Repository navigation

All pull requests

Labels (38)Milestones (0)New pull request

fix(bailingmoe3): load upstream-converted Ling 3.0 GGUFs
#95notforu777 opened Oct 12, 2026
docs: ccache scope with dev as the default branch
#94notforu777 opened Oct 6, 2026
feat(clef): -m <Clef repo dir> points to the GGUF
#92notforu777 opened Oct 5, 2026
vulkan : fix unclosed brace in turbo type guard (backend never compiled)
#73skulze opened Aug 13, 2026
1
fix gemma4_assistant: arch name, tensor names, MTP dimension alignment
#72limitlessstoner opened Aug 12, 2026
🔨 feat(quants): use the importance matrix when encoding NVFP4
#66worthant opened Aug 5, 2026
HIP/turbo3: graph-safe decode + inline-dequant TILE prefill on gfx1201 (RDNA4)
#28KaiFelixBennett 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
#25PhilEgly opened Jun 6, 2026
gemma4_assistant: protect n_layer_kv_from_start against shared_kv_layers == n_layer
#24PhilEgly opened Jun 4, 2026
[codex] Add a reserve-only Gemma4 MTP reserve context
#23nycdubliner opened Jun 1, 2026
[codex] Add speculative draft Prometheus metrics
#22nycdubliner opened May 31, 2026
docs(turbo-quant): TURBO_LAYER_ADAPTIVE mode 7 validation on Pi16 ARM
#21WillowOneVision opened May 25, 20265 of 6 tasks
docs: Gemma-4-E4B-It perplexity benchmark — turbo4 vs F16 cross-corpus
#20WillowOneVision opened May 24, 20263 tasks
Phase C.2 dispatch behavior: MTP+mmproj coexistence behind --allow-mtp-with-mmproj (5th first-in-world)
#19WillowOneVision opened May 21, 20264 of 9 tasks
1
Phase C.2 foundational APIs: server_tokens coexistence + common_speculative_reset
#18WillowOneVision opened May 21, 20265 of 6 tasks
fix(server): SEGV when --mtp-head + --mmproj are both passed
#17WillowOneVision opened May 21, 20264 of 5 tasks
ggml: ARM NEON dequant kernel for turbo4 (vqtbl4q_u8 4-bit PolarQuant)
#16WillowOneVision opened May 21, 20265 tasks done
fix: add missing prototype for turbo_cpu_fwht_inverse to resolve -Wmissing-prototypes CI error
#12sujitvasanth opened May 13, 2026
feat: one-sided target probability acceptance for MTP drafts increases acceptance rate and throughput compared to argmax alone
#8sujitvasanth opened May 11, 2026
2
Enhance CUDA flash attention kernel selection for DKQ=512 with low gq…
#6Ooooze 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
#5jameseiten opened May 8, 2026
3