Oct 14 29 days ago 20:00 One arithmetic on every device: mixed CPU+GPU optimizer, replica hash check, FTZ pin, subnormal gate, live f32 audit, probe/fragment fixes (2026-10-14) Oct 1 05:39 Logic pass: CPU attention mirrors get the same int32 bound as the GPU kernels (peers refuse identically); gpuMlpChain returns pooled buffers when the row-max guard throws 05:01 Reductions refuse what they used to accept: GPU row-max non-finite guard, averageGrads length/NaN refusal, int32 accumulator bound, capacity-weighted cluster loss, finite reduced-gradient check, empty CUDA_VISIBLE_DEVICES Sep 27 12:26 web: restore LF line endings on verified_core.js / transformer.js The previous upload carried CRLF line endings on these two files (a local git checkout artifact). Content is otherwise unchanged; this restores the repo's LF convention so the files diff cleanly. 12:19 Refuse NaN/Inf at every int8 quantize; requant rounds half-up (retrained); bounded units; README for September Four local commits in one upload (ec432c8, c8db1cf, 47724cd, ae9cfa2): bounded elementwise units and LUT build (407 MB -> 2 MB, 277 MB -> 1 MB at startup); requant16 retrained from floor (-0.4981 LSB bias) to round-half-up, N/N 65,536/65,536, browser requant LUT re-exported; qat.py and every browser quantizer now refuse NaN/Inf instead of silently quantizing the whole layer to 0 (new tests fail on the old code); README September section. python test_verified_units.py 25/25, npm test 12/12 suites green. Jul 18 08:40 Upload web/public/app.js with huggingface_hub 08:40 Upload web/public/webgpu.js with huggingface_hub 08:29 Upload web/public/webgpu.js with huggingface_hub 08:29 Upload web/public/transformer.js with huggingface_hub 08:29 Upload web/public/verified_core.js with huggingface_hub 07:59 Upload web/public/app.js with huggingface_hub 07:59 Upload web/public/verified_core.js with huggingface_hub 07:47 Upload web/public/webgpu.js with huggingface_hub 07:42 Upload web/public/webgpu.js with huggingface_hub Jul 17 22:19 Upload web/public/webgpu.js with huggingface_hub 08:30 Upload web/public/webgpu.js with huggingface_hub 02:06 Upload web/public/webgpu.js with huggingface_hub 01:52 Upload web/public/verified_core.js with huggingface_hub 01:52 Upload web/public/webgpu.js with huggingface_hub Jul 16 20:38 web: gate-the-gate boundary discrimination in the B2B init gate 10:50 web: logic-pass fixes (portable test_b2b, symmetric repair patience) 09:56 web: B2B MLP chain (CUTLASS ex. 13 two-GEMM fusion + ex. 23 epilogue reduction), test_b2b.js, ten-suite npm test 05:36 web: QKV dual-GEMM fusion (bit-identical, CUTLASS ex. 45) 03:57 web: sync-guard gradient repair, FineWeb-Edu hardcoded (parquet /data streaming), dispatch-optimized backward (bit-identical), IEEE-754 oracle + corpus mutation tests, TEST_RESULTS.md Jul 15 23:39 Web demo: lr default 0.01, batched per-head backward GEMMs 19:26 Web demo: exact kernel gates + continuous cross-device kernel verification Jul 14 07:52 Web demo: split-K backward + gather-fused attention 03:28 Web demo: weight tying + WebRTC resilience (relay fallback, keepalive) 03:17 Web demo: block-scaled INT8 + batched/fused GEMM kernels Jul 13 21:32 Web demo: leader failover + dataset picker + reconnect policy 05:07 Web demo: mid-run joins + auto-reconnect 04:32 Web demo: multi-device sync + connection fixes 03:32 Web trainer upgrade: sync guard (leader roster + weight-hash divergence check), 3xINT8 fast-accurate GEMM, DP4A hardware path (verified vs units), Spikewhale tokenizer (16.5k vocab), FineWeb-Edu streaming, gradient/checkpoint fragmentation, inference kit, generation tester, contribution logs Jul 12 22:31 Web demo: mini transformer LM + scaling sliders 22:17 Web demo: room-first lobby + group training settings sliders 21:48 Release: SpikeWhale slider panel (HF dataset picker, stop/back), DaisyChain-Web (P2P WebRTC training, DaisyAdam, checkpoints, room host approval, verified-units-only) 
