Eval bug: Qwen2.5-vl在AMD GPU上做图像识别时崩溃（分辨率1242*881）

[cdwuchun]

Name and Version

llama-server --version
load_backend: loaded RPC backend from D:\AI\app\llama.cpp\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 6600M (AMD proprietary driver) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from D:\AI\app\llama.cpp\ggml-vulkan.dll
load_backend: loaded CPU backend from D:\AI\app\llama.cpp\ggml-cpu-haswell.dll
version: 5342 (0208355)
built with MSVC 19.43.34808.0 for x64

Operating systems

Windows

GGML backends

Vulkan

Hardware

AMD Ryzen 7 5800H/AMD Radeon RX 6600M

Models

https://www.modelscope.cn/models/bartowski/Qwen_Qwen2.5-VL-7B-Instruct-GGUF/Qwen_Qwen2.5-VL-7B-Instruct-Q5_K_S.gguf 和https://www.modelscope.cn/models/bartowski/Qwen_Qwen2.5-VL-7B-Instruct-GGUF/mmproj-Qwen_Qwen2.5-VL-7B-Instruct-bf16.gguf
也试过：https://www.modelscope.cn/models/ggml-org/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf 和 https://www.modelscope.cn/models/ggml-org/Qwen2.5-VL-7B-Instruct-GGUF/mmproj-Qwen2.5-VL-7B-Instruct-f16.gguf

Problem description & steps to reproduce

Qwen2.5-VL-7B-Instruct 在做图像识别时：分辨率为328409，没有问题；分辨率为1242881 时 模型会崩溃，但是如果在加载模型的参数中添加--no-mmproj-offload，可以正常工作（很慢）。Gemma-3-4B、12B都没有问题。

First Bad Commit

从llama-server可以支持多模态开始就存在该错误

Relevant log output

PS D:\AI> ./llama-swap -config config.yaml
llama-swap listening on :8080

  "20.Qwen2.5-VL-7B-Instruct":
    cmd: >
      llama-server
      --host 0.0.0.0
      --port ${PORT}
      --model models/Bartowski/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct-Q5_K_S.gguf
      --mmproj models/Bartowski/Qwen2.5-VL-7B-Instruct-GGUF/mmproj-Qwen2.5-VL-7B-Instruct-f16.gguf

      -ngl 99
      
    proxy: 
    # list of model name aliases this llama.cpp instance can serve
    aliases:
    - Qwen2.5-VL-7B-Instruct
    # `useModelName` overrides the model name in the request
    # and sends a specific name to the upstream server
    useModelName: "Qwen2.5-VL-7B-Instruct"
    # check this path for a HTTP 200 response for the server to be ready
    checkEndpoint: /health

load_backend: loaded RPC backend from D:\AI\app\llama.cpp\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 6600M (AMD proprietary driver) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from D:\AI\app\llama.cpp\ggml-vulkan.dll
load_backend: loaded CPU backend from D:\AI\app\llama.cpp\ggml-cpu-haswell.dll
build: 5342 (0208355f) with MSVC 19.43.34808.0 for x64
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 |

main: binding port with default address family
main: HTTP server is listening, hostname: 0.0.0.0, port: 8090, http threads: 15
main: loading model
srv    load_model: loading model 'models/Bartowski/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct-Q5_K_S.gguf'
llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon RX 6600M) - 8176 MiB free
llama_model_loader: loaded meta data with 31 key-value pairs and 339 tensors from models/Bartowski/Qwen2.5-VL-7B-Instruct-GGUF/Qwen2.5-VL-7B-Instruct-Q5_K_S.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen2vl
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 VL 7B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5-VL
llama_model_loader: - kv   5:                         general.size_label str              = 7B
llama_model_loader: - kv   6:                        qwen2vl.block_count u32              = 28
llama_model_loader: - kv   7:                     qwen2vl.context_length u32              = 128000
llama_model_loader: - kv   8:                   qwen2vl.embedding_length u32              = 3584
llama_model_loader: - kv   9:                qwen2vl.feed_forward_length u32              = 18944
llama_model_loader: - kv  10:               qwen2vl.attention.head_count u32              = 28
llama_model_loader: - kv  11:            qwen2vl.attention.head_count_kv u32              = 4
llama_model_loader: - kv  12:                     qwen2vl.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  13:   qwen2vl.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:            qwen2vl.rope.dimension_sections arr[i32,4]       = [16, 24, 24, 0]
llama_model_loader: - kv  15:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  16:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  17:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  18:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
srv  log_server_r: request: GET /health 127.0.0.1 503
[INFO] <20.Qwen2.5-VL-7B-Instruct> Health check error on http://localhost:8090/health, status code: 503
llama_model_loader: - kv  19:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  20:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  21:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  22:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  23:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  24:                    tokenizer.chat_template str              = {% set image_count = namespace(value=...
llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llama_model_loader: - kv  26:                          general.file_type u32              = 16
llama_model_loader: - kv  27:                      quantize.imatrix.file str              = /models_out/Qwen2.5-VL-7B-Instruct-GG...
llama_model_loader: - kv  28:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  29:             quantize.imatrix.entries_count i32              = 196
llama_model_loader: - kv  30:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  141 tensors
llama_model_loader: - type q5_K:  197 tensors
llama_model_loader: - type q6_K:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q5_K - Small
print_info: file size   = 4.94 GiB (5.58 BPW)
load: special tokens cache size = 22
load: token to piece cache size = 0.9310 MB
print_info: arch             = qwen2vl
print_info: vocab_only       = 0
print_info: n_ctx_train      = 128000
print_info: n_embd           = 3584
print_info: n_layer          = 28
print_info: n_head           = 28
print_info: n_head_kv        = 4
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 7
print_info: n_embd_k_gqa     = 512
print_info: n_embd_v_gqa     = 512
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-06
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 18944
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = -1
print_info: rope type        = 8
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 128000
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 7B
print_info: model params     = 7.62 B
print_info: general.name     = Qwen2.5 VL 7B Instruct
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 28 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 29/29 layers to GPU
load_tensors:      Vulkan0 model buffer size =  4705.94 MiB
load_tensors:   CPU_Mapped model buffer size =   357.33 MiB
......................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (128000) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     0.58 MiB
llama_kv_cache_unified: kv_size = 4096, type_k = 'f16', type_v = 'f16', n_layer = 28, can_shift = 1, padding = 32
llama_kv_cache_unified:    Vulkan0 KV buffer size =   224.00 MiB
llama_kv_cache_unified: KV self size  =  224.00 MiB, K (f16):  112.00 MiB, V (f16):  112.00 MiB
llama_context:    Vulkan0 compute buffer size =   304.00 MiB
llama_context: Vulkan_Host compute buffer size =    15.01 MiB
llama_context: graph nodes  = 1042
llama_context: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
clip_ctx: CLIP using Vulkan0 backend
clip_model_loader: model name:   Qwen2.5 VL 7B Instruct
clip_model_loader: description:
clip_model_loader: GGUF version: 3
clip_model_loader: alignment:    32
clip_model_loader: n_tensors:    519
clip_model_loader: n_kv:         22

load_hparams: projector:          qwen2.5vl_merger
load_hparams: n_embd:             1280
load_hparams: n_head:             16
load_hparams: n_ff:               3420
load_hparams: n_layer:            32
load_hparams: projection_dim:     3584
load_hparams: image_size:         3584
load_hparams: patch_size:         14

load_hparams: has_llava_proj:     0
load_hparams: minicpmv_version:   0
load_hparams: proj_scale_factor:  0
load_hparams: n_wa_pattern:       8
load_hparams: ffn_op:             silu
load_hparams: model size:         1291.40 MiB
load_hparams: metadata size:      0.18 MiB
srv  log_server_r: request: GET /health 127.0.0.1 503
[INFO] <20.Qwen2.5-VL-7B-Instruct> Health check error on http://localhost:8090/health, status code: 503
alloc_compute_meta:    Vulkan0 compute buffer size =     2.77 MiB
alloc_compute_meta:        CPU compute buffer size =     0.16 MiB
srv    load_model: loaded multimodal model, 'models/Bartowski/Qwen2.5-VL-7B-Instruct-GGUF/mmproj-Qwen2.5-VL-7B-Instruct-f16.gguf'
srv    load_model: ctx_shift is not supported by multimodal, it will be disabled
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 4096
main: model loaded
main: chat template, chat_template: {% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system
You are a helpful assistant.<|im_end|>
{% endif %}<|im_start|>{{ message['role'] }}
{% if message['content'] is string %}{{ message['content'] }}<|im_end|>
{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>
{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant
{% endif %}, example_format: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
'
main: server is listening on http://0.0.0.0:8090 - starting the main loop
srv  update_slots: all slots are idle
srv  log_server_r: request: GET /health 127.0.0.1 200
[INFO] <20.Qwen2.5-VL-7B-Instruct> Health check passed on http://localhost:8090/health
srv  log_server_r: request: GET / 127.0.0.1 200
[INFO] Request ::1 "GET /upstream/20.Qwen2.5-VL-7B-Instruct/ HTTP/1.1" 200 1290053 "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/136.0.0.0 Safari/537.36 Edg/136.0.0.0" 10.3580702s
srv  log_server_r: request: GET /props 127.0.0.1 200
[INFO] Request ::1 "GET /upstream/20.Qwen2.5-VL-7B-Instruct/props HTTP/1.1" 200 2535 "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/136.0.0.0 Safari/537.36 Edg/136.0.0.0" 2.0878ms
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, n_prompt_tokens = 24
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 14, n_tokens = 14, progress = 0.583333
slot update_slots: id  0 | task 0 | kv cache rm [14, end)
srv  process_chun: processing image...
ggml_vulkan: Device memory allocation of size 2417184000 failed.
ggml_vulkan: Requested buffer size exceeds device memory allocation limit: ErrorOutOfDeviceMemory
ggml_gallocr_reserve_n: failed to allocate Vulkan0 buffer of size 2417184000
D:\a\llama.cpp\llama.cpp\ggml\src\ggml-backend.cpp:1663: GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <= (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer)) failed
[INFO] Request ::1 "POST /upstream/20.Qwen2.5-VL-7B-Instruct/v1/chat/completions HTTP/1.1" 502 114 "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/136.0.0.0 Safari/537.36 Edg/136.0.0.0" 757.1853ms