biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

0 Releases 17 Tags

RSS Feed
  • v0.1-e2e-working acc20dffd7
    Compare

    Single-shot E2E inference working: DSV4-Pro NVFP4 full 61-layer pipeline

    biondizzle released this 2026-05-31 22:27:27 +00:00 | 446 commits to master since this release

    • Correct NVFP4 two-level scale dequant (ws * ws2, NO input_scale)
    • mHC with Sinkhorn-Knopp, B^T in post_block
    • CSA/HCA compressor producing compressed KV
    • CSA indexer with top-k selection
    • Full attention over [compressed_kv, swa_kv]
    • MoE with hash (L0-2) + noaux_tc (L3+) routing
    • Shared expert, YaRN RoPE + inverse RoPE
    • Attention sinks, grouped output projection
    • HcHead readout, thinking mode enabled

    Output: 'The capital of France is Paris' — coherent reasoning response
    Known issues: residual growth in L56-60, decode degradation after ~10 steps

    Downloads