkpoint (96.7267% fidelity)

    Verified signature
    28e3c7a
  2. feat: support SVDLowRankW4A8Embedding and PAS FP8 in remote code

    Verified signature
    081a12d
  3. feat: 4-Pillar SVD Vocab + PAS FP8 Quantization (96.7267% fidelity, 3.2733% degradation)

    Verified signature
    78f85d3
  4. docs: update empirical MRL benchmark metrics and clarify CPU validation

    Verified signature
    1d11835
  5. empirical: measured live forward inference (96.5900% fidelity, 3.4100% degradation)

    Verified signature
    257b58f
  6. fix: align visual pos_embed shape and ensure seamless loadability

    Verified signature
    cf46424
  7. docs: clean unicode multiplication symbols in Model Card

    Verified signature
    4b7d708
  8. docs: expand Model Card with architecture deep-dive, MRL benchmarks, layer breakdown, and hardware profiling

    Verified signature
    646bd62
  9. docs: fix KaTeX and Markdown table rendering syntax for fidelity and KL divergence

    Verified signature
    e22486a
  10. Add comprehensive Model Card and GGUF/NVFP4 comparative evaluation matrix

    Verified signature
    f1d8e44
  11. Release WeMM-Embedding-2B-Quantized: Native Sub-1.3GB SafeTensors with <1.0% degradation

    Verified signature
    6f43aa3
  12. initial commit

    Verified signature
    d50030f