Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

🐛 Bug Report: Local model is recreated for each request, causing delay and out of memory errors #945

Open
2 tasks done
starkgate opened this issue May 10, 2024 · 0 comments
Open
2 tasks done

🐛 Bug Report: Local model is recreated for each request, causing delay and out of memory errors #945

starkgate opened this issue May 10, 2024 · 0 comments
Assignees
@starkgate

Comments

@starkgate
Copy link

📜 Description

I am using the docsgpt-7b-mistral.Q8_0.gguf model on a 1080 Ti. The model fits in the VRAM with 300MB to spare. I can query the model and it works. However, I noticed that querying the model a second time causes an out of memory error. After investigation, this is caused by DocsGPT creating a new instance of the model to answer my second query, instead of reusing the previous one or at least removing the first one before creating a new instance. Even if I had enough memory, creating a new instance causes an unnecessary delay with each query.

👟 Reproduction steps

I am running DocsGPT 0.9.0 locally (./setup.sh with option 2). The environment is configured to use llama-cpp with huggingface_sentence-transformers/all-mpnet-base-v2 embeddings. To reproduce the bug, monitor the VRAM with nvidia-smi then ask at least 2 questions with this environment. Notice how the VRAM usage either doubles (if you have enough VRAM) or DocsGPT runs out of memory.

👍 Expected behavior

The LLM instance should be cached for future queries.

👎 Actual Behavior with Screenshots

DocsGPT runs out of memory

💻 Operating system

Windows

What browsers are you seeing the problem on?

Firefox, Chrome

🤖 What development environment are you experiencing this bug on?

Docker

🔒 Did you set the correct environment variables in the right path? List the environment variable names (not values please!)

CELERY_BROKER_URL
CELERY_RESULT_BACKEND
EMBEDDINGS_NAME
FLASK_APP
FLASK_DEBUG
LLM_NAME
VITE_API_STREAMING

📃 Provide any additional context for the Bug.

I have fixed this issue on my side by modifying the llama_cpp.py script to cache and reuse LLM instances. This requires the flask-caching package and some modifications to llms, retrievers and the answer API.

Here is the implementation in llama_cpp.py:

singleton_llm = {
    'type': None,
    'llm': None
}

def create_llm(self, type, api_key, user_api_key, *args, **kwargs):
    llm_class = self.llms.get(type.lower())
    if not llm_class:
        raise ValueError(f"No LLM class found for type {type}")

    # do not create a new LLM (and allocate memory again) for each request for local models
    if self.singleton_llm['type'] != llm_class or self.singleton_llm['type'] != LlamaCpp:
        llm = llm_class(api_key, user_api_key, *args, **kwargs)
        self.singleton_llm['type'] = llm_class
        self.singleton_llm['llm'] = llm

    return self.singleton_llm['llm']

📖 Relevant log output

* Restarting with stat
 * Debugger is active!
 * Debugger PIN: 760-653-407
192.168.2.98 - - [10/May/2024 15:34:15] "GET /api/combine HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:34:15] "GET /api/combine HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:34:15] "GET /api/get_conversations HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:34:15] "GET /api/get_conversations HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:34:34] "GET /api/get_single_conversation?id=663e1d2c8f046fbf993c849b HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:34:37] "OPTIONS /stream HTTP/1.1" 200 -
llama_model_loader: loaded meta data with 25 key-value pairs and 291 tensors from /home/jstark/DocsGPT/models/docsgpt-7b-mistral.Q8_0.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              = arc53_docsgpt-7b-mistral
llama_model_loader: - kv   2:                       llama.context_length u32              = 32768
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              = 7
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  16:                      tokenizer.ggml.merges arr[str,58980]   = ["▁ t", "i n", "e r", "▁ a", "h e...
llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  19:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 2
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% for message in messages %}\n{% if m...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q8_0:  226 tensors
llm_load_vocab: special tokens definition check successful ( 259/32000 ).
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          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 32768
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  = 32768
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       = 8B
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 7.24 B
llm_load_print_meta: model size       = 7.17 GiB (8.50 BPW)
llm_load_print_meta: general.name     = arc53_docsgpt-7b-mistral
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 2 '</s>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
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 GeForce GTX 1080 Ti, compute capability 6.1, VMM: yes
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:        CPU buffer size =   132.81 MiB
llm_load_tensors:      CUDA0 buffer size =  7205.83 MiB
...................................................................................................
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      CUDA0 KV buffer size =   256.00 MiB
llama_new_context_with_model: KV self size  =  256.00 MiB, K (f16):  128.00 MiB, V (f16):  128.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.12 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =   164.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    12.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2
AVX = 1 | AVX_VNNI = 0 | 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 = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |
Model metadata: {'tokenizer.chat_template': "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n'  + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}", 'tokenizer.ggml.add_eos_token': 'false', 'tokenizer.ggml.padding_token_id': '2', 'tokenizer.ggml.unknown_token_id': '0', 'tokenizer.ggml.eos_token_id': '2', 'general.architecture': 'llama', 'llama.rope.freq_base': '10000.000000', 'llama.context_length': '32768', 'general.name': 'arc53_docsgpt-7b-mistral', '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': '7'}
Using gguf chat template: {% for message in messages %}
{% if message['role'] == 'user' %}
{{ '<|user|>
' + message['content'] + eos_token }}
{% elif message['role'] == 'system' %}
{{ '<|system|>
' + message['content'] + eos_token }}
{% elif message['role'] == 'assistant' %}
{{ '<|assistant|>
'  + message['content'] + eos_token }}
{% endif %}
{% if loop.last and add_generation_prompt %}
{{ '<|assistant|>' }}
{% endif %}
{% endfor %}
Using chat eos_token: </s>
Using chat bos_token: <s>
192.168.2.98 - - [10/May/2024 15:34:44] "POST /stream HTTP/1.1" 200 -

llama_print_timings:        load time =     470.03 ms
llama_print_timings:      sample time =     190.46 ms /   369 runs   (    0.52 ms per token,  1937.37 tokens per second)
llama_print_timings: prompt eval time =     469.84 ms /   121 tokens (    3.88 ms per token,   257.54 tokens per second)
llama_print_timings:        eval time =   10895.33 ms /   368 runs   (   29.61 ms per token,    33.78 tokens per second)
llama_print_timings:       total time =   12657.91 ms /   489 tokens
192.168.2.98 - - [10/May/2024 15:34:56] "OPTIONS /api/search HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:34:56] "GET /api/get_conversations HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:34:58] "POST /api/search HTTP/1.1" 200 -
192.168.2.98 - - [10/May/2024 15:35:01] "OPTIONS /stream HTTP/1.1" 200 -
llama_model_loader: loaded meta data with 25 key-value pairs and 291 tensors from /home/jstark/DocsGPT/models/docsgpt-7b-mistral.Q8_0.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              = arc53_docsgpt-7b-mistral
llama_model_loader: - kv   2:                       llama.context_length u32              = 32768
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              = 7
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  16:                      tokenizer.ggml.merges arr[str,58980]   = ["▁ t", "i n", "e r", "▁ a", "h e...
llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  19:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 2
llama_model_loader: - kv  21:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  22:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  23:                    tokenizer.chat_template str              = {% for message in messages %}\n{% if m...
llama_model_loader: - kv  24:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q8_0:  226 tensors
llm_load_vocab: special tokens definition check successful ( 259/32000 ).
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          = 32000
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 32768
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  = 32768
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       = 8B
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 7.24 B
llm_load_print_meta: model size       = 7.17 GiB (8.50 BPW)
llm_load_print_meta: general.name     = arc53_docsgpt-7b-mistral
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 2 '</s>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 2 '</s>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
llm_load_tensors: ggml ctx size =    0.30 MiB
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 7205.83 MiB on device 0: cudaMalloc failed: out of memory
llama_model_load: error loading model: unable to allocate backend buffer
llama_load_model_from_file: failed to load model
Debugging middleware caught exception in streamed response at a point where response headers were already sent.
Traceback (most recent call last):
  File "/home/jstark/.local/lib/python3.10/site-packages/werkzeug/wsgi.py", line 256, in __next__
    return self._next()
  File "/home/jstark/.local/lib/python3.10/site-packages/werkzeug/wrappers/response.py", line 32, in _iter_encoded
    for item in iterable:
  File "/home/jstark/DocsGPT/application/api/answer/routes.py", line 183, in complete_stream
    for line in answer:
  File "/home/jstark/DocsGPT/application/retriever/classic_rag.py", line 99, in gen
    llm = cache.get("llm_creator").create_llm(
  File "/home/jstark/DocsGPT/application/llm/llm_creator.py", line 35, in create_llm
    llm = llm_class(api_key, user_api_key, *args, **kwargs)
  File "/home/jstark/DocsGPT/application/llm/llama_cpp.py", line 26, in __init__
    llama = Llama(model_path=llm_name, n_ctx=2048, n_gpu_layers=-1)
  File "/home/jstark/.conda/envs/docsgpt/lib/python3.10/site-packages/llama_cpp/llama.py", line 333, in __init__
    self._model = _LlamaModel(
  File "/home/jstark/.conda/envs/docsgpt/lib/python3.10/site-packages/llama_cpp/_internals.py", line 57, in __init__
    raise ValueError(f"Failed to load model from file: {path_model}")
ValueError: Failed to load model from file: /home/jstark/DocsGPT/models/docsgpt-7b-mistral.Q8_0.gguf

👀 Have you spent some time to check if this bug has been raised before?

  • I checked and didn't find similar issue

🔗 Are you willing to submit PR?

Yes, I am willing to submit a PR!

🧑‍⚖️ Code of Conduct

  • I agree to follow this project's Code of Conduct
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@starkgate starkgate

Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
1 participant
@starkgate