Eval bug: Llama.cpp 40% slower than VLLM + high CPU usage when running Qwen Coder Next #19345
Description
Name and Version
Author: Xuan-Son Nguyen <son@huggingface.co>
Date: Wed Feb 4 17:55:31 2026 +0100
debug: make common_debug_print_tensor readable (#19331)
* debug: make common_debug_print_tensor readable
* editorconfigOperating systems
Linux
GGML backends
CUDA
Hardware
Ryzen 7950X3D + RTX 6000 Blackwell
Models
Qwen 3 Coder Next, official GGUF at Q8_0
https://huggingface.co/Qwen/Qwen3-Coder-Next-GGUF
Problem description & steps to reproduce
When running Qwen 3 Coder Next at FP8 on VLLM, token generation speed is ~ 120 tokens per second.
When using llama.cpp, token generation is 70-80 tokens per second
First Bad Commit
No response
Relevant log output
Launch options:
./build/bin/llama-server \
-hf Qwen/Qwen3-Coder-Next-GGUF:Q8_0 \
-a qwen/qwen3coder-next \
--jinja \
--host 0.0.0.0 \
--port 11434 \
--no-mmap \
-fa on \
-kvu
Console Output:
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, compute capability 12.0, VMM: yes
common_download_file_single_online: no previous model file found /mnt/media/llm-cache/llama.cpp/Qwen_Qwen3-Coder-Next-GGUF_preset.ini
common_download_file_single_online: HEAD invalid http status code received: 404
no remote preset found, skipping
common_download_file_single_online: using cached file: /mnt/media/llm-cache/llama.cpp/Qwen_Qwen3-Coder-Next-GGUF_Qwen3-Coder-Next-Q8_0_Qwen3-Coder-Next-Q8_0-00001-of-00004.gguf
common_download_file_single_online: using cached file: /mnt/media/llm-cache/llama.cpp/Qwen_Qwen3-Coder-Next-GGUF_Qwen3-Coder-Next-Q8_0_Qwen3-Coder-Next-Q8_0-00004-of-00004.gguf
common_download_file_single_online: using cached file: /mnt/media/llm-cache/llama.cpp/Qwen_Qwen3-Coder-Next-GGUF_Qwen3-Coder-Next-Q8_0_Qwen3-Coder-Next-Q8_0-00002-of-00004.gguf
common_download_file_single_online: using cached file: /mnt/media/llm-cache/llama.cpp/Qwen_Qwen3-Coder-Next-GGUF_Qwen3-Coder-Next-Q8_0_Qwen3-Coder-Next-Q8_0-00003-of-00004.gguf
main: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true
build: 7938 (e0c93af2a) with GNU 14.2.0 for Linux x86_64
system info: n_threads = 16, n_threads_batch = 16, total_threads = 32
system_info: n_threads = 16 (n_threads_batch = 16) / 32 | CUDA : ARCHS = 1200 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | BLACKWELL_NATIVE_FP4 = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 31 threads for HTTP server
start: binding port with default address family
main: loading model
srv load_model: loading model '/mnt/media/llm-cache/llama.cpp/Qwen_Qwen3-Coder-Next-GGUF_Qwen3-Coder-Next-Q8_0_Qwen3-Coder-Next-Q8_0-00001-of-00004.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 87811 MiB of device memory vs. 96365 MiB of free device memory
llama_params_fit_impl: will leave 8553 >= 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 0.24 seconds
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition) (0000:01:00.0) - 96667 MiB free
llama_model_loader: additional 3 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 44 key-value pairs and 807 tensors from /mnt/media/llm-cache/llama.cpp/Qwen_Qwen3-Coder-Next-GGUF_Qwen3-Coder-Next-Q8_0_Qwen3-Coder-Next-Q8_0-00001-of-00004.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 = qwen3next
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.sampling.top_k i32 = 40
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 Coder Next 0129
llama_model_loader: - kv 6: general.version str = 0129
llama_model_loader: - kv 7: general.basename str = Qwen3-Coder-Next
llama_model_loader: - kv 8: general.size_label str = 512x2.5B
llama_model_loader: - kv 9: qwen3next.block_count u32 = 48
llama_model_loader: - kv 10: qwen3next.context_length u32 = 262144
llama_model_loader: - kv 11: qwen3next.embedding_length u32 = 2048
llama_model_loader: - kv 12: qwen3next.feed_forward_length u32 = 5120
llama_model_loader: - kv 13: qwen3next.attention.head_count u32 = 16
llama_model_loader: - kv 14: qwen3next.attention.head_count_kv u32 = 2
llama_model_loader: - kv 15: qwen3next.rope.freq_base f32 = 5000000.000000
llama_model_loader: - kv 16: qwen3next.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 17: qwen3next.expert_used_count u32 = 10
llama_model_loader: - kv 18: qwen3next.attention.key_length u32 = 256
llama_model_loader: - kv 19: qwen3next.attention.value_length u32 = 256
llama_model_loader: - kv 20: qwen3next.expert_count u32 = 512
llama_model_loader: - kv 21: qwen3next.expert_feed_forward_length u32 = 512
llama_model_loader: - kv 22: qwen3next.expert_shared_feed_forward_length u32 = 512
llama_model_loader: - kv 23: qwen3next.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 24: qwen3next.ssm.state_size u32 = 128
llama_model_loader: - kv 25: qwen3next.ssm.group_count u32 = 16
llama_model_loader: - kv 26: qwen3next.ssm.time_step_rank u32 = 32
llama_model_loader: - kv 27: qwen3next.ssm.inner_size u32 = 4096
llama_model_loader: - kv 28: qwen3next.rope.dimension_count u32 = 64
llama_model_loader: - kv 29: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 30: tokenizer.ggml.pre str = qwen2
llama_model_loader: - kv 31: tokenizer.ggml.tokens arr[str,151936] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 32: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 33: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 34: tokenizer.ggml.eos_token_id u32 = 151645
llama_model_loader: - kv 35: tokenizer.ggml.padding_token_id u32 = 151643
llama_model_loader: - kv 36: tokenizer.ggml.bos_token_id u32 = 151643
llama_model_loader: - kv 37: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 38: tokenizer.chat_template str = {% macro render_extra_keys(json_dict,...
llama_model_loader: - kv 39: general.quantization_version u32 = 2
llama_model_loader: - kv 40: general.file_type u32 = 7
llama_model_loader: - kv 41: split.no u16 = 0
llama_model_loader: - kv 42: split.count u16 = 4
llama_model_loader: - kv 43: split.tensors.count i32 = 807
llama_model_loader: - type f32: 313 tensors
llama_model_loader: - type f16: 48 tensors
llama_model_loader: - type q8_0: 446 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q8_0
print_info: file size = 78.98 GiB (8.52 BPW)
load: 0 unused tokens
load: printing all EOG tokens:
load: - 151643 ('<|endoftext|>')
load: - 151645 ('<|im_end|>')
load: - 151662 ('<|fim_pad|>')
load: - 151663 ('<|repo_name|>')
load: - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch = qwen3next
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 = 48
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 = 5120
print_info: n_expert = 512
print_info: n_expert_used = 10
print_info: n_expert_groups = 0
print_info: n_group_used = 0
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 5000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 262144
print_info: rope_yarn_log_mul = 0.0000
print_info: rope_finetuned = unknown
print_info: 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 = 80B.A3B
print_info: model params = 79.67 B
print_info: general.name = Qwen3 Coder Next 0129
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, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 47 repeating layers to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors: CUDA0 model buffer size = 80562.07 MiB
load_tensors: CUDA_Host model buffer size = 315.30 MiB
....................................................................................................
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max = 4
llama_context: n_ctx = 262144
llama_context: n_ctx_seq = 262144
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = enabled
llama_context: kv_unified = true
llama_context: freq_base = 5000000.0
llama_context: freq_scale = 1
llama_context: CUDA_Host output buffer size = 2.32 MiB
llama_kv_cache: CUDA0 KV buffer size = 6144.00 MiB
llama_kv_cache: size = 6144.00 MiB (262144 cells, 12 layers, 4/1 seqs), K (f16): 3072.00 MiB, V (f16): 3072.00 MiB
llama_memory_recurrent: CUDA0 RS buffer size = 301.50 MiB
llama_memory_recurrent: size = 301.50 MiB ( 4 cells, 48 layers, 4 seqs), R (f32): 13.50 MiB, S (f32): 288.00 MiB
sched_reserve: reserving ...
sched_reserve: CUDA0 compute buffer size = 804.06 MiB
sched_reserve: CUDA_Host compute buffer size = 520.01 MiB
sched_reserve: graph nodes = 9590 (with bs=512), 6242 (with bs=1)
sched_reserve: graph splits = 2
sched_reserve: reserve took 202.02 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 = 4
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 = 262144
no implementations specified for speculative decoding
slot load_model: id 1 | task -1 | speculative decoding context not initialized
slot load_model: id 1 | task -1 | new slot, n_ctx = 262144
no implementations specified for speculative decoding
slot load_model: id 2 | task -1 | speculative decoding context not initialized
slot load_model: id 2 | task -1 | new slot, n_ctx = 262144
no implementations specified for speculative decoding
slot load_model: id 3 | task -1 | speculative decoding context not initialized
slot load_model: id 3 | task -1 | new slot, n_ctx = 262144
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 = 0
main: model loaded
main: server is listening on http://0.0.0.0:11434
main: starting the main loop...
srv update_slots: all slots are idle
srv params_from_: Chat format: Qwen3 Coder
slot get_availabl: id 3 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id 3 | 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 3 | task 0 | processing task, is_child = 0
slot update_slots: id 3 | task 0 | new prompt, n_ctx_slot = 262144, n_keep = 0, task.n_tokens = 15
slot update_slots: id 3 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id 3 | task 0 | prompt processing progress, n_tokens = 15, batch.n_tokens = 15, progress = 1.000000
slot update_slots: id 3 | task 0 | prompt done, n_tokens = 15, batch.n_tokens = 15
slot init_sampler: id 3 | task 0 | init sampler, took 0.00 ms, tokens: text = 15, total = 15
slot print_timing: id 3 | task 0 |
prompt eval time = 64.43 ms / 15 tokens ( 4.30 ms per token, 232.82 tokens per second)
eval time = 54353.28 ms / 4601 tokens ( 11.81 ms per token, 84.65 tokens per second)
total time = 54417.71 ms / 4616 tokens
slot release: id 3 | task 0 | stop processing: n_tokens = 4615, truncated = 0
srv update_slots: all slots are idle
srv log_server_r: done request: POST /v1/chat/completions 192.168.15.105 200
srv params_from_: Chat format: Qwen3 Coder
slot get_availabl: id 2 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id 2 | 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 2 | task 4602 | processing task, is_child = 0
slot update_slots: id 2 | task 4602 | new prompt, n_ctx_slot = 262144, n_keep = 0, task.n_tokens = 5320
slot update_slots: id 2 | task 4602 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id 2 | task 4602 | prompt processing progress, n_tokens = 2048, batch.n_tokens = 2048, progress = 0.384962
slot update_slots: id 2 | task 4602 | n_tokens = 2048, memory_seq_rm [2048, end)
slot update_slots: id 2 | task 4602 | prompt processing progress, n_tokens = 4096, batch.n_tokens = 2048, progress = 0.769925
slot update_slots: id 2 | task 4602 | n_tokens = 4096, memory_seq_rm [4096, end)
slot update_slots: id 2 | task 4602 | prompt processing progress, n_tokens = 5256, batch.n_tokens = 1160, progress = 0.987970
slot update_slots: id 2 | task 4602 | n_tokens = 5256, memory_seq_rm [5256, end)
slot update_slots: id 2 | task 4602 | prompt processing progress, n_tokens = 5320, batch.n_tokens = 64, progress = 1.000000
slot update_slots: id 2 | task 4602 | prompt done, n_tokens = 5320, batch.n_tokens = 64
slot init_sampler: id 2 | task 4602 | init sampler, took 0.38 ms, tokens: text = 5320, total = 5320
slot update_slots: id 2 | task 4602 | created context checkpoint 1 of 8 (pos_min = 5255, pos_max = 5255, size = 75.376 MiB)
slot print_timing: id 2 | task 4602 |
prompt eval time = 2168.33 ms / 5320 tokens ( 0.41 ms per token, 2453.51 tokens per second)
eval time = 79.00 ms / 7 tokens ( 11.29 ms per token, 88.61 tokens per second)
total time = 2247.33 ms / 5327 tokens
slot release: id 2 | task 4602 | stop processing: n_tokens = 5326, truncated = 0
srv update_slots: all slots are idle
srv log_server_r: done request: POST /v1/chat/completions 192.168.15.105 200
srv params_from_: Chat format: Qwen3 Coder
slot get_availabl: id 3 | task -1 | selected slot by LCP similarity, sim_best = 1.000 (> 0.100 thold), f_keep = 0.003
srv get_availabl: updating prompt cache
srv prompt_save: - saving prompt with length 4615, total state size = 183.593 MiB
srv load: - looking for better prompt, base f_keep = 0.003, sim = 1.000
srv update: - cache state: 1 prompts, 183.593 MiB (limits: 8192.000 MiB, 262144 tokens, 262144 est)
srv update: - prompt 0x56158786d870: 4615 tokens, checkpoints: 0, 183.593 MiB
srv get_availabl: prompt cache update took 71.80 ms
slot launch_slot_: id 3 | 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 3 | task 4613 | processing task, is_child = 0
slot update_slots: id 3 | task 4613 | new prompt, n_ctx_slot = 262144, n_keep = 0, task.n_tokens = 15
slot update_slots: id 3 | task 4613 | n_past = 15, slot.prompt.tokens.size() = 4615, seq_id = 3, pos_min = 4614, n_swa = 1
slot update_slots: id 3 | task 4613 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
slot update_slots: id 3 | task 4613 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id 3 | task 4613 | prompt processing progress, n_tokens = 15, batch.n_tokens = 15, progress = 1.000000
slot update_slots: id 3 | task 4613 | prompt done, n_tokens = 15, batch.n_tokens = 15
slot init_sampler: id 3 | task 4613 | init sampler, took 0.00 ms, tokens: text = 15, total = 15
slot print_timing: id 3 | task 4613 |
prompt eval time = 41.36 ms / 15 tokens ( 2.76 ms per token, 362.64 tokens per second)
eval time = 68872.90 ms / 5407 tokens ( 12.74 ms per token, 78.51 tokens per second)
total time = 68914.26 ms / 5422 tokens
slot release: id 3 | task 4613 | stop processing: n_tokens = 5421, truncated = 0
srv update_slots: all slots are idle
^[[A^Csrv operator(): operator(): cleaning up before exit...
llama_memory_breakdown_print: | memory breakdown [MiB] | total free self model context compute unaccounted |
llama_memory_breakdown_print: | - CUDA0 (RTX PRO 6000 Blackwell Max-Q Workstation Edition) | 97250 = 8691 + (87811 = 80562 + 6445 + 804) + 747 |
llama_memory_breakdown_print: | - Host | 835 = 315 + 0 + 520 |
VLLM Performance:
(APIServer pid=1508569) INFO 02-04 12:20:46 [loggers.py:257] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 127.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 0.0%
(APIServer pid=1508569) INFO 02-04 12:20:56 [loggers.py:257] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 126.9 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 0.0%
(APIServer pid=1508569) INFO 02-04 12:21:06 [loggers.py:257] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 126.7 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 0.0%
Screenshot of 100% CPU usage:
