All pull requests

Pull requests

• vulkan : fix unclosed brace in turbo type guard (backend never compiled)

  ggmlVulkan
  #73·skulze opened Aug 13, 2026·

  ·
  1

• fix gemma4_assistant: arch name, tensor names, MTP dimension alignment

  model
  #72·limitlessstoner opened Aug 12, 2026·

  ·

• 🔨 feat(quants): use the importance matrix when encoding NVFP4

  ggml
  #66·worthant opened Aug 5, 2026·

  ·

• HIP/turbo3: graph-safe decode + inline-dequant TILE prefill on gfx1201 (RDNA4)

  CUDAggmltesting
  #28·KaiFelixBennett opened Jun 14, 2026·

  ·

• Gemma-4 MTP: skip pooling/sampling epilogue for MTP graphs (fixes decode crash) + widen DKQ=512 TILE routing to all gqa_ratio

  ggmlNvidia GPUpython
  #25·PhilEgly opened Jun 6, 2026·

  ·

• gemma4_assistant: protect n_layer_kv_from_start against shared_kv_layers == n_layer

  #24·PhilEgly opened Jun 4, 2026·

  ·

• [codex] Add a reserve-only Gemma4 MTP reserve context

  documentationexamplesserver
  #23·nycdubliner opened Jun 1, 2026·

  ·

• [codex] Add speculative draft Prometheus metrics

  examplesserver
  #22·nycdubliner opened May 31, 2026·

  ·

• docs(turbo-quant): TURBO_LAYER_ADAPTIVE mode 7 validation on Pi16 ARM

  documentation
  #21·WillowOneVision opened May 25, 2026·

  ·

• docs: Gemma-4-E4B-It perplexity benchmark — turbo4 vs F16 cross-corpus

  documentation
  #20·WillowOneVision opened May 24, 2026·

  ·

• Phase C.2 dispatch behavior: MTP+mmproj coexistence behind --allow-mtp-with-mmproj (5th first-in-world)

  examplesservertesting
  #19·WillowOneVision opened May 21, 2026·

  ·
  1

• Phase C.2 foundational APIs: server_tokens coexistence + common_speculative_reset

  examplesservertesting
  #18·WillowOneVision opened May 21, 2026·

  ·

• fix(server): SEGV when --mtp-head + --mmproj are both passed

  examplesserver
  #17·WillowOneVision opened May 21, 2026·

  ·

• ggml: ARM NEON dequant kernel for turbo4 (vqtbl4q_u8 4-bit PolarQuant)

  ggml
  #16·WillowOneVision opened May 21, 2026·

  ·

• fix: add missing prototype for turbo_cpu_fwht_inverse to resolve -Wmissing-prototypes CI error

  ggml
  #12·sujitvasanth opened May 13, 2026·

  ·

• feat: one-sided target probability acceptance for MTP drafts increases acceptance rate and throughput compared to argmax alone

  examplesggmlserver
  #8·sujitvasanth opened May 11, 2026·

  ·
  2

• Enhance CUDA flash attention kernel selection for DKQ=512 with low gq…

  ggmlNvidia GPU
  #6·Ooooze opened May 8, 2026·

  ·
  2

• Repro: MTP path on CUDA aborts at fattn.cu:109 (DKQ=512) for Gemma 4 — Blackwell sm_120 + Ampere sm_86

  documentation
  #5·jameseiten opened May 8, 2026·

  ·
  3