Segmentation fault (core dumped) - 0.2.58

[anakin87]

Prerequisites

Please answer the following questions for yourself before submitting an issue.

• I am running the latest code. Development is very rapid so there are no tagged versions as of now.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new bug or useful enhancement to share.

Expected Behavior

I'm trying to create a completion using a GGUF model

Current Behavior

from llama_cpp import Llama

model = Llama(model_path="./tests/models/openchat-3.5-1210.Q3_K_S.gguf",  n_ctx=128, n_batch=128)

questions_and_answers = [
    ("What's the capital of France?", "Paris"),
    ("What is the capital of Canada?", "Ottawa"),
    ("What is the capital of Ghana?", "Accra"),
]

for i, (question, answer) in enumerate(questions_and_answers):
    prompt = f"GPT4 Correct User: Answer in a single word. {question} <|end_of_turn|>\n GPT4 Correct Assistant:"
    result = model.create_completion(prompt=prompt)

    print(i)
    print(result)

Segmentation fault (core dumped) on 0.2.58
(works well on 0.2.57)

Environment and Context

• all GithHub runners: Linux, macOS, Windows https://github.com/deepset-ai/haystack-core-integrations/actions/runs/8515449565
• my Ubuntu laptop

Failure Information (for bugs)

Segmentation fault (core dumped)

[riedgar-ms]

We have seen this issue in the guidance project.

[anakin87]

Thanks!
Related: guidance-ai/guidance#735

[sepcnt]

Here is a rough bug trace. It seems that any context longer than a certain bounder would cause the crash down

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA A100-PCIE-40GB, compute capability 8.0, VMM: yes
Generating Responses:   0%|                                                                                                                                                       | 0/96 [00:00<?, ?it/s]Fatal Python error: Segmentation fault

Thread 0x00007f8851089700 (most recent call first):
  File "~/miniforge3/lib/python3.10/threading.py", line 324 in wait
  File "~/miniforge3/lib/python3.10/threading.py", line 607 in wait
  File "~/miniforge3/lib/python3.10/site-packages/tqdm/_monitor.py", line 60 in run
  File "~/miniforge3/lib/python3.10/threading.py", line 1016 in _bootstrap_inner
  File "~/miniforge3/lib/python3.10/threading.py", line 973 in _bootstrap

Current thread 0x00007f8a06f84740 (most recent call first):
  File "~/miniforge3/lib/python3.10/site-packages/guidance/models/llama_cpp/_llama_cpp.py", line 165 in get_logits
  File "~/miniforge3/lib/python3.10/site-packages/guidance/models/_model.py", line 624 in __call__
  File "~/miniforge3/lib/python3.10/site-packages/guidance/models/_model.py", line 1188 in _run_stateless
  File "~/miniforge3/lib/python3.10/site-packages/guidance/models/_model.py", line 992 in __add__
  File "~/miniforge3/lib/python3.10/site-packages/guidance/models/_model.py", line 984 in __add__
  File "~/workspace/models/inference.py", line 100 in predict


Extension modules: (total: 224)
Segmentation fault (core dumped)

Probably relevant to ggml-org/llama.cpp#6017 (comment)

[abetlen]

I'll work on a fix once I get a chance to repro, in the meantime let me share the recipe for debugging this (on Linux at least).

# install the package in debug mode to maintain symbols (may need to add additional cmake flags for specific backends)
python3 -m pip install \
  --verbose \
  --config-settings cmake.args='-DCMAKE_BUILD_TYPE=Debug;-DCMAKE_CXX_FLAGS=-g3;-DCMAKE_C_FLAGS=-g3' \
  --config-settings cmake.verbose=true \
  --config-settings logging.level=INFO \
  --config-settings install.strip=false \
  --editable .
# run test script with gdb
gdb --args python3 test_script.py

[anakin87]

(gdb) run try_llama.py 
Starting program: /home/anakin87/apps/haystack-core-integrations/integrations/llama_cpp/.hatch/llama-cpp-haystack/bin/python try_llama.py
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[New Thread 0x7ffff2dff640 (LWP 25295)]
[New Thread 0x7ffff25fe640 (LWP 25296)]
[New Thread 0x7fffeddfd640 (LWP 25297)]
[New Thread 0x7fffed5fc640 (LWP 25298)]
[New Thread 0x7fffeadfb640 (LWP 25299)]
[New Thread 0x7fffe65fa640 (LWP 25300)]
[New Thread 0x7fffe3df9640 (LWP 25301)]
[New Thread 0x7fffe35f8640 (LWP 25302)]
[New Thread 0x7fffe0df7640 (LWP 25303)]
[New Thread 0x7fffdc5f6640 (LWP 25304)]
[New Thread 0x7fffdbdf5640 (LWP 25305)]
[New Thread 0x7fffd75f4640 (LWP 25306)]
[New Thread 0x7fffd4df3640 (LWP 25307)]
[New Thread 0x7fffd25f2640 (LWP 25308)]
[New Thread 0x7fffcfdf1640 (LWP 25309)]
[New Thread 0x7fffcd5f0640 (LWP 25310)]
[New Thread 0x7fffcadef640 (LWP 25311)]
[New Thread 0x7fffca5ee640 (LWP 25312)]
[New Thread 0x7fffc7ded640 (LWP 25313)]
llama_model_loader: loaded meta data with 23 key-value pairs and 291 tensors from ./tests/models/openchat-3.5-1210.Q3_K_S.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              = llama
llama_model_loader: - kv   1:                               general.name str              = openchat_openchat-3.5-1210
llama_model_loader: - kv   2:                       llama.context_length u32              = 8192
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   4:                          llama.block_count u32              = 32
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  11:                          general.file_type u32              = 11
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,32002]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,32002]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,32002]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 32000
llama_model_loader: - kv  18:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  19:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  20:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  21:                    tokenizer.chat_template str              = {{ bos_token }}{% for message in mess...
llama_model_loader: - kv  22:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q3_K:  225 tensors
llama_model_loader: - type q6_K:    1 tensors
llm_load_vocab: special tokens definition check successful ( 261/32002 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32002
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 8192
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 4
llm_load_print_meta: n_embd_k_gqa     = 1024
llm_load_print_meta: n_embd_v_gqa     = 1024
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 14336
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 8192
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: model type       = 7B
llm_load_print_meta: model ftype      = Q3_K - Small
llm_load_print_meta: model params     = 7.24 B
llm_load_print_meta: model size       = 2.95 GiB (3.50 BPW) 
llm_load_print_meta: general.name     = openchat_openchat-3.5-1210
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 32000 '<|end_of_turn|>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
llm_load_tensors: ggml ctx size =    0.11 MiB
llm_load_tensors:        CPU buffer size =  3017.28 MiB
.................................................................................................
llama_new_context_with_model: n_ctx      = 128
llama_new_context_with_model: n_batch    = 128
llama_new_context_with_model: n_ubatch   = 128
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =    16.00 MiB
llama_new_context_with_model: KV self size  =   16.00 MiB, K (f16):    8.00 MiB, V (f16):    8.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.12 MiB
llama_new_context_with_model:        CPU compute buffer size =    20.06 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 1
AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | 
Model metadata: {'tokenizer.chat_template': "{{ bos_token }}{% for message in messages %}{{ 'GPT4 Correct ' + message['role'].title() + ': ' + message['content'] + '<|end_of_turn|>'}}{% endfor %}{% if add_generation_prompt %}{{ 'GPT4 Correct Assistant:' }}{% endif %}", 'tokenizer.ggml.add_eos_token': 'false', 'tokenizer.ggml.padding_token_id': '0', 'tokenizer.ggml.eos_token_id': '32000', 'general.architecture': 'llama', 'llama.rope.freq_base': '10000.000000', 'llama.context_length': '8192', 'general.name': 'openchat_openchat-3.5-1210', 'tokenizer.ggml.add_bos_token': 'true', 'llama.embedding_length': '4096', 'llama.feed_forward_length': '14336', 'llama.attention.layer_norm_rms_epsilon': '0.000010', 'llama.rope.dimension_count': '128', 'tokenizer.ggml.bos_token_id': '1', 'llama.attention.head_count': '32', 'llama.block_count': '32', 'llama.attention.head_count_kv': '8', 'general.quantization_version': '2', 'tokenizer.ggml.model': 'llama', 'general.file_type': '11'}
Using gguf chat template: {{ bos_token }}{% for message in messages %}{{ 'GPT4 Correct ' + message['role'].title() + ': ' + message['content'] + '<|end_of_turn|>'}}{% endfor %}{% if add_generation_prompt %}{{ 'GPT4 Correct Assistant:' }}{% endif %}
Using chat eos_token: <|end_of_turn|>
Using chat bos_token: <s>
[New Thread 0x7fff01f3a640 (LWP 25314)]
[New Thread 0x7fff01739640 (LWP 25315)]
[New Thread 0x7fff00f38640 (LWP 25316)]
[New Thread 0x7fff00737640 (LWP 25317)]
[New Thread 0x7ffefff36640 (LWP 25318)]
[New Thread 0x7ffeff735640 (LWP 25319)]
[New Thread 0x7ffefef34640 (LWP 25320)]
[New Thread 0x7ffefe733640 (LWP 25321)]
[New Thread 0x7ffefdf32640 (LWP 25322)]
[Thread 0x7ffefdf32640 (LWP 25322) exited]
[Thread 0x7ffefe733640 (LWP 25321) exited]
[Thread 0x7ffefef34640 (LWP 25320) exited]
[Thread 0x7ffeff735640 (LWP 25319) exited]
[Thread 0x7ffefff36640 (LWP 25318) exited]
[Thread 0x7fff00737640 (LWP 25317) exited]
[Thread 0x7fff00f38640 (LWP 25316) exited]
[Thread 0x7fff01739640 (LWP 25315) exited]
[Thread 0x7fff01f3a640 (LWP 25314) exited]

Thread 1 "python" received signal SIGSEGV, Segmentation fault.
0x00007ffff7ba7d83 in ?? () from /usr/lib/python3.10/lib-dynload/_ctypes.cpython-310-x86_64-linux-gnu.so

[abetlen]

@anakin87 ah okay I think I see what it is, can you set Llama(..., logits_all=True) in your test, almost certain that's it, I'll work on a fix.

[abetlen]

Long-term I'll set up some tests that use Qwen1.5 0.5B or some other small model to smoke test for issues like this.

[anakin87]

@abetlen setting logits_all=True fixes the problem.