Eval bug: Major performance drop since b7406

[Kaspur2012]

Name and Version

ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 2 CUDA devices:
Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
Device 1: NVIDIA GeForce RTX 2070 SUPER, compute capability 7.5, VMM: yes
load_backend: loaded CUDA backend from D:\llamacpp\llama-b7495-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from D:\llamacpp\llama-b7495-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from D:\llamacpp\llama-b7495-bin-win-cuda-12.4-x64\ggml-cpu-haswell.dll
version: 7495 (4117ae5)
built with Clang 19.1.5 for Windows x86_64

Operating systems

Windows

GGML backends

CUDA

Hardware

Ryzen 7 3700x
RTX 3090

Models

fully offload to a 3090

Qwen3-VL-32B-Thinking.Q4_K_M_TEXT_DRAFT:

• llama-server.exe -m D:/lm_studio/mradermacher/Qwen3-VL-32B-Thinking-GGUF/Qwen3-VL-32B-Thinking.Q4_K_M.gguf --jinja -c 20000 -ngl 999 -fa on --temp 1.0 --top-k 40 --top-p 1.0 --min-p 0.05 --repeat-penalty 1.0 --no-mmap --cache-type-k q8_0 --cache-type-v q8_0 --no-warmup --main-gpu 0 -md D:/lm_studio/lmstudio-community/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf --cache-type-k-draft q8_0 --cache_type-v-draft q8_0 --device-draft CUDA0 --split-mode none

Magistral-Small-2509-UD-Q5_K_XL:

• llama-server.exe -m D:\lm_studio\unsloth\Magistral-Small-2509-GGUF\Magistral-Small-2509-UD-Q5_K_XL.gguf --jinja -c 85000 -ngl 999 -fa on --temp 0.7 --top-k 40 --top-p 0.95 --min-p 0.01 --repeat-penalty 1.1 --no-mmap --split-mode none --main-gpu 0 --cache-type-k q8_0 --cache-type-v q8_0 --no-warmup

Found similar result for Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL but I do not recorded t/s for it.

Problem description & steps to reproduce

Lately I have major performance reduction on these 2 models: Qwen3-VL-32B-Thinking.Q4_K_M_TEXT_DRAFT and Magistral-Small-2509-UD-Q5_K_XL.

I have downloaded numerous binary build and located where the reduction in performance started.

Below are what I found:

b7406 :

• Qwen3-VL-32B-Thinking.Q4_K_M_TEXT_DRAFT - 45 t/s
• Magistral-Small-2509-UD-Q5_K_XL - 45 t/s

b7410 :

• Qwen3-VL-32B-Thinking.Q4_K_M_TEXT_DRAFT - 10 t/s
• Magistral-Small-2509-UD-Q5_K_XL - 20 t/s

b7495:

• Qwen3-VL-32B-Thinking.Q4_K_M_TEXT_DRAFT - failed to load, log below
• Magistral-Small-2509-UD-Q5_K_XL - 20 t/s

First Bad Commit

b7410 is when the performance started.

I am not sure when the loading problem started for Qwen3-VL-32B-Thinking.Q4_K_M.

Relevant log output

log for Qwen3-VL-32B-Thinking.Q4_K_M with draft that fail to load on b7495:

Working Dir: D:/llamacpp/llama-b7495-bin-win-cuda-12.4-x64
Executing Command: llama-server.exe -m D:/lm_studio/mradermacher/Qwen3-VL-32B-Thinking-GGUF/Qwen3-VL-32B-Thinking.Q4_K_M.gguf --jinja -c 20000 -ngl 999 -fa on --temp 1.0 --top-k 40 --top-p 1.0 --min-p 0.05 --repeat-penalty 1.0 --no-mmap --cache-type-k q8_0 --cache-type-v q8_0 --no-warmup --main-gpu 0 -md D:/lm_studio/lmstudio-community/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf --cache-type-k-draft q8_0 --cache_type-v-draft q8_0 --device-draft CUDA0 --split-mode none

================================================================================

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
  Device 1: NVIDIA GeForce RTX 2070 SUPER, compute capability 7.5, VMM: yes
load_backend: loaded CUDA backend from D:\llamacpp\llama-b7495-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from D:\llamacpp\llama-b7495-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from D:\llamacpp\llama-b7495-bin-win-cuda-12.4-x64\ggml-cpu-haswell.dll
main: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true
build: 7495 (4117ae555) with Clang 19.1.5 for Windows x86_64
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 500,610,700,750,800,860,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model 'D:/lm_studio/mradermacher/Qwen3-VL-32B-Thinking-GGUF/Qwen3-VL-32B-Thinking.Q4_K_M.gguf'
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on

llama_params_fit_impl: projected to use 21416 MiB of device memory vs. 24575 MiB of free device memory
llama_params_fit_impl: will leave 1887 >= 1024 MiB of free device memory, no changes needed
llama_params_fit: successfully fit params to free device memory
llama_params_fit: fitting params to free memory took 0.46 seconds
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) (0000:0b:00.0) - 23304 MiB free
llama_model_loader: loaded meta data with 39 key-value pairs and 707 tensors from D:/lm_studio/mradermacher/Qwen3-VL-32B-Thinking-GGUF/Qwen3-VL-32B-Thinking.Q4_K_M.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              = qwen3vl
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3 VL 32B Thinking
llama_model_loader: - kv   3:                           general.finetune str              = Thinking
llama_model_loader: - kv   4:                           general.basename str              = Qwen3-VL
llama_model_loader: - kv   5:                         general.size_label str              = 32B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                               general.tags arr[str,1]       = ["image-text-to-text"]
llama_model_loader: - kv   8:                        qwen3vl.block_count u32              = 64
llama_model_loader: - kv   9:                     qwen3vl.context_length u32              = 262144
llama_model_loader: - kv  10:                   qwen3vl.embedding_length u32              = 5120
llama_model_loader: - kv  11:                qwen3vl.feed_forward_length u32              = 25600
llama_model_loader: - kv  12:               qwen3vl.attention.head_count u32              = 64
llama_model_loader: - kv  13:            qwen3vl.attention.head_count_kv u32              = 8
llama_model_loader: - kv  14:                     qwen3vl.rope.freq_base f32              = 5000000.000000
llama_model_loader: - kv  15:   qwen3vl.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  16:               qwen3vl.attention.key_length u32              = 128
llama_model_loader: - kv  17:             qwen3vl.attention.value_length u32              = 128
llama_model_loader: - kv  18:            qwen3vl.rope.dimension_sections arr[i32,4]       = [24, 20, 20, 0]
llama_model_loader: - kv  19:                 qwen3vl.n_deepstack_layers u32              = 3
llama_model_loader: - kv  20:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  21:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  22:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  23:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...

llama_model_loader: - kv  24:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  25:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  26:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  27:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  28:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  29:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
llama_model_loader: - kv  30:               general.quantization_version u32              = 2
llama_model_loader: - kv  31:                          general.file_type u32              = 15
llama_model_loader: - kv  32:                                general.url str              = https://huggingface.co/mradermacher/Q...
llama_model_loader: - kv  33:              mradermacher.quantize_version str              = 2
llama_model_loader: - kv  34:                  mradermacher.quantized_by str              = mradermacher
llama_model_loader: - kv  35:                  mradermacher.quantized_at str              = 2025-10-30T18:34:42+01:00
llama_model_loader: - kv  36:                  mradermacher.quantized_on str              = nico1
llama_model_loader: - kv  37:                         general.source.url str              = https://huggingface.co/Qwen/Qwen3-VL-...
llama_model_loader: - kv  38:                  mradermacher.convert_type str              = hf
llama_model_loader: - type  f32:  257 tensors
llama_model_loader: - type q4_K:  385 tensors
llama_model_loader: - type q6_K:   65 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 18.40 GiB (4.82 BPW) 

load: printing all EOG tokens:
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3vl
print_info: vocab_only       = 0
print_info: no_alloc         = 0
print_info: n_ctx_train      = 262144
print_info: n_embd           = 5120
print_info: n_embd_inp       = 20480
print_info: n_layer          = 64
print_info: n_head           = 64
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
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             = 25600
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 40
print_info: rope scaling     = linear
print_info: freq_base_train  = 5000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 262144
print_info: rope_yarn_log_mul= 0.0000
print_info: rope_finetuned   = unknown
print_info: mrope sections   = [24, 20, 20, 0]
print_info: model type       = 32B
print_info: model params     = 32.76 B
print_info: general.name     = Qwen3 VL 32B Thinking
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
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 = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 63 repeating layers to GPU
load_tensors: offloaded 65/65 layers to GPU
load_tensors:          CPU model buffer size =   417.30 MiB
load_tensors:        CUDA0 model buffer size = 18423.65 MiB

.
....
...
....
...
...
....
....
...
...
....
...
...
....
...
...
....
...
....
...
....
...
...
...
....
...
...
...
....
.
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 4
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 5000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     2.32 MiB
llama_kv_cache:      CUDA0 KV buffer size =  2686.00 MiB
llama_kv_cache: size = 2686.00 MiB ( 20224 cells,  64 layers,  4/1 seqs), K (q8_0): 1343.00 MiB, V (q8_0): 1343.00 MiB

llama_context:      CUDA0 compute buffer size =   306.75 MiB
llama_context:  CUDA_Host compute buffer size =    49.52 MiB
llama_context: graph nodes  = 2247
llama_context: graph splits = 2
srv    load_model: loading draft model 'D:/lm_studio/lmstudio-community/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf'
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on

llama_params_fit_impl: projected to use 2078 MiB of device memory vs. 24575 MiB of free device memory
llama_params_fit_impl: cannot fulfill margin of 1024 MiB, need to reduce device memory by 1216 MiB
llama_params_fit_impl: context size set by user to 20224 -> no change

llama_params_fit_impl: filling dense layers back-to-front:

llama_params_fit_impl:   - CUDA0 (NVIDIA GeForce RTX 3090):  8 layers,    861 MiB used,   1024 MiB free
llama_params_fit: successfully fit params to free device memory
llama_params_fit: fitting params to free memory took -3.75 seconds
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) (0000:0b:00.0) - 1886 MiB free
llama_model_loader: loaded meta data with 27 key-value pairs and 311 tensors from D:/lm_studio/lmstudio-community/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.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              = qwen3
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3 0.6B
llama_model_loader: - kv   3:                           general.basename str              = Qwen3
llama_model_loader: - kv   4:                         general.size_label str              = 0.6B
llama_model_loader: - kv   5:                          qwen3.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen3.context_length u32              = 32768
llama_model_loader: - kv   7:                     qwen3.embedding_length u32              = 1024
llama_model_loader: - kv   8:                  qwen3.feed_forward_length u32              = 3072
llama_model_loader: - kv   9:                 qwen3.attention.head_count u32              = 16
llama_model_loader: - kv  10:              qwen3.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen3.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen3.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                 qwen3.attention.key_length u32              = 128
llama_model_loader: - kv  14:               qwen3.attention.value_length u32              = 128
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,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  18:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...

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              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llama_model_loader: - kv  26:                          general.file_type u32              = 7
llama_model_loader: - type  f32:  113 tensors
llama_model_loader: - type q8_0:  198 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q8_0
print_info: file size   = 761.80 MiB (8.50 BPW) 
load: printing all EOG tokens:
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')
load: special tokens cache size = 26

load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3
print_info: vocab_only       = 0
print_info: no_alloc         = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 1024
print_info: n_embd_inp       = 1024
print_info: n_layer          = 28
print_info: n_head           = 16
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 2
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
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             = 3072
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = -1
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
print_info: rope_yarn_log_mul= 0.0000
print_info: rope_finetuned   = unknown
print_info: model type       = 0.6B
print_info: model params     = 751.63 M
print_info: general.name     = Qwen3 0.6B
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
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 = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 7 repeating layers to GPU
load_tensors: offloaded 8/29 layers to GPU
load_tensors:        CUDA0 model buffer size =   269.28 MiB
load_tensors:    CUDA_Host model buffer size =   492.52 MiB

................................................
.............
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:        CPU KV buffer size =   881.34 MiB

llama_kv_cache:      CUDA0 KV buffer size =   293.78 MiB
llama_kv_cache: size = 1175.12 MiB ( 20224 cells,  28 layers,  1/1 seqs), K (q8_0):  587.56 MiB, V (q8_0):  587.56 MiB

llama_context:      CUDA0 compute buffer size =   298.75 MiB
llama_context:  CUDA_Host compute buffer size =    43.51 MiB
llama_context: graph nodes  = 987
llama_context: graph splits = 275 (with bs=512), 2 (with bs=1)
srv    load_model: initializing slots, n_slots = 4
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 20224
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:        CPU KV buffer size =   881.34 MiB

llama_kv_cache:      CUDA0 KV buffer size =   293.78 MiB

llama_kv_cache: size = 1175.12 MiB ( 20224 cells,  28 layers,  1/1 seqs), K (q8_0):  587.56 MiB, V (q8_0):  587.56 MiB
llama_context:      CUDA0 compute buffer size =   298.75 MiB
llama_context:  CUDA_Host compute buffer size =    43.51 MiB
llama_context: graph nodes  = 987
llama_context: graph splits = 275 (with bs=512), 2 (with bs=1)
slot   load_model: id  0 | task -1 | new slot, n_ctx = 20224
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 20224
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:        CPU KV buffer size =   881.34 MiB

llama_kv_cache:      CUDA0 KV buffer size =   293.78 MiB
llama_kv_cache: size = 1175.12 MiB ( 20224 cells,  28 layers,  1/1 seqs), K (q8_0):  587.56 MiB, V (q8_0):  587.56 MiB

llama_context:      CUDA0 compute buffer size =   298.75 MiB
llama_context:  CUDA_Host compute buffer size =    43.51 MiB
llama_context: graph nodes  = 987
llama_context: graph splits = 275 (with bs=512), 2 (with bs=1)
slot   load_model: id  1 | task -1 | new slot, n_ctx = 20224
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 20224
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:        CPU KV buffer size =   881.34 MiB

llama_kv_cache:      CUDA0 KV buffer size =   293.78 MiB
llama_kv_cache: size = 1175.12 MiB ( 20224 cells,  28 layers,  1/1 seqs), K (q8_0):  587.56 MiB, V (q8_0):  587.56 MiB
llama_context:      CUDA0 compute buffer size =   298.75 MiB
llama_context:  CUDA_Host compute buffer size =    43.51 MiB
llama_context: graph nodes  = 987
llama_context: graph splits = 275 (with bs=512), 2 (with bs=1)
slot   load_model: id  2 | task -1 | new slot, n_ctx = 20224
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 20224
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:        CPU KV buffer size =   881.34 MiB

llama_kv_cache:      CUDA0 KV buffer size =   293.78 MiB
llama_kv_cache: size = 1175.12 MiB ( 20224 cells,  28 layers,  1/1 seqs), K (q8_0):  587.56 MiB, V (q8_0):  587.56 MiB
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 298.75 MiB on device 0: cudaMalloc failed: out of memory
ggml_gallocr_reserve_n_impl: failed to allocate CUDA0 buffer of size 313262080
graph_reserve: failed to allocate compute buffers
llama_init_from_model: failed to initialize the context: failed to allocate compute pp buffers
srv    load_model: failed to create draft context
srv    operator(): operator(): cleaning up before exit...
main: exiting due to model loading error

log for Qwen3-VL-32B-Thinking.Q4_K_M with draft that able to load on b7406:

Working Dir: D:/llamacpp/llama-b7406-bin-win-cuda-12.4-x64
Executing Command: llama-server.exe -m D:/lm_studio/mradermacher/Qwen3-VL-32B-Thinking-GGUF/Qwen3-VL-32B-Thinking.Q4_K_M.gguf --jinja -c 20000 -ngl 999 -fa on --temp 1.0 --top-k 40 --top-p 1.0 --min-p 0.05 --repeat-penalty 1.0 --no-mmap --cache-type-k q8_0 --cache-type-v q8_0 --no-warmup --main-gpu 0 -md D:/lm_studio/lmstudio-community/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf --cache-type-k-draft q8_0 --cache_type-v-draft q8_0 --device-draft CUDA0 --split-mode none

================================================================================

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
  Device 1: NVIDIA GeForce RTX 2070 SUPER, compute capability 7.5, VMM: yes
load_backend: loaded CUDA backend from D:\llamacpp\llama-b7406-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from D:\llamacpp\llama-b7406-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from D:\llamacpp\llama-b7406-bin-win-cuda-12.4-x64\ggml-cpu-haswell.dll
build: 7406 (4aced7a63) with Clang 19.1.5 for Windows x86_64
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 500,610,700,750,800,860,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model 'D:/lm_studio/mradermacher/Qwen3-VL-32B-Thinking-GGUF/Qwen3-VL-32B-Thinking.Q4_K_M.gguf'

llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) (0000:0b:00.0) - 23304 MiB free
llama_model_loader: loaded meta data with 39 key-value pairs and 707 tensors from D:/lm_studio/mradermacher/Qwen3-VL-32B-Thinking-GGUF/Qwen3-VL-32B-Thinking.Q4_K_M.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              = qwen3vl
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3 VL 32B Thinking
llama_model_loader: - kv   3:                           general.finetune str              = Thinking
llama_model_loader: - kv   4:                           general.basename str              = Qwen3-VL
llama_model_loader: - kv   5:                         general.size_label str              = 32B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                               general.tags arr[str,1]       = ["image-text-to-text"]
llama_model_loader: - kv   8:                        qwen3vl.block_count u32              = 64
llama_model_loader: - kv   9:                     qwen3vl.context_length u32              = 262144
llama_model_loader: - kv  10:                   qwen3vl.embedding_length u32              = 5120
llama_model_loader: - kv  11:                qwen3vl.feed_forward_length u32              = 25600
llama_model_loader: - kv  12:               qwen3vl.attention.head_count u32              = 64
llama_model_loader: - kv  13:            qwen3vl.attention.head_count_kv u32              = 8
llama_model_loader: - kv  14:                     qwen3vl.rope.freq_base f32              = 5000000.000000
llama_model_loader: - kv  15:   qwen3vl.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  16:               qwen3vl.attention.key_length u32              = 128
llama_model_loader: - kv  17:             qwen3vl.attention.value_length u32              = 128
llama_model_loader: - kv  18:            qwen3vl.rope.dimension_sections arr[i32,4]       = [24, 20, 20, 0]
llama_model_loader: - kv  19:                 qwen3vl.n_deepstack_layers u32              = 3
llama_model_loader: - kv  20:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  21:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  22:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  23:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...

llama_model_loader: - kv  24:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  25:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  26:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  27:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  28:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  29:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
llama_model_loader: - kv  30:               general.quantization_version u32              = 2
llama_model_loader: - kv  31:                          general.file_type u32              = 15
llama_model_loader: - kv  32:                                general.url str              = https://huggingface.co/mradermacher/Q...
llama_model_loader: - kv  33:              mradermacher.quantize_version str              = 2
llama_model_loader: - kv  34:                  mradermacher.quantized_by str              = mradermacher
llama_model_loader: - kv  35:                  mradermacher.quantized_at str              = 2025-10-30T18:34:42+01:00
llama_model_loader: - kv  36:                  mradermacher.quantized_on str              = nico1
llama_model_loader: - kv  37:                         general.source.url str              = https://huggingface.co/Qwen/Qwen3-VL-...
llama_model_loader: - kv  38:                  mradermacher.convert_type str              = hf
llama_model_loader: - type  f32:  257 tensors
llama_model_loader: - type q4_K:  385 tensors
llama_model_loader: - type q6_K:   65 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 18.40 GiB (4.82 BPW) 
load: printing all EOG tokens:
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')

load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3vl
print_info: vocab_only       = 0
print_info: n_ctx_train      = 262144
print_info: n_embd           = 5120
print_info: n_embd_inp       = 20480
print_info: n_layer          = 64
print_info: n_head           = 64
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
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             = 25600
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 40
print_info: rope scaling     = linear
print_info: freq_base_train  = 5000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 262144
print_info: rope_yarn_log_mul= 0.0000
print_info: rope_finetuned   = unknown
print_info: mrope sections   = [24, 20, 20, 0]
print_info: model type       = 32B
print_info: model params     = 32.76 B
print_info: general.name     = Qwen3 VL 32B Thinking
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
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 = false)

load_tensors: offloading 64 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 65/65 layers to GPU
load_tensors:          CPU model buffer size =   417.30 MiB
load_tensors:        CUDA0 model buffer size = 18423.65 MiB

.
...
.
..
.
..
..
.
..
..
.
..
..
..
..
.
..
..
.
..
..
..
..
.
..
..
.
..
.
..
..
..
.
..
.
..
..
.
..
..
..
..
.
..
..
..
.
..
..
..
.
..
..
.
..
..
.
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = false
llama_context: freq_base     = 5000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:      CUDA0 KV buffer size =  2686.00 MiB

llama_kv_cache: size = 2686.00 MiB ( 20224 cells,  64 layers,  1/1 seqs), K (q8_0): 1343.00 MiB, V (q8_0): 1343.00 MiB
llama_context:      CUDA0 compute buffer size =   306.75 MiB
llama_context:  CUDA_Host compute buffer size =    49.52 MiB
llama_context: graph nodes  = 2247
llama_context: graph splits = 2
srv    load_model: loading draft model 'D:/lm_studio/lmstudio-community/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3090) (0000:0b:00.0) - 1886 MiB free
llama_model_loader: loaded meta data with 27 key-value pairs and 311 tensors from D:/lm_studio/lmstudio-community/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.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              = qwen3
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3 0.6B
llama_model_loader: - kv   3:                           general.basename str              = Qwen3
llama_model_loader: - kv   4:                         general.size_label str              = 0.6B
llama_model_loader: - kv   5:                          qwen3.block_count u32              = 28
llama_model_loader: - kv   6:                       qwen3.context_length u32              = 32768
llama_model_loader: - kv   7:                     qwen3.embedding_length u32              = 1024
llama_model_loader: - kv   8:                  qwen3.feed_forward_length u32              = 3072
llama_model_loader: - kv   9:                 qwen3.attention.head_count u32              = 16
llama_model_loader: - kv  10:              qwen3.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       qwen3.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  12:     qwen3.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  13:                 qwen3.attention.key_length u32              = 128
llama_model_loader: - kv  14:               qwen3.attention.value_length u32              = 128
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,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  18:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
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              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llama_model_loader: - kv  26:                          general.file_type u32              = 7
llama_model_loader: - type  f32:  113 tensors
llama_model_loader: - type q8_0:  198 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q8_0
print_info: file size   = 761.80 MiB (8.50 BPW) 

load: printing all EOG tokens:
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 1024
print_info: n_embd_inp       = 1024
print_info: n_layer          = 28
print_info: n_head           = 16
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 2
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
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             = 3072
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = -1
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
print_info: rope_yarn_log_mul= 0.0000
print_info: rope_finetuned   = unknown
print_info: model type       = 0.6B
print_info: model params     = 751.63 M
print_info: general.name     = Qwen3 0.6B
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
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 = false)

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:        CUDA0 model buffer size =   604.15 MiB
load_tensors:    CUDA_Host model buffer size =   157.65 MiB
.............
............................
...................
.
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = false
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:      CUDA0 KV buffer size =  1175.12 MiB
llama_kv_cache: size = 1175.12 MiB ( 20224 cells,  28 layers,  1/1 seqs), K (q8_0):  587.56 MiB, V (q8_0):  587.56 MiB
llama_context:      CUDA0 compute buffer size =   298.75 MiB
llama_context:  CUDA_Host compute buffer size =    41.51 MiB
llama_context: graph nodes  = 987
llama_context: graph splits = 2

srv          init: initializing slots, n_slots = 1
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20224
llama_context: n_ctx_seq     = 20224
llama_context: n_batch       = 20224
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = false
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (20224) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     0.58 MiB
llama_kv_cache:      CUDA0 KV buffer size =  1175.12 MiB
llama_kv_cache: size = 1175.12 MiB ( 20224 cells,  28 layers,  1/1 seqs), K (q8_0):  587.56 MiB, V (q8_0):  587.56 MiB
llama_context:      CUDA0 compute buffer size =   298.75 MiB
llama_context:  CUDA_Host compute buffer size =    41.51 MiB
llama_context: graph nodes  = 987
llama_context: graph splits = 2
slot         init: id  0 | task -1 | new slot, n_ctx = 20224
srv          init: prompt cache is enabled, size limit: 8192 MiB
srv          init: use `--cache-ram 0` to disable the prompt cache
srv          init: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: thinking = 1
init: chat template, chat_template: {%- set image_count = namespace(value=0) %}
{%- set video_count = namespace(value=0) %}
{%- macro render_content(content, do_vision_count) %}
    {%- if content is string %}
        {{- content }}
    {%- else %}
        {%- for item in content %}
            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
                {%- if do_vision_count %}
                    {%- set image_count.value = image_count.value + 1 %}
                {%- endif %}
                {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
                <|vision_start|><|image_pad|><|vision_end|>
            {%- elif 'video' in item or item.type == 'video' %}
                {%- if do_vision_count %}
                    {%- set video_count.value = video_count.value + 1 %}
                {%- endif %}
                {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
                <|vision_start|><|video_pad|><|vision_end|>
            {%- elif 'text' in item %}
                {{- item.text }}
            {%- endif %}
        {%- endfor %}
    {%- endif %}
{%- endmacro %}
{%- if tools %}
    {{- '<|im_start|>system\n' }}
    {%- if messages[0].role == 'system' %}
        {{- render_content(messages[0].content, false) + '\n\n' }}
    {%- endif %}
    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
    {%- for tool in tools %}
        {{- "\n" }}
        {{- tool | tojson }}
    {%- endfor %}
    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
{%- else %}
    {%- if messages[0].role == 'system' %}
        {{- '<|im_start|>system\n' + render_content(messages[0].content, false) + '<|im_end|>\n' }}
    {%- endif %}
{%- endif %}
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
{%- for message in messages[::-1] %}
    {%- set index = (messages|length - 1) - loop.index0 %}
    {%- if ns.multi_step_tool and message.role == "user" %}
        {%- set content = render_content(message.content, false) %}
        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
            {%- set ns.multi_step_tool = false %}
            {%- set ns.last_query_index = index %}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- for message in messages %}
    {%- set content = render_content(message.content, True) %}
    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
    {%- elif message.role == "assistant" %}
        {%- set reasoning_content = '' %}
        {%- if message.reasoning_content is string %}
            {%- set reasoning_content = message.reasoning_content %}
        {%- else %}
            {%- if '</think>' in content %}
                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
            {%- endif %}
        {%- endif %}
        {%- if loop.index0 > ns.last_query_index %}
            {%- if loop.last or (not loop.last and reasoning_content) %}
                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
            {%- else %}
                {{- '<|im_start|>' + message.role + '\n' + content }}
            {%- endif %}
        {%- else %}
            {{- '<|im_start|>' + message.role + '\n' + content }}
        {%- endif %}
        {%- if message.tool_calls %}
            {%- for tool_call in message.tool_calls %}
                {%- if (loop.first and content) or (not loop.first) %}
                    {{- '\n' }}
                {%- endif %}
                {%- if tool_call.function %}
                    {%- set tool_call = tool_call.function %}
                {%- endif %}
                {{- '<tool_call>\n{"name": "' }}
                {{- tool_call.name }}
                {{- '", "arguments": ' }}
                {%- if tool_call.arguments is string %}
                    {{- tool_call.arguments }}
                {%- else %}
                    {{- tool_call.arguments | tojson }}
                {%- endif %}
                {{- '}\n</tool_call>' }}
            {%- endfor %}
        {%- endif %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "tool" %}
        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
            {{- '<|im_start|>user' }}
        {%- endif %}
        {{- '\n<tool_response>\n' }}
        {{- content }}
        {{- '\n</tool_response>' }}
        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
            {{- '<|im_end|>\n' }}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {{- '<|im_start|>assistant\n<think>\n' }}
{%- 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
<think>
'
main: model loaded
main: server is listening on http://127.0.0.1:8080
main: starting the main loop...
srv  update_slots: all slots are idle


[INFO] Model is fully loaded.
[DIAGNOSTICS] Detected 'model loaded' string.
srv  params_from_: Chat format: Hermes 2 Pro
slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
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 
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 20224, n_keep = 0, task.n_tokens = 3692
slot update_slots: id  0 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 2048, batch.n_tokens = 2048, progress = 0.554713

slot update_slots: id  0 | task 0 | n_tokens = 2048, memory_seq_rm [2048, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 3692, batch.n_tokens = 1644, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_tokens = 3692, batch.n_tokens = 1644

slot print_timing: id  0 | task 0 | 
prompt eval time =    3342.05 ms /  3692 tokens (    0.91 ms per token,  1104.71 tokens per second)
       eval time =   91573.07 ms /  4522 tokens (   20.25 ms per token,    49.38 tokens per second)
      total time =   94915.12 ms /  8214 tokens
draft acceptance rate = 0.71409 ( 3574 accepted /  5005 generated)
slot      release: id  0 | task 0 | stop processing: n_tokens = 8213, truncated = 0
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
'''

[grapevine-AI]

Yes, me too.
In my environment, Gemma3 27B and Qwen3-30B-A3B with b7410 become slower than b7406.
But gpt-oss-20b keeps performance.
I don't seem that All model drop speed.

env: Windows, llama.cpp CUDA13.1 binary, RTX3090
Gemma3 27B token-generation 40t/s -> 20t/s
Qwen3-30B-A3B token-generation 150t/s -> 65t/s

[Kaspur2012]

yead only some models seem to be affected.
I just tested:

• Qwen3-VL-30B-A3B-Instruct-UD-Q4_K_XL since I deleted the old non VL version. the t/s remain somewhat the same for me on the VL version around 130 t/s

• Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft also reduced in t/s - from (b7406) 50 t/s -> (b7495) 16 t/s

• gemma-3-27b-it-qat-UD-Q4_K_XL with draft: 65 t/k (b7406) -> (b7495) failed to load, similar error to qwen 30b vl above, OOM. I removed the draft model and it worked going around 39 t/s (b7495)

• went back and test Qwen3-VL-32B-Thinking.Q4_K_M with no draft model on b7495. loaded fine - around 30 t/s. So on b7406 with draft, works fine. b7495 with draft, failed to load, oom error, without draft loaded fine. all model loaded with same command beside stated.

[pwilkin]

@Kaspur2012 Can you please retry the tests with -fit off and see if that fixes the performance regression?

This is most likely tied to #16653

[Kaspur2012]

I have tried -fit off, here are the result for b7495:

• Qwen3-VL-32B-Thinking.Q4_K_M - with draft model failed to load due to OOM. Without draft loaded just fine with fit on or off

• Magistral-Small-2509-UD-Q5_K_XL - 45 t/s, so -fit off seem to fix the issue

• Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft result in OOM. Without draft loaded just fine with fit on or off

So seem like -fit was causing the issue with the performance reduction on Magistral-Small-2509-UD-Q5_K_XL. I am assuming -fit default is on. The out of memory on b7495 on models with draft model I am not sure what is causing it.

Magistral-Small-2509-UD-Q5_K_XL with fit on - 20 t/s
Magistral-Small-2509-UD-Q5_K_XL with fit off - 45 t/s
Magistral-Small-2509-UD-Q5_K_XL with b7406 - 45 t/s

Qwen3-VL-32B-Thinking.Q4_K_M with draft model on b7406 - 45 t/s
Qwen3-VL-32B-Thinking.Q4_K_M with draft model on b7495 with fit on and off - OOM

Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft on b7495 - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft on b7495 with -fit off - OOM

Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft on b7406 - 50 t/s

[Kaspur2012]

so I was curious on when it started to give OOM error on Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL when I have draft model and here are the results:

llama-server.exe -m D:\lm_studio\unsloth\Magistral-Small-2509-GGUF\Magistral-Small-2509-UD-Q5_K_XL.gguf --jinja -c 85000 -ngl 999 -fa on --temp 0.7 --top-k 40 --top-p 0.95 --min-p 0.01 --repeat-penalty 1.1 --no-mmap --split-mode none --main-gpu 0 --cache-type-k q8_0 --cache-type-v q8_0 --no-warmup

b7406:
fit not avai
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft - 50 t/s

b7410:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - 50 t/s

b7418:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - 50 t/s

b7426:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - 50 t/s

b7429:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - 50 t/s

b7432:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - 50 t/s

b7433:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - OOM

b7434:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - OOM

b7438:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - OOM

b7495:
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit on - 15 t/s
Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off - OOM

As you can see starting with b7433 is when I starting to get OOM error when loading Mistral-Small-3.2-24B-Instruct-2506-UD-Q5_K_XL with draft with -fit off.

I went back and re-test Qwen3-VL-32B-Thinking.Q4_K_M based on my findings, results:

b7432:
Qwen3-VL-32B-Thinking.Q4_K_M with draft with -fit on - 10 t/s
Qwen3-VL-32B-Thinking.Q4_K_M with draft with -fit off - 45 t/s

b7433:
Qwen3-VL-32B-Thinking.Q4_K_M with draft with -fit on - OOM
Qwen3-VL-32B-Thinking.Q4_K_M with draft with -fit off - OOM

I'd prefer -fit default to be off at this point

[xwildfyre]

Hey all, can confirm this is not just Windows/NVIDIA - it's happening to me on b7499 on Apple silicon as well (Apple M3, 24GB, 10 GPU cores)

Using gemma-3-27b-it-UD-Q3_K_XL

--fit enabled automatically:

~/llama/llama-server --log-verbose --log-colors on --log-file 'llama/logfile.log' --log-prefix --log-timestamps --n-gpu-layers 99 --flash-attn enabled --model 'models/gemma-3-27b-it-UD-Q3_K_XL.gguf' --alias 'Gemma 3 27B' --top-p 0.95 --top-k 64 --temp 1.0 --min-p 0 --ctx-size 12288 --prio 2 --threads 4 --cpu-strict 1 --cpu-range 4-7 --mmproj 'models/mmproj-BF16-gemma-3-27b-it-UD.gguf'

Gives me 1 t/s before gradually slowing down to 0.01 t/s and then stopping completely.

As suggested, --fit off fixes it:

~/llama/llama-server --log-verbose --log-colors on --log-file 'llama/logfile.log' --log-prefix --log-timestamps --n-gpu-layers 99 --flash-attn enabled --model 'models/gemma-3-27b-it-UD-Q3_K_XL.gguf' --alias 'Gemma 3 27B' --top-p 0.95 --top-k 64 --temp 1.0 --min-p 0 --ctx-size 12288 --prio 2 --threads 4 --cpu-strict 1 --cpu-range 4-7 --mmproj 'models/mmproj-BF16-gemma-3-27b-it-UD.gguf' --fit off

Gives me about 4 t/s, which is more or less what I got with b6993.