| Qwen3.6 MTP can now run in Unsloth Studio! | danielhanchen | open | 0 | disc |
| Is -np 1 / --parallel 1 still required for llama.cpp? I don't see it in examples for MTP any more...or am I missing something? | vijaygill | closed | 0 | disc |
| UD-Q6_K_XL or Q8_0? | XaviCampa | open | 0 | disc |
| DGX Spark --linear-backend flashinfer_b12x issue | algorithmcookies | closed | 0 | disc |
| Question for @unsloth: mobile deployment? | 3morixd | open | 0 | disc |
| Qwen3.6 27B MTP with Vision | zygoat | open | 0 | disc |
| Duplicate tensor name: blk.40.ssm_out.weight for BF16 files | jpetrina | closed | 0 | disc |
| 1000000 Context. How exactly? | BigDeeper | open | 0 | disc |
| Qwen3.6-27B-MTP-UD-IQ4_XS | abdo1948 | open | 0 | disc |
| mmproj support for MTP | hitchhiker3010 | closed | 0 | disc |
| random chinese characters issue with unsloth GGUF for Qwen3.6-27B for the pre and post MTP | Malmuk1 | closed | 0 | disc |
| Qwen3.6-27b-1M上下文需求。 | kelei999999 | closed | 0 | disc |
| Update README.md | dyoung | merged | 0 | disc |
| IQ3XXS giving garbage output | mbhagya | closed | 0 | disc |
| UD-Q4_K_XL produces 0% MTP draft acceptance — blk.64.attn_*/ffn_* quantized below the threshold for MTP rejection sampling | revise | open | 0 | disc |
| inference using vllm | kuopching | open | 0 | disc |
| Official llama.cpp slower than the https://github.com/am17an/llama.cpp.git | adr-insc | open | 0 | disc |
| Official llama.cpp now has '--spec-type draft-mtp' | hellork | closed | 0 | disc |
| the trade off is not good (new update) | rosspanda0 | closed | 0 | disc |
| llama.cpp MTP crash with Unsloth sampler flags | MrWootz | closed | 0 | disc |
| ADD A WAY TO CHANGE THE CONTEXT length of a model in unsloth studio before loading please | curtis1969 | closed | 0 | disc |
| Report: 56 t/s on RTX 4090D (48GB VRAM) with UD-Q6_K_XL | SlavikF | open | 0 | disc |
| Update README.md | dyoung | merged | 0 | disc |
| Noticeable Performance Decrease | WebWeaverWraith | closed | 0 | disc |
| Starts with 50% speedup, but speed very fast decreases | seleznyov | open | 0 | disc |
| Number guessing inconsistency | Enigrand | open | 0 | disc |
| Are you going to release NVFP4 MTP GGUF | sharon8811 | open | 0 | disc |
| One question about this GGUF | alexloops | closed | 0 | disc |
| Low VRAM<=8GB best result (40t/s WITHOUT MTP) | hellork | open | 0 | disc |
| FAST!!!! 39tps! | mazuj2 | open | 0 | disc |
| Does the latest LM Studio not support it? | nk789 | closed | 0 | disc |
| FYI : --spec-type mtp syntax has changed to --spec-type draft-mtp | qenme | open | 0 | disc |
| 50% bump in speed. | BrokeAmerican | open | 0 | disc |
| --n-gpu-layers or -ngl | owao | open | 0 | disc |
| Does it support --spec-draft-device? | dmiche | open | 0 | disc |
| This model is a bulldozer | owao | open | 0 | disc |
| is it work for gemma4? | koyukira | closed | 0 | disc |
| presence-penalty | owao | closed | 0 | disc |
| no GGUF files found in repository unsloth/Qwen3.6-27B-GGUF-MTP | mohamedemam | closed | 0 | disc |
| Stable MTP first release! | danielhanchen | closed | 0 | disc |
| support openbmb/MiniCPM-V-4.6 | beyoru | closed | 0 | disc |
| Don't forget to add MLX too, please! ;-) | egodoyca | closed | 0 | disc |
| These shoudl work on ik_llama.cpp too | ubergarm | closed | 0 | disc |
| QWEN35_MTP requires nextn_predict_layers > 0. | jidaigeist | closed | 0 | disc |
| No nextn_predict_layers? | saldanhacl | open | 0 | disc |