Issues

• vulkan: Q6_K rows / Q8_0 multi decay above ~10 columns (VGPR-bound); needs ISA/RGP

  jamesburton/dotLLM#1523 · Filter by author jamesburton opened on Oct 31, 2026

• vulkan: Q5_K decode (1 column) on the deterministic multi-column shader so verify == decode bits

  jamesburton/dotLLM#1522 · Filter by author jamesburton opened on Oct 31, 2026

• vulkan: Q6_K rows GEMV 5-column NR 4 cliff in the dense qwen35 host, e2e unresolved

  jamesburton/dotLLM#1521 · Filter by author jamesburton opened on Oct 31, 2026

• CLI model resolver: org/file.gguf path for a model stored in a repo subfolder triggers an HF pull 404 instead of resolving locally

  jamesburton/dotLLM#1518 · Filter by author jamesburton opened on Oct 31, 2026

• Vulkan gemma-embedding: FFN-down F16 overflow margin (peak ~6.2e4 vs 65504) and comparison vs the CPU inter-sequence batcher

  jamesburton/dotLLM#1509 · Filter by author jamesburton opened on Oct 31, 2026

• telemetry: add instruments for prefix-cache hit/miss by tier, snapshot bytes, MTP drafted/accepted, device memory (expose via /metrics)

  jamesburton/dotLLM#1505 · Filter by author jamesburton opened on Oct 31, 2026

• vulkan: validate q8_1_exact_div.glsl (q81Recip fma refinement) on Arc / other vendors (from #1340)

  jamesburton/dotLLM#1501 · Filter by author jamesburton opened on Oct 31, 2026

• vulkan: KV-cache quantize shaders (kv_quantize_q8_0/q4_0) use GLSL round() and amax/127.0 - same rounding-contract gap as the Q8_1 activation quantizers (from #1340)

  jamesburton/dotLLM#1500 · Filter by author jamesburton opened on Oct 31, 2026

• cuda: activation quantizers use inv_scale = 127/amax but store d = amax/127 (mixed convention vs CPU/Vulkan Q8_1) - T5500 validation (from #1340)

  jamesburton/dotLLM#1499 · Filter by author jamesburton opened on Oct 31, 2026

• flaky test: GraniteToolCallParserTests.RealOutput_Streamed_NoMarkerOrJsonInContent_CallsInFinalChunk(granite-toolcall-parallel) JsonReaderException on CI

  jamesburton/dotLLM#1498 · Filter by author jamesburton opened on Oct 31, 2026

• Logits rows-only: remaining CPU hybrid states, CUDA (G3L) and Vulkan (V6L) (follow-up of #1137)

  jamesburton/dotLLM#1485 · Filter by author jamesburton opened on Oct 31, 2026
  1

• NPU (XDNA2) background embedder for BERT-class models

  jamesburton/dotLLM#1476 · Filter by author jamesburton opened on Oct 31, 2026

• CUDA BERT encoder embedding backend

  jamesburton/dotLLM#1475 · Filter by author jamesburton opened on Oct 31, 2026

• Vulkan encoder: F16 activations, fused QKV, fewer per-document attention dispatches

  jamesburton/dotLLM#1474 · Filter by author jamesburton opened on Oct 31, 2026
  1

• DFlash draft-vocab limit (#1360 follow-up): measure on CUDA (3060) and CPU, test CJK prompts, consider a non-prefix top-K set

  jamesburton/dotLLM#1468 · Filter by author jamesburton opened on Oct 31, 2026

• embeddings: packed Vulkan forward follow-ups (single varlen attention dispatch, other decoder archs, mean-pool tail)

  jamesburton/dotLLM#1465 · Filter by author jamesburton opened on Oct 31, 2026
  1

• vulkan: dense GEMM path (17+ rows) intermittently runs 2-3x slower within one session; cause unattributed

  jamesburton/dotLLM#1462 · Filter by author jamesburton opened on Oct 31, 2026

• [perf] Small-row GEMV leftovers after #1420: Q2_K/Q3_K/Q4_1/BF16 multi, Nemotron-H 5-row NR tuning, dense decode-on-rows identity, scheduler for plain Vulkan dense

  jamesburton/dotLLM#1459 · Filter by author jamesburton opened on Oct 30, 2026
  1

• CudaQwen3HybridDensePerRowKvqVerifyBench depth 32768 stalled and the T5500 dropped off the network

  jamesburton/dotLLM#1444 · Filter by author jamesburton opened on Oct 30, 2026

• perplexity layer-window/cycling: CUDA paths unvalidated on GPU; gaps with --kv-cache, safetensors, --device auto

  jamesburton/dotLLM#1443 · Filter by author jamesburton opened on Oct 30, 2026
  1

• embeddings: CUDA models do not implement IEmbeddingModel (/v1/embeddings returns 501 on CUDA)

  jamesburton/dotLLM#1438 · Filter by author jamesburton opened on Oct 30, 2026

• perf(cuda): evaluate guarded flash prefill in hybrid model

  jamesburton/dotLLM#1429 · Filter by author jamesburton opened on Oct 30, 2026
  1

• server: explicit auto KV selection can inherit the previous model's unsupported Q8 cache

  jamesburton/dotLLM#1416 · Filter by author jamesburton opened on Oct 29, 2026
  1

• flaky (unreproduced): OllamaApiEndpointTests.Tags_ListsProfilesAndLocalAndOllamaStoreModels + ModelSwapRobustnessTests keep-alive

  jamesburton/dotLLM#1395 · Filter by author jamesburton opened on Oct 28, 2026
  2

• cuda: multi-column GEMVs cap at 8 columns (Pq2_0GemvMultiMaxColumns, CudaQ8_0RbGemv, CudaSmallSGemvDispatch) - 9..16-row verify cliff not measured

  jamesburton/dotLLM#1375 · Filter by author jamesburton opened on Oct 28, 2026