Skip to content

Eval bug: Dflash2: vision support is still broken #27862

Description

@showgood163

Name and Version

llama-server --version
version: 0.3.0-dev (build 10665, commit ca3d5a3)
built with GNU 15.3.0 for Linux x86_64

Operating systems

Linux

GGML backends

CUDA

Hardware

Ryzen 9700x + rtx 5090

Models

https://huggingface.co/incoai/Qwen3.8-27B-DFlash2-GGUF/tree/main?show_file_info=Qwen3.8-27B-DFlash2-Q4_K_M.gguf

Problem description & steps to reproduce

Previous report: #27342 (comment)

After some testing in with qwen 3.8 in llama.cpp webui, I found out that vision support (with PDFs) is broken, while the native mtp is not.

checkpoint from https://huggingface.co/incoai/Qwen3.8-27B-DFlash2-GGUF
parameters (as suggested):
model-draft = Qwen3.8-27B-DFlash2-Q4_K_M.gguf
spec-type = draft-dflash
spec-draft-n-max = 7

First Bad Commit

No response

Relevant log output

Logs
[ 6483] 0.06.244.043 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
[ 6483] 0.06.244.050 I slot   operator(): id  0 | task 0 | new prompt, n_ctx_slot = 131072, n_keep = 0, task.n_tokens = 16853
[ 6483] 0.06.244.063 I slot   operator(): id  0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end)
[ 6483] 0.06.244.450 I sched_reserve: reserving ...
[ 6483] 0.06.273.694 I sched_reserve:      CUDA0 compute buffer size =   300.02 MiB
[ 6483] 0.06.273.696 I sched_reserve:  CUDA_Host compute buffer size =   148.02 MiB
[ 6483] 0.06.273.697 I sched_reserve: graph nodes  = 4663
[ 6483] 0.06.273.697 I sched_reserve: graph splits = 2
[ 6483] 0.06.273.697 I sched_reserve: reserve took 29.23 ms, sched copies = 1
[ 6483] 0.06.503.782 I slot   operator(): id  0 | task 0 | cached n_tokens = 512, memory_seq_rm [512, end)
[ 6483] 0.06.560.675 I slot   operator(): id  0 | task 0 | cached n_tokens = 690, memory_seq_rm [690, end)
[ 6483] 0.06.576.548 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 689, pos_max = 689, n_tokens = 690, size = 163.111 MiB)
[ 6483] 0.06.598.228 I slot   operator(): id  0 | task 0 | cached n_tokens = 722, memory_seq_rm [722, end)
[ 6483] 0.06.598.347 I slot process_mtmd: id  0 | task 0 | encoding mtmd batch from idx = 722, n_chunks = 1
[ 6483] 0.06.717.374 I decoding image batch 1/3, n_tokens_batch = 512
[ 6483] 0.06.718.166 W find_slot: non-consecutive token position 722 after 721 for sequence 0 with 512 new tokens
[ 6483] 0.06.718.315 W find_slot: non-consecutive token position 722 after 721 for sequence 0 with 512 new tokens
[ 6483] 0.06.849.657 I image decoded (batch 1/3) in 132 ms
[ 6483] 0.06.849.663 I decoding image batch 2/3, n_tokens_batch = 512
[ 6483] 0.06.850.590 W find_slot: non-consecutive token position 722 after 722 for sequence 0 with 512 new tokens
[ 6483] 0.06.850.730 W find_slot: non-consecutive token position 722 after 722 for sequence 0 with 512 new tokens
[ 6483] 0.06.981.035 E init: the tokens of sequence 0 in the input batch have inconsistent sequence positions:
[ 6483]  - the last position stored in the memory module of the context (i.e. the KV cache) for sequence 0 is X = 722
[ 6483]  - the tokens for sequence 0 in the input batch have a starting position of Y = 722
[ 6483]  it is required that the sequence positions remain consecutive: Y = X + 1
[ 6483] 0.06.981.038 E decode: failed to initialize batch
[ 6483] 0.06.981.038 E llama_decode: failed to decode, ret = -1
[ 6483] 0.06.981.039 E process: llama_decode(ctx_dft) failed rc=-1 (n_tokens=512, offset=0)
[ 6483] 0.06.981.041 E post-decode callback failed
[ 6483] 0.06.981.044 E slot   operator(): id  0 | task 0 | failed to decode mtmd chunk, idx = 722, res = 1
[ 6483] 0.06.981.066 E slot   operator(): id  0 | task 0 | failed to process mtmd chunk, res = -1
[ 6483] 0.06.981.067 E srv    send_error: task id = 0, error: failed to process mtmd chunk
[ 6483] 0.06.981.069 I slot      release: id  0 | task 0 | stop processing: n_tokens = 722, truncated = 0
[ 6483] 0.06.981.079 W srv          stop: cancel task, id_task = 0

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions