Issues

• MiMo-V2.6-Flash REAP50 Q2_K on CPU: real-file validation (quality, perplexity, tasks)

  architecture
  jamesburton/dotLLM#943 · Filter by author jamesburton opened on Oct 13, 2026

• perf(vulkan): nemotron_h_moe prefill - grouped coopmat Q5_1/Q8_0 expert path (0.41x llama.cpp pp512)

  perfvulkan
  jamesburton/dotLLM#942 · Filter by author jamesburton opened on Oct 12, 2026

• [#825 stage 2] Expert tiering design doc + small-budget prototype (provider abstraction, slot allocator, id->slot remap)

  jamesburton/dotLLM#939 · Filter by author jamesburton opened on Oct 12, 2026

• [#825 stage 1] Expert routing traces + LRU/static hit-rate gate (qwen4exp)

  jamesburton/dotLLM#938 · Filter by author jamesburton opened on Oct 12, 2026

• [#825 stage 0] Expert tiering I/O primitives: NVMe -> GPU-visible RAM throughput, latency, cache-bloat (measure on Strix Halo)

  jamesburton/dotLLM#937 · Filter by author jamesburton opened on Oct 12, 2026

• MiMo-V2.6: 9B works today; Flash (mimo2, 310B MoE) feasibility, gap list and plan

  architecture
  jamesburton/dotLLM#935 · Filter by author jamesburton opened on Oct 12, 2026
  1

• Chat template override missing (--chat-template-file): cannot use ggml-org's patched MiMo template / prefilled <think>

  enhancement
  jamesburton/dotLLM#934 · Filter by author jamesburton opened on Oct 12, 2026

• perf(cpu): Gemma-4 E2B CPU is ~9x (pp) / ~3-4x (tg) behind llama.cpp - attribute and close

  jamesburton/dotLLM#933 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] CPU MoE: gpt-oss prefill is per-token with a float MXFP4 kernel (pp128 21.6 vs llama.cpp 120.6); GLM MLA on host F32 (tg 2.6 vs 20)

  cpuperfperf-survey-2026-10-12
  jamesburton/dotLLM#922 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] Vulkan has no Q4_0/Q4_1/Q5_1 dense kernels (weights widened to F32); QUANTIZATION.md loader table stale

  perfperf-survey-2026-10-12vulkan
  jamesburton/dotLLM#921 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] Gemma-4 grouped-MoE prefill arm is gated on exactly Q4_K gate/up + Q8_0/Q5_1 down: other quants of 26B-A4B fall off it silently

  perfperf-survey-2026-10-12vulkan
  jamesburton/dotLLM#920 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] Vulkan keeps a full vocab x hidden F32 device embedding table for Q4_K/Q6_K embeds (Llama-8B 2.0 GiB = 43% of the file)

  perfperf-survey-2026-10-12vulkan
  jamesburton/dotLLM#919 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] Dense Vulkan KV cache is F32, eager and full-length: Llama-8B decode 32 -> 21.5 tok/s at depth 8192 (llama.cpp 29-31), Gemma-4 26B 34.6 -> 25.7 at 6K

  perfperf-survey-2026-10-12vulkan
  jamesburton/dotLLM#918 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] Dense Vulkan models: any 2-8 row forward costs 3.1x a 1-row forward (Llama-8B 83 ms vs 27 ms) - no small-row GEMV

  perfperf-survey-2026-10-12vulkan
  jamesburton/dotLLM#917 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] gpt-oss sink attention scans every KV tile and has no split-KV/flash path: tg 40 -> 22 at depth 1024, 26% of pp512

  perfperf-survey-2026-10-12vulkan
  jamesburton/dotLLM#915 · Filter by author jamesburton opened on Oct 12, 2026

• [perf-survey] Vulkan MLA runs on F32-widened weights + scalar matmul + expanded KV: GLM-4.7-Flash pp 53 vs 343-470, 84% of prefill in attn_block_mla

  perfperf-survey-2026-10-12vulkan
  jamesburton/dotLLM#913 · Filter by author jamesburton opened on Oct 12, 2026

• Streaming tool calls omit indexes and corrupt parallel calls in OpenAI clients

  jamesburton/dotLLM#894 · Filter by author jamesburton opened on Oct 11, 2026

• qwen4exp Vulkan model load takes 158-169 s for 69 GiB (~0.44 GB/s): profile and speed up

  jamesburton/dotLLM#874 · Filter by author jamesburton opened on Oct 9, 2026
  1

• README identity cleanup (CI badge, clone URL, author, contributing links point at upstream; News mixes kkokosa/jamesburton links)

  jamesburton/dotLLM#859 · Filter by author jamesburton opened on Oct 8, 2026

• Track: hybrid CPU-spill placement for DeepSeek-V4-Flash / Qwen3-235B-class models

  jamesburton/dotLLM#858 · Filter by author jamesburton opened on Oct 8, 2026

• Deferred test: old (96 GB carve-out) vs dynamic (512 MB) BIOS split A/B on same build

  jamesburton/dotLLM#857 · Filter by author jamesburton opened on Oct 8, 2026

• [qwen4exp B1] Benchmark harness with Strata as target (M)

  jamesburton/dotLLM#826 · Filter by author jamesburton opened on Oct 8, 2026
  1

• [qwen4exp T1] Expert tiering / cache (VRAM <-> RAM <-> SSD) (L, defer)

  jamesburton/dotLLM#825 · Filter by author jamesburton opened on Oct 8, 2026
  1

• [qwen4exp C1] CUDA port (build-verify here, tiny fixtures on T5500/Framework) (L, low priority)

  jamesburton/dotLLM#824 · Filter by author jamesburton opened on Oct 8, 2026

• [qwen4exp N1] N-gram table prefetch/offload service (M)

  jamesburton/dotLLM#822 · Filter by author jamesburton opened on Oct 8, 2026
  1