Eval bug: MCP Tool Lock/Crash with large (write) payload

Aberta
#200 0 comentários 0 reações 0 responsáveis Ver no GitHub

Ninguém assumiu esta issue ainda.

Avaliação

Dificuldade
4/5
Tempo estimado
3-5 dias
Facilidade para iniciantes
48/100
Tipo de issue
Bug
Clareza
Razoavelmente clara
Status de atividade
Ativa
Stack de tecnologia
cpp

Direção de pesquisa

Start by tracing the MCP tool request path in the server, focusing on how large tool arguments are handled and where requests can time out. The report reproduces a failed Desktop-Commander write with an approximately 8 KB payload; compare that with a successful write using another llama-server and the reported workaround of keeping writes below 4 KB. Done means large writes complete and subsequent MCP tool calls remain usable.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

Name and Version

version: 0.4.9 (build 12930, commit b46c53e54)
built with Clang 20.1.8 for Windows x86_64
ggml op count: 106

Operating systems

Windows

GGML backends

CUDA

Hardware

Ryzen 7800X3D + RTX 5080

Models

https://huggingface.co/DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NM-DAU-NEO-MTP-GGUF/blob/main/Qwen3.8-27B-TTURBO-Fable-C-Fusion-709-L-Uncen-NM-DAU-NEO-MAX-MTP-Q4_K_M.gguf

Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NM-DAU-NEO-MTP-GGUF Q4_K_M

Problem description & steps to reproduce

Server launched with batch file:

set SERVER=C:\Tools\AI\LlamaCustom\beellamawin64\llama-server.exe
set MODEL_DIR=C:\Tools\AI\LlamaCustom\llama-server

%SERVER% --models-preset "%MODEL_DIR%\models.ini" ^
    --host 0.0.0.0 ^
    --port 3337 ^
    -fit off ^
    -ngl 64 ^
    -sm none ^
    -b 512 ^
    -ub 512 ^
    -ubd 256 ^
    --timeout 7200 ^
    --kv-tail-tokens 1024 ^
    --spec-type draft-mtp ^
    --spec-draft-n-max 3 ^
    --spec-draft-p-min 0.0 ^
    --spec-draft-type-k q5_0 ^
    --spec-draft-type-v q4_1 ^
    -fa 1 ^
    --cache-type-k kvarn5 ^
    --cache-type-v kvarn5 ^
    --override-tensor "blk\.(63|62|61|60|59|58|57|56|55|25|54|53|52|50|26|24|38|51|40|27|35|22|41|39|36|21|3|42|34|30|20|6|4|49|47|43|37|32|23|10|8|7|5|2|1|48|46|45|44)\.ffn_.*=CPU" ^
    --no-mmproj-offload ^
    --image-min-tokens 1024 ^
    --parallel 1 ^
    -lm none ^
    -t 8 ^
    --jinja ^
    --reasoning on ^
    --chat-template-kwargs "{\"preserve_thinking\":true,\"reasoning_effort\":\"xhigh\"}" ^
    --temp 0.6 ^
    --mcp-servers-config mcp_config.json ^
    --top-p 0.95 ^
    --top-k 20 ^
    --min-p 0.0 ^
    --repeat-penalty 1.0 ^
    --presence-penalty 0.0

pause

https://github.com/wonderwhy-er/DesktopCommanderMCP Add Desktop-Commander MCP to mcp_config.json

Open a new chat in WebUI and, paste a block of text (~8Kb of text was chosen for testing) and prompt to write this text to a file.

The write will fail, task won't complete.

Try using any other Desktop-Commander MCP tool, it will fail. Other MCP tools will function fine.

It is presumed that the issue is with beellama itself rather than the MCP used, as this same MCP was used with a different llama-server to write a file over 80Kb in size without issue. Requesting (in prompt) to keep writes/edits small (below 4kb payload) appears to serve as a workaround.

First Bad Commit

No response

Relevant log output
Write this to a new txt file in DIRECTORY Logs
<think>
The user wants me to write the DS4Windows log content to a new txt file in the specified directory. Let me use the write_file tool to create this file.

</think>


{
  "name": "Desktop Commander_write_file",
  "arguments": {
    "path": "TEXTFILE",
    "content": "TEXTCONTENT"
  }
}


[Tool Result]
request timed out


<think>
The write timed out. Let me try again - this could be a transient issue.

</think>
[52815] 5.12.823.587 I slot print_timing: id  0 | task 0 | n_gen =   3231, tg =  11.23 t/s, tg_3s =  11.17 t/s
[52815] 5.13.227.651 I sched_graph_realloc: graph reallocation: nodes = 4876 (changed = 0), splits = 102, n_same = 13, n_diff = 0
[52815] 5.16.045.924 I slot print_timing: id  0 | task 0 | n_gen =   3267, tg =  11.23 t/s, tg_3s =  11.17 t/s
[52815] 5.19.279.273 I slot print_timing: id  0 | task 0 | n_gen =   3303, tg =  11.23 t/s, tg_3s =  11.13 t/s
[52815] 5.22.471.619 I slot print_timing: id  0 | task 0 | n_gen =   3339, tg =  11.23 t/s, tg_3s =  11.28 t/s
[52815] 5.25.746.654 I slot print_timing: id  0 | task 0 | n_gen =   3375, tg =  11.23 t/s, tg_3s =  10.99 t/s
[52815] 5.29.010.866 I slot print_timing: id  0 | task 0 | n_gen =   3411, tg =  11.22 t/s, tg_3s =  11.03 t/s
[52815] 5.32.210.782 I slot print_timing: id  0 | task 0 | n_gen =   3447, tg =  11.22 t/s, tg_3s =  11.25 t/s
[52815] 5.35.449.515 I slot print_timing: id  0 | task 0 | n_gen =   3483, tg =  11.22 t/s, tg_3s =  11.12 t/s
[52815] 5.36.220.903 I sched_graph_realloc: graph reallocation: nodes = 4876 (changed = 0), splits = 102, n_same = 14, n_diff = 0
[52815] 5.38.685.029 I slot print_timing: id  0 | task 0 | n_gen =   3519, tg =  11.22 t/s, tg_3s =  11.13 t/s
[52815] 5.41.917.543 I slot print_timing: id  0 | task 0 | n_gen =   3555, tg =  11.22 t/s, tg_3s =  11.14 t/s
[52815] 5.45.209.639 I slot print_timing: id  0 | task 0 | n_gen =   3591, tg =  11.22 t/s, tg_3s =  10.94 t/s
[52815] 5.48.429.915 I slot print_timing: id  0 | task 0 | n_gen =   3627, tg =  11.22 t/s, tg_3s =  11.18 t/s
[52815] 5.51.682.857 I slot print_timing: id  0 | task 0 | n_gen =   3663, tg =  11.22 t/s, tg_3s =  11.07 t/s
[52815] 5.54.927.188 I slot print_timing: id  0 | task 0 | n_gen =   3699, tg =  11.22 t/s, tg_3s =  11.10 t/s
[52815] 5.58.144.801 I slot print_timing: id  0 | task 0 | n_gen =   3735, tg =  11.21 t/s, tg_3s =  11.19 t/s
[52815] 5.59.231.258 I sched_graph_realloc: graph reallocation: nodes = 4876 (changed = 0), splits = 102, n_same = 15, n_diff = 0
[52815] 6.01.347.417 I slot print_timing: id  0 | task 0 | n_gen =   3770, tg =  11.21 t/s, tg_3s =  10.93 t/s
[52815] 6.04.651.995 I slot print_timing: id  0 | task 0 | n_gen =   3806, tg =  11.21 t/s, tg_3s =  10.89 t/s
[52815] 6.07.834.057 I slot print_timing: id  0 | task 0 | n_gen =   3842, tg =  11.21 t/s, tg_3s =  11.31 t/s
[52815] 6.09.255.874 I slot print_timing: id  0 | task 0 | prompt eval time =   17803.44 ms / 20572 tokens (    0.87 ms per token,  1155.51 tokens per second)
[52815] 6.09.255.879 I slot print_timing: id  0 | task 0 |        eval time =  344059.38 ms /  3856 tokens (   89.25 ms per token,    11.20 tokens per second)
[52815] 6.09.255.880 I slot print_timing: id  0 | task 0 |       total time =  361862.82 ms / 24428 tokens
[52815] 6.09.255.880 I slot print_timing: id  0 | task 0 |    graphs reused =        951
[52815] 6.09.255.883 I slot print_timing: id  0 | task 0 |   graph reallocs =         15    /     0        (same size) / (changed size)
[52815] 6.09.255.892 I slot print_timing: id  0 | task 0 | draft acceptance = 0.99552 ( 2888 accepted /  2901 generated), mean len =  3.99
[52815] 6.09.255.892 I slot print_timing: id  0 | task 0 |      acc per pos = (1.000, 0.999, 0.988)
[52815] 6.09.255.915 I spec common_specu: statistics        draft-mtp: #calls(b,g,a) =    1    967    967, #gen drafts =    967, #acc drafts =   967, #gen tokens =   2901, #acc tokens =  2888, #mean acc len = 3.99, #acc rate/pos = (1.000, 0.999, 0.988), dur(b,g,a) = 0.005, 21677.743, 1.151 ms
[52815] 6.09.256.376 I slot      release: id  0 | task 0 | stop processing: n_tokens = 24427, truncated = 0
[52815] 6.09.256.392 I srv  update_slots: all slots are idle
6.59.067.612 I srv  proxy_reques: proxying request to model Qwen3.8-27B-TTurbo on port 52815
[52815] 6.40.966.430 I srv  server_strea: conv_id=5f702c58-83b0-42d1-a7cc-1627fe41cd84::Qwen3.8-27B-TTurbo (empty=0)
[52815] 6.40.991.373 I slot get_availabl: id  0 | task -1 |  - checking restorable sim = 0.999 (24427 restorable, 24427 lexical/24446) > 0.100
[52815] 6.40.991.378 I slot get_availabl: id  0 | task -1 | selected slot by restorable prefix, n_restorable = 24427, f_sim_best = 0.999 (> 0.100 thold), f_keep = 1.000
[52815] 6.40.992.451 I cmn  common_reaso: activated, budget=2147483647 tokens
[52815] 6.40.992.602 I slot launch_slot_: id  0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> ?min-p -> ?xtc -> temp-ext -> dist
[52815] 6.40.992.612 I slot launch_slot_: id  0 | task -1 | sampler params:
[52815]         repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
[52815]         dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64
[52815]         top_k = 20, top_p = 0.950, min_p = 0.000, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.600
[52815]         mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
[52815] 6.40.992.615 I slot launch_slot_: id  0 | task 1010 | processing task, is_child = 0
[52815] 6.40.992.620 I slot   operator(): id  0 | task 1010 | new prompt, n_ctx_slot = 170240, n_keep = 0, task.n_tokens = 24446
[52815] 6.40.992.656 I slot   operator(): id  0 | task 1010 | checking checkpoint with [20479, 20479] against 24427...
[52815] 6.40.993.731 I llama_kv_cache: standard KV tail plan: storage=compact_overlay requested=1024 effective=1024 window=170240 body=f16/f16 actual=f16/f16 physical_rows=170240 arena=1536 sink=0 overlay_rows=1536 promotion=0.00 MiB overlay=0.00 MiB
[52815] 6.40.993.737 I llama_kv_cache: size =    0.00 MiB (170240 cells,   0 layers,  1/1 seqs), K (none):    0.00 MiB, V (none):    0.00 MiB
[52815] 6.40.993.741 I llama_kv_cache: n_rot_k = 0, n_embd_head_k_all = 0
[52815] 6.40.993.742 I llama_kv_cache: n_rot_v = 0, n_embd_head_k_all = 0
[52815] 6.41.157.813 I slot   operator(): id  0 | task 1010 | restored context checkpoint (pos_min = 20479, pos_max = 20479, n_tokens = 20480, n_past = 20480, size = 238.673 MiB)
[52815] 6.41.157.818 I slot   operator(): id  0 | task 1010 | cached n_tokens = 20480, memory_seq_rm [20480, end)
[52815] 6.41.606.297 I slot   operator(): id  0 | task 1010 | cached n_tokens = 20992, memory_seq_rm [20992, end)
[52815] 6.41.609.181 I sched_graph_realloc: graph reallocation: nodes = 5000 (changed = 0), splits = 180, n_same = 16, n_diff = 0
[52815] 6.42.062.532 I slot   operator(): id  0 | task 1010 | cached n_tokens = 21504, memory_seq_rm [21504, end)
[52815] 6.42.065.392 I sched_graph_realloc: graph reallocation: nodes = 5000 (changed = 0), splits = 180, n_same = 17, n_diff = 0
[52815] 6.42.518.379 I slot   operator(): id  0 | task 1010 | cached n_tokens = 22016, memory_seq_rm [22016, end)
[52815] 6.42.521.242 I sched_graph_realloc: graph reallocation: nodes = 5000 (changed = 0), splits = 180, n_same = 18, n_diff = 0
[52815] 6.42.966.742 I slot   operator(): id  0 | task 1010 | cached n_tokens = 22528, memory_seq_rm [22528, end)
[52815] 6.42.969.641 I sched_graph_realloc: graph reallocation: nodes = 5000 (changed = 0), splits = 180, n_same = 19, n_diff = 0
[52815] 6.43.431.723 I slot   operator(): id  0 | task 1010 | cached n_tokens = 23040, memory_seq_rm [23040, end)
[52815] 6.43.435.018 I sched_graph_realloc: graph reallocation: nodes = 5000 (changed = 0), splits = 180, n_same = 20, n_diff = 0
[52815] 6.43.919.056 I slot   operator(): id  0 | task 1010 | cached n_tokens = 23552, memory_seq_rm [23552, end)
[52815] 6.44.281.408 I slot print_timing: id  0 | task 1010 | prompt processing, n_tokens =   3328, progress = 0.97, t =   3.28 s / 1014.28 tokens per second
[52815] 6.44.281.412 I slot   operator(): id  0 | task 1010 | cached n_tokens = 23808, memory_seq_rm [23808, end)
[52815] 6.44.346.309 I slot create_check: id  0 | task 1010 | created context checkpoint 4 of 32 (pos_min = 23807, pos_max = 23807, n_tokens = 23808, size = 238.787 MiB)
[52815] 6.44.348.956 I sched_graph_realloc: graph reallocation: nodes = 5000 (changed = 0), splits = 180, n_same = 21, n_diff = 0
[52815] 6.44.836.873 I slot print_timing: id  0 | task 1010 | prompt processing, n_tokens =   3840, progress = 0.99, t =   3.83 s / 1002.87 tokens per second
[52815] 6.44.836.876 I slot   operator(): id  0 | task 1010 | cached n_tokens = 24320, memory_seq_rm [24320, end)
[52815] 6.44.839.602 I slot init_sampler: id  0 | task 1010 | init sampler, took 2.29 ms, tokens: text = 24446, total = 24446
[52815] 6.44.904.789 I slot create_check: id  0 | task 1010 | created context checkpoint 5 of 32 (pos_min = 24319, pos_max = 24319, n_tokens = 24320, size = 238.805 MiB)
[52815] 6.44.906.880 I sched_graph_realloc: graph reallocation: nodes = 5000 (changed = 0), splits = 180, n_same = 22, n_diff = 0
[52815] 6.47.616.124 I cmn  common_reaso: deactivated (natural end)
7.07.507.245 E srv    operator(): http client error: Failed to read connection
[52815] 6.49.541.952 I srv    operator(): DELETE /v1/stream conv_id=5f702c58-83b0-42d1-a7cc-1627fe41cd84::Qwen3.8-27B-TTurbo -> evict_and_cancel
[52815] 6.49.717.519 W srv          stop: cancel task, id_task = 1010
[52815] 6.50.054.001 I slot      release: id  0 | task 1010 | stop processing: n_tokens = 24492, truncated = 0
[52815] 6.50.054.015 I srv  update_slots: all slots are idle
Linguagem predominante
C++
Estrelas
1.2k
Forks
80
Merge médio
4d 20h
PRs com merge (30d)
2

Preparar o ambiente

Este projeto não oferece contêiner de desenvolvimento, Dockerfile nem guia de contribuição, então a configuração fica por sua conta: comece pelo README e veja nosso guia da primeira contribuição para os passos gerais.

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de Anbeeld/beellama.cpp

Todas as issues de Anbeeld/beellama.cpp

Issues semelhantes

Mais issues de C++

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.