Issues

• Why did you omit the trivial switch for censorship that strata-nvfp4 has out of the box when all the engine machinery already exists?

  jesdga95/basalt#210 · Filter by author lampymoth opened on Nov 2, 2026

• can the coder weights from strata work on basalt?

  jesdga95/basalt#209 · Filter by author mmoalem opened on Nov 1, 2026
  1

• Will video support be added?

  jesdga95/basalt#205 · Filter by author TingTingin opened on Nov 1, 2026
  1

• Q8_0 pack: engine exits under concurrent requests, "batch segments of several rows need the PLE window kernels"

  bugre-test
  jesdga95/basalt#196 · Filter by author paulp83 opened on Nov 1, 2026
  1

  

• verify: a hung window survives release_gpu_waits ("the GPU did not finish within 5 s") and the engine exit TDRs the card - Xid 153, 12.67 GiB ResourceTimeout

  jesdga95/basalt#192 · Filter by author beastrobin opened on Nov 1, 2026
  2

• decode: lazy L2 prefetch should also take the next layer's FFN read matrices

  jesdga95/basalt#189 · Filter by author kaushikvira opened on Nov 1, 2026
  1

• Xid 13 (Illegal Instruction Encoding) during prefill with --kv-resident + --spec on RTX 5060 (sm_120)

  bug
  jesdga95/basalt#161 · Filter by author sucubus666 opened on Oct 30, 2026
  7

  

• Codex CLI and /v1/responses: two behaviours we hit (informative)

  jesdga95/basalt#152 · Filter by author hagudelom opened on Oct 30, 2026
  1

• Feature request: kv-resident with --batch/parallel, and report the practical elastic-KV ceiling

  jesdga95/basalt#148 · Filter by author Jacou opened on Oct 30, 2026
  1

• Windows: greedy output differs between two runs of the same binary on one card (greedy_identity)

  jesdga95/basalt#146 · Filter by author hagudelom opened on Oct 29, 2026
  2

• Windows: iq_avx512_test frozen hashes differ under MSVC for IQ1_S/IQ2_XXS/IQ2_XS/IQ2_S/IQ4_XS (59/60 cpu tests pass)

  jesdga95/basalt#140 · Filter by author taylor-shift opened on Oct 29, 2026
  1

• Windows: exit code 3221226505

  jesdga95/basalt#137 · Filter by author HetagKoroev opened on Oct 29, 2026

• tests: prefill_fused_iq_test's "v2 vs fused" checks compare v2 with itself (the fused arm runs v2 by default)

  jesdga95/basalt#136 · Filter by author xyehya opened on Oct 29, 2026

• tests: dec_moe_parity_kq fails intermittently in "combine, own wait (landing late)" (16 of 2560 values differ)

  jesdga95/basalt#134 · Filter by author xyehya opened on Oct 29, 2026

• tests: dense_tc_parity exits on cards with fewer than 160 SMs (out IQ3_S needs the one-tile path), skipping the rest of the test

  jesdga95/basalt#133 · Filter by author xyehya opened on Oct 29, 2026

• prefill: the FP16/BF16 cuBLAS prompt products are not width-invariant (CUBLAS_GEMM_DEFAULT picks kernels per T)

  jesdga95/basalt#132 · Filter by author xyehya opened on Oct 29, 2026

• mtp: the proxy draft head keeps the first 64 candidates to arrive, so a draft probability can change run to run

  jesdga95/basalt#131 · Filter by author xyehya opened on Oct 29, 2026

• verify: on cards with fewer than 160 SMs, a token's logits depend on the window width (the mid chain covers T<=4, the unfused out projection rounds differently)

  jesdga95/basalt#130 · Filter by author xyehya opened on Oct 29, 2026

• IQ3_S / IQ2_S CUDA kernels compute wrong dot products in optimized builds (byte-pointer unpack) — model loops on IQ3_XXS pack

  jesdga95/basalt#121 · Filter by author extremecodetv opened on Oct 29, 2026
  1

• ram: since 8d94e805, decode at ~140K context is ~3% slower on a 16 GB card (lent slots lose their RAM copies)

  jesdga95/basalt#108 · Filter by author xyehya opened on Oct 28, 2026

• Vision on Windows: cpu aborts on .img unlink (WinError 32); dynamic cannot write the .sve

  jesdga95/basalt#94 · Filter by author jerry78424 opened on Oct 27, 2026

• Two bitwise speedups from an 82-SM card: Q2_0 expert down in 64-row blocks (763 -> 491 us/window) and pdep sign masks in the AVX-512 i-quant rows (15-23%)

  jesdga95/basalt#86 · Filter by author mitalbert opened on Oct 26, 2026
  3

• sm_120 cluster launches keep at most 2 CTAs per SM: dec_gdn_rec's 192 CTAs run in two waves under 96 SMs; two heads per block takes rec from 587 to 365 us/window (82 SMs, 175 W)

  jesdga95/basalt#85 · Filter by author mitalbert opened on Oct 26, 2026
  1

• PCIe staging is off for single requests and mixed-type packs; enabled, single-request decode is 14% faster on a 16 GB card (300 W)

  jesdga95/basalt#82 · Filter by author arytek opened on Oct 26, 2026
  1

• Dead code: nineteen functions nothing calls, and the ones worth wiring instead

  jesdga95/basalt#78 · Filter by author giveen opened on Oct 26, 2026
  1