Code

History

Commits on main30
  1. README: don't claim the 262K three-tier window until a boot of that config is observed healthykacper-daftcode committed 4 months ago591250b
  2. README: lead rewritten around MoE on consumer Blackwell; GLM-5.2 first-class (incl. 2x PRO 6000 three-tier); drop stale 16K/8K context figureskacper-daftcode committed 4 months ago1477b84
  3. Three-tier hierarchy shipped: host 2-bit -> GPU 2-bit cache -> gate-filled FP4kacper-daftcode committed 4 months agocf02998
  4. GLM-5.2 on 2x96GB: MTP works (33 tok/s, acceptance 3.0), needle 4/4 to 27Kkacper-daftcode committed 4 months ago1d2fb5b
  5. GLM-5.2 (753B) live on two 96 GB cards via the host-resident basekacper-daftcode committed 4 months ago22f0862
  6. Sync the shipped patch with worktree HEAD (base-cache miss-tolerance knob)kacper-daftcode committed 4 months ago1b1fd28
  7. README: GLM-5.2 section with measured numbers (MTP 105 tok/s, quality tier 83-85, needle 126K/276K, PP determinism)kacper-daftcode committed 4 months ago28532d5
  8. Single-5090: MTP validated on the host-resident base (38 tok/s), determinism 6/6 on PP4kacper-daftcode committed 4 months ago7127056
  9. docs: NVFP4 KV cache + deterministic unpermute / BASE cache / AFRAG in the v0.24 port noteskacper-daftcode committed 4 months ago765310a
  10. 159B on one RTX 5090: host-resident 2-bit base with the GPU as an expert cachekacper-daftcode committed 4 months ago6da9beb
  11. Sync the shipped patch with HEAD (AFRAG default-on) + honest README claimskacper-daftcode committed 4 months ago68e38f6
  12. README: drop the benchmark provenance footnote — current numbers onlykacper-daftcode committed 4 months ago51cce3e
  13. README/docs: AFRAG prefill numbers (5 340 / 6 100 tok/s) + patch size 22 files/+4.2kkacper-daftcode committed 4 months ago2fc804d
  14. Ship the AFRAG prefill cubins (all 5 K) — measured +12% e2e prefill, on by defaultkacper-daftcode committed 4 months ago4fb5363
  15. NOTICE: correct the DeepGEMM attribution for the v0.24 portkacper-daftcode committed 4 months ago20700ae
  16. GLM-5.2-NVFP4 live on TP4 + MTP-under-PP: loader, runner fixes, first numberskacper-daftcode committed 4 months ago4747bf4
  17. Delta pool auto-sizing: VLLM_MOE_W2_DELTA_GB=auto sizes the FP4 pool from free VRAM after KV allocationkacper-daftcode committed 4 months agof445898
  18. Dockerfile.sm120-v024: e2e-validated build (3 fixes) — smoke PASS on a fresh imagekacper-daftcode committed 4 months agod81f31e
  19. README: batched-serving numbers as a first-class table in What you getkacper-daftcode committed 4 months agoed03e0d
  20. Batch-scaling on 4x5090 TP4: 198/460/762/1006/1560 tok/s aggregate @ conc 1/4/8/16/32kacper-daftcode committed 4 months agoe9316bc
  21. Batch-scaling bench on 1x PRO 6000: 156/290/493/659/933 tok/s aggregate @ conc 1/4/8/16/32kacper-daftcode committed 4 months ago73fcdd9
  22. Remove legacy v0.19 fork files (old patch, Dockerfile, BUILD.md, serve.sh); NOTICE -> v0.24.0 basekacper-daftcode committed 4 months ago3d845fe
  23. 512K context live-validated on v0.24 (needle PASS @102K/256K/453K on 1x PRO 6000)kacper-daftcode committed 4 months ago119ab5b
  24. README: drop stale long-context/evals TODO from Qualitykacper-daftcode committed 4 months ago1d97a31
  25. README/docs: drop all references to the retired v0.19 base — current numbers onlykacper-daftcode committed 4 months ago7556aa2
  26. Confidence-gate driver re-integrated on v0.24 (patch now 20 files, +3.7k)kacper-daftcode committed 4 months ago04c52c9
  27. README: single-era rewrite — current v0.24 numbers only, legacy fork demoted to BUILD.mdkacper-daftcode committed 4 months ago442544d
  28. Dockerfile.sm120-v024: reproducible image for the v0.24.0 basekacper-daftcode committed 4 months ago3897a73
  29. v0.24.0 port: rebase onto official vLLM, GLM-5.x prep, same-day A/B benchmarkskacper-daftcode committed 4 months ago406fe04
  30. vLLM-Moet: DeepSeek-V4-Flash on SM120 (Blackwell)Kacper committed 4 months ago9093a5c