Skip to content

Eval bug: [CUDA] MoE model (Qwen3-30B-A3B) loads to GPU but does not utilize CUDA for inference in build b5466 #13729

New issue
New issue
Open
Open
Eval bug: [CUDA] MoE model (Qwen3-30B-A3B) loads to GPU but does not utilize CUDA for inference in build b5466#13729
Labels
bug-unconfirmed
@JaydenChao101

Description

@JaydenChao101
JaydenChao101
opened on May 23, 2025

Name and Version

ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 4080 SUPER, compute capability 8.9, VMM: yes
load_backend: loaded CUDA backend from C:\Users\Jayden\Downloads\Compressed\llama-b5466-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Users\Jayden\Downloads\Compressed\llama-b5466-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Users\Jayden\Downloads\Compressed\llama-b5466-bin-win-cuda-12.4-x64\ggml-cpu-alderlake.dll
version: 5466 (9ecf3e6)
built with clang version 18.1.8 for x86_64-pc-windows-msvc

Operating systems

Windows

GGML backends

CUDA

Hardware

CPU: Intel(R) Core(TM) 17-14700KF
RAM: 64.0GB
GPU: NVIDIA GeForce RTX 4080 SUPER

Models

https://huggingface.co/Qwen/Qwen3-30B-A3B-GGUF
Qwen3-30B-A3B-Q4_K_M.gguf

Problem description & steps to reproduce

🧾 Problem Description

When running the Qwen3-30B-A3B-Q4_K_M.gguf MoE-based model using llama-server.exe from the llama-b5466-bin-win-cuda-12.4-x64 build on an NVIDIA RTX 4080 SUPER (Ampere architecture), the following issues occur:

  • The model successfully loads onto the GPU (CUDA0) and allocates memory.
  • However, no actual CUDA kernel computation is triggered, leading to:
    • Extremely slow inference
    • Unstable or incomplete generation output
    • No visible CUDA-related activity in logs (e.g., no ggml_cuda_assign_buffers, kernel launched, etc.)

In contrast, the same model runs correctly using the llama-b5333 build, with proper GPU acceleration and stable output.

This suggests a regression or compatibility issue in the newer b5466 build, especially with how it handles MoE models or GPU offloading logic.

📋 Steps to Reproduce

✅ Affected Version

llama-b5466-bin-win-cuda-12.4-x64

🔧 Command Used

llama-server.exe --model "C:\Users\Jayden\Downloads\Qwen3-30B-A3B-Q4_K_M.gguf" --n-gpu-layers 30 --threads 16 --repeat-penalty 1.3 --temp 0.7 --mirostat 2 --mlock

🧪 Reproduction Steps

  1. Download and extract the llama-b5466-bin-win-cuda-12.4-x64 build.
  2. Place the Qwen3-30B-A3B-Q4_K_M.gguf model file in a known location.
  3. Run the above command in CMD/PowerShell.
  4. Wait for the model to load (you'll see logs indicating successful loading to CUDA).
  5. Send a chat completion request via the API endpoint: 
    POST http://127.0.0.1:8080/v1/chat/completions
    With body: 
    {
  "model": "Qwen3-30B-A3B",
  "messages": [{"role": "user", "content": "Hello"}]
}
  6. Observe:
    • Very slow response time or timeout
    • Incomplete or unstable output
    • No evidence of CUDA kernel usage in logs

✅ Control Test (Working Version)

Repeat the exact steps above using llama-b5333. You will observe:

  • Fast inference
  • Coherent output
  • Logs clearly indicate CUDA kernel execution

⚠️ Observed Behavior in b5466

Despite seeing these log entries:

llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4080 SUPER) - 15035 MiB free
load_tensors: offloaded 30/49 layers to GPU

There is no further CUDA activity during inference.
No ggml_cuda_assign_buffers or kernel launched messages are shown.

🛠 Potential Causes

  • Regression in how MoE layers are offloaded or executed in b5466.
  • Changes in graph execution logic that prevent CUDA kernels from being used.
  • Incomplete or incorrect support for MoE models in this particular build.
  • Possible missing or broken CUDA kernel bindings for certain ops used by MoE models.

✅ Suggested Fixes / Investigations

  • Add better detection of MoE model structures
  • Ensure CUDA kernels are properly triggered for MoE layers
  • Consider adding a warning when certain features are not supported in current builds
  • Compare source code diffs between b5333 and b5466 focusing on:
    • src/ggml-cuda.cu
    • src/llama.cpp (offloading logic)
    • MoE-specific handling in llm_load_tensors()

First Bad Commit

No response

Relevant log output

go.bat
@echo off
llama-server.exe --model "C:\Users\Jayden\Downloads\Qwen3-30B-A3B-Q4_K_M.gguf" --n-gpu-layers 30 --threads 16 --repeat-penalty 1.3 --temp 0.7 --mirostat 2 --mlock

C:\Users\Jayden\Downloads\Compressed\llama-b5466-bin-win-cuda-12.4-x64>go
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4080 SUPER, compute capability 8.9, VMM: yes
load_backend: loaded CUDA backend from C:\Users\Jayden\Downloads\Compressed\llama-b5466-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Users\Jayden\Downloads\Compressed\llama-b5466-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Users\Jayden\Downloads\Compressed\llama-b5466-bin-win-cuda-12.4-x64\ggml-cpu-alderlake.dll
build: 5466 (9ecf3e66) with clang version 18.1.8 for x86_64-pc-windows-msvc
system info: n_threads = 16, n_threads_batch = 16, total_threads = 28

system_info: n_threads = 16 (n_threads_batch = 16) / 28 | CUDA : ARCHS = 500,610,700,750,800,860,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 |

main: binding port with default address family
main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 27
main: loading model
srv    load_model: loading model 'C:\Users\Jayden\Downloads\Qwen3-30B-A3B-Q4_K_M.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4080 SUPER) - 15035 MiB free
llama_model_loader: loaded meta data with 31 key-value pairs and 579 tensors from C:\Users\Jayden\Downloads\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 Gptq Fp16
llama_model_loader: - kv   3:                           general.finetune str              = gptq
llama_model_loader: - kv   4:                           general.basename str              = Qwen3
llama_model_loader: - kv   5:                         general.size_label str              = 30B
llama_model_loader: - kv   6:                       qwen3moe.block_count u32              = 48
llama_model_loader: - kv   7:                    qwen3moe.context_length u32              = 40960
llama_model_loader: - kv   8:                  qwen3moe.embedding_length u32              = 2048
llama_model_loader: - kv   9:               qwen3moe.feed_forward_length u32              = 6144
llama_model_loader: - kv  10:              qwen3moe.attention.head_count u32              = 32
llama_model_loader: - kv  11:           qwen3moe.attention.head_count_kv u32              = 4
llama_model_loader: - kv  12:                    qwen3moe.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  13:  qwen3moe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:                 qwen3moe.expert_used_count u32              = 8
llama_model_loader: - kv  15:              qwen3moe.attention.key_length u32              = 128
llama_model_loader: - kv  16:            qwen3moe.attention.value_length u32              = 128
llama_model_loader: - kv  17:                      qwen3moe.expert_count u32              = 128
llama_model_loader: - kv  18:        qwen3moe.expert_feed_forward_length u32              = 768
llama_model_loader: - kv  19:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  20:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  21:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  22:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  23:                      tokenizer.ggml.merges arr[str,151387]  = ["? ?", "?? ??", "i n", "? t",...
llama_model_loader: - kv  24:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  25:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  26:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  27:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  28:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  29:               general.quantization_version u32              = 2
llama_model_loader: - kv  30:                          general.file_type u32              = 15
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 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   = 17.28 GiB (4.86 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      = 40960
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  = 40960
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 Gptq Fp16
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 '?'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 30 repeating layers to GPU
load_tensors: offloaded 30/49 layers to GPU
load_tensors:        CUDA0 model buffer size = 10750.86 MiB
load_tensors:   CPU_Mapped model buffer size =  6940.48 MiB
...................................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.58 MiB
llama_kv_cache_unified:      CUDA0 KV buffer size =   240.00 MiB
llama_kv_cache_unified:        CPU KV buffer size =   144.00 MiB
llama_kv_cache_unified: size =  384.00 MiB (  4096 cells,  48 layers,  1 seqs), K (f16):  192.00 MiB, V (f16):  192.00 MiB
llama_context:      CUDA0 compute buffer size =   544.18 MiB
llama_context:  CUDA_Host compute buffer size =    12.01 MiB
llama_context: graph nodes  = 3222
llama_context: graph splits = 256 (with bs=512), 39 (with bs=1)
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 4096
main: model loaded
main: chat template, chat_template: {%- if tools %}
    {{- '<|im_start|>system\n' }}
    {%- if messages[0].role == 'system' %}
        {{- messages[0].content + '\n\n' }}
    {%- endif %}
    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
    {%- for tool in tools %}
        {{- "\n" }}
        {{- tool | tojson }}
    {%- endfor %}
    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
{%- else %}
    {%- if messages[0].role == 'system' %}
        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
    {%- endif %}
{%- endif %}
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
{%- for index in range(ns.last_query_index, -1, -1) %}
    {%- set message = messages[index] %}
    {%- if ns.multi_step_tool and message.role == "user" and not('<tool_response>' in message.content and '</tool_response>' in message.content) %}
        {%- set ns.multi_step_tool = false %}
        {%- set ns.last_query_index = index %}
    {%- endif %}
{%- endfor %}
{%- for message in messages %}
    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
    {%- elif message.role == "assistant" %}
        {%- set content = message.content %}
        {%- set reasoning_content = '' %}
        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
            {%- set reasoning_content = message.reasoning_content %}
        {%- else %}
            {%- if '</think>' in message.content %}
                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
            {%- endif %}
        {%- endif %}
        {%- if loop.index0 > ns.last_query_index %}
            {%- if loop.last or (not loop.last and reasoning_content) %}
                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
            {%- else %}
                {{- '<|im_start|>' + message.role + '\n' + content }}
            {%- endif %}
        {%- else %}
            {{- '<|im_start|>' + message.role + '\n' + content }}
        {%- endif %}
        {%- if message.tool_calls %}
            {%- for tool_call in message.tool_calls %}
                {%- if (loop.first and content) or (not loop.first) %}
                    {{- '\n' }}
                {%- endif %}
                {%- if tool_call.function %}
                    {%- set tool_call = tool_call.function %}
                {%- endif %}
                {{- '<tool_call>\n{"name": "' }}
                {{- tool_call.name }}
                {{- '", "arguments": ' }}
                {%- if tool_call.arguments is string %}
                    {{- tool_call.arguments }}
                {%- else %}
                    {{- tool_call.arguments | tojson }}
                {%- endif %}
                {{- '}\n</tool_call>' }}
            {%- endfor %}
        {%- endif %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "tool" %}
        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
            {{- '<|im_start|>user' }}
        {%- endif %}
        {{- '\n<tool_response>\n' }}
        {{- message.content }}
        {{- '\n</tool_response>' }}
        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
            {{- '<|im_end|>\n' }}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {{- '<|im_start|>assistant\n' }}
    {%- if enable_thinking is defined and enable_thinking is false %}
        {{- '<think>\n\n</think>\n\n' }}
    {%- endif %}
{%- 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://127.0.0.1:8080 - starting the main loop
srv  update_slots: all slots are idle
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, n_prompt_tokens = 383
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 383, n_tokens = 383, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 383, n_tokens = 383
srv  cancel_tasks: cancel task, id_task = 0
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
slot      release: id  0 | task 0 | stop processing: n_past = 418, truncated = 0
srv  update_slots: all slots are idle
srv    operator(): operator(): cleaning up before exit...

Metadata

Metadata

Assignees

No one assigned

    Labels

    bug-unconfirmed

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions