Eval bug: QWQ generates repeated text when running with reduced context length

[remixer-dec]

Name and Version

$./llama-server
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 4090, compute capability 8.9, VMM: yes
version: 0 (unknown)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

Not sure why it shows unknown version, possibly because the source was downloaded as tgz from github, I built it manually today, probably at this commit ea00281
I tried earlier versions from a week and a month ago, they have the same issue

Operating systems

Linux

GGML backends

CUDA

Hardware

RTX 4090

Models

Tried a lot of different quants from
https://huggingface.co/Qwen/QwQ-32B-GGUF and
https://huggingface.co/bartowski/Qwen_QwQ-32B-GGUF
for example
https://huggingface.co/bartowski/Qwen_QwQ-32B-GGUF/blob/main/Qwen_QwQ-32B-IQ4_XS.gguf

Problem description & steps to reproduce

When running llama-server with fixed (reduced) context length, the model starts to repeat itself, the less is context length, the earlier it starts to repeat itself.

To reproduce:

1. Run the server with --ctx-size 1024, I used some of the settings from here to test if it helps, except that I didn't touch repetition penalty since it should work just fine without it. Setting -n -2 does not help.

2. Send chat completions request

curl --request POST \
  --url http://localhost:40000/v1/chat/completions \
  --header 'Content-Type: application/json' \
  --data '{
"model":"/models/llms/Qwen_QwQ-32B-IQ4_XS.gguf",
"messages": [
             {"role": "user", "content": "Hello. How many Rs are in strawberry?"}
],
"max_tokens": 1024,
"temperature": 0.6
}'

3. See this:

Expected result: the model thinks as much as it wants hitting the context length limit and stopping if reaching it. This is what happens in AWQ+SGlang stack.

First Bad Commit

No response

Relevant log output

/llama-server --host \
127.0.0.1 \
--port \
40000 \
--model \
/models/llms/Qwen_QwQ-32B-IQ4_XS.gguf \
--ctx-size \
1024 \
--gpu-layers \
99 \
--parallel \
10 \
--mlock \
--samplers \
"top_k;top_p;min_p;temperature;dry;typ_p;xtc" \
--temp \
0.6 \
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 4090, compute capability 8.9, VMM: yes
warn: LLAMA_ARG_HOST environment variable is set, but will be overwritten by command line argument --host
build: 0 (unknown) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
system info: n_threads = 12, n_threads_batch = 12, total_threads = 24

system_info: n_threads = 12 (n_threads_batch = 12) / 24 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: HTTP server is listening, hostname: 127.0.0.1, port: 40000, http threads: 23
main: loading model
srv    load_model: loading model '/models/llms/Qwen_QwQ-32B-IQ4_XS.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4090) - 23725 MiB free
llama_model_loader: loaded meta data with 37 key-value pairs and 771 tensors from /models/llms/Qwen_QwQ-32B-IQ4_XS.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              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = QwQ 32B
llama_model_loader: - kv   3:                           general.basename str              = QwQ
llama_model_loader: - kv   4:                         general.size_label str              = 32B
llama_model_loader: - kv   5:                            general.license str              = apache-2.0
llama_model_loader: - kv   6:                       general.license.link str              = https://huggingface.co/Qwen/QWQ-32B/b...
llama_model_loader: - kv   7:                   general.base_model.count u32              = 1
llama_model_loader: - kv   8:                  general.base_model.0.name str              = Qwen2.5 32B
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/Qwen2.5-32B
llama_model_loader: - kv  11:                               general.tags arr[str,2]       = ["chat", "text-generation"]
llama_model_loader: - kv  12:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  13:                          qwen2.block_count u32              = 64
llama_model_loader: - kv  14:                       qwen2.context_length u32              = 131072
llama_model_loader: - kv  15:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv  16:                  qwen2.feed_forward_length u32              = 27648
llama_model_loader: - kv  17:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  18:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  19:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  20:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  21:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  22:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  23:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  24:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  25:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  26:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  27:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  28:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  29:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  30:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  31:               general.quantization_version u32              = 2
llama_model_loader: - kv  32:                          general.file_type u32              = 30
llama_model_loader: - kv  33:                      quantize.imatrix.file str              = /models_out/QwQ-32B-GGUF/Qwen_QwQ-32B...
llama_model_loader: - kv  34:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  35:             quantize.imatrix.entries_count i32              = 448
llama_model_loader: - kv  36:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  321 tensors
llama_model_loader: - type q5_K:   64 tensors
llama_model_loader: - type q6_K:    1 tensors
llama_model_loader: - type iq4_xs:  385 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ4_XS - 4.25 bpw
print_info: file size   = 16.47 GiB (4.32 BPW) 
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 5120
print_info: n_layer          = 64
print_info: n_head           = 40
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 5
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-05
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: n_ff             = 27648
print_info: n_expert         = 0
print_info: n_expert_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  = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 131072
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       = 32B
print_info: model params     = 32.76 B
print_info: general.name     = QwQ 32B
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
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)
load_tensors: offloading 64 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 65/65 layers to GPU
load_tensors:        CUDA0 model buffer size = 16473.35 MiB
load_tensors:   CPU_Mapped model buffer size =   394.45 MiB
...............................................................................................warning: failed to mlock 1058283520-byte buffer (after previously locking 0 bytes): Cannot allocate memory
Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).
.
llama_init_from_model: n_seq_max     = 10
llama_init_from_model: n_ctx         = 1024
llama_init_from_model: n_ctx_per_seq = 102
llama_init_from_model: n_batch       = 1024
llama_init_from_model: n_ubatch      = 512
llama_init_from_model: flash_attn    = 0
llama_init_from_model: freq_base     = 1000000.0
llama_init_from_model: freq_scale    = 1
llama_init_from_model: n_ctx_per_seq (102) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init: kv_size = 1024, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 64, can_shift = 1
llama_kv_cache_init:      CUDA0 KV buffer size =   256.00 MiB
llama_init_from_model: KV self size  =  256.00 MiB, K (f16):  128.00 MiB, V (f16):  128.00 MiB
llama_init_from_model:  CUDA_Host  output buffer size =     5.80 MiB
llama_init_from_model:      CUDA0 compute buffer size =   307.00 MiB
llama_init_from_model:  CUDA_Host compute buffer size =    12.01 MiB
llama_init_from_model: graph nodes  = 2246
llama_init_from_model: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 1024
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 10
slot         init: id  0 | task -1 | new slot n_ctx_slot = 102
slot         init: id  1 | task -1 | new slot n_ctx_slot = 102
slot         init: id  2 | task -1 | new slot n_ctx_slot = 102
slot         init: id  3 | task -1 | new slot n_ctx_slot = 102
slot         init: id  4 | task -1 | new slot n_ctx_slot = 102
slot         init: id  5 | task -1 | new slot n_ctx_slot = 102
slot         init: id  6 | task -1 | new slot n_ctx_slot = 102
slot         init: id  7 | task -1 | new slot n_ctx_slot = 102
slot         init: id  8 | task -1 | new slot n_ctx_slot = 102
slot         init: id  9 | task -1 | new slot n_ctx_slot = 102
main: model loaded
main: chat template, chat_template: {%- if tools %}
    {{- '<|im_start|>system\n' }}
    {%- if messages[0]['role'] == 'system' %}
        {{- messages[0]['content'] }}
    {%- else %}
        {{- '' }}
    {%- endif %}
    {{- "\n\n# 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 %}
{%- 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" and not message.tool_calls %}
        {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
    {%- elif message.role == "assistant" %}
        {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
        {{- '<|im_start|>' + message.role }}
        {%- if message.content %}
            {{- '\n' + content }}
        {%- endif %}
        {%- for tool_call in message.tool_calls %}
            {%- if tool_call.function is defined %}
                {%- set tool_call = tool_call.function %}
            {%- endif %}
            {{- '\n<tool_call>\n{"name": "' }}
            {{- tool_call.name }}
            {{- '", "arguments": ' }}
            {{- tool_call.arguments | tojson }}
            {{- '}\n</tool_call>' }}
        {%- endfor %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "tool" %}
        {%- if (loop.index0 == 0) 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<think>\n' }}
{%- 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://127.0.0.1:40000 - starting the main loop
srv  update_slots: all slots are idle
srv  log_server_r: request: GET /v1/models 127.0.0.1 200
srv  log_server_r: request: GET /v1/models 127.0.0.1 200
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 = 102, n_keep = 0, n_prompt_tokens = 18
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 18, n_tokens = 18, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 18, n_tokens = 18
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
slot update_slots: id  0 | task 0 | slot context shift, n_keep = 0, n_left = 101, n_discard = 50
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 = 71, truncated = 1
srv  update_slots: all slots are idle

[ggerganov]

Use --no-context-shift

[remixer-dec]

@ggerganov with this option, it responds with {"error":{"code":500,"message":"context shift is disabled","type":"server_error"}}, in all cases except for when I run it with -n -2 and remove max_tokens from request, in this case it returns 1 token

[aviallon]

@remixer-dec you can also try to reduce the batch-size: --ctx-size 1024 -b 512 would probably work.

[remixer-dec]

@aviallon thanks for suggestion, tried it, didn't work.

[remixer-dec]

Update: I think it is related to the --parallel argument, when I change it from 10 to 1, I don't see any repetitions, but people reported this behavior with different settings

[ggerganov]

Increasing -np will reduce the per-slot context, so make sure to adjust the context size. With --ctx-size 1024 --parallel 10 you have only 100 tokens per slot which is very small for any meaningful task.

Also, it's recommended to use greedy sampling (e.g. --top-k 1) for these kind of tasks. Other sampling settings will likely degrade the overall quality.

[aviallon]

Increasing -np will reduce the per-slot context, so make sure to adjust the context size. With --ctx-size 1024 --parallel 10 you have only 100 tokens per slot which is very small for any meaningful task.

Also, it's recommended to use greedy sampling (e.g. --top-k 1) for these kind of tasks. Other sampling settings will likely degrade the overall quality.

I remarked that model tend to repeat themselves more if ctx-size-per-slot < batch-size.
Perhaps outputting an error in that case would make sense?

[remixer-dec]

Is there any documentation about how slots are implemented and what is the correct use for them?

I feel like it is a very useful feature, but I am misunderstanding how it works. In case of SGlang it’s pretty obvious, you hit the server with a bunch of requests with similar tokens,they are processed together reusing the parts that are the same, with llama.cpp it’s a bit more tricky from my understanding, does each slot have it’s own kv cache? Who decides to which slot does a request go? Do idle slots guarantee immediate responses when others are busy? Is this why idle slots are affecting the output quality of a single used slot?