Repository navigation
DFlash2 is not working with --split-mode tensor #27819
Description
Assertion is failed on llama-server startup:
ggml\src\ggml-backend-meta.cpp:543: GGML_ASSERT(src_ss[0].axis != GGML_BACKEND_SPLIT_AXIS_0) failed
Minimal command options to reproduce assert:
llama-server.exe ^
-m Qwen3.8-27B-UD-Q4_K_XL.gguf ^
--model-draft Qwen3.8-27B-DFlash2-Q4_K_M.gguf ^
--spec-type draft-dflash ^
--spec-draft-n-max 2 ^
--ctx-size 16384 ^
--split-mode tensor
It is working with --split-mode layer (no assertion on startup)
More information:
- Windows 10
- Cuda
- RTX 5060 Ti 16 Gb x2
Activity
aki1-asr commentedon Aug 27, 2026
I'm experiencing the
exactsame issue and it's causing me trouble.By the way, I've confirmed that the same problem occurs with DSpark as well.
Additionally, there's no issue with regular DFlash (not v2) or MTP.
Edit:
While the triggering conditions are the same, the specific location where the error occurs differs slightly in my case.
Environment:
art-den commentedon Aug 28, 2026
@aki1-asr I did PR. Try this branch https://github.com/art-den/llama.cpp/tree/fix/dflash2-tensor-split
Anbeeld commentedon Aug 28, 2026
Sharing a separate tensor-split runtime report that may remain relevant after the DFlash2 initialization problem is fixed:
#27750
The separate runtime findings are:
ggml_cuda_ar_acquire_slot allreduce.cu:368 cudaEventSynchronize(p->ev_pool[i][slot].ker)aki1-asr commentedon Aug 29, 2026
@art-den
I fetched and tested PR #27858.
As a result, DSpark loaded correctly and inference worked fine.
However, for DFlash2, while it loads successfully, it crashes during inference with the following error:
The draft model I used is: https://huggingface.co/incoai/Qwen3.8-27B-DFlash2-GGUF
art-den commentedon Aug 29, 2026
@aki1-asr It looks like this is already a problem for ROCm.
Studio: keep tensor split and MTP when Auto would pick a DFlash draft…