All issues

Search results

Eval bug: --spec-type draft-mtp segfaults on Qwen3.6-35B-A3B (MoE) — works fine on dense Qwen3.8-27B

Status: Open.

#12 In mxxm-t/mx-llama.cpp;

· zhou51320 opened on Sep 8, 2026

Working 4x MI50 (gfx906) recipe for Qwen3.8-Flash-Next + note: tensor-split needs -fa on

Status: Open.

#11 In mxxm-t/mx-llama.cpp;

· JackDanger opened on Sep 5, 2026

Small-batch matmul cost curve on 2x MI50 (Q8_0 repack, tensor split)

Status: Open.

#10 In mxxm-t/mx-llama.cpp;

· JCraigWasTaken opened on Sep 4, 2026

Eval bug: 2x MI50 tensor split aborts in RCCL on the first AllReduce once peer access works, no fallback to butterfly

Status: Open.

#7 In mxxm-t/mx-llama.cpp;

· JCraigWasTaken opened on Sep 2, 2026