Issues

• Autotuner selections are unstable on SM121; one retune changes 42 of 93 winners and costs about 1.2% decode

  local-inference-lab/b12x#463 · Filter by author sjug opened on Oct 2, 2026

• DeepSeek V4.1 Flash at TP3 on three GB10s: report, and an mcg K3 trellis checkpoint I'm starting on

  local-inference-lab/b12x#460 · Filter by author Defilan opened on Oct 2, 2026

• deepseek-ai/DeepSeek-V4-Flash-Vision-Exp won't load on dual DGX Spark with latest commit

  local-inference-lab/b12x#445 · Filter by author xor2k opened on Sep 29, 2026
  1

• crash when using more than 3 parallel requests and nvidia/Qwen3.8-Flash-Next-NVFP4

  local-inference-lab/b12x#444 · Filter by author xor2k opened on Sep 29, 2026
  2

• mxfp8_linear produces corrupted tokens on GB10 (sm_121) with DeepSeek-V4.1-Flash MXFP8 dense layers via vLLM

  local-inference-lab/b12x#430 · Filter by author tsw2k opened on Sep 25, 2026

• QSA: breakable PIECEWISE graphs bake a 3-wide prefill block table → OOB table read in _compress_completed_groups_kernel (IMA / silent KV corruption) when capture size > max_num_seqs*(1+spec)

  local-inference-lab/b12x#416 · Filter by author randomvariable opened on Sep 23, 2026
  1

• CUDA_ERROR_ILLEGAL_ADDRESS in B12X GDN prefill kernel during CUDA Graph replay

  local-inference-lab/b12x#399 · Filter by author keeratita opened on Sep 20, 2026
  4

• Tracking: port DeepGEMM SM12x feature gaps

  local-inference-lab/b12x#378 · Filter by author lucifer1004 opened on Sep 16, 2026

• Upgrade to nvidia-cutlass-dsl 4.7

  local-inference-lab/b12x#377 · Filter by author lucifer1004 opened on Sep 16, 2026

• perf(moe): dynamic NVFP4 split admission, tile geometry, and capacity are re-resolved on every launch — resolve once at planning

  local-inference-lab/b12x#357 · Filter by author L14nY1Wang opened on Sep 11, 2026

• fix(moe): W4A16 route-pack capacity bounds are compile-time constexprs — recompile-per-capacity and hard ceiling on packed routes

  local-inference-lab/b12x#351 · Filter by author malaiwah opened on Sep 9, 2026

• KDA prefill: export multiple checkpoints on SM120 to preserve fine-grained prefix caching

  local-inference-lab/b12x#340 · Filter by author voipmonitor opened on Sep 8, 2026

• loader: direct checkpoint transfer rejects TP>1 K-sharded (column-sliced) weights — needs a column-narrow transform

  local-inference-lab/b12x#333 · Filter by author original-el8 opened on Sep 6, 2026

• [comm.roce] graph-replayed collective can wedge one rank indefinitely despite spin limit

  local-inference-lab/b12x#313 · Filter by author tobymao opened on Sep 5, 2026
  1

• [Bug]: K6 Trellis route-128 planner undercounts shared memory

  local-inference-lab/b12x#255 · Filter by author malaiwah opened on Aug 29, 2026

• BTX format extension for projection-tiered MCG K3/K4/K5 checkpoints

  local-inference-lab/b12x#242 · Filter by author brandonmmusic-max opened on Aug 23, 2026

• W4A16 c_tmp sizing contract is undocumented: route_slots rounding (64x under-alloc) and a shape-independent cap that should be shared

  local-inference-lab/b12x#235 · Filter by author malaiwah opened on Aug 19, 2026
  2

• Feature: 16-row skinny-M tile for the NVFP4 dense GEMM decode path

  local-inference-lab/b12x#234 · Filter by author malaiwah opened on Aug 19, 2026

• Feature: extend fused_quant_a to the NVFP4 (W4A4) dense GEMM

  local-inference-lab/b12x#233 · Filter by author malaiwah opened on Aug 19, 2026
  1

• Feature: W4A8 dense GEMM (A=e4m3, B=e2m1) for SM120

  local-inference-lab/b12x#232 · Filter by author malaiwah opened on Aug 19, 2026
  1

• Define reference-oracle precedence for degenerate Kimi-K3 top-k routing

  local-inference-lab/b12x#225 · Filter by author voipmonitor opened on Aug 16, 2026
  1

• [perf] Fused NSA indexer merge selector left at 8-bit coarse radix / 2496-slot buffer after #77, so 48-SM (GB10) long-context decode routinely takes the exact-rescan fallback

  local-inference-lab/b12x#195 · Filter by author malaiwah opened on Aug 14, 2026
  1

• FC2 'ultra' tile upgrade selects a tile that _candidate_tile_fits then rejects (breaks DeepSeek-V4 on GB10/sm_121)

  local-inference-lab/b12x#182 · Filter by author mechramc opened on Aug 13, 2026
  2

• security: coordinate IPC unmap completion before freeing Top-K exports

  local-inference-lab/b12x#181 · Filter by author malaiwah opened on Aug 13, 2026
  1

• security: collectively agree DMA all-reduce IPC layout

  local-inference-lab/b12x#180 · Filter by author malaiwah opened on Aug 13, 2026
  1