Eval bug: Qwen3-Coder-Next Poor Outputs #19305
Description
Name and Version
$ ./build/bin/llama-server --version
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GB10, compute capability 12.1, VMM: yes
version: 7925 (8bece2e)
built with GNU 13.3.0 for Linux aarch64
Operating systems
Linux
GGML backends
CUDA
Hardware
DGX Spark
Models
Same behavior observed on all quants I tested
- lmstudio-community/Qwen3-Coder-Next-GGUF @ Q4_K_M and Q8_0
- Qwen/Qwen3-Coder-Next-GGUF @ Q4_K_M
Problem description & steps to reproduce
Qwen3-Coder-Next gets tripped up on "syntax issues" when run in llama-server. I run the attached prompt.txt, which asks if server.cpp has any syntax errors.
This specific prompt is not the most "realistic," but it clearly demonstrates an issue I see when I ask the model to generally "review the changes in file X." It smells to me like an implementation bug, given that the output is noticeably worse in llama-server at both 4 bit and 8 bit quants compared to vLLM running FP8 or MLX running int 4.
llama-server Outputs:
- llama-server-out.txt: uses lmstudio-community/Qwen3-Coder-Next-GGUF @ Q4_K_M. Shows poor output with misidentified syntax errors, seemingly stumbling over itself.
- llama-server-out-q8.txt: uses lmstudio-community/Qwen3-Coder-Next-GGUF @ Q8_0 . Shows similarly poor outputs to Q4.
- llama-server-out-qwen-quant.txt: Uses Qwen/Qwen3-Coder-Next-GGUF @ Q4_K_M . Shows similar outputs to the lmstudio-community quants.
Outputs from other engines:
- vllm-fp8.txt : vLLM running the Qwen/Qwen3-Coder-Next-FP8 model. Does not report any incorrect errors.
- lms-mlx-engine-out.txt: mlx-engine running the lmstudio-community/Qwen3-Coder-Next-MLX-4bit model LM Studio. Does report any incorrect errors.
First Bad Commit
I did not observe a commit where this was previously working.
Relevant log output
Logs
./build/bin/llama-server -m ~/.lmstudio/models/lmstudio-community/Qwen3-Coder-Next-GGUF/Qwen3-Coder-Next-Q8_0-00001-of-00003.gguf -c 65536
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GB10, compute capability 12.1, VMM: yes
main: n_parallel is set to auto, using n_parallel = 4 and kv
_unified = true
build: 7925 (8bece2eb2) with GNU 13.3.0 for Linux aarch64
system info: n_threads = 20, n_threads_batch = 20, total_threads = 20
system_info: n_threads = 20 (n_threads_batch = 20) / 20 | CUDA : ARCHS = 1210 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | BLACKWELL_NATIVE_FP4 = 1 | CPU : NEON = 1 | ARM_FMA = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 19 threads for HTTP server
start: binding port with default address family
main: loading model
srv load_model: loading model '/home/lms/.lmstudio/models/lmstudio-community/Qwen3-Coder-Next-GGUF/Qwen3-Coder-Next-Q8_0-00001-of-00003.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 82704 MiB of device memory vs. 114941 MiB of free device memory
llama_params_fit_impl: will leave 32236 >= 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 0.19 seconds
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GB10) (000f:01:00.0) - 115242 MiB free
llama_model_loader: additional 2 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 42 key-value pairs and 843 tensors from /home/lms/.lmstudio/models/lmstudio-community/Qwen3-Coder-Next-GGUF/Qwen3-Coder-Next-Q8_0-00001-of-00003.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 = Qwen_Qwen3 Coder Next
llama_model_loader: - kv 6: general.size_label str = 80B
llama_model_loader: - kv 7: qwen3next.block_count u32 = 48
llama_model_loader: - kv 8: qwen3next.context_length u32 = 262144
llama_model_loader: - kv 9: qwen3next.embedding_length u32 = 2048
llama_model_loader: - kv 10: qwen3next.feed_forward_length u32 = 5120
llama_model_loader: - kv 11: qwen3next.attention.head_count u32 = 16
llama_model_loader: - kv 12: qwen3next.attention.head_count_kv u32 = 2
llama_model_loader: - kv 13: qwen3next.rope.freq_base f32 = 5000000.000000
llama_model_loader: - kv 14: qwen3next.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 15: qwen3next.expert_used_count u32 = 10
llama_model_loader: - kv 16: qwen3next.attention.key_length u32 = 256
llama_model_loader: - kv 17: qwen3next.attention.value_length u32 = 256
llama_model_loader: - kv 18: qwen3next.expert_count u32 = 512
llama_model_loader: - kv 19: qwen3next.expert_feed_forward_length u32 = 512
llama_model_loader: - kv 20: qwen3next.expert_shared_feed_forward_length u32 = 512
llama_model_loader: - kv 21: qwen3next.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 22: qwen3next.ssm.state_size u32 = 128
llama_model_loader: - kv 23: qwen3next.ssm.group_count u32 = 16
llama_model_loader: - kv 24: qwen3next.ssm.time_step_rank u32 = 32
llama_model_loader: - kv 25: qwen3next.ssm.inner_size u32 = 4096
llama_model_loader: - kv 26: qwen3next.rope.dimension_count u32 = 64
llama_model_loader: - kv 27: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 28: tokenizer.ggml.pre str = qwen2
llama_model_loader: - kv 29: tokenizer.ggml.tokens arr[str,151936] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 30: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 31: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 32: tokenizer.ggml.eos_token_id u32 = 151645
llama_model_loader: - kv 33: tokenizer.ggml.padding_token_id u32 = 151643
llama_model_loader: - kv 34: tokenizer.ggml.bos_token_id u32 = 151643
llama_model_loader: - kv 35: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 36: tokenizer.chat_template str = {% macro render_extra_keys(json_dict,...
llama_model_loader: - kv 37: general.quantization_version u32 = 2
llama_model_loader: - kv 38: general.file_type u32 = 7
llama_model_loader: - kv 39: split.no u16 = 0
llama_model_loader: - kv 40: split.tensors.count i32 = 843
llama_model_loader: - kv 41: split.count u16 = 3
llama_model_loader: - type f32: 313 tensors
llama_model_loader: - type q8_0: 482 tensors
llama_model_loader: - type bf16: 48 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 = Qwen_Qwen3 Coder Next
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 = true, 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: CPU_Mapped model buffer size = 315.30 MiB
load_tensors: CUDA0 model buffer size = 80562.07 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 = 65536
llama_context: n_ctx_seq = 65536
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = auto
llama_context: kv_unified = true
llama_context: freq_base = 5000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: CUDA_Host output buffer size = 2.32 MiB
llama_kv_cache: CUDA0 KV buffer size = 1536.00 MiB
llama_kv_cache: size = 1536.00 MiB ( 65536 cells, 12 layers, 4/1 seqs), K (f16): 768.00 MiB, V (f16): 768.00 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: Flash Attention was auto, set to enabled
sched_reserve: CUDA0 compute buffer size = 304.75 MiB
sched_reserve: CUDA_Host compute buffer size = 136.01 MiB
sched_reserve: graph nodes = 9374 (with bs=512), 5918 (with bs=1)
sched_reserve: graph splits = 2
sched_reserve: reserve took 168.35 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 = 65536
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 = 65536
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 = 65536
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 = 65536
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://127.0.0.1:8080
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 = 65536, n_keep = 0, task.n_tokens = 3051
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 = 2048, batch.n_tokens = 2048, progress = 0.671255
slot update_slots: id 3 | task 0 | n_tokens = 2048, memory_seq_rm [2048, end)
slot update_slots: id 3 | task 0 | prompt processing progress, n_tokens = 2987, batch.n_tokens = 939, progress = 0.979023
slot update_slots: id 3 | task 0 | n_tokens = 2987, memory_seq_rm [2987, end)
slot update_slots: id 3 | task 0 | prompt processing progress, n_tokens = 3051, batch.n_tokens = 64, progress = 1.000000
slot update_slots: id 3 | task 0 | prompt done, n_tokens = 3051, batch.n_tokens = 64
slot init_sampler: id 3 | task 0 | init sampler, took 0.33 ms, tokens: text = 3051, total = 3051
slot update_slots: id 3 | task 0 | created context checkpoint 1 of 8 (pos_min = 2986, pos_max = 2986, size = 75.376 MiB)
slot print_timing: id 3 | task 0 |
prompt eval time = 4135.85 ms / 3051 tokens ( 1.36 ms per token, 737.70 tokens per second)
eval time = 112321.23 ms / 3501 tokens ( 32.08 ms per token, 31.17 tokens per second)
total time = 116457.08 ms / 6552 tokens
slot release: id 3 | task 0 | stop processing: n_tokens = 6551, truncated = 0
srv update_slots: all slots are idle
srv log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200
^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 (GB10) | 122548 = 30640 + (82704 = 80562 + 1837 + 304) + 9203 |
llama_memory_breakdown_print: | - Host | 451 = 315 + 0 + 136 |