Vulkan backend crashes when reasoning is disabled with Qwen3.5-35B-A3B

[Zirconium419122]

When running the Qwen3.5-35B-A3B GGUF model with the Vulkan backend in llama.cpp, inference succeeds when reasoning (thinking) is enabled, but consistently crashes when reasoning is disabled.

This behavior is reproducible across both llama-cli and llama-server. The failure manifests as a Vulkan device loss (vk::DeviceLostError) during inference when reasoning is turned off, while identical configurations with reasoning enabled run without issues.

Notably, disabling reasoning via --reasoning off triggers the crash in llama-cli, whereas using chat template kwargs (--chat-template-kwargs '{"enable_thinking":false}') does not exhibit the same failure in llama-cli, suggesting a discrepancy in how reasoning modes are applied or handled internally.

llama-bench -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -p 0 -n 128,256,512
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg128 |         12.39 ± 0.12 |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg256 |         12.46 ± 0.10 |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg512 |         12.30 ± 0.14 |

build: 9bcb4ef (8548)

llama-cli with reasoning:

llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja -c 32768 -n 100 -p "Hello\!"
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b8548-9bcb4ef
model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read               add a text file


> Hello!

[Start thinking]

Thinking Process:

1.  **Analyze the Input:**
    *   Input: "Hello!"
    *   Intent: Greeting.
    *   Tone: Friendly, casual.
    *   Context: Initial interaction.

2.  **Determine the Appropriate Response:**
    *   Acknowledge the greeting.
    *   Offer assistance.
    *   Maintain a friendly and helpful tone.
    *   Keep it concise (since it's

[ Prompt: 23,8 t/s | Generation: 9,8 t/s ]

llama-cli without reasoning:

llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja -c 32768 --reasoning off -n 100 -p "Hello\!"
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b8548-9bcb4ef
model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read               add a text file


> Hello!

-/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(+0x17c5a) [0x7f0a9b6b7c5a]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(ggml_print_backtrace+0x204) [0x7f0a9b6b8114]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(+0x2c5d9) [0x7f0a9b6cc5d9]
/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xc539a) [0x7f0a970c539a]
/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f0a970b286e]
/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xc5637) [0x7f0a970c5637]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-vulkan.so.0(+0x89426) [0x7f0a97489426]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-vulkan.so.0(+0x193945) [0x7f0a97593945]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-vulkan.so.0(+0x193c28) [0x7f0a97593c28]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x92c) [0x7f0a9b6d5ecc]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1) [0x7f0a9accd2f1]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0x11f) [0x7f0a9accfd7f]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3c9) [0x7f0a9acd6aa9]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(llama_decode+0x11) [0x7f0a9acd8781]
llama-cli(+0x1a7fcb) [0x55cc2e4affcb]
llama-cli(+0x139fc6) [0x55cc2e441fc6]
/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xf2fa4) [0x7f0a970f2fa4]
/nix/store/jms7zxzm7w1whczwny5m3gkgdjghmi2r-glibc-2.42-51/lib/libc.so.6(+0x9dd53) [0x7f0a96c9dd53]
/nix/store/jms7zxzm7w1whczwny5m3gkgdjghmi2r-glibc-2.42-51/lib/libc.so.6(+0x12563c) [0x7f0a96d2563c]
terminate called after throwing an instance of 'vk::DeviceLostError'
  what():  vk::Device::waitForFences: ErrorDeviceLost
zsh: IOT instruction (core dumped)  llama-cli -m  --jinja -c 32768 --reasoning off -n 100 -p "Hello!"

Works using --chat-template-kwargs '{"enable_thinking":false}' for some reason:

llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja -c 32768 --chat-template-kwargs '{"enable_thinking":false}' -n 100 -p "Hello\!"
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b8548-9bcb4ef
model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read               add a text file


> Hello!

[Start thinking]

Thinking Process:

1.  **Analyze the Input:**
    *   Input: "Hello!"
    *   Intent: Greeting.
    *   Tone: Friendly, casual.

2.  **Determine the appropriate response:**
    *   Acknowledge the greeting.
    *   Offer assistance.
    *   Maintain a friendly and helpful tone.
    *   Keep it concise (since it's just a greeting).

3.  **Draft

[ Prompt: 24,0 t/s | Generation: 9,5 t/s ]

llama-server fails with --chat-template-kwargs '{"enable_thinking":false}' (prompt was Hello!):

llama-server -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -np 1 --jinja -c 32768 --chat-template-kwargs '{"enable_thinking":false}' -lv 2 --host 127.0.0.1 --port 8088
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
Setting 'enable_thinking' via --chat-template-kwargs is deprecated. Use --reasoning on / --reasoning off instead.
llama_context: n_ctx_seq (32768) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
common_speculative_is_compat: the target context does not support partial sequence removal
srv    load_model: speculative decoding not supported by this context
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(+0x17c5a) [0x7f2385365c5a]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(ggml_print_backtrace+0x204) [0x7f2385366114]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(+0x2c5d9) [0x7f238537a5d9]
/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xc539a) [0x7f23814c539a]
/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f23814b286e]
/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xc5637) [0x7f23814c5637]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-vulkan.so.0(+0x89426) [0x7f2381889426]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-vulkan.so.0(+0x193945) [0x7f2381993945]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-vulkan.so.0(+0x193c28) [0x7f2381993c28]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x92c) [0x7f2385383ecc]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1) [0x7f23850cd2f1]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0x11f) [0x7f23850cfd7f]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3c9) [0x7f23850d6aa9]
/nix/store/g120d4bqgw63ashvjda55a1z9dgl8bd7-llama-cpp-8548/lib/libllama.so.0(llama_decode+0x11) [0x7f23850d8781]
llama-server(+0x186feb) [0x5653b7877feb]
llama-server(+0x1d4e66) [0x5653b78c5e66]
llama-server(+0xd4ced) [0x5653b77c5ced]
/nix/store/jms7zxzm7w1whczwny5m3gkgdjghmi2r-glibc-2.42-51/lib/libc.so.6(+0x2b285) [0x7f238102b285]
/nix/store/jms7zxzm7w1whczwny5m3gkgdjghmi2r-glibc-2.42-51/lib/libc.so.6(__libc_start_main+0x88) [0x7f238102b338]
llama-server(+0xdbfe5) [0x5653b77ccfe5]
terminate called after throwing an instance of 'vk::DeviceLostError'
  what():  vk::Device::waitForFences: ErrorDeviceLost
zsh: IOT instruction (core dumped)  llama-server -m  -np 1 --jinja -c 32768 --chat-template-kwargs  -lv 2 --host

llama-server with reasoning:

Details

llama-server -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -np 1 --jinja -c 32768 -lv 3 --host 127.0.0.1 --port 8088
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
build: 8548 (9bcb4ef) with GNU 15.2.0 for Linux x86_64
system info: n_threads = 3, n_threads_batch = 3, total_threads = 16
system_info: n_threads = 3 (n_threads_batch = 3) / 16 | CPU : LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL

init: using 15 threads for HTTP server

start: binding port with default address family

main: loading model

srv    load_model: loading model '/home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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 12526 MiB of device memory vs. 21316 MiB of free device memory

llama_params_fit_impl: will leave 8790 >= 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 0.78 seconds

llama_model_load_from_file_impl: using device Vulkan0 (Intel(R) Graphics (ARL)) (0000:00:02.0) - 21317 MiB free

llama_model_loader: loaded meta data with 52 key-value pairs and 733 tensors from /home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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              = qwen35moe

llama_model_loader: - kv   1:                               general.type str              = model

llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 20

llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000

llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000

llama_model_loader: - kv   5:                               general.name str              = Qwen3.5-35B-A3B

llama_model_loader: - kv   6:                           general.basename str              = Qwen3.5-35B-A3B

llama_model_loader: - kv   7:                       general.quantized_by str              = Unsloth

llama_model_loader: - kv   8:                         general.size_label str              = 35B-A3B

llama_model_loader: - kv   9:                            general.license str              = apache-2.0

llama_model_loader: - kv  10:                       general.license.link str              = https://huggingface.co/Qwen/Qwen3.5-3...

llama_model_loader: - kv  11:                           general.repo_url str              = https://huggingface.co/unsloth

llama_model_loader: - kv  12:                   general.base_model.count u32              = 1

llama_model_loader: - kv  13:                  general.base_model.0.name str              = Qwen3.5 35B A3B

llama_model_loader: - kv  14:          general.base_model.0.organization str              = Qwen

llama_model_loader: - kv  15:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen3.5-3...

llama_model_loader: - kv  16:                               general.tags arr[str,2]       = ["unsloth", "image-text-to-text"]

llama_model_loader: - kv  17:                      qwen35moe.block_count u32              = 40

llama_model_loader: - kv  18:                   qwen35moe.context_length u32              = 262144

llama_model_loader: - kv  19:                 qwen35moe.embedding_length u32              = 2048

llama_model_loader: - kv  20:             qwen35moe.attention.head_count u32              = 16

llama_model_loader: - kv  21:          qwen35moe.attention.head_count_kv u32              = 2

llama_model_loader: - kv  22:          qwen35moe.rope.dimension_sections arr[i32,4]       = [11, 11, 10, 0]

llama_model_loader: - kv  23:                   qwen35moe.rope.freq_base f32              = 10000000.000000

llama_model_loader: - kv  24: qwen35moe.attention.layer_norm_rms_epsilon f32              = 0.000001

llama_model_loader: - kv  25:                     qwen35moe.expert_count u32              = 256

llama_model_loader: - kv  26:                qwen35moe.expert_used_count u32              = 8

llama_model_loader: - kv  27:             qwen35moe.attention.key_length u32              = 256

llama_model_loader: - kv  28:           qwen35moe.attention.value_length u32              = 256

llama_model_loader: - kv  29:       qwen35moe.expert_feed_forward_length u32              = 512

llama_model_loader: - kv  30: qwen35moe.expert_shared_feed_forward_length u32              = 512

llama_model_loader: - kv  31:                  qwen35moe.ssm.conv_kernel u32              = 4

llama_model_loader: - kv  32:                   qwen35moe.ssm.state_size u32              = 128

llama_model_loader: - kv  33:                  qwen35moe.ssm.group_count u32              = 16

llama_model_loader: - kv  34:               qwen35moe.ssm.time_step_rank u32              = 32

llama_model_loader: - kv  35:                   qwen35moe.ssm.inner_size u32              = 4096

llama_model_loader: - kv  36:          qwen35moe.full_attention_interval u32              = 4

llama_model_loader: - kv  37:             qwen35moe.rope.dimension_count u32              = 64

llama_model_loader: - kv  38:                       tokenizer.ggml.model str              = gpt2

llama_model_loader: - kv  39:                         tokenizer.ggml.pre str              = qwen35

llama_model_loader: - kv  40:                      tokenizer.ggml.tokens arr[str,248320]  = ["!", """, "#", "$", "%", "&", "'", ...

llama_model_loader: - kv  41:                  tokenizer.ggml.token_type arr[i32,248320]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...

llama_model_loader: - kv  42:                      tokenizer.ggml.merges arr[str,247587]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...

llama_model_loader: - kv  43:                tokenizer.ggml.eos_token_id u32              = 248046

llama_model_loader: - kv  44:            tokenizer.ggml.padding_token_id u32              = 248055

llama_model_loader: - kv  45:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...

llama_model_loader: - kv  46:               general.quantization_version u32              = 2

llama_model_loader: - kv  47:                          general.file_type u32              = 10

llama_model_loader: - kv  48:                      quantize.imatrix.file str              = Qwen3.5-35B-A3B-GGUF/imatrix_unsloth....

llama_model_loader: - kv  49:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3.5-35B-A3B.txt

llama_model_loader: - kv  50:             quantize.imatrix.entries_count u32              = 510

llama_model_loader: - kv  51:              quantize.imatrix.chunks_count u32              = 76

llama_model_loader: - type  f32:  301 tensors

llama_model_loader: - type q8_0:   61 tensors

llama_model_loader: - type q4_K:    1 tensors

llama_model_loader: - type q5_K:  177 tensors

llama_model_loader: - type q6_K:   73 tensors

llama_model_loader: - type iq2_xs:   78 tensors

llama_model_loader: - type iq3_xxs:   41 tensors

llama_model_loader: - type iq4_xs:    1 tensors

print_info: file format = GGUF V3 (latest)

print_info: file type   = Q2_K - Medium

print_info: file size   = 11.31 GiB (2.80 BPW)

load: 0 unused tokens

load: printing all EOG tokens:

load:   - 248044 ('<|endoftext|>')

load:   - 248046 ('<|im_end|>')

load:   - 248063 ('<|fim_pad|>')

load:   - 248064 ('<|repo_name|>')

load:   - 248065 ('<|file_sep|>')

load: special tokens cache size = 33

load: token to piece cache size = 1.7581 MB

print_info: arch                  = qwen35moe

print_info: vocab_only            = 0

print_info: no_alloc              = 0

print_info: n_ctx_train           = 262144

print_info: n_embd                = 2048

print_info: n_embd_inp            = 2048

print_info: n_layer               = 40

print_info: n_head                = 16

print_info: n_head_kv             = 2

print_info: n_rot                 = 64

print_info: n_swa                 = 0

print_info: is_swa_any            = 0

print_info: n_embd_head_k         = 256

print_info: n_embd_head_v         = 256

print_info: n_gqa                 = 8

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                  = 0

print_info: n_expert              = 256

print_info: n_expert_used         = 8

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             = 40

print_info: rope scaling          = linear

print_info: freq_base_train       = 10000000.0

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        = [11, 11, 10, 0]

print_info: ssm_d_conv            = 4

print_info: ssm_d_inner           = 4096

print_info: ssm_d_state           = 128

print_info: ssm_dt_rank           = 32

print_info: ssm_n_group           = 16

print_info: ssm_dt_b_c_rms        = 0

print_info: model type            = 35B.A3B

print_info: model params          = 34.66 B

print_info: general.name          = Qwen3.5-35B-A3B

print_info: vocab type            = BPE

print_info: n_vocab               = 248320

print_info: n_merges              = 247587

print_info: BOS token             = 11 ','

print_info: EOS token             = 248046 '<|im_end|>'

print_info: EOT token             = 248046 '<|im_end|>'

print_info: PAD token             = 248055 '<|vision_pad|>'

print_info: LF token              = 198 'Ċ'

print_info: FIM PRE token         = 248060 '<|fim_prefix|>'

print_info: FIM SUF token         = 248062 '<|fim_suffix|>'

print_info: FIM MID token         = 248061 '<|fim_middle|>'

print_info: FIM PAD token         = 248063 '<|fim_pad|>'

print_info: FIM REP token         = 248064 '<|repo_name|>'

print_info: FIM SEP token         = 248065 '<|file_sep|>'

print_info: EOG token             = 248044 '<|endoftext|>'

print_info: EOG token             = 248046 '<|im_end|>'

print_info: EOG token             = 248063 '<|fim_pad|>'

print_info: EOG token             = 248064 '<|repo_name|>'

print_info: EOG token             = 248065 '<|file_sep|>'

print_info: max token length      = 256

load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)

load_tensors: offloading output layer to GPU

load_tensors: offloading 39 repeating layers to GPU

load_tensors: offloaded 41/41 layers to GPU

load_tensors:   CPU_Mapped model buffer size =   333.44 MiB

load_tensors:      Vulkan0 model buffer size = 11249.67 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         = 32768

llama_context: n_ctx_seq     = 32768

llama_context: n_batch       = 2048

llama_context: n_ubatch      = 512

llama_context: causal_attn   = 1

llama_context: flash_attn    = auto

llama_context: kv_unified    = false

llama_context: freq_base     = 10000000.0

llama_context: freq_scale    = 1

llama_context: n_ctx_seq (32768) < n_ctx_train (262144) -- the full capacity of the model will not be utilized

llama_context: Vulkan_Host  output buffer size =     0.95 MiB

llama_kv_cache:    Vulkan0 KV buffer size =   640.00 MiB

llama_kv_cache: size =  640.00 MiB ( 32768 cells,  10 layers,  1/1 seqs), K (f16):  320.00 MiB, V (f16):  320.00 MiB

llama_memory_recurrent:    Vulkan0 RS buffer size =    62.81 MiB

llama_memory_recurrent: size =   62.81 MiB (     1 cells,  40 layers,  1 seqs), R (f32):    2.81 MiB, S (f32):   60.00 MiB

sched_reserve: reserving ...

sched_reserve: Flash Attention was auto, set to enabled

sched_reserve: resolving fused Gated Delta Net support:

sched_reserve: fused Gated Delta Net (autoregressive) enabled

sched_reserve: fused Gated Delta Net (chunked) enabled

sched_reserve:    Vulkan0 compute buffer size =   574.02 MiB

sched_reserve: Vulkan_Host compute buffer size =    72.03 MiB

sched_reserve: graph nodes  = 3729

sched_reserve: graph splits = 2

sched_reserve: reserve took 45.75 ms, sched copies = 1

common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)

srv    load_model: initializing slots, n_slots = 1

common_speculative_is_compat: the target context does not support partial sequence removal

srv    load_model: speculative decoding not supported by this context

slot   load_model: id  0 | task -1 | new slot, n_ctx = 32768

srv    load_model: prompt cache is enabled, size limit: 8192 MiB

srv    load_model: use --cache-ram 0 to disable the prompt cache

srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391

init: chat template, 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



'

srv          init: init: chat template, thinking = 1

main: model loaded

main: server is listening on http://127.0.0.1:8088

main: starting the main loop...

srv  update_slots: all slots are idle

srv  log_server_r: done request: GET / 127.0.0.1 200

srv  log_server_r: done request: HEAD /cors-proxy 127.0.0.1 404

srv  params_from_: Chat format: peg-native

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, is_child = 0

slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, task.n_tokens = 12

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 = 8, batch.n_tokens = 8, progress = 0.666667

srv  log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200

slot update_slots: id  0 | task 0 | n_tokens = 8, memory_seq_rm [8, end)

slot init_sampler: id  0 | task 0 | init sampler, took 0.01 ms, tokens: text = 12, total = 12

slot update_slots: id  0 | task 0 | prompt processing done, n_tokens = 12, batch.n_tokens = 4

slot print_timing: id  0 | task 0 |

prompt eval time =     548.88 ms /    12 tokens (   45.74 ms per token,    21.86 tokens per second)

eval time =    7830.80 ms /    78 tokens (  100.39 ms per token,     9.96 tokens per second)

total time =    8379.68 ms /    90 tokens

slot      release: id  0 | task 0 | stop processing: n_tokens = 89, truncated = 0

srv  update_slots: all slots are idle

[Zirconium419122]

Built llama-cpp using nix shell github:ggml-org/llama.cpp#packages.x86_64-linux.vulkan, commit 97508acb17ff933c67edb3a2f0997bc19a6abb98.

llama-bench -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -p 0 -n 128,256,512
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 49152 | int dot: 0 | matrix cores: none
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg128 |         10.30 ± 0.10 |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg256 |         10.29 ± 0.03 |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg512 |         10.22 ± 0.06 |

build: unknown (0)

llama-server with reasoning:

Details

llama-server -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -np 1 --jinja -c 32768 --host 127.0.0.1 --port 8088
build_info: b0-unknown
system_info: n_threads = 3 (n_threads_batch = 3) / 16 | CPU : LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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 12526 MiB of device memory vs. 21279 MiB of free device memory
llama_params_fit_impl: will leave 8753 >= 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 0.80 seconds
llama_model_load_from_file_impl: using device Vulkan0 (Intel(R) Graphics (ARL)) (0000:00:02.0) - 21317 MiB free
llama_model_loader: loaded meta data with 52 key-value pairs and 733 tensors from /home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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              = qwen35moe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 20
llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000
llama_model_loader: - kv   5:                               general.name str              = Qwen3.5-35B-A3B
llama_model_loader: - kv   6:                           general.basename str              = Qwen3.5-35B-A3B
llama_model_loader: - kv   7:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   8:                         general.size_label str              = 35B-A3B
llama_model_loader: - kv   9:                            general.license str              = apache-2.0
llama_model_loader: - kv  10:                       general.license.link str              = https://huggingface.co/Qwen/Qwen3.5-3...
llama_model_loader: - kv  11:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  12:                   general.base_model.count u32              = 1
llama_model_loader: - kv  13:                  general.base_model.0.name str              = Qwen3.5 35B A3B
llama_model_loader: - kv  14:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  15:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen3.5-3...
llama_model_loader: - kv  16:                               general.tags arr[str,2]       = ["unsloth", "image-text-to-text"]
llama_model_loader: - kv  17:                      qwen35moe.block_count u32              = 40
llama_model_loader: - kv  18:                   qwen35moe.context_length u32              = 262144
llama_model_loader: - kv  19:                 qwen35moe.embedding_length u32              = 2048
llama_model_loader: - kv  20:             qwen35moe.attention.head_count u32              = 16
llama_model_loader: - kv  21:          qwen35moe.attention.head_count_kv u32              = 2
llama_model_loader: - kv  22:          qwen35moe.rope.dimension_sections arr[i32,4]       = [11, 11, 10, 0]
llama_model_loader: - kv  23:                   qwen35moe.rope.freq_base f32              = 10000000.000000
llama_model_loader: - kv  24: qwen35moe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  25:                     qwen35moe.expert_count u32              = 256
llama_model_loader: - kv  26:                qwen35moe.expert_used_count u32              = 8
llama_model_loader: - kv  27:             qwen35moe.attention.key_length u32              = 256
llama_model_loader: - kv  28:           qwen35moe.attention.value_length u32              = 256
llama_model_loader: - kv  29:       qwen35moe.expert_feed_forward_length u32              = 512
llama_model_loader: - kv  30: qwen35moe.expert_shared_feed_forward_length u32              = 512
llama_model_loader: - kv  31:                  qwen35moe.ssm.conv_kernel u32              = 4
llama_model_loader: - kv  32:                   qwen35moe.ssm.state_size u32              = 128
llama_model_loader: - kv  33:                  qwen35moe.ssm.group_count u32              = 16
llama_model_loader: - kv  34:               qwen35moe.ssm.time_step_rank u32              = 32
llama_model_loader: - kv  35:                   qwen35moe.ssm.inner_size u32              = 4096
llama_model_loader: - kv  36:          qwen35moe.full_attention_interval u32              = 4
llama_model_loader: - kv  37:             qwen35moe.rope.dimension_count u32              = 64
llama_model_loader: - kv  38:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  39:                         tokenizer.ggml.pre str              = qwen35
llama_model_loader: - kv  40:                      tokenizer.ggml.tokens arr[str,248320]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  41:                  tokenizer.ggml.token_type arr[i32,248320]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  42:                      tokenizer.ggml.merges arr[str,247587]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  43:                tokenizer.ggml.eos_token_id u32              = 248046
llama_model_loader: - kv  44:            tokenizer.ggml.padding_token_id u32              = 248055
llama_model_loader: - kv  45:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
llama_model_loader: - kv  46:               general.quantization_version u32              = 2
llama_model_loader: - kv  47:                          general.file_type u32              = 10
llama_model_loader: - kv  48:                      quantize.imatrix.file str              = Qwen3.5-35B-A3B-GGUF/imatrix_unsloth....
llama_model_loader: - kv  49:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3.5-35B-A3B.txt
llama_model_loader: - kv  50:             quantize.imatrix.entries_count u32              = 510
llama_model_loader: - kv  51:              quantize.imatrix.chunks_count u32              = 76
llama_model_loader: - type  f32:  301 tensors
llama_model_loader: - type q8_0:   61 tensors
llama_model_loader: - type q4_K:    1 tensors
llama_model_loader: - type q5_K:  177 tensors
llama_model_loader: - type q6_K:   73 tensors
llama_model_loader: - type iq2_xs:   78 tensors
llama_model_loader: - type iq3_xxs:   41 tensors
llama_model_loader: - type iq4_xs:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 11.31 GiB (2.80 BPW)
load: 0 unused tokens
load: printing all EOG tokens:
load:   - 248044 ('<|endoftext|>')
load:   - 248046 ('<|im_end|>')
load:   - 248063 ('<|fim_pad|>')
load:   - 248064 ('<|repo_name|>')
load:   - 248065 ('<|file_sep|>')
load: special tokens cache size = 33
load: token to piece cache size = 1.7581 MB
print_info: arch                  = qwen35moe
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 2048
print_info: n_embd_inp            = 2048
print_info: n_layer               = 40
print_info: n_head                = 16
print_info: n_head_kv             = 2
print_info: n_rot                 = 64
print_info: n_swa                 = 0
print_info: is_swa_any            = 0
print_info: n_embd_head_k         = 256
print_info: n_embd_head_v         = 256
print_info: n_gqa                 = 8
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                  = 0
print_info: n_expert              = 256
print_info: n_expert_used         = 8
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             = 40
print_info: rope scaling          = linear
print_info: freq_base_train       = 10000000.0
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        = [11, 11, 10, 0]
print_info: ssm_d_conv            = 4
print_info: ssm_d_inner           = 4096
print_info: ssm_d_state           = 128
print_info: ssm_dt_rank           = 32
print_info: ssm_n_group           = 16
print_info: ssm_dt_b_c_rms        = 0
print_info: model type            = 35B.A3B
print_info: model params          = 34.66 B
print_info: general.name          = Qwen3.5-35B-A3B
print_info: vocab type            = BPE
print_info: n_vocab               = 248320
print_info: n_merges              = 247587
print_info: BOS token             = 11 ','
print_info: EOS token             = 248046 '<|im_end|>'
print_info: EOT token             = 248046 '<|im_end|>'
print_info: PAD token             = 248055 '<|vision_pad|>'
print_info: LF token              = 198 'Ċ'
print_info: FIM PRE token         = 248060 '<|fim_prefix|>'
print_info: FIM SUF token         = 248062 '<|fim_suffix|>'
print_info: FIM MID token         = 248061 '<|fim_middle|>'
print_info: FIM PAD token         = 248063 '<|fim_pad|>'
print_info: FIM REP token         = 248064 '<|repo_name|>'
print_info: FIM SEP token         = 248065 '<|file_sep|>'
print_info: EOG token             = 248044 '<|endoftext|>'
print_info: EOG token             = 248046 '<|im_end|>'
print_info: EOG token             = 248063 '<|fim_pad|>'
print_info: EOG token             = 248064 '<|repo_name|>'
print_info: EOG token             = 248065 '<|file_sep|>'
print_info: max token length      = 256
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 39 repeating layers to GPU
load_tensors: offloaded 41/41 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   333.44 MiB
load_tensors:      Vulkan0 model buffer size = 11249.67 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         = 32768
llama_context: n_ctx_seq     = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 10000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (32768) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     0.95 MiB
llama_kv_cache:    Vulkan0 KV buffer size =   640.00 MiB
llama_kv_cache: size =  640.00 MiB ( 32768 cells,  10 layers,  1/1 seqs), K (f16):  320.00 MiB, V (f16):  320.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
llama_memory_recurrent:    Vulkan0 RS buffer size =    62.81 MiB
llama_memory_recurrent: size =   62.81 MiB (     1 cells,  40 layers,  1 seqs), R (f32):    2.81 MiB, S (f32):   60.00 MiB
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:    Vulkan0 compute buffer size =   574.02 MiB
sched_reserve: Vulkan_Host compute buffer size =    72.03 MiB
sched_reserve: graph nodes  = 3729
sched_reserve: graph splits = 2
sched_reserve: reserve took 59.55 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv    load_model: initializing slots, n_slots = 1
common_speculative_is_compat: the target context does not support partial sequence removal
srv    load_model: speculative decoding not supported by this context
slot   load_model: id  0 | task -1 | new slot, n_ctx = 32768
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: init: --clear-idle requires --kv-unified, disabling
init: chat template, 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

'
srv          init: init: chat template, thinking = 1
main: model loaded
main: server is listening on http://127.0.0.1:8088
main: starting the main loop...
srv  update_slots: all slots are idle
srv  params_from_: Chat format: peg-native
slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
srv  get_availabl: updating prompt cache
srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 32768 tokens, 8589934592 est)
srv  get_availabl: prompt cache update took 0.01 ms
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, is_child = 0
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, task.n_tokens = 12
slot update_slots: id  0 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
srv  log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 8, batch.n_tokens = 8, progress = 0.666667
slot update_slots: id  0 | task 0 | n_tokens = 8, memory_seq_rm [8, end)
reasoning-budget: activated, budget=2147483647 tokens
slot init_sampler: id  0 | task 0 | init sampler, took 0.01 ms, tokens: text = 12, total = 12                                                                              srv          stop: cancel task, id_task = 0 processing done, n_tokens = 12, batch.n_tokens = 4
slot      release: id  0 | task 0 | stop processing: n_tokens = 129, truncated = 0
srv  update_slots: all slots are idle
srv    operator(): operator(): cleaning up before exit...
llama_memory_breakdown_print: | memory breakdown [MiB]                | total    free     self   model   context   compute    unaccounted |
llama_memory_breakdown_print: |   - Vulkan0 (Intel(R) Graphics (ARL)) | 23686 = 10763 + (12526 = 11249 +     702 +     574) +         396 |
llama_memory_breakdown_print: |   - Host                              |                    405 =   333 +       0 +      72                |

llama-cpp without reasoning:

Details

llama-server -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -np 1 --jinja -c 32768 --chat-template-kwargs '{"enable_thinking":false}' --host 127.0.0.1 --port 8088
Setting 'enable_thinking' via --chat-template-kwargs is deprecated. Use --reasoning on / --reasoning off instead.
build_info: b0-unknown
system_info: n_threads = 3 (n_threads_batch = 3) / 16 | CPU : LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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 12526 MiB of device memory vs. 21316 MiB of free device memory
llama_params_fit_impl: will leave 8790 >= 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 0.79 seconds
llama_model_load_from_file_impl: using device Vulkan0 (Intel(R) Graphics (ARL)) (0000:00:02.0) - 21317 MiB free
llama_model_loader: loaded meta data with 52 key-value pairs and 733 tensors from /home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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              = qwen35moe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 20
llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000
llama_model_loader: - kv   5:                               general.name str              = Qwen3.5-35B-A3B
llama_model_loader: - kv   6:                           general.basename str              = Qwen3.5-35B-A3B
llama_model_loader: - kv   7:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   8:                         general.size_label str              = 35B-A3B
llama_model_loader: - kv   9:                            general.license str              = apache-2.0
llama_model_loader: - kv  10:                       general.license.link str              = https://huggingface.co/Qwen/Qwen3.5-3...
llama_model_loader: - kv  11:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  12:                   general.base_model.count u32              = 1
llama_model_loader: - kv  13:                  general.base_model.0.name str              = Qwen3.5 35B A3B
llama_model_loader: - kv  14:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  15:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen3.5-3...
llama_model_loader: - kv  16:                               general.tags arr[str,2]       = ["unsloth", "image-text-to-text"]
llama_model_loader: - kv  17:                      qwen35moe.block_count u32              = 40
llama_model_loader: - kv  18:                   qwen35moe.context_length u32              = 262144
llama_model_loader: - kv  19:                 qwen35moe.embedding_length u32              = 2048
llama_model_loader: - kv  20:             qwen35moe.attention.head_count u32              = 16
llama_model_loader: - kv  21:          qwen35moe.attention.head_count_kv u32              = 2
llama_model_loader: - kv  22:          qwen35moe.rope.dimension_sections arr[i32,4]       = [11, 11, 10, 0]
llama_model_loader: - kv  23:                   qwen35moe.rope.freq_base f32              = 10000000.000000
llama_model_loader: - kv  24: qwen35moe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  25:                     qwen35moe.expert_count u32              = 256
llama_model_loader: - kv  26:                qwen35moe.expert_used_count u32              = 8
llama_model_loader: - kv  27:             qwen35moe.attention.key_length u32              = 256
llama_model_loader: - kv  28:           qwen35moe.attention.value_length u32              = 256
llama_model_loader: - kv  29:       qwen35moe.expert_feed_forward_length u32              = 512
llama_model_loader: - kv  30: qwen35moe.expert_shared_feed_forward_length u32              = 512
llama_model_loader: - kv  31:                  qwen35moe.ssm.conv_kernel u32              = 4
llama_model_loader: - kv  32:                   qwen35moe.ssm.state_size u32              = 128
llama_model_loader: - kv  33:                  qwen35moe.ssm.group_count u32              = 16
llama_model_loader: - kv  34:               qwen35moe.ssm.time_step_rank u32              = 32
llama_model_loader: - kv  35:                   qwen35moe.ssm.inner_size u32              = 4096
llama_model_loader: - kv  36:          qwen35moe.full_attention_interval u32              = 4
llama_model_loader: - kv  37:             qwen35moe.rope.dimension_count u32              = 64
llama_model_loader: - kv  38:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  39:                         tokenizer.ggml.pre str              = qwen35
llama_model_loader: - kv  40:                      tokenizer.ggml.tokens arr[str,248320]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  41:                  tokenizer.ggml.token_type arr[i32,248320]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  42:                      tokenizer.ggml.merges arr[str,247587]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  43:                tokenizer.ggml.eos_token_id u32              = 248046
llama_model_loader: - kv  44:            tokenizer.ggml.padding_token_id u32              = 248055
llama_model_loader: - kv  45:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
llama_model_loader: - kv  46:               general.quantization_version u32              = 2
llama_model_loader: - kv  47:                          general.file_type u32              = 10
llama_model_loader: - kv  48:                      quantize.imatrix.file str              = Qwen3.5-35B-A3B-GGUF/imatrix_unsloth....
llama_model_loader: - kv  49:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3.5-35B-A3B.txt
llama_model_loader: - kv  50:             quantize.imatrix.entries_count u32              = 510
llama_model_loader: - kv  51:              quantize.imatrix.chunks_count u32              = 76
llama_model_loader: - type  f32:  301 tensors
llama_model_loader: - type q8_0:   61 tensors
llama_model_loader: - type q4_K:    1 tensors
llama_model_loader: - type q5_K:  177 tensors
llama_model_loader: - type q6_K:   73 tensors
llama_model_loader: - type iq2_xs:   78 tensors
llama_model_loader: - type iq3_xxs:   41 tensors
llama_model_loader: - type iq4_xs:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 11.31 GiB (2.80 BPW)
load: 0 unused tokens
load: printing all EOG tokens:
load:   - 248044 ('<|endoftext|>')
load:   - 248046 ('<|im_end|>')
load:   - 248063 ('<|fim_pad|>')
load:   - 248064 ('<|repo_name|>')
load:   - 248065 ('<|file_sep|>')
load: special tokens cache size = 33
load: token to piece cache size = 1.7581 MB
print_info: arch                  = qwen35moe
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 2048
print_info: n_embd_inp            = 2048
print_info: n_layer               = 40
print_info: n_head                = 16
print_info: n_head_kv             = 2
print_info: n_rot                 = 64
print_info: n_swa                 = 0
print_info: is_swa_any            = 0
print_info: n_embd_head_k         = 256
print_info: n_embd_head_v         = 256
print_info: n_gqa                 = 8
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                  = 0
print_info: n_expert              = 256
print_info: n_expert_used         = 8
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             = 40
print_info: rope scaling          = linear
print_info: freq_base_train       = 10000000.0
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        = [11, 11, 10, 0]
print_info: ssm_d_conv            = 4
print_info: ssm_d_inner           = 4096
print_info: ssm_d_state           = 128
print_info: ssm_dt_rank           = 32
print_info: ssm_n_group           = 16
print_info: ssm_dt_b_c_rms        = 0
print_info: model type            = 35B.A3B
print_info: model params          = 34.66 B
print_info: general.name          = Qwen3.5-35B-A3B
print_info: vocab type            = BPE
print_info: n_vocab               = 248320
print_info: n_merges              = 247587
print_info: BOS token             = 11 ','
print_info: EOS token             = 248046 '<|im_end|>'
print_info: EOT token             = 248046 '<|im_end|>'
print_info: PAD token             = 248055 '<|vision_pad|>'
print_info: LF token              = 198 'Ċ'
print_info: FIM PRE token         = 248060 '<|fim_prefix|>'
print_info: FIM SUF token         = 248062 '<|fim_suffix|>'
print_info: FIM MID token         = 248061 '<|fim_middle|>'
print_info: FIM PAD token         = 248063 '<|fim_pad|>'
print_info: FIM REP token         = 248064 '<|repo_name|>'
print_info: FIM SEP token         = 248065 '<|file_sep|>'
print_info: EOG token             = 248044 '<|endoftext|>'
print_info: EOG token             = 248046 '<|im_end|>'
print_info: EOG token             = 248063 '<|fim_pad|>'
print_info: EOG token             = 248064 '<|repo_name|>'
print_info: EOG token             = 248065 '<|file_sep|>'
print_info: max token length      = 256
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 39 repeating layers to GPU
load_tensors: offloaded 41/41 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   333.44 MiB
load_tensors:      Vulkan0 model buffer size = 11249.67 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         = 32768
llama_context: n_ctx_seq     = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 10000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (32768) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     0.95 MiB
llama_kv_cache:    Vulkan0 KV buffer size =   640.00 MiB
llama_kv_cache: size =  640.00 MiB ( 32768 cells,  10 layers,  1/1 seqs), K (f16):  320.00 MiB, V (f16):  320.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
llama_memory_recurrent:    Vulkan0 RS buffer size =    62.81 MiB
llama_memory_recurrent: size =   62.81 MiB (     1 cells,  40 layers,  1 seqs), R (f32):    2.81 MiB, S (f32):   60.00 MiB
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:    Vulkan0 compute buffer size =   574.02 MiB
sched_reserve: Vulkan_Host compute buffer size =    72.03 MiB
sched_reserve: graph nodes  = 3729
sched_reserve: graph splits = 2
sched_reserve: reserve took 42.58 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv    load_model: initializing slots, n_slots = 1
common_speculative_is_compat: the target context does not support partial sequence removal
srv    load_model: speculative decoding not supported by this context
slot   load_model: id  0 | task -1 | new slot, n_ctx = 32768
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: init: --clear-idle requires --kv-unified, disabling
init: chat template, 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


'

srv          init: init: chat template, thinking = 1

main: model loaded

main: server is listening on http://127.0.0.1:8088

main: starting the main loop...

srv  update_slots: all slots are idle

srv  log_server_r: done request: GET / 127.0.0.1 200

srv  log_server_r: done request: GET /bundle.css 127.0.0.1 200

srv  log_server_r: done request: GET /bundle.js 127.0.0.1 200

srv  log_server_r: done request: HEAD /cors-proxy 127.0.0.1 404

srv  params_from_: Chat format: peg-native

slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1

srv  get_availabl: updating prompt cache

srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000

srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 32768 tokens, 8589934592 est)

srv  get_availabl: prompt cache update took 0.02 ms

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, is_child = 0

slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, task.n_tokens = 14

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 = 10, batch.n_tokens = 10, progress = 0.714286

srv  log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200

Batch offset=0x2000 len=0x0 on queue 0 (aperture: 13202.9Mb, 256.0Mb VRAM only)

BO: addr=0x00000005f8b00000-0x00000005f8b41fff map=           (nil) size=    264KB handle=00034 capture=0 vram_only=0 name=user

BO: addr=0x00000001bfe00000-0x00000001bfffffff map=           (nil) size=   2048KB handle=00030 capture=1 vram_only=0 name=slab_parent

BO: addr=0x00000000c0000000-0x00000000c01fffff map=  0x7fe5261fd000 size=   2048KB handle=00005 capture=1 vram_only=0 name=binding table pool

BO: addr=0xffffeffeeec00000-0xffffeffefebfffff map=           (nil) size= 262144KB handle=00037 capture=0 vram_only=1 name=scratch

BO: addr=0x0000000100000000-0x00000001001fffff map=  0x7fe5265fd000 size=   2048KB handle=00003 capture=1 vram_only=0 name=scratch surface state pool

BO: addr=0x0000000100800000-0x00000001009fffff map=  0x7fe5263fd000 size=   2048KB handle=00004 capture=1 vram_only=0 name=internal surface state pool

BO: addr=0x00000002c0000000-0x00000002c01fffff map=  0x7fe544b70000 size=   2048KB handle=00002 capture=1 vram_only=0 name=dynamic pool

BO: addr=0x0000000000200000-0x00000000003fffff map=  0x7fe544d70000 size=   2048KB handle=00001 capture=1 vram_only=0 name=general pool

BO: addr=0x0000000200000000-0x00000002001fffff map=  0x7fe525ffd000 size=   2048KB handle=00012 capture=1 vram_only=0 name=shaders

BO: addr=0x0000000200200000-0x00000002003fffff map=  0x7fe5255fc000 size=   2048KB handle=00036 capture=1 vram_only=0 name=shaders

BO: addr=0x0000000200400000-0x00000002007fffff map=  0x7fe5251fc000 size=   4096KB handle=00038 capture=1 vram_only=0 name=shaders

BO: addr=0x00000003c0000000-0x00000003c01fffff map=  0x7fe524d59000 size=   2048KB handle=00006 capture=1 vram_only=0 name=aux-tt pool

BO: addr=0x0000000440000000-0x000000047fd36fff map=           (nil) size=1045724KB handle=00010 capture=0 vram_only=0 name=user

BO: addr=0x000000047fe00000-0x00000004bf8aefff map=           (nil) size=1043132KB handle=00011 capture=0 vram_only=0 name=user

BO: addr=0x00000004bfa00000-0x00000004fdfc9fff map=           (nil) size=1021736KB handle=00013 capture=0 vram_only=0 name=user

BO: addr=0x00000004fe000000-0x000000053cb96fff map=           (nil) size=1027676KB handle=00014 capture=0 vram_only=0 name=user

BO: addr=0x000000053cc00000-0x000000057ae96fff map=           (nil) size=1018460KB handle=00015 capture=0 vram_only=0 name=user

BO: addr=0x000000057b000000-0x00000005b8f08fff map=           (nil) size=1014820KB handle=00016 capture=0 vram_only=0 name=user

BO: addr=0x00000005b9000000-0x00000005f8aaefff map=           (nil) size=1043132KB handle=00017 capture=0 vram_only=0 name=user

BO: addr=0x00000005f8c00000-0x0000000636e96fff map=           (nil) size=1018460KB handle=00018 capture=0 vram_only=0 name=user

BO: addr=0x0000000637000000-0x0000000674f08fff map=           (nil) size=1014820KB handle=00019 capture=0 vram_only=0 name=user

BO: addr=0x0000000675000000-0x00000006b4aaefff map=           (nil) size=1043132KB handle=00020 capture=0 vram_only=0 name=user

BO: addr=0x00000006b4c00000-0x00000006f2e96fff map=           (nil) size=1018460KB handle=00021 capture=0 vram_only=0 name=user

BO: addr=0x00000006f3000000-0x000000070292cfff map=           (nil) size= 255156KB handle=00022 capture=0 vram_only=0 name=user

BO: addr=0x0000000702a00000-0x0000000713be0fff map=           (nil) size= 280452KB handle=00023 capture=0 vram_only=0 name=user

BO: addr=0x00000004bf900000-0x00000004bf9f3fff map=           (nil) size=    976KB handle=00024 capture=0 vram_only=0 name=user

BO: addr=0x0000000713c00000-0x000000073be7ffff map=           (nil) size= 657920KB handle=00025 capture=0 vram_only=0 name=user

BO: addr=0x000000073c000000-0x000000073ff0efff map=           (nil) size=  64572KB handle=00026 capture=0 vram_only=0 name=user

BO: addr=0x0000000740000000-0x0000000764043fff map=           (nil) size= 590096KB handle=00027 capture=0 vram_only=0 name=user

BO: addr=0x0000000764200000-0x0000000768a50fff map=           (nil) size=  74052KB handle=00028 capture=0 vram_only=0 name=user

BO: addr=0x000000057af00000-0x000000057af01fff map=           (nil) size=      8KB handle=00029 capture=0 vram_only=0 name=user

BO: addr=0xffffeffeffe00000-0xffffeffeffffffff map=           (nil) size=   2048KB handle=00007 capture=1 vram_only=0 name=slab_parent

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x185eb) [0x7fe54a5cf5eb]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_print_backtrace+0x21c) [0x7fe54a5cfa9c]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x2c4a9) [0x7fe54a5e34a9]

/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc20a) [0x7fe546abc20a]

/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc275) [0x7fe546abc275]

/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc4c7) [0x7fe546abc4c7]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x847a7) [0x7fe546e847a7]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x16ecbb) [0x7fe546f6ecbb]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x16f9d3) [0x7fe546f6f9d3]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x837) [0x7fe54a5ed317]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1) [0x7fe549cbb061]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xf7) [0x7fe549cbd0f7]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x390) [0x7fe549cc4510]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(llama_decode+0xb) [0x7fe549cc5fcb]

llama-server() [0x54acc7]

llama-server() [0x5ca176]

llama-server() [0x4a3596]

/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(+0x2a27e) [0x7fe54683127e]

/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(__libc_start_main+0x89) [0x7fe546831339]

llama-server() [0x4aa665]

terminate called after throwing an instance of 'vk::DeviceLostError'

what():  vk::Queue::submit: ErrorDeviceLost

zsh: IOT instruction (core dumped)  llama-server -m  -np 1 --jinja -c 32768 --chat-template-kwargs  --host  --por

llama-cpp wihout reasoning (--reasoning off):

Details

llama-server -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -np 1 --jinja -c 32768 --reasoning off --host 127.0.0.1 --port 8088
build_info: b0-unknown
system_info: n_threads = 3 (n_threads_batch = 3) / 16 | CPU : LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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 12526 MiB of device memory vs. 21316 MiB of free device memory
llama_params_fit_impl: will leave 8790 >= 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 0.84 seconds
llama_model_load_from_file_impl: using device Vulkan0 (Intel(R) Graphics (ARL)) (0000:00:02.0) - 21317 MiB free
llama_model_loader: loaded meta data with 52 key-value pairs and 733 tensors from /home/user/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.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              = qwen35moe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 20
llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000
llama_model_loader: - kv   5:                               general.name str              = Qwen3.5-35B-A3B
llama_model_loader: - kv   6:                           general.basename str              = Qwen3.5-35B-A3B
llama_model_loader: - kv   7:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   8:                         general.size_label str              = 35B-A3B
llama_model_loader: - kv   9:                            general.license str              = apache-2.0
llama_model_loader: - kv  10:                       general.license.link str              = https://huggingface.co/Qwen/Qwen3.5-3...
llama_model_loader: - kv  11:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  12:                   general.base_model.count u32              = 1
llama_model_loader: - kv  13:                  general.base_model.0.name str              = Qwen3.5 35B A3B
llama_model_loader: - kv  14:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  15:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen3.5-3...
llama_model_loader: - kv  16:                               general.tags arr[str,2]       = ["unsloth", "image-text-to-text"]
llama_model_loader: - kv  17:                      qwen35moe.block_count u32              = 40
llama_model_loader: - kv  18:                   qwen35moe.context_length u32              = 262144
llama_model_loader: - kv  19:                 qwen35moe.embedding_length u32              = 2048
llama_model_loader: - kv  20:             qwen35moe.attention.head_count u32              = 16
llama_model_loader: - kv  21:          qwen35moe.attention.head_count_kv u32              = 2
llama_model_loader: - kv  22:          qwen35moe.rope.dimension_sections arr[i32,4]       = [11, 11, 10, 0]
llama_model_loader: - kv  23:                   qwen35moe.rope.freq_base f32              = 10000000.000000
llama_model_loader: - kv  24: qwen35moe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  25:                     qwen35moe.expert_count u32              = 256
llama_model_loader: - kv  26:                qwen35moe.expert_used_count u32              = 8
llama_model_loader: - kv  27:             qwen35moe.attention.key_length u32              = 256
llama_model_loader: - kv  28:           qwen35moe.attention.value_length u32              = 256
llama_model_loader: - kv  29:       qwen35moe.expert_feed_forward_length u32              = 512
llama_model_loader: - kv  30: qwen35moe.expert_shared_feed_forward_length u32              = 512
llama_model_loader: - kv  31:                  qwen35moe.ssm.conv_kernel u32              = 4
llama_model_loader: - kv  32:                   qwen35moe.ssm.state_size u32              = 128
llama_model_loader: - kv  33:                  qwen35moe.ssm.group_count u32              = 16
llama_model_loader: - kv  34:               qwen35moe.ssm.time_step_rank u32              = 32
llama_model_loader: - kv  35:                   qwen35moe.ssm.inner_size u32              = 4096
llama_model_loader: - kv  36:          qwen35moe.full_attention_interval u32              = 4
llama_model_loader: - kv  37:             qwen35moe.rope.dimension_count u32              = 64
llama_model_loader: - kv  38:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  39:                         tokenizer.ggml.pre str              = qwen35
llama_model_loader: - kv  40:                      tokenizer.ggml.tokens arr[str,248320]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  41:                  tokenizer.ggml.token_type arr[i32,248320]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  42:                      tokenizer.ggml.merges arr[str,247587]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  43:                tokenizer.ggml.eos_token_id u32              = 248046
llama_model_loader: - kv  44:            tokenizer.ggml.padding_token_id u32              = 248055
llama_model_loader: - kv  45:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
llama_model_loader: - kv  46:               general.quantization_version u32              = 2
llama_model_loader: - kv  47:                          general.file_type u32              = 10
llama_model_loader: - kv  48:                      quantize.imatrix.file str              = Qwen3.5-35B-A3B-GGUF/imatrix_unsloth....
llama_model_loader: - kv  49:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3.5-35B-A3B.txt
llama_model_loader: - kv  50:             quantize.imatrix.entries_count u32              = 510
llama_model_loader: - kv  51:              quantize.imatrix.chunks_count u32              = 76
llama_model_loader: - type  f32:  301 tensors
llama_model_loader: - type q8_0:   61 tensors
llama_model_loader: - type q4_K:    1 tensors
llama_model_loader: - type q5_K:  177 tensors
llama_model_loader: - type q6_K:   73 tensors
llama_model_loader: - type iq2_xs:   78 tensors
llama_model_loader: - type iq3_xxs:   41 tensors
llama_model_loader: - type iq4_xs:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 11.31 GiB (2.80 BPW)
load: 0 unused tokens
load: printing all EOG tokens:
load:   - 248044 ('<|endoftext|>')
load:   - 248046 ('<|im_end|>')
load:   - 248063 ('<|fim_pad|>')
load:   - 248064 ('<|repo_name|>')
load:   - 248065 ('<|file_sep|>')
load: special tokens cache size = 33
load: token to piece cache size = 1.7581 MB
print_info: arch                  = qwen35moe
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 2048
print_info: n_embd_inp            = 2048
print_info: n_layer               = 40
print_info: n_head                = 16
print_info: n_head_kv             = 2
print_info: n_rot                 = 64
print_info: n_swa                 = 0
print_info: is_swa_any            = 0
print_info: n_embd_head_k         = 256
print_info: n_embd_head_v         = 256
print_info: n_gqa                 = 8
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                  = 0
print_info: n_expert              = 256
print_info: n_expert_used         = 8
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             = 40
print_info: rope scaling          = linear
print_info: freq_base_train       = 10000000.0
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        = [11, 11, 10, 0]
print_info: ssm_d_conv            = 4
print_info: ssm_d_inner           = 4096
print_info: ssm_d_state           = 128
print_info: ssm_dt_rank           = 32
print_info: ssm_n_group           = 16
print_info: ssm_dt_b_c_rms        = 0
print_info: model type            = 35B.A3B
print_info: model params          = 34.66 B
print_info: general.name          = Qwen3.5-35B-A3B
print_info: vocab type            = BPE
print_info: n_vocab               = 248320
print_info: n_merges              = 247587
print_info: BOS token             = 11 ','
print_info: EOS token             = 248046 '<|im_end|>'
print_info: EOT token             = 248046 '<|im_end|>'
print_info: PAD token             = 248055 '<|vision_pad|>'
print_info: LF token              = 198 'Ċ'
print_info: FIM PRE token         = 248060 '<|fim_prefix|>'
print_info: FIM SUF token         = 248062 '<|fim_suffix|>'
print_info: FIM MID token         = 248061 '<|fim_middle|>'
print_info: FIM PAD token         = 248063 '<|fim_pad|>'
print_info: FIM REP token         = 248064 '<|repo_name|>'
print_info: FIM SEP token         = 248065 '<|file_sep|>'
print_info: EOG token             = 248044 '<|endoftext|>'
print_info: EOG token             = 248046 '<|im_end|>'
print_info: EOG token             = 248063 '<|fim_pad|>'
print_info: EOG token             = 248064 '<|repo_name|>'
print_info: EOG token             = 248065 '<|file_sep|>'
print_info: max token length      = 256
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 39 repeating layers to GPU
load_tensors: offloaded 41/41 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   333.44 MiB
load_tensors:      Vulkan0 model buffer size = 11249.67 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         = 32768
llama_context: n_ctx_seq     = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 10000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (32768) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     0.95 MiB
llama_kv_cache:    Vulkan0 KV buffer size =   640.00 MiB
llama_kv_cache: size =  640.00 MiB ( 32768 cells,  10 layers,  1/1 seqs), K (f16):  320.00 MiB, V (f16):  320.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
llama_memory_recurrent:    Vulkan0 RS buffer size =    62.81 MiB
llama_memory_recurrent: size =   62.81 MiB (     1 cells,  40 layers,  1 seqs), R (f32):    2.81 MiB, S (f32):   60.00 MiB
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:    Vulkan0 compute buffer size =   574.02 MiB
sched_reserve: Vulkan_Host compute buffer size =    72.03 MiB
sched_reserve: graph nodes  = 3729
sched_reserve: graph splits = 2
sched_reserve: reserve took 32.51 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv    load_model: initializing slots, n_slots = 1
common_speculative_is_compat: the target context does not support partial sequence removal
srv    load_model: speculative decoding not supported by this context
slot   load_model: id  0 | task -1 | new slot, n_ctx = 32768
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: init: --clear-idle requires --kv-unified, disabling
init: chat template, 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


'

srv          init: init: chat template, thinking = 0

main: model loaded

main: server is listening on http://127.0.0.1:8088

main: starting the main loop...

srv  update_slots: all slots are idle

srv  log_server_r: done request: GET / 127.0.0.1 200

srv  log_server_r: done request: GET /bundle.css 127.0.0.1 200

srv  log_server_r: done request: GET /bundle.js 127.0.0.1 200

srv  log_server_r: done request: HEAD /cors-proxy 127.0.0.1 404

srv  params_from_: Chat format: peg-native

slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1

srv  get_availabl: updating prompt cache

srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000

srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 32768 tokens, 8589934592 est)

srv  get_availabl: prompt cache update took 0.01 ms

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, is_child = 0

slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, task.n_tokens = 14

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 = 10, batch.n_tokens = 10, progress = 0.714286

srv  log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200

Batch offset=0x2c000 len=0x0 on queue 0 (aperture: 13214.9Mb, 258.0Mb VRAM only)

BO: addr=0x00000005f8b00000-0x00000005f8b41fff map=           (nil) size=    264KB handle=00034 capture=0 vram_only=0 name=user

BO: addr=0x00000001bfe00000-0x00000001bfffffff map=           (nil) size=   2048KB handle=00030 capture=1 vram_only=0 name=slab_parent

BO: addr=0x00000000c0000000-0x00000000c01fffff map=  0x7fd75f1ff000 size=   2048KB handle=00005 capture=1 vram_only=0 name=binding table pool

BO: addr=0xffffeffeff200000-0xffffeffeff3fffff map=           (nil) size=   2048KB handle=00031 capture=0 vram_only=1 name=scratch

BO: addr=0xffffeffeeec00000-0xffffeffefebfffff map=           (nil) size= 262144KB handle=00037 capture=0 vram_only=1 name=scratch

BO: addr=0xffffeffefec00000-0xffffeffefedfffff map=           (nil) size=   2048KB handle=00033 capture=1 vram_only=0 name=slab_parent

BO: addr=0x0000000100000000-0x00000001001fffff map=  0x7fd75f5ff000 size=   2048KB handle=00003 capture=1 vram_only=0 name=scratch surface state pool

BO: addr=0x0000000100800000-0x00000001009fffff map=  0x7fd75f3ff000 size=   2048KB handle=00004 capture=1 vram_only=0 name=internal surface state pool

BO: addr=0x00000002c0000000-0x00000002c01fffff map=  0x7fd7641fd000 size=   2048KB handle=00002 capture=1 vram_only=0 name=dynamic pool

BO: addr=0x0000000000200000-0x00000000003fffff map=  0x7fd779f8b000 size=   2048KB handle=00001 capture=1 vram_only=0 name=general pool

BO: addr=0x0000000200000000-0x00000002001fffff map=  0x7fd75efff000 size=   2048KB handle=00012 capture=1 vram_only=0 name=shaders

BO: addr=0x0000000200200000-0x00000002003fffff map=  0x7fd75e5fe000 size=   2048KB handle=00036 capture=1 vram_only=0 name=shaders

BO: addr=0x0000000200400000-0x00000002007fffff map=  0x7fd75e1fe000 size=   4096KB handle=00038 capture=1 vram_only=0 name=shaders

BO: addr=0x0000000200800000-0x0000000200ffffff map=  0x7fd75cd5a000 size=   8192KB handle=00039 capture=1 vram_only=0 name=shaders

BO: addr=0x00000003c0000000-0x00000003c01fffff map=  0x7fd75dd5b000 size=   2048KB handle=00006 capture=1 vram_only=0 name=aux-tt pool

BO: addr=0x0000000440000000-0x000000047fd36fff map=           (nil) size=1045724KB handle=00010 capture=0 vram_only=0 name=user

BO: addr=0x000000047fe00000-0x00000004bf8aefff map=           (nil) size=1043132KB handle=00011 capture=0 vram_only=0 name=user

BO: addr=0x00000004bfa00000-0x00000004fdfc9fff map=           (nil) size=1021736KB handle=00013 capture=0 vram_only=0 name=user

BO: addr=0x00000004fe000000-0x000000053cb96fff map=           (nil) size=1027676KB handle=00014 capture=0 vram_only=0 name=user

BO: addr=0x000000053cc00000-0x000000057ae96fff map=           (nil) size=1018460KB handle=00015 capture=0 vram_only=0 name=user

BO: addr=0x000000057b000000-0x00000005b8f08fff map=           (nil) size=1014820KB handle=00016 capture=0 vram_only=0 name=user

BO: addr=0x00000005b9000000-0x00000005f8aaefff map=           (nil) size=1043132KB handle=00017 capture=0 vram_only=0 name=user

BO: addr=0x00000005f8c00000-0x0000000636e96fff map=           (nil) size=1018460KB handle=00018 capture=0 vram_only=0 name=user

BO: addr=0x0000000637000000-0x0000000674f08fff map=           (nil) size=1014820KB handle=00019 capture=0 vram_only=0 name=user

BO: addr=0x0000000675000000-0x00000006b4aaefff map=           (nil) size=1043132KB handle=00020 capture=0 vram_only=0 name=user

BO: addr=0x00000006b4c00000-0x00000006f2e96fff map=           (nil) size=1018460KB handle=00021 capture=0 vram_only=0 name=user

BO: addr=0x00000006f3000000-0x000000070292cfff map=           (nil) size= 255156KB handle=00022 capture=0 vram_only=0 name=user

BO: addr=0x0000000702a00000-0x0000000713be0fff map=           (nil) size= 280452KB handle=00023 capture=0 vram_only=0 name=user

BO: addr=0x00000004bf900000-0x00000004bf9f3fff map=           (nil) size=    976KB handle=00024 capture=0 vram_only=0 name=user

BO: addr=0x0000000713c00000-0x000000073be7ffff map=           (nil) size= 657920KB handle=00025 capture=0 vram_only=0 name=user

BO: addr=0x000000073c000000-0x000000073ff0efff map=           (nil) size=  64572KB handle=00026 capture=0 vram_only=0 name=user

BO: addr=0x0000000740000000-0x0000000764043fff map=           (nil) size= 590096KB handle=00027 capture=0 vram_only=0 name=user

BO: addr=0x0000000764200000-0x0000000768a50fff map=           (nil) size=  74052KB handle=00028 capture=0 vram_only=0 name=user

BO: addr=0x000000057af00000-0x000000057af01fff map=           (nil) size=      8KB handle=00029 capture=0 vram_only=0 name=user

BO: addr=0xffffeffeffe00000-0xffffeffeffffffff map=           (nil) size=   2048KB handle=00007 capture=1 vram_only=0 name=slab_parent

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x185eb) [0x7fd77f16b5eb]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_print_backtrace+0x21c) [0x7fd77f16ba9c]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x2c4a9) [0x7fd77f17f4a9]

/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc20a) [0x7fd77bcbc20a]

/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc275) [0x7fd77bcbc275]

/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc4c7) [0x7fd77bcbc4c7]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x847a7) [0x7fd77c0847a7]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x16ecbb) [0x7fd77c16ecbb]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x16f9d3) [0x7fd77c16f9d3]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x837) [0x7fd77f189317]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1) [0x7fd77eebb061]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xf7) [0x7fd77eebd0f7]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x390) [0x7fd77eec4510]

/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(llama_decode+0xb) [0x7fd77eec5fcb]

llama-server() [0x54acc7]

llama-server() [0x5ca176]

llama-server() [0x4a3596]

/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(+0x2a27e) [0x7fd77ba3127e]

/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(__libc_start_main+0x89) [0x7fd77ba31339]

llama-server() [0x4aa665]

terminate called after throwing an instance of 'vk::DeviceLostError'

what():  vk::Queue::submit: ErrorDeviceLost

zsh: IOT instruction (core dumped)  llama-server -m  -np 1 --jinja -c 32768 --reasoning off --host 127.0.0.1  808

[Zirconium419122]

Probably related to: #19327

[Zirconium419122]

Gemma-4-26B-4B-it fails with and without reasoning.

llama-bench -m ~/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.gguf -p 0 -n 128,256,512
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 49152 | int dot: 0 | matrix cores: none
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| gemma4 ?B Q3_K - Medium        |  11.98 GiB |    25.23 B | Vulkan     |  99 |           tg128 |          5.96 ± 0.02 |
| gemma4 ?B Q3_K - Medium        |  11.98 GiB |    25.23 B | Vulkan     |  99 |           tg256 |          6.02 ± 0.07 |
| gemma4 ?B Q3_K - Medium        |  11.98 GiB |    25.23 B | Vulkan     |  99 |           tg512 |          6.04 ± 0.04 |

build: unknown (0)

llama-server with reasoning:

Details

llama-server -m ~/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.gguf -np 1 --jinja -c 32768 --chat-template-kwargs '{"enable_thinking":true}' --host 127.0.0.1 --port 8088
Setting 'enable_thinking' via --chat-template-kwargs is deprecated. Use --reasoning on / --reasoning off instead.
build_info: b0-unknown
system_info: n_threads = 3 (n_threads_batch = 3) / 16 | CPU : LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/user/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.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 13809 MiB of device memory vs. 21316 MiB of free device memory
llama_params_fit_impl: will leave 7507 >= 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 1.34 seconds
llama_model_load_from_file_impl: using device Vulkan0 (Intel(R) Graphics (ARL)) (0000:00:02.0) - 21317 MiB free
llama_model_loader: loaded meta data with 60 key-value pairs and 658 tensors from /home/user/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.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              = gemma4
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 64
llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000
llama_model_loader: - kv   5:                               general.name str              = Gemma-4-26B-A4B-It
llama_model_loader: - kv   6:                           general.finetune str              = it
llama_model_loader: - kv   7:                           general.basename str              = Gemma-4-26B-A4B-It
llama_model_loader: - kv   8:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   9:                         general.size_label str              = 26B-A4B
llama_model_loader: - kv  10:                            general.license str              = apache-2.0
llama_model_loader: - kv  11:                       general.license.link str              = https://ai.google.dev/gemma/docs/gemm...
llama_model_loader: - kv  12:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  13:                   general.base_model.count u32              = 1
llama_model_loader: - kv  14:                  general.base_model.0.name str              = Gemma 4 26B A4B It
llama_model_loader: - kv  15:          general.base_model.0.organization str              = Google
llama_model_loader: - kv  16:              general.base_model.0.repo_url str              = https://huggingface.co/google/gemma-4...
llama_model_loader: - kv  17:                               general.tags arr[str,2]       = ["unsloth", "image-text-to-text"]
llama_model_loader: - kv  18:                         gemma4.block_count u32              = 30
llama_model_loader: - kv  19:                      gemma4.context_length u32              = 262144
llama_model_loader: - kv  20:                    gemma4.embedding_length u32              = 2816
llama_model_loader: - kv  21:                 gemma4.feed_forward_length u32              = 2112
llama_model_loader: - kv  22:                gemma4.attention.head_count u32              = 16
llama_model_loader: - kv  23:             gemma4.attention.head_count_kv arr[i32,30]      = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, ...
llama_model_loader: - kv  24:                      gemma4.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  25:                  gemma4.rope.freq_base_swa f32              = 10000.000000
llama_model_loader: - kv  26:    gemma4.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  27:                        gemma4.expert_count u32              = 128
llama_model_loader: - kv  28:                   gemma4.expert_used_count u32              = 8
llama_model_loader: - kv  29:                gemma4.attention.key_length u32              = 512
llama_model_loader: - kv  30:              gemma4.attention.value_length u32              = 512
llama_model_loader: - kv  31:             gemma4.final_logit_softcapping f32              = 30.000000
llama_model_loader: - kv  32:            gemma4.attention.sliding_window u32              = 1024
llama_model_loader: - kv  33:          gemma4.attention.shared_kv_layers u32              = 0
llama_model_loader: - kv  34:    gemma4.embedding_length_per_layer_input u32              = 0
llama_model_loader: - kv  35:    gemma4.attention.sliding_window_pattern arr[bool,30]     = [true, true, true, true, true, false,...
llama_model_loader: - kv  36:            gemma4.attention.key_length_swa u32              = 256
llama_model_loader: - kv  37:          gemma4.attention.value_length_swa u32              = 256
llama_model_loader: - kv  38:          gemma4.expert_feed_forward_length u32              = 704
llama_model_loader: - kv  39:                gemma4.rope.dimension_count u32              = 512
llama_model_loader: - kv  40:            gemma4.rope.dimension_count_swa u32              = 256
llama_model_loader: - kv  41:                       tokenizer.ggml.model str              = gemma4
llama_model_loader: - kv  42:                      tokenizer.ggml.tokens arr[str,262144]  = ["", "", "", "", ...
llama_model_loader: - kv  43:                      tokenizer.ggml.scores arr[f32,262144]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  44:                  tokenizer.ggml.token_type arr[i32,262144]  = [3, 1, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  45:                      tokenizer.ggml.merges arr[str,514906]  = ["\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n \n", ...
llama_model_loader: - kv  46:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  47:                tokenizer.ggml.eos_token_id u32              = 106
llama_model_loader: - kv  48:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  49:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  50:               tokenizer.ggml.mask_token_id u32              = 4
llama_model_loader: - kv  51:                    tokenizer.chat_template str              = {%- macro format_parameters(propertie...
llama_model_loader: - kv  52:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  53:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  54:               general.quantization_version u32              = 2
llama_model_loader: - kv  55:                          general.file_type u32              = 12
llama_model_loader: - kv  56:                      quantize.imatrix.file str              = gemma-4-26B-A4B-it-GGUF/imatrix_unslo...
llama_model_loader: - kv  57:                   quantize.imatrix.dataset str              = unsloth_calibration_gemma-4-26B-A4B-i...
llama_model_loader: - kv  58:             quantize.imatrix.entries_count u32              = 295
llama_model_loader: - kv  59:              quantize.imatrix.chunks_count u32              = 141
llama_model_loader: - type  f32:  392 tensors
llama_model_loader: - type q5_1:    2 tensors
llama_model_loader: - type q8_0:  206 tensors
llama_model_loader: - type iq3_xxs:   29 tensors
llama_model_loader: - type iq4_nl:   28 tensors
llama_model_loader: - type iq4_xs:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q3_K - Medium
print_info: file size   = 11.98 GiB (4.08 BPW)
load: 0 unused tokens
load: control-looking token:    212 '' was not control-type; this is probably a bug in the model. its type will be overridden
load: control-looking token:     50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
load: printing all EOG tokens:
load:   - 50 ('<|tool_response>')
load:   - 106 ('')
load:   - 212 ('')
load: special tokens cache size = 24
load: token to piece cache size = 1.9445 MB
print_info: arch                  = gemma4
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 2816
print_info: n_embd_inp            = 2816
print_info: n_layer               = 30
print_info: n_head                = 16
print_info: n_head_kv             = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2]
print_info: n_rot                 = 512
print_info: n_swa                 = 1024
print_info: is_swa_any            = 1
print_info: n_embd_head_k         = 512
print_info: n_embd_head_v         = 512
print_info: n_gqa                 = [2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8]
print_info: n_embd_k_gqa          = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024]
print_info: n_embd_v_gqa          = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 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          = 1.0e+00
print_info: n_ff                  = 2112
print_info: n_expert              = 128
print_info: n_expert_used         = 8
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: freq_base_swa         = 10000.0
print_info: freq_scale_swa        = 1
print_info: n_embd_head_k_swa     = 256
print_info: n_embd_head_v_swa     = 256
print_info: n_rot_swa             = 256
print_info: n_ctx_orig_yarn       = 262144
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: model type            = ?B
print_info: model params          = 25.23 B
print_info: general.name          = Gemma-4-26B-A4B-It
print_info: vocab type            = BPE
print_info: n_vocab               = 262144
print_info: n_merges              = 514906
print_info: BOS token             = 2 ''
print_info: EOS token             = 106 ''
print_info: UNK token             = 3 ''
print_info: PAD token             = 0 ''
print_info: MASK token            = 4 ''
print_info: LF token              = 107 '
'
print_info: EOG token             = 50 '<|tool_response>'
print_info: EOG token             = 106 ''
print_info: EOG token             = 212 ''
print_info: max token length      = 93
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 29 repeating layers to GPU
load_tensors: offloaded 31/31 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   748.00 MiB
load_tensors:      Vulkan0 model buffer size = 12264.00 MiB
.............................................................................
common_init_result: added <|tool_response> logit bias = -inf
common_init_result: added  logit bias = -inf
common_init_result: added  logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 32768
llama_context: n_ctx_seq     = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (32768) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     1.00 MiB
llama_kv_cache_iswa: creating non-SWA KV cache, size = 32768 cells
llama_kv_cache:    Vulkan0 KV buffer size =   640.00 MiB
llama_kv_cache: size =  640.00 MiB ( 32768 cells,   5 layers,  1/1 seqs), K (f16):  320.00 MiB, V (f16):  320.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 512
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 512
llama_kv_cache_iswa: creating     SWA KV cache, size = 1536 cells
llama_kv_cache:    Vulkan0 KV buffer size =   300.00 MiB
llama_kv_cache: size =  300.00 MiB (  1536 cells,  25 layers,  1/1 seqs), K (f16):  150.00 MiB, V (f16):  150.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:    Vulkan0 compute buffer size =   605.50 MiB
sched_reserve: Vulkan_Host compute buffer size =    78.03 MiB
sched_reserve: graph nodes  = 2647
sched_reserve: graph splits = 2
sched_reserve: reserve took 14.63 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv    load_model: initializing slots, n_slots = 1
no implementations specified for speculative decoding
slot   load_model: id  0 | task -1 | speculative decoding context not initialized
slot   load_model: id  0 | task -1 | new slot, n_ctx = 32768
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: init: --clear-idle requires --kv-unified, disabling
init: chat template, example_format: '<|turn>system
<|think|>You are a helpful assistant
<|turn>user
Hello
<|turn>model
Hi there
<|turn>user
How are you?
<|turn>model
'
srv          init: init: chat template, thinking = 1
main: model loaded
main: server is listening on http://127.0.0.1:8088
main: starting the main loop...
srv  update_slots: all slots are idle
srv  log_server_r: done request: GET / 127.0.0.1 200
srv  log_server_r: done request: GET /bundle.css 127.0.0.1 200
srv  log_server_r: done request: GET /bundle.js 127.0.0.1 200
srv  log_server_r: done request: HEAD /cors-proxy 127.0.0.1 404
srv  params_from_: Chat format: peg-gemma4
slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
srv  get_availabl: updating prompt cache
srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 32768 tokens, 8589934592 est)
srv  get_availabl: prompt cache update took 0.01 ms
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, is_child = 0
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, task.n_tokens = 17
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 = 13, batch.n_tokens = 13, progress = 0.764706
srv  log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200
slot update_slots: id  0 | task 0 | n_tokens = 13, memory_seq_rm [13, end)
slot init_sampler: id  0 | task 0 | init sampler, took 0.01 ms, tokens: text = 17, total = 17
slot update_slots: id  0 | task 0 | prompt processing done, n_tokens = 17, batch.n_tokens = 4
Batch offset=0x16000 len=0x0 on queue 0 (aperture: 14740.3Mb, 0.0Mb VRAM only)
   BO: addr=0x00000004fc500000-0x00000004fc581fff map=           (nil) size=    520KB handle=00032 capture=0 vram_only=0 name=user
   BO: addr=0x00000007aca00000-0x00000007d29ddfff map=           (nil) size= 622456KB handle=00028 capture=0 vram_only=0 name=user
   BO: addr=0x00000007d2a00000-0x00000007d7856fff map=           (nil) size=  80220KB handle=00029 capture=0 vram_only=0 name=user
   BO: addr=0x00000002c0000000-0x00000002c01fffff map=  0x7fb27db70000 size=   2048KB handle=00002 capture=1 vram_only=0 name=dynamic pool
   BO: addr=0x00000000c0000000-0x00000000c01fffff map=  0x7fb2631ff000 size=   2048KB handle=00005 capture=1 vram_only=0 name=binding table pool
   BO: addr=0x0000000100000000-0x00000001001fffff map=  0x7fb2635ff000 size=   2048KB handle=00003 capture=1 vram_only=0 name=scratch surface state pool
   BO: addr=0x0000000100800000-0x00000001009fffff map=  0x7fb2633ff000 size=   2048KB handle=00004 capture=1 vram_only=0 name=internal surface state pool
   BO: addr=0x0000000000200000-0x00000000003fffff map=  0x7fb27dd70000 size=   2048KB handle=00001 capture=1 vram_only=0 name=general pool
   BO: addr=0x0000000200000000-0x00000002001fffff map=  0x7fb262dff000 size=   2048KB handle=00024 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000200200000-0x00000002003fffff map=  0x7fb262bff000 size=   2048KB handle=00035 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000200400000-0x00000002007fffff map=  0x7fb2627ff000 size=   4096KB handle=00038 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000200800000-0x0000000200ffffff map=  0x7fb260ffb000 size=   8192KB handle=00040 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000201000000-0x0000000201ffffff map=  0x7fb23f000000 size=  16384KB handle=00041 capture=1 vram_only=0 name=shaders
   BO: addr=0x00000003c0000000-0x00000003c01fffff map=  0x7fb262fff000 size=   2048KB handle=00006 capture=1 vram_only=0 name=aux-tt pool
   BO: addr=0x0000000440000000-0x000000047dfc1fff map=           (nil) size=1015560KB handle=00010 capture=0 vram_only=0 name=user
   BO: addr=0x000000047e000000-0x00000004be2e7fff map=           (nil) size=1051552KB handle=00011 capture=0 vram_only=0 name=user
   BO: addr=0x00000004be400000-0x00000004fc4d7fff map=           (nil) size=1016672KB handle=00012 capture=0 vram_only=0 name=user
   BO: addr=0x00000004fc600000-0x000000053786ffff map=           (nil) size= 969152KB handle=00013 capture=0 vram_only=0 name=user
   BO: addr=0x0000000537a00000-0x00000005726f7fff map=           (nil) size= 963552KB handle=00014 capture=0 vram_only=0 name=user
   BO: addr=0x0000000572800000-0x00000005ae91afff map=           (nil) size= 984172KB handle=00015 capture=0 vram_only=0 name=user
   BO: addr=0x00000005aea00000-0x00000005e96f7fff map=           (nil) size= 963552KB handle=00016 capture=0 vram_only=0 name=user
   BO: addr=0x00000005e9800000-0x000000062591afff map=           (nil) size= 984172KB handle=00017 capture=0 vram_only=0 name=user
   BO: addr=0x0000000625a00000-0x00000006606f7fff map=           (nil) size= 963552KB handle=00018 capture=0 vram_only=0 name=user
   BO: addr=0x0000000660800000-0x000000069c91afff map=           (nil) size= 984172KB handle=00019 capture=0 vram_only=0 name=user
   BO: addr=0x000000069ca00000-0x00000006d76f7fff map=           (nil) size= 963552KB handle=00020 capture=0 vram_only=0 name=user
   BO: addr=0x00000006d7800000-0x0000000712a6ffff map=           (nil) size= 969152KB handle=00021 capture=0 vram_only=0 name=user
   BO: addr=0x0000000712c00000-0x0000000742461fff map=           (nil) size= 778632KB handle=00022 capture=0 vram_only=0 name=user
   BO: addr=0x0000000742600000-0x00000007714ebfff map=           (nil) size= 768944KB handle=00023 capture=0 vram_only=0 name=user
   BO: addr=0x0000000771600000-0x0000000771702fff map=           (nil) size=   1036KB handle=00025 capture=0 vram_only=0 name=user
   BO: addr=0x0000000771800000-0x0000000799a7ffff map=           (nil) size= 657920KB handle=00026 capture=0 vram_only=0 name=user
   BO: addr=0x0000000799c00000-0x00000007ac92bfff map=           (nil) size= 308400KB handle=00027 capture=0 vram_only=0 name=user
   BO: addr=0x00000004be300000-0x00000004be301fff map=           (nil) size=      8KB handle=00030 capture=0 vram_only=0 name=user
   BO: addr=0xffffeffeffe00000-0xffffeffeffffffff map=           (nil) size=   2048KB handle=00007 capture=1 vram_only=0 name=slab_parent
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x185eb) [0x7fb2835d85eb]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_print_backtrace+0x21c) [0x7fb2835d8a9c]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x2c4a9) [0x7fb2835ec4a9]
/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc20a) [0x7fb27fabc20a]
/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc275) [0x7fb27fabc275]
/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc4c7) [0x7fb27fabc4c7]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x847a7) [0x7fb27fe847a7]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x14b156) [0x7fb27ff4b156]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x14bb3f) [0x7fb27ff4bb3f]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x206) [0x7fb2835f5ce6]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1) [0x7fb282cbb061]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xf7) [0x7fb282cbd0f7]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x390) [0x7fb282cc4510]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(llama_decode+0xb) [0x7fb282cc5fcb]
llama-server() [0x54acc7]
llama-server() [0x5ca176]
llama-server() [0x4a3596]
/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(+0x2a27e) [0x7fb27f83127e]
/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(__libc_start_main+0x89) [0x7fb27f831339]
llama-server() [0x4aa665]
terminate called after throwing an instance of 'vk::DeviceLostError'
  what():  vk::Queue::submit: ErrorDeviceLost
zsh: IOT instruction (core dumped)  llama-server -m ~/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.gguf -np 1

llama-server without reasoning:

Details

llama-server -m ~/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.gguf -np 1 --jinja -c 32768 --chat-template-kwargs '{"enable_thinking":false}' --host 127.0.0.1 --port 8088
Setting 'enable_thinking' via --chat-template-kwargs is deprecated. Use --reasoning on / --reasoning off instead.
build_info: b0-unknown
system_info: n_threads = 3 (n_threads_batch = 3) / 16 | CPU : LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/user/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.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 13809 MiB of device memory vs. 21316 MiB of free device memory
llama_params_fit_impl: will leave 7507 >= 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 1.21 seconds
llama_model_load_from_file_impl: using device Vulkan0 (Intel(R) Graphics (ARL)) (0000:00:02.0) - 21317 MiB free
llama_model_loader: loaded meta data with 60 key-value pairs and 658 tensors from /home/user/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.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              = gemma4
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 64
llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000
llama_model_loader: - kv   5:                               general.name str              = Gemma-4-26B-A4B-It
llama_model_loader: - kv   6:                           general.finetune str              = it
llama_model_loader: - kv   7:                           general.basename str              = Gemma-4-26B-A4B-It
llama_model_loader: - kv   8:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   9:                         general.size_label str              = 26B-A4B
llama_model_loader: - kv  10:                            general.license str              = apache-2.0
llama_model_loader: - kv  11:                       general.license.link str              = https://ai.google.dev/gemma/docs/gemm...
llama_model_loader: - kv  12:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  13:                   general.base_model.count u32              = 1
llama_model_loader: - kv  14:                  general.base_model.0.name str              = Gemma 4 26B A4B It
llama_model_loader: - kv  15:          general.base_model.0.organization str              = Google
llama_model_loader: - kv  16:              general.base_model.0.repo_url str              = https://huggingface.co/google/gemma-4...
llama_model_loader: - kv  17:                               general.tags arr[str,2]       = ["unsloth", "image-text-to-text"]
llama_model_loader: - kv  18:                         gemma4.block_count u32              = 30
llama_model_loader: - kv  19:                      gemma4.context_length u32              = 262144
llama_model_loader: - kv  20:                    gemma4.embedding_length u32              = 2816
llama_model_loader: - kv  21:                 gemma4.feed_forward_length u32              = 2112
llama_model_loader: - kv  22:                gemma4.attention.head_count u32              = 16
llama_model_loader: - kv  23:             gemma4.attention.head_count_kv arr[i32,30]      = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, ...
llama_model_loader: - kv  24:                      gemma4.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  25:                  gemma4.rope.freq_base_swa f32              = 10000.000000
llama_model_loader: - kv  26:    gemma4.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  27:                        gemma4.expert_count u32              = 128
llama_model_loader: - kv  28:                   gemma4.expert_used_count u32              = 8
llama_model_loader: - kv  29:                gemma4.attention.key_length u32              = 512
llama_model_loader: - kv  30:              gemma4.attention.value_length u32              = 512
llama_model_loader: - kv  31:             gemma4.final_logit_softcapping f32              = 30.000000
llama_model_loader: - kv  32:            gemma4.attention.sliding_window u32              = 1024
llama_model_loader: - kv  33:          gemma4.attention.shared_kv_layers u32              = 0
llama_model_loader: - kv  34:    gemma4.embedding_length_per_layer_input u32              = 0
llama_model_loader: - kv  35:    gemma4.attention.sliding_window_pattern arr[bool,30]     = [true, true, true, true, true, false,...
llama_model_loader: - kv  36:            gemma4.attention.key_length_swa u32              = 256
llama_model_loader: - kv  37:          gemma4.attention.value_length_swa u32              = 256
llama_model_loader: - kv  38:          gemma4.expert_feed_forward_length u32              = 704
llama_model_loader: - kv  39:                gemma4.rope.dimension_count u32              = 512
llama_model_loader: - kv  40:            gemma4.rope.dimension_count_swa u32              = 256
llama_model_loader: - kv  41:                       tokenizer.ggml.model str              = gemma4
llama_model_loader: - kv  42:                      tokenizer.ggml.tokens arr[str,262144]  = ["", "", "", "", ...
llama_model_loader: - kv  43:                      tokenizer.ggml.scores arr[f32,262144]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  44:                  tokenizer.ggml.token_type arr[i32,262144]  = [3, 1, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  45:                      tokenizer.ggml.merges arr[str,514906]  = ["\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n \n", ...
llama_model_loader: - kv  46:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  47:                tokenizer.ggml.eos_token_id u32              = 106
llama_model_loader: - kv  48:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  49:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  50:               tokenizer.ggml.mask_token_id u32              = 4
llama_model_loader: - kv  51:                    tokenizer.chat_template str              = {%- macro format_parameters(propertie...
llama_model_loader: - kv  52:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  53:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  54:               general.quantization_version u32              = 2
llama_model_loader: - kv  55:                          general.file_type u32              = 12
llama_model_loader: - kv  56:                      quantize.imatrix.file str              = gemma-4-26B-A4B-it-GGUF/imatrix_unslo...
llama_model_loader: - kv  57:                   quantize.imatrix.dataset str              = unsloth_calibration_gemma-4-26B-A4B-i...
llama_model_loader: - kv  58:             quantize.imatrix.entries_count u32              = 295
llama_model_loader: - kv  59:              quantize.imatrix.chunks_count u32              = 141
llama_model_loader: - type  f32:  392 tensors
llama_model_loader: - type q5_1:    2 tensors
llama_model_loader: - type q8_0:  206 tensors
llama_model_loader: - type iq3_xxs:   29 tensors
llama_model_loader: - type iq4_nl:   28 tensors
llama_model_loader: - type iq4_xs:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q3_K - Medium
print_info: file size   = 11.98 GiB (4.08 BPW)
load: 0 unused tokens
load: control-looking token:    212 '' was not control-type; this is probably a bug in the model. its type will be overridden
load: control-looking token:     50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
load: printing all EOG tokens:
load:   - 50 ('<|tool_response>')
load:   - 106 ('')
load:   - 212 ('')
load: special tokens cache size = 24
load: token to piece cache size = 1.9445 MB
print_info: arch                  = gemma4
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 2816
print_info: n_embd_inp            = 2816
print_info: n_layer               = 30
print_info: n_head                = 16
print_info: n_head_kv             = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2]
print_info: n_rot                 = 512
print_info: n_swa                 = 1024
print_info: is_swa_any            = 1
print_info: n_embd_head_k         = 512
print_info: n_embd_head_v         = 512
print_info: n_gqa                 = [2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8]
print_info: n_embd_k_gqa          = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024]
print_info: n_embd_v_gqa          = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 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          = 1.0e+00
print_info: n_ff                  = 2112
print_info: n_expert              = 128
print_info: n_expert_used         = 8
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: freq_base_swa         = 10000.0
print_info: freq_scale_swa        = 1
print_info: n_embd_head_k_swa     = 256
print_info: n_embd_head_v_swa     = 256
print_info: n_rot_swa             = 256
print_info: n_ctx_orig_yarn       = 262144
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: model type            = ?B
print_info: model params          = 25.23 B
print_info: general.name          = Gemma-4-26B-A4B-It
print_info: vocab type            = BPE
print_info: n_vocab               = 262144
print_info: n_merges              = 514906
print_info: BOS token             = 2 ''
print_info: EOS token             = 106 ''
print_info: UNK token             = 3 ''
print_info: PAD token             = 0 ''
print_info: MASK token            = 4 ''
print_info: LF token              = 107 '
'
print_info: EOG token             = 50 '<|tool_response>'
print_info: EOG token             = 106 ''
print_info: EOG token             = 212 ''
print_info: max token length      = 93
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 29 repeating layers to GPU
load_tensors: offloaded 31/31 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   748.00 MiB
load_tensors:      Vulkan0 model buffer size = 12264.00 MiB
.............................................................................
common_init_result: added <|tool_response> logit bias = -inf
common_init_result: added  logit bias = -inf
common_init_result: added  logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 32768
llama_context: n_ctx_seq     = 32768
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (32768) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     1.00 MiB
llama_kv_cache_iswa: creating non-SWA KV cache, size = 32768 cells
llama_kv_cache:    Vulkan0 KV buffer size =   640.00 MiB
llama_kv_cache: size =  640.00 MiB ( 32768 cells,   5 layers,  1/1 seqs), K (f16):  320.00 MiB, V (f16):  320.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 512
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 512
llama_kv_cache_iswa: creating     SWA KV cache, size = 1536 cells
llama_kv_cache:    Vulkan0 KV buffer size =   300.00 MiB
llama_kv_cache: size =  300.00 MiB (  1536 cells,  25 layers,  1/1 seqs), K (f16):  150.00 MiB, V (f16):  150.00 MiB
llama_kv_cache: attn_rot_k = 0, n_embd_head_k_all = 256
llama_kv_cache: attn_rot_v = 0, n_embd_head_k_all = 256
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:    Vulkan0 compute buffer size =   605.50 MiB
sched_reserve: Vulkan_Host compute buffer size =    78.03 MiB
sched_reserve: graph nodes  = 2647
sched_reserve: graph splits = 2
sched_reserve: reserve took 19.63 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv    load_model: initializing slots, n_slots = 1
no implementations specified for speculative decoding
slot   load_model: id  0 | task -1 | speculative decoding context not initialized
slot   load_model: id  0 | task -1 | new slot, n_ctx = 32768
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: init: --clear-idle requires --kv-unified, disabling
init: chat template, example_format: '<|turn>system
You are a helpful assistant
<|turn>user
Hello
<|turn>model
Hi there
<|turn>user
How are you?
<|turn>model
<|channel>thought
'
srv          init: init: chat template, thinking = 1
main: model loaded
main: server is listening on http://127.0.0.1:8088
main: starting the main loop...
srv  update_slots: all slots are idle
srv  log_server_r: done request: GET / 127.0.0.1 200
srv  log_server_r: done request: GET /bundle.css 127.0.0.1 200
srv  log_server_r: done request: GET /bundle.js 127.0.0.1 200
srv  log_server_r: done request: HEAD /cors-proxy 127.0.0.1 404
srv  params_from_: Chat format: peg-gemma4
slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
srv  get_availabl: updating prompt cache
srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 32768 tokens, 8589934592 est)
srv  get_availabl: prompt cache update took 0.01 ms
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, is_child = 0
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 32768, n_keep = 0, task.n_tokens = 15
slot update_slots: id  0 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
srv  log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 11, batch.n_tokens = 11, progress = 0.733333
slot update_slots: id  0 | task 0 | n_tokens = 11, memory_seq_rm [11, end)
slot init_sampler: id  0 | task 0 | init sampler, took 0.01 ms, tokens: text = 15, total = 15
slot update_slots: id  0 | task 0 | prompt processing done, n_tokens = 15, batch.n_tokens = 4
Batch offset=0x16000 len=0x0 on queue 0 (aperture: 14740.3Mb, 0.0Mb VRAM only)
   BO: addr=0x00000004fc500000-0x00000004fc581fff map=           (nil) size=    520KB handle=00032 capture=0 vram_only=0 name=user
   BO: addr=0x00000007aca00000-0x00000007d29ddfff map=           (nil) size= 622456KB handle=00028 capture=0 vram_only=0 name=user
   BO: addr=0x00000007d2a00000-0x00000007d7856fff map=           (nil) size=  80220KB handle=00029 capture=0 vram_only=0 name=user
   BO: addr=0x00000002c0000000-0x00000002c01fffff map=  0x7f3b8a5fc000 size=   2048KB handle=00002 capture=1 vram_only=0 name=dynamic pool
   BO: addr=0x00000000c0000000-0x00000000c01fffff map=  0x7f3b89ffc000 size=   2048KB handle=00005 capture=1 vram_only=0 name=binding table pool
   BO: addr=0x0000000100000000-0x00000001001fffff map=  0x7f3b8a3fc000 size=   2048KB handle=00003 capture=1 vram_only=0 name=scratch surface state pool
   BO: addr=0x0000000100800000-0x00000001009fffff map=  0x7f3b8a1fc000 size=   2048KB handle=00004 capture=1 vram_only=0 name=internal surface state pool
   BO: addr=0x0000000000200000-0x00000000003fffff map=  0x7f3ba0b8b000 size=   2048KB handle=00001 capture=1 vram_only=0 name=general pool
   BO: addr=0x0000000200000000-0x00000002001fffff map=  0x7f3b89dfc000 size=   2048KB handle=00024 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000200200000-0x00000002003fffff map=  0x7f3b89bfc000 size=   2048KB handle=00035 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000200400000-0x00000002007fffff map=  0x7f3b897fc000 size=   4096KB handle=00038 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000200800000-0x0000000200ffffff map=  0x7f3b83800000 size=   8192KB handle=00040 capture=1 vram_only=0 name=shaders
   BO: addr=0x0000000201000000-0x0000000201ffffff map=  0x7f3b82800000 size=  16384KB handle=00041 capture=1 vram_only=0 name=shaders
   BO: addr=0x00000003c0000000-0x00000003c01fffff map=  0x7f3b895fa000 size=   2048KB handle=00006 capture=1 vram_only=0 name=aux-tt pool
   BO: addr=0x0000000440000000-0x000000047dfc1fff map=           (nil) size=1015560KB handle=00010 capture=0 vram_only=0 name=user
   BO: addr=0x000000047e000000-0x00000004be2e7fff map=           (nil) size=1051552KB handle=00011 capture=0 vram_only=0 name=user
   BO: addr=0x00000004be400000-0x00000004fc4d7fff map=           (nil) size=1016672KB handle=00012 capture=0 vram_only=0 name=user
   BO: addr=0x00000004fc600000-0x000000053786ffff map=           (nil) size= 969152KB handle=00013 capture=0 vram_only=0 name=user
   BO: addr=0x0000000537a00000-0x00000005726f7fff map=           (nil) size= 963552KB handle=00014 capture=0 vram_only=0 name=user
   BO: addr=0x0000000572800000-0x00000005ae91afff map=           (nil) size= 984172KB handle=00015 capture=0 vram_only=0 name=user
   BO: addr=0x00000005aea00000-0x00000005e96f7fff map=           (nil) size= 963552KB handle=00016 capture=0 vram_only=0 name=user
   BO: addr=0x00000005e9800000-0x000000062591afff map=           (nil) size= 984172KB handle=00017 capture=0 vram_only=0 name=user
   BO: addr=0x0000000625a00000-0x00000006606f7fff map=           (nil) size= 963552KB handle=00018 capture=0 vram_only=0 name=user
   BO: addr=0x0000000660800000-0x000000069c91afff map=           (nil) size= 984172KB handle=00019 capture=0 vram_only=0 name=user
   BO: addr=0x000000069ca00000-0x00000006d76f7fff map=           (nil) size= 963552KB handle=00020 capture=0 vram_only=0 name=user
   BO: addr=0x00000006d7800000-0x0000000712a6ffff map=           (nil) size= 969152KB handle=00021 capture=0 vram_only=0 name=user
   BO: addr=0x0000000712c00000-0x0000000742461fff map=           (nil) size= 778632KB handle=00022 capture=0 vram_only=0 name=user
   BO: addr=0x0000000742600000-0x00000007714ebfff map=           (nil) size= 768944KB handle=00023 capture=0 vram_only=0 name=user
   BO: addr=0x0000000771600000-0x0000000771702fff map=           (nil) size=   1036KB handle=00025 capture=0 vram_only=0 name=user
   BO: addr=0x0000000771800000-0x0000000799a7ffff map=           (nil) size= 657920KB handle=00026 capture=0 vram_only=0 name=user
   BO: addr=0x0000000799c00000-0x00000007ac92bfff map=           (nil) size= 308400KB handle=00027 capture=0 vram_only=0 name=user
   BO: addr=0x00000004be300000-0x00000004be301fff map=           (nil) size=      8KB handle=00030 capture=0 vram_only=0 name=user
   BO: addr=0xffffeffeffe00000-0xffffeffeffffffff map=           (nil) size=   2048KB handle=00007 capture=1 vram_only=0 name=slab_parent
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x185eb) [0x7f3ba5d6b5eb]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_print_backtrace+0x21c) [0x7f3ba5d6ba9c]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(+0x2c4a9) [0x7f3ba5d7f4a9]
/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc20a) [0x7f3ba28bc20a]
/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc275) [0x7f3ba28bc275]
/nix/store/97f3gw9vpyxvwjv2i673isvg92q65mwn-gcc-13.3.0-lib/lib/libstdc++.so.6(+0xbc4c7) [0x7f3ba28bc4c7]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x847a7) [0x7f3ba2c847a7]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x14b156) [0x7f3ba2d4b156]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-vulkan.so.0(+0x14bb3f) [0x7f3ba2d4bb3f]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x206) [0x7f3ba5d88ce6]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1) [0x7f3ba5abb061]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xf7) [0x7f3ba5abd0f7]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x390) [0x7f3ba5ac4510]
/nix/store/6imwj6d815cb67cmcfkmz0fdbdbcfxrh-llama-cpp-vulkan-0.0.0/lib/libllama.so.0(llama_decode+0xb) [0x7f3ba5ac5fcb]
llama-server() [0x54acc7]
llama-server() [0x5ca176]
llama-server() [0x4a3596]
/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(+0x2a27e) [0x7f3ba263127e]
/nix/store/pacbfvpzqz2mksby36awvbcn051zcji3-glibc-2.40-36/lib/libc.so.6(__libc_start_main+0x89) [0x7f3ba2631339]
llama-server() [0x4aa665]
terminate called after throwing an instance of 'vk::DeviceLostError'
  what():  vk::Queue::submit: ErrorDeviceLost
zsh: IOT instruction (core dumped)  llama-server -m ~/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.gguf -np 1

[Zirconium419122]

Updating nixpkgs input to the flake. Yields significantly better performance.

Gemma-4-26B-A4B-it

result/bin/llama-bench -m ~/.cache/llama.cpp/gemma-4-26B-A4B-it-UD-Q3_K_XL.gguf -p 0 -n 128,256,512
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| gemma4 ?B Q3_K - Medium        |  11.98 GiB |    25.23 B | Vulkan     |  99 |           tg128 |         10.79 ± 0.28 |
| gemma4 ?B Q3_K - Medium        |  11.98 GiB |    25.23 B | Vulkan     |  99 |           tg256 |         11.23 ± 0.08 |
| gemma4 ?B Q3_K - Medium        |  11.98 GiB |    25.23 B | Vulkan     |  99 |           tg512 |         10.50 ± 0.21 |

build: unknown (0)

Qwen3.5-35B-A3B

result/bin/llama-bench -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf -p 0 -n 128,256,512
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Intel(R) Graphics (ARL) (Intel open-source Mesa driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg128 |         11.30 ± 0.16 |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg256 |         11.24 ± 0.16 |
| qwen35moe 35B.A3B Q2_K - Medium |  11.31 GiB |    34.66 B | Vulkan     |  99 |           tg512 |         11.34 ± 0.30 |

build: unknown (0)

[Zirconium419122]

Reducing amount of layers on the GPU to near zero does work:

One layer of the GPU (--ngl 1):

Details

./result/bin/llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja -ngl 1 -c 32768 --reasoning off -p "Hello\!"
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.
Loading model...
▄▄ ▄▄

██ ██

██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄

██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██

██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀

██    ██

▀▀    ▀▀
build      : b8864-ff6b106

model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf

modalities : text
available commands:

/exit or Ctrl+C     stop or exit

/regen              regenerate the last response

/clear              clear the chat history

/read         add a text file

/glob      add text files using globbing pattern
> Hello!
Hello! How can I help you today?
[ Prompt: 1,7 t/s | Generation: 1,5 t/s ]
>
Exiting...

llama_memory_breakdown_print: | memory breakdown [MiB]                | total    free     self   model   context   compute    unaccounted |

llama_memory_breakdown_print: |   - Vulkan0 (Intel(R) Graphics (ARL)) | 23686 = 21132 + (  771 =   272 +       0 +     498) +        1783 |

llama_memory_breakdown_print: |   - Host                              |                  12097 = 11310 +     702 +      84                |

Two layers on the GPU (--ngl 2):

Details

./result/bin/llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja -ngl 2 -c 32768 --reasoning off -p "Hello\!"
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.
Loading model...
▄▄ ▄▄

██ ██

██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄

██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██

██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀

██    ██

▀▀    ▀▀
build      : b8864-ff6b106

model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf

modalities : text
available commands:

/exit or Ctrl+C     stop or exit

/regen              regenerate the last response

/clear              clear the chat history

/read         add a text file

/glob      add text files using globbing pattern
> Hello!
Hello! How can I help you today?
[ Prompt: 1,6 t/s | Generation: 1,4 t/s ]
>
Exiting...

llama_memory_breakdown_print: | memory breakdown [MiB]                | total    free     self   model   context   compute    unaccounted |

llama_memory_breakdown_print: |   - Vulkan0 (Intel(R) Graphics (ARL)) | 23686 = 18774 + ( 1102 =   540 +      64 +     498) +        3808 |

llama_memory_breakdown_print: |   - Host                              |                  11765 = 11042 +     638 +      84                |

Three layers on the GPU fails:

Details

./result/bin/llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja -ngl 3 -c 32768 --reasoning off -p "Hello\!"
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.
Loading model...
▄▄ ▄▄

██ ██

██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄

██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██

██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀

██    ██

▀▀    ▀▀
build      : b8864-ff6b106

model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf

modalities : text
available commands:

/exit or Ctrl+C     stop or exit

/regen              regenerate the last response

/clear              clear the chat history

/read         add a text file

/glob      add text files using globbing pattern
> Hello!
//nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libggml-base.so.0(+0x1979a) [0x7f383d00179a]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libggml-base.so.0(ggml_print_backtrace+0x204) [0x7f383d001c64]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libggml-base.so.0(+0x2e149) [0x7f383d016149]

/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xc539a) [0x7f38390c539a]

/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(_ZSt10unexpectedv+0x0) [0x7f38390b286e]

/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xc5637) [0x7f38390c5637]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libggml-vulkan.so.0(+0xa12ca) [0x7f38394a12ca]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libggml-vulkan.so.0(+0x1a3ec7) [0x7f38395a3ec7]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libggml-vulkan.so.0(+0x1a4268) [0x7f38395a4268]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libggml-base.so.0(ggml_backend_sched_graph_compute_async+0x92c) [0x7f383d0202bc]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libllama.so.0(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1) [0x7f383d2bc7f1]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libllama.so.0(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0x11f) [0x7f383d2bf27f]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libllama.so.0(_ZN13llama_context6decodeERK11llama_batch+0x3c9) [0x7f383d2c6029]

/nix/store/1lp1x6qd4108yh3rnkhx6038nwl6pv9h-llama-cpp-8864/lib/libllama.so.0(llama_decode+0x11) [0x7f383d2c7d01]

./result/bin/llama-cli(+0x123268) [0x56282dcaf268]

./result/bin/llama-cli(+0xbb582) [0x56282dc47582]

/nix/store/ab3753m6i7isgvzphlar0a8xb84gl96i-gcc-15.2.0-lib/lib/libstdc++.so.6(+0xf2fa4) [0x7f38390f2fa4]

/nix/store/jms7zxzm7w1whczwny5m3gkgdjghmi2r-glibc-2.42-51/lib/libc.so.6(+0x9dd53) [0x7f3838c9dd53]

/nix/store/jms7zxzm7w1whczwny5m3gkgdjghmi2r-glibc-2.42-51/lib/libc.so.6(+0x12563c) [0x7f3838d2563c]

terminate called after throwing an instance of 'vk::DeviceLostError'

what():  vk::Queue::submit: ErrorDeviceLost

zsh: IOT instruction (core dumped)  ./result/bin/llama-cli -m  --jinja -ngl 3 -c 32768 --reasoning off -p "Hello!

Running with --cpu-moe works:

Details

./result/bin/llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja --cpu-moe -c 32768 --reasoning off -p "Hello\!"
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.
Loading model...
▄▄ ▄▄

██ ██

██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄

██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██

██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀

██    ██

▀▀    ▀▀
build      : b8864-ff6b106

model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf

modalities : text
available commands:

/exit or Ctrl+C     stop or exit

/regen              regenerate the last response

/clear              clear the chat history

/read         add a text file

/glob      add text files using globbing pattern
> Hello!
Hello! How can I help you today?
[ Prompt: 2,4 t/s | Generation: 1,8 t/s ]
>
Exiting...

llama_memory_breakdown_print: | memory breakdown [MiB]                | total    free     self   model   context   compute    unaccounted |

llama_memory_breakdown_print: |   - Vulkan0 (Intel(R) Graphics (ARL)) | 23686 = 18884 + ( 2527 =  1323 +     702 +     501) +        2274 |

llama_memory_breakdown_print: |   - Host                              |                  11381 = 11309 +       0 +      72                |

[Zirconium419122]

Running llama-cli and llama-server with --reasoning-budget 0 works and has the same effect as disabling reasoning.

Running with --reasoning-budget 0:

llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja -c 32768 --no-mmap --reasoning-budget 0 -p "Hello\!"
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b8770-82764d8
model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read <file>        add a text file
  /glob <pattern>     add text files using globbing pattern


> Hello!

Hello! How can I help you today?

[ Prompt: 22,0 t/s | Generation: 15,7 t/s ]

>

Exiting...
llama_memory_breakdown_print: | memory breakdown [MiB]                | total    free     self   model   context   compute    unaccounted |
llama_memory_breakdown_print: |   - Vulkan0 (Intel(R) Graphics (ARL)) | 23688 = 10389 + (12526 = 11249 +     702 +     574) +         772 |
llama_memory_breakdown_print: |   - Host                              |                    405 =   333 +       0 +      72                |

Running with --reasoning-budget 0 on latest release:

./result/bin/llama-cli -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.5-35B-A3B-GGUF/snapshots/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf --jinja --no-mmap -c 32768 --reasoning-budget 0 -p "Hello\!"
MESA: warning: Support for this platform is experimental with Xe KMD, bug reports may be ignored.

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b8893-6217b49
model      : Qwen3.5-35B-A3B-UD-Q2_K_XL.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read <file>        add a text file
  /glob <pattern>     add text files using globbing pattern


> Hello!

Hello! How can I help you today?

[ Prompt: 20,9 t/s | Generation: 15,1 t/s ]

>

Exiting...
common_memory_breakdown_print: | memory breakdown [MiB]                | total    free     self   model   context   compute    unaccounted |
common_memory_breakdown_print: |   - Vulkan0 (Intel(R) Graphics (ARL)) | 23688 = 10684 + (12526 = 11249 +     702 +     574) +         477 |
common_memory_breakdown_print: |   - Host                              |                    405 =   333 +       0 +      72                |