Eval bug: Qwen3-30B-A3B-Q4_K_M: Vulkan ~10% slower than AVX2

[ZUIcat]

Name and Version

version: b5233 (ceda28e) with MSVC

Operating systems

Windows

GGML backends

Vulkan

Hardware

AMD 780m (8840u)

Models

Qwen3-30B-A3B-Q4_K_M.gguf
https://huggingface.co/bartowski/Qwen_Qwen3-30B-A3B-GGUF/blob/main/Qwen_Qwen3-30B-A3B-Q4_K_M.gguf

Problem description & steps to reproduce

The AVX2 llama.cpp gives ~15 t/s, but Vulkan drops to ~9 t/s—unexpected slowdown.
Here are the command lines I used:
llama-server.exe -m "bartowski_Qwen3-30B-A3B-Q4_K_M.gguf" --host 0.0.0.0 --port 8090 --slots --props --metrics -np 1 -c 20480 -ngl 999 -ctk f16 -ctv f16 -fa --no-mmap --keep 0

First Bad Commit

No response

Relevant log output

ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon 780M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
build: 5233 (ceda28ef) with MSVC 19.43.34810.0 for x64
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

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

main: binding port with default address family
main: HTTP server is listening, hostname: 0.0.0.0, port: 8090, http threads: 15
main: loading model
srv    load_model: loading model 'bartowski_Qwen3-30B-A3B-Q4_K_M.gguf'
llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon 780M Graphics) - 16128 MiB free
llama_model_loader: loaded meta data with 41 key-value pairs and 579 tensors from bartowski_Qwen3-30B-A3B-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen3moe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3 30B A3B
llama_model_loader: - kv   3:                           general.basename str              = Qwen3
llama_model_loader: - kv   4:                         general.size_label str              = 30B-A3B
llama_model_loader: - kv   5:                            general.license str              = apache-2.0
llama_model_loader: - kv   6:                       general.license.link str              = https://huggingface.co/Qwen/Qwen3-30B...
llama_model_loader: - kv   7:                   general.base_model.count u32              = 1
llama_model_loader: - kv   8:                  general.base_model.0.name str              = Qwen3 30B A3B Base
llama_model_loader: - kv   9:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  10:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen3-30B...
llama_model_loader: - kv  11:                               general.tags arr[str,1]       = ["text-generation"]
llama_model_loader: - kv  12:                       qwen3moe.block_count u32              = 48
llama_model_loader: - kv  13:                    qwen3moe.context_length u32              = 32768
llama_model_loader: - kv  14:                  qwen3moe.embedding_length u32              = 2048
llama_model_loader: - kv  15:               qwen3moe.feed_forward_length u32              = 6144
llama_model_loader: - kv  16:              qwen3moe.attention.head_count u32              = 32
llama_model_loader: - kv  17:           qwen3moe.attention.head_count_kv u32              = 4
llama_model_loader: - kv  18:                    qwen3moe.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  19:  qwen3moe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  20:                 qwen3moe.expert_used_count u32              = 8
llama_model_loader: - kv  21:              qwen3moe.attention.key_length u32              = 128
llama_model_loader: - kv  22:            qwen3moe.attention.value_length u32              = 128
llama_model_loader: - kv  23:                      qwen3moe.expert_count u32              = 128
llama_model_loader: - kv  24:        qwen3moe.expert_feed_forward_length u32              = 768
llama_model_loader: - kv  25:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  26:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  27:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  28:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  29:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  30:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  31:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  32:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  33:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  35:               general.quantization_version u32              = 2
llama_model_loader: - kv  36:                          general.file_type u32              = 15
llama_model_loader: - kv  37:                      quantize.imatrix.file str              = /models_out/Qwen3-30B-A3B-GGUF/Qwen_Q...
llama_model_loader: - kv  38:                   quantize.imatrix.dataset str              = /training_data/calibration_datav3.txt
llama_model_loader: - kv  39:             quantize.imatrix.entries_count i32              = 384
llama_model_loader: - kv  40:              quantize.imatrix.chunks_count i32              = 209
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q8_0:   48 tensors
llama_model_loader: - type q4_K:  193 tensors
llama_model_loader: - type q5_K:   48 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 17.35 GiB (4.88 BPW)
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3moe
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 2048
print_info: n_layer          = 48
print_info: n_head           = 32
print_info: n_head_kv        = 4
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 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             = 6144
print_info: n_expert         = 128
print_info: n_expert_used    = 8
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 30B.A3B
print_info: model params     = 30.53 B
print_info: general.name     = Qwen3 30B A3B
print_info: n_ff_exp         = 768
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = false)
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors:      Vulkan0 model buffer size = 17596.42 MiB
load_tensors:          CPU model buffer size =   166.92 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 20480
llama_context: n_ctx_per_seq = 20480
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 1
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (20480) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     0.58 MiB
init: kv_size = 20480, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
init:    Vulkan0 KV buffer size =  1920.00 MiB
llama_context: KV self size  = 1920.00 MiB, K (f16):  960.00 MiB, V (f16):  960.00 MiB
llama_context:    Vulkan0 compute buffer size =   300.75 MiB
llama_context: Vulkan_Host compute buffer size =    80.01 MiB
llama_context: graph nodes  = 2935
llama_context: graph splits = 98
common_init_from_params: setting dry_penalty_last_n to ctx_size = 20480
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
common_chat_templates_init: failed to parse chat template (defaulting to chatml): Expected value expression at row 18, column 30:
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
{%- for message in messages[::-1] %}
                             ^
    {%- set index = (messages|length - 1) - loop.index0 %}

srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 20480
main: model loaded
main: chat template, chat_template: {%- for message in messages -%}
  {{- '<|im_start|>' + message.role + '
' + message.content + '<|im_end|>
' -}}
{%- endfor -%}
{%- if add_generation_prompt -%}
  {{- '<|im_start|>assistant
' -}}
{%- endif -%}, example_format: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
'
main: server is listening on http://0.0.0.0:8090 - starting the main loop
srv  update_slots: all slots are idle

[ggerganov]

Try to remove the -fa flag.

[ZUIcat]

Try to remove the -fa flag.

Surprisingly effective. Removing the -fa flag, Vulkan version reached ~25 t/s.

---

Here are my previous test results for others to test and reference:

• AV2 Version with -fa: ~15t/s
• Vulkan Version with -fa: ~9t/s
• Vulkan Version without -fa: ~25t/s

[0cc4m]

This is because Flash Attention on Vulkan is only supported on Nvidia GPUs with coopmat2 on an up-to-date driver. For any other device it will fall back to CPU, which is slow.

[ZUIcat]

This is because Flash Attention on Vulkan is only supported on Nvidia GPUs with coopmat2 on an up-to-date driver. For any other device it will fall back to CPU, which is slow.

Thank you. I hadn't noticed this before. I just tested with other models (e.g., gemma-2-27b-it-Q4_K_M) and confirmed that enabling -fa actually slows down performance when using Vulkan.