parallel.cpp exits when encountering a long prompt.

[littlebai3618]

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

In ./examples/parallel/parallel.cpp, I added the following two lines to the final output:

int cache_count = llama_get_kv_cache_token_count(ctx);
LOG_TEE("Cache KV size %d", cache_count);

I believe that the logic in line 221 of parallel.cpp:

// all sequences have ended - clear the entire KV cache
for (int i = 0; i < n_clients; ++i) {
   llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
}

should release all the occupied cache when the entire task is completed. However, in reality, it does not release the cache.

Current Behavior

I expected the value of cache_count to be 0, but in reality, it is 1153.

Environment and Context

1. GPU info

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.85.05    Driver Version: 525.85.05    CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Quadro RTX 8000     Off  | 00000000:3B:00.0 Off |                  N/A |
| 33%   29C    P8    22W / 260W |      0MiB / 49152MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

2. make cmd

LLAMA_CUDA_NVCC=/usr/local/cuda-12/bin/nvcc make LLAMA_CUBLAS=1 -Wdeprecated-declarations

3. test cmd

./parallel -m ./CodeLlama-7B/ggml-model-q8_0.gguf -n -1 -c 16324 -b 4096 --cont_batching --parallel 10 --sequences 600 --n-gpu-layers 1000

4. model

It appears that you are using the "CodeLlama-7B-HF" model from the repository you mentioned (https://huggingface.co/codellama/CodeLlama-7b-hf). You mentioned that you performed the conversion using the "convert.py" script included in the repository.

$ lscpu

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   46 bits physical, 48 bits virtual
CPU(s):                          64
On-line CPU(s) list:             0-63
Thread(s) per core:              2
Core(s) per socket:              16
Socket(s):                       2
NUMA node(s):                    2
Vendor ID:                       GenuineIntel
CPU family:                      6
Model:                           85
Model name:                      Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz
Stepping:                        7
CPU MHz:                         1693.308
CPU max MHz:                     3900.0000
CPU min MHz:                     1000.0000
BogoMIPS:                        4600.00
Virtualization:                  VT-x
L1d cache:                       1 MiB
L1i cache:                       1 MiB
L2 cache:                        32 MiB
L3 cache:                        44 MiB

• Operating System, e.g. for Linux:

$ uname -a

Linux studio-0 4.19.96 #1 SMP Tue Mar 10 10:34:01 CST 2020 x86_64 x86_64 x86_64 GNU/Linux

• SDK version, e.g. for Linux:

$ python3 --version 3.11.5
$ make --version  GNU Make 4.2.1
$ g++ --version  g++ (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0

Failure Information (for bugs)

I expected the value of cache_count to be 0, but in reality, it is 1153.

Steps to Reproduce

please download model from https://huggingface.co/codellama/CodeLlama-7b-hf

1. python convert.py ./CodeLlama-7B/ --outtype q8_0
2. after parallel.cpp's LOG_TEE("Cache misses: %6d\n", n_cache_miss); add

int cache_count = llama_get_kv_cache_token_count(ctx);
LOG_TEE("Cache KV size %d", cache_count);

3. ./parallel -m ./CodeLlama-7B/ggml-model-q8_0.gguf -n -1 -c 16324 -b 4096 --cont_batching --parallel 10 --sequences 600 --n-gpu-layers 1000

Failure Logs

Example run with the Linux command perf

ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 CUDA devices:
  Device 0: Quadro RTX 8000, compute capability 7.5
llama_model_loader: loaded meta data with 19 key-value pairs and 291 tensors from /aistudio/workspace/system-default/models/CodeLlama-7B/ggml-model-q8_0.gguf (version GGUF V3 (latest))
llama_model_loader: - tensor    0:                token_embd.weight q8_0     [  4096, 32016,     1,     1 ]
llama_model_loader: - tensor    1:           blk.0.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor    2:            blk.0.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor    3:            blk.0.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor    4:              blk.0.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor    5:            blk.0.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor    6:              blk.0.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor    7:         blk.0.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor    8:              blk.0.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor    9:              blk.0.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   10:           blk.1.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   11:            blk.1.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   12:            blk.1.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   13:              blk.1.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   14:            blk.1.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   15:              blk.1.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   16:         blk.1.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   17:              blk.1.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   18:              blk.1.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   19:          blk.10.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   20:           blk.10.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   21:           blk.10.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   22:             blk.10.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   23:           blk.10.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   24:             blk.10.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   25:        blk.10.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   26:             blk.10.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   27:             blk.10.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   28:          blk.11.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   29:           blk.11.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   30:           blk.11.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   31:             blk.11.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   32:           blk.11.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   33:             blk.11.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   34:        blk.11.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   35:             blk.11.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   36:             blk.11.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   37:          blk.12.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   38:           blk.12.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   39:           blk.12.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   40:             blk.12.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   41:           blk.12.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   42:             blk.12.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   43:        blk.12.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   44:             blk.12.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   45:             blk.12.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   46:          blk.13.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   47:           blk.13.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   48:           blk.13.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   49:             blk.13.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   50:           blk.13.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   51:             blk.13.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   52:        blk.13.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   53:             blk.13.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   54:             blk.13.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   55:          blk.14.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   56:           blk.14.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   57:           blk.14.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   58:             blk.14.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   59:           blk.14.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   60:             blk.14.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   61:        blk.14.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   62:             blk.14.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   63:             blk.14.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   64:          blk.15.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   65:           blk.15.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   66:           blk.15.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   67:             blk.15.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   68:           blk.15.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   69:             blk.15.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   70:        blk.15.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   71:             blk.15.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   72:             blk.15.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   73:          blk.16.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   74:           blk.16.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   75:           blk.16.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   76:             blk.16.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   77:           blk.16.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   78:             blk.16.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   79:        blk.16.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   80:             blk.16.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   81:             blk.16.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   82:          blk.17.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   83:           blk.17.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   84:           blk.17.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   85:             blk.17.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   86:           blk.17.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   87:             blk.17.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   88:        blk.17.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   89:             blk.17.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   90:             blk.17.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   91:          blk.18.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   92:           blk.18.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor   93:           blk.18.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   94:             blk.18.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor   95:           blk.18.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   96:             blk.18.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   97:        blk.18.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   98:             blk.18.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   99:             blk.18.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  100:          blk.19.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  101:           blk.19.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  102:           blk.19.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  103:             blk.19.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  104:           blk.19.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  105:             blk.19.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  106:        blk.19.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  107:             blk.19.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  108:             blk.19.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  109:           blk.2.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  110:            blk.2.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  111:            blk.2.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  112:              blk.2.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  113:            blk.2.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  114:              blk.2.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  115:         blk.2.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  116:              blk.2.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  117:              blk.2.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  118:          blk.20.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  119:           blk.20.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  120:           blk.20.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  121:             blk.20.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  122:           blk.20.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  123:             blk.20.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  124:        blk.20.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  125:             blk.20.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  126:             blk.20.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  127:          blk.21.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  128:           blk.21.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  129:           blk.21.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  130:             blk.21.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  131:           blk.21.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  132:             blk.21.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  133:        blk.21.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  134:             blk.21.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  135:             blk.21.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  136:          blk.22.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  137:           blk.22.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  138:           blk.22.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  139:             blk.22.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  140:           blk.22.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  141:             blk.22.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  142:        blk.22.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  143:             blk.22.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  144:             blk.22.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  145:          blk.23.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  146:           blk.23.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  147:           blk.23.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  148:             blk.23.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  149:           blk.23.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  150:             blk.23.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  151:        blk.23.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  152:             blk.23.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  153:             blk.23.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  154:           blk.3.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  155:            blk.3.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  156:            blk.3.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  157:              blk.3.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  158:            blk.3.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  159:              blk.3.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  160:         blk.3.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  161:              blk.3.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  162:              blk.3.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  163:           blk.4.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  164:            blk.4.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  165:            blk.4.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  166:              blk.4.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  167:            blk.4.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  168:              blk.4.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  169:         blk.4.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  170:              blk.4.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  171:              blk.4.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  172:           blk.5.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  173:            blk.5.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  174:            blk.5.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  175:              blk.5.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  176:            blk.5.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  177:              blk.5.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  178:         blk.5.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  179:              blk.5.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  180:              blk.5.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  181:           blk.6.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  182:            blk.6.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  183:            blk.6.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  184:              blk.6.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  185:            blk.6.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  186:              blk.6.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  187:         blk.6.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  188:              blk.6.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  189:              blk.6.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  190:           blk.7.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  191:            blk.7.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  192:            blk.7.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  193:              blk.7.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  194:            blk.7.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  195:              blk.7.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  196:         blk.7.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  197:              blk.7.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  198:              blk.7.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  199:           blk.8.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  200:            blk.8.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  201:            blk.8.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  202:              blk.8.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  203:            blk.8.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  204:              blk.8.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  205:         blk.8.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  206:              blk.8.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  207:              blk.8.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  208:           blk.9.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  209:            blk.9.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  210:            blk.9.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  211:              blk.9.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  212:            blk.9.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  213:              blk.9.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  214:         blk.9.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  215:              blk.9.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  216:              blk.9.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  217:                    output.weight q8_0     [  4096, 32016,     1,     1 ]
llama_model_loader: - tensor  218:          blk.24.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  219:           blk.24.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  220:           blk.24.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  221:             blk.24.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  222:           blk.24.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  223:             blk.24.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  224:        blk.24.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  225:             blk.24.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  226:             blk.24.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  227:          blk.25.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  228:           blk.25.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  229:           blk.25.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  230:             blk.25.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  231:           blk.25.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  232:             blk.25.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  233:        blk.25.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  234:             blk.25.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  235:             blk.25.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  236:          blk.26.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  237:           blk.26.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  238:           blk.26.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  239:             blk.26.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  240:           blk.26.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  241:             blk.26.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  242:        blk.26.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  243:             blk.26.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  244:             blk.26.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  245:          blk.27.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  246:           blk.27.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  247:           blk.27.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  248:             blk.27.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  249:           blk.27.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  250:             blk.27.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  251:        blk.27.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  252:             blk.27.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  253:             blk.27.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  254:          blk.28.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  255:           blk.28.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  256:           blk.28.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  257:             blk.28.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  258:           blk.28.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  259:             blk.28.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  260:        blk.28.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  261:             blk.28.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  262:             blk.28.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  263:          blk.29.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  264:           blk.29.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  265:           blk.29.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  266:             blk.29.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  267:           blk.29.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  268:             blk.29.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  269:        blk.29.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  270:             blk.29.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  271:             blk.29.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  272:          blk.30.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  273:           blk.30.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  274:           blk.30.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  275:             blk.30.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  276:           blk.30.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  277:             blk.30.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  278:        blk.30.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  279:             blk.30.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  280:             blk.30.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  281:          blk.31.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  282:           blk.31.ffn_down.weight q8_0     [ 11008,  4096,     1,     1 ]
llama_model_loader: - tensor  283:           blk.31.ffn_gate.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  284:             blk.31.ffn_up.weight q8_0     [  4096, 11008,     1,     1 ]
llama_model_loader: - tensor  285:           blk.31.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  286:             blk.31.attn_k.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  287:        blk.31.attn_output.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  288:             blk.31.attn_q.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  289:             blk.31.attn_v.weight q8_0     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  290:               output_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - kv   0:                       general.architecture str     
llama_model_loader: - kv   1:                               general.name str     
llama_model_loader: - kv   2:                       llama.context_length u32     
llama_model_loader: - kv   3:                     llama.embedding_length u32     
llama_model_loader: - kv   4:                          llama.block_count u32     
llama_model_loader: - kv   5:                  llama.feed_forward_length u32     
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32     
llama_model_loader: - kv   7:                 llama.attention.head_count u32     
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32     
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32     
llama_model_loader: - kv  10:                       llama.rope.freq_base f32     
llama_model_loader: - kv  11:                          general.file_type u32     
llama_model_loader: - kv  12:                       tokenizer.ggml.model str     
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr     
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr     
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr     
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32     
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32     
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32     
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q8_0:  226 tensors
llm_load_vocab: mismatch in special tokens definition ( 264/32016 vs 259/32016 ).
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          = 32016
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 16384
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_gqa            = 1
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: n_ff             = 11008
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 16384
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: model type       = 7B
llm_load_print_meta: model ftype      = mostly Q8_0
llm_load_print_meta: model params     = 6.74 B
llm_load_print_meta: model size       = 6.67 GiB (8.50 BPW) 
llm_load_print_meta: general.name   = models
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: LF token  = 13 '<0x0A>'
llm_load_tensors: ggml ctx size =    0.11 MB
llm_load_tensors: using CUDA for GPU acceleration
llm_load_tensors: mem required  =  132.99 MB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 35/35 layers to GPU
llm_load_tensors: VRAM used: 6695.89 MB
....................................................................................................
llama_new_context_with_model: n_ctx      = 16324
llama_new_context_with_model: freq_base  = 1000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: offloading v cache to GPU
llama_kv_cache_init: offloading k cache to GPU
llama_kv_cache_init: VRAM kv self = 8162.00 MB
llama_new_context_with_model: kv self size  = 8162.00 MB
llama_build_graph: non-view tensors processed: 740/740
llama_new_context_with_model: compute buffer total size = 8615.72 MB
llama_new_context_with_model: VRAM scratch buffer: 8609.09 MB
llama_new_context_with_model: total VRAM used: 23466.99 MB (model: 6695.89 MB, context: 16771.09 MB)

No new questions so proceed with build-in defaults.


main: Simulating parallel requests from clients:
main: n_parallel = 10, n_sequences = 600, cont_batching = 1, system tokens = 305


。。。。。

main: clearing the KV cache

run parameters as at 2023-11-15 10:43:39

main: n_parallel = 10, n_sequences = 600, cont_batching = 1, system tokens = 305
External prompt file: used built-in defaults
Model and path used:  ./CodeLlama-7B/ggml-model-q8_0.gguf

Total prompt tokens:   8752, speed: 36.53 t/s
Total gen tokens:     29266, speed: 122.15 t/s
Total speed (AVG):           speed: 158.68 t/s
Cache misses:             0
Cache KV size 628

llama_print_timings:        load time =   37371.88 ms
llama_print_timings:      sample time =   12907.57 ms / 29866 runs   (    0.43 ms per token,  2313.84 tokens per second)
llama_print_timings: prompt eval time =  214845.96 ms / 38277 tokens (    5.61 ms per token,   178.16 tokens per second)
llama_print_timings:        eval time =     936.39 ms /    46 runs   (   20.36 ms per token,    49.13 tokens per second)
llama_print_timings:       total time =  239590.67 ms

[KerfuffleV2]

Your cache actually got cleared correctly (probably, anyway!). The problem is that llama_get_kv_cache_token_count isn't doing what the name suggests. It's actually doing something like returning the index of the last KV cache cell that's populated.

[littlebai3618]

Your cache actually got cleared correctly (probably, anyway!). The problem is that llama_get_kv_cache_token_count isn't doing what the name suggests. It's actually doing something like returning the index of the last KV cache cell that's populated.

Thank you for your response. I am using the low-level API of llama-cpp-python (which is completely consistent with llama.h) to implement dynamic batch processing in Python, following the example in parallel.cpp. However, I am encountering a strange phenomenon where running the program with parallel > 1 causes an additional increase of 30-200MB in GPU memory when reaching the llama_decode() method. This increase accumulates over time until the program crashes. Yesterday, I suspected that it might be an issue with KV_cache. Currently, I am reviewing the differences between my implementation and parallel.cpp. Do you have any suggestions? For example, which parts might be causing the memory leak? Today, I observed that after the abnormal increase in GPU memory, llama_get_state_size also shows a significant increase. I'm not sure if there is any correlation. It seems that this might be related to a variable in the context.

[KerfuffleV2]

Do you have any suggestions?

Well, I can't tell you what the problem is but I can basically rule out the KV cache for you. You can pretty much trust that clearing the KV cache works correctly, but even if it didn't, it wouldn't matter for the purposes of memory leaks. As far as I know, all the KV cache memory gets allocated up front based on the context size you set. So for the most part, it just doesn't matter what's in it.

This increase accumulates over time until the program crashes.

Are you saying that the memory increases while running decode on a large parallel sequence, or that the memory continues increasing in between calls to decode with the parallel sequence?

In other words, when the call to decode ends and you get your result, does the memory usage go back down?

I am using the low-level API of llama-cpp-python

I'd say you'd probably have better luck asking in their repo. These other projects that build on llama.cpp aren't necessarily using the latest version, they may have their own patches applied, they may be tweaking settings, etc.

Debugging problems is basically a process of elimination, and there are too many unknowns for someone who just knows about llama.cpp to deal with in this case. Or, you can try reproducing the issue using the latest version of llama.cpp directly.

[ggerganov]

Diagnosing the llama-cpp-python problem would be difficult, and I don't have anything to add to @KerfuffleV2's comment.

Regarding llama_get_kv_cache_token_count() - as mentioned, it currently does not work and is also deprecated. However, I think we can easily restore it's functionality by counting the number of tokens in the kv cache functions. Might be useful for debugging stuff, so we should probably de-deprecate it

[KerfuffleV2]

I think we can easily restore it's functionality by counting the number of tokens in the kv cache functions.

Definitely would be easy. Slightly harder is answering the question "what's a token"? Should it return the number of populated cells or the sum of the cells sequence lengths? Like if a cell belongs to 10 sequences, is that 10 tokens?

[ggerganov]

I guess, we can have 2 counters - "number of tokens" and "number of occupied cells". And add API for the cells

[KerfuffleV2]

If we're changing the API, how about something that basically just exports the whole KV cache state so people can extract whatever information is useful? Maybe even add the token id to it. Even for a 200,000 context size, that's only 800k size it's a 32bit type.

Something like the batch api when you create/destroy a batch, then another function could copy state into it (since you wouldn't want to allocate every time it was fetched probably).

#4035 wanted that functionality, I think (and also having the count function fixed as well).

[littlebai3618]

Are you saying that the memory increases while running decode on a large parallel sequence, or that the memory continues increasing in between calls to decode with the parallel sequence?

In other words, when the call to decode ends and you get your result, does the memory usage go back down?

The memory usage increases during large-scale parallel decoding, but it doesn't increase with every decoding operation. Once the memory usage increases, it doesn't decrease even as the number of completed sentences increases. It keeps accumulating.

1. I have raised a new issue on llama-cpp-python: #924.
2. I suspect that the issue might be with my code or with the Python C-bindings.
   Lastly, thank you very much for your response. My current work goal is to implement continuous batch processing in Python, so if you have any guesses or insights regarding the potential memory leak, please feel free to let me know.

Additionally, could you provide more detailed comments for the parallel.cpp example? For instance, explaining which parts of the code might have potential issues to facilitate understanding. I'm not proficient in the C++ language, so it would be helpful to have clearer explanations.

[littlebai3618]

It's my apologies for disturbing you again. @KerfuffleV2 @ggerganov

My English is a little rusty, and I used translation software to help translate some parts.

I think I have found the reason. When I changed the example question in parallel.cpp to a longer prompt of about 1500-2000 tokens, I can consistently reproduce the issue using the following command:

./parallel -m ./CodeLlama-7B/ggml-model-q8_0.gguf -n -1 -c 16324 -b 4096 --cont_batching --parallel 2 --sequences 600 --n-gpu-layers 1000

Failure Information

GPU memory suddenly increases.

Then the program throws an error.

cuBLAS error 13 at ggml-cuda.cu:6464: the function failed to launch on the GPU
current device: 0

I am running only one task of parallel.cpp on the GPU. Clearly, I have sufficient resources to run it, but the program throws an error.

When running parallel.cpp with a shorter example prompt, the GPU memory does not increase, and the program can terminate without errors.

What configurations should I adjust or what actions should I take to avoid this problem? Or is this a bug?

Steps to Reproduce

Using the latest llama.cpp repository code: commit_id: dae06c0
I am testing with the CodeLlama-7B-hf model：https://huggingface.co/codellama/CodeLlama-7b-hf
convert it:

python convert.py ./CodeLlama-7B/ --outtype q8_0 

1. Please use my code.
   parallel.cpp.zip
   only change k_prompts、k_system
2. run LLAMA_CUDA_NVCC=/usr/local/cuda-12/bin/nvcc make LLAMA_CUBLAS=1
3. exectue：./parallel -m ./CodeLlama-7B/ggml-model-q8_0.gguf -n -1 -c 16324 -b 4096 --cont_batching --parallel 2 --sequences 600 --n-gpu-layers 1000

[littlebai3618]

The same situation is happening on the V100S. I reduced the batch parameter from 4096 to 1024, but the same error still persists.

[KerfuffleV2]

It's my apologies for disturbing you again.

Sorry, you're not disturbing me but I don't really have much to add at this point. I don't really know enough about your specific problem to say something helpful. 4096 (or even 1024) sounds like a very, very high batch size though. Maybe it's normal for those cards.

Hopefully GG will be able to help you, he wrote the parallel example. (By the way, your English seemed fine to me.)

[littlebai3618]

I found that the comments on this pull request are similar to the issue I encountered, but verifying this problem is beyond my ability.

#3776