Issues

• 欢迎进群！能看到这里的，都是懂技术，感兴趣，深入了解moe4all的大佬。欢迎进群讨论！

  Headmaster218/MoE4All#159 · Filter by author Headmaster218 opened on Nov 2, 2026

• [Other] poorseek: an independent, single-config (Arc Pro B70 + 192 GB) SYCL line that started here

  Headmaster218/MoE4All#151 · Filter by author dingyiyuan opened on Nov 1, 2026

• 0xC0000005 on first chat prefill — coopmat MoE prefill crashes on RX 9070 (RDNA4 reported as AmdRdna3)

  Headmaster218/MoE4All#144 · Filter by author DSOTMRHW opened on Oct 30, 2026

• infr 1.1.0 预填充段错误：-p 在 513–800 崩溃，但 1024/2048 正常

  Headmaster218/MoE4All#127 · Filter by author CYCLC opened on Oct 27, 2026
  1

• 使用gemma4模型进行对话时会陷入循环

  Headmaster218/MoE4All#126 · Filter by author TONYBTK opened on Oct 27, 2026
  1

• backend: moe pager: routed hit set contains a missing expert

  Headmaster218/MoE4All#120 · Filter by author kaiwater opened on Oct 25, 2026
  1

• 【Linux/RADV】RX 9070 (gfx1201) + Ornith 1.5 35B MiniPlus v2.1：batch-1 decode 只有 2.4 tok/s —— host 单线程 100% 用户态，GPU 时钟全程停在 25 MHz

  Headmaster218/MoE4All#118 · Filter by author luffffyz opened on Oct 23, 2026
  1

• 欢迎进群！能看到这里的，都是懂技术，感兴趣，深入了解moe4all的大佬。欢迎进群讨论！

  Headmaster218/MoE4All#83 · Filter by author Headmaster218 opened on Oct 13, 2026

• 【KV cache 优化】每 query 只读 2,051 个 cell（262K 下 0.78%）：常驻环从 verify window 反推、graph 内的 CLOCK 淘汰，与 expert cache 的 1:1 兑换（实测 50.9→62.6 tok/s）

  Headmaster218/MoE4All#71 · Filter by author not-zen-girl opened on Oct 8, 2026

• 【MTP 优化】MTP 固定资源先于 expert arena 扣除、draft 层 K/V 只留 window：一次 ~5,500 次 host 往返的教训，与它在 AMD 上的反例

  Headmaster218/MoE4All#70 · Filter by author not-zen-girl opened on Oct 8, 2026

• 【CUDA graph 捕获】单张 graph 承载 48 层动态分页：录制期冻结参数的三个坑，与零 host 往返的门铃握手（对照 #31 / #66 / INC-20260925）

  Headmaster218/MoE4All#69 · Filter by author not-zen-girl opened on Oct 8, 2026

• 【CPU 为主】Strata(C++/CUDA) 源码逐行分析：CPU 专家路径 663.6 MB/token 的完整工作量核算，与 4 项可移植机制（为 #66 补上目标刻度）

  Headmaster218/MoE4All#68 · Filter by author not-zen-girl opened on Oct 8, 2026

• ### RFC: 探讨在现代硬件（AVX-512 + DDR5）上重新评估 CPU 辅助专家计算的可行性

  Headmaster218/MoE4All#66 · Filter by author weiplwei opened on Oct 4, 2026
  12

• deepseek4: two #[ignore]d Vulkan GPU tests fail on RDNA3

  Headmaster218/MoE4All#64 · Filter by author HuangShanglong opened on Oct 4, 2026

• Performance on Linux, Turing, MiltiGPU.

  Headmaster218/MoE4All#57 · Filter by author shrisha108 opened on Sep 29, 2026

• Windows NVIDIA 4080 SUPER: first-prefill device lost; GPU-AV splitk_reduce OOB writes and 47.57GiB heap accounting

  Headmaster218/MoE4All#52 · Filter by author F4ult-a11 opened on Sep 26, 2026

• NVIDIA/Windows: qwen4exp vision off-grid K (4304) assert panic; after relaxing, LLM prefill hangs at 0% CPU

  Headmaster218/MoE4All#51 · Filter by author rogerhik999-arch opened on Sep 26, 2026
  1

• Support PrismML Ternary-Bonsai-2 weight quants (ggml type 142 / 143) - Bonsai-2-27B fails to load

  Headmaster218/MoE4All#48 · Filter by author longge1213812138 opened on Sep 24, 2026

• [A770 / Xe1] v0.7.0 decode is ~24% slower than v0.6.0-beta.1 on an identical command line (and loses half its lead over llama.cpp Vulkan)

  Headmaster218/MoE4All#41 · Filter by author skiyo0177-lgtm opened on Sep 20, 2026
  8

• [MTP] Restore MTP (INFR_MTP parked since ~2026-07-31) — tool-call parser request withdrawn

  Headmaster218/MoE4All#36 · Filter by author skiyo0177-lgtm opened on Sep 9, 2026
  1

• 【Perf】分页 decode 瓶颈在单线程 paging 循环（~2GB/s 车道），NVMe 几乎闲置——resolve/refill 并行化建议

  Headmaster218/MoE4All#31 · Filter by author StarBalll opened on Sep 2, 2026
  1

• 【验证报告】Strix Halo (gfx1151/RADV/Linux, 40CU RDNA3.5) Phase-1 端到端验证 + 构建注意事项

  Headmaster218/MoE4All#30 · Filter by author StarBalll opened on Sep 2, 2026

• 【APU/UMA 假死】device.ram_budget 不约束 GPU 侧分配，UMA 上 RAM/VRAM 两层预算物理重复计数导致整机 swap 抖动假死

  Headmaster218/MoE4All#28 · Filter by author StarBalll opened on Sep 2, 2026
  1

• CPU报错“双重故障”

  Headmaster218/MoE4All#25 · Filter by author dyyuan01-spec opened on Sep 1, 2026

• 能不能支持rocm10呢？

  Headmaster218/MoE4All#24 · Filter by author buxlc opened on Sep 1, 2026
  3