Bug: Memory Leak in llama-server after exit

edwin0cheng commented 3 weeks ago

What happened?

I found that after running :

free -m > before.log
./llama-server -ngl 10 -m models/Llama-3.2-3B.Q4_K_M.gguf 
free -m > after.log

the memory used is increased a lot:

# before.log
               total        used        free      shared  buff/cache   available
Mem:           64042       23546       35229         427        6358       40496
Swap:           4095           0        4095

# after.log
               total        used        free      shared  buff/cache   available
Mem:           64042       41949       16819         425        6364       22093
Swap:           4095           0        4095

Note that the buff/cache are not increased.

And here are `/proc/meminfo' if needed:

`cat /proc/meminfo` before `llama-server`

```log MemTotal: 65579952 kB MemFree: 26684720 kB MemAvailable: 32084964 kB Buffers: 3228 kB Cached: 6354208 kB SwapCached: 0 kB Active: 33631472 kB Inactive: 4406440 kB Active(anon): 32120904 kB Inactive(anon): 0 kB Active(file): 1510568 kB Inactive(file): 4406440 kB Unevictable: 17308 kB Mlocked: 17308 kB SwapTotal: 4194300 kB SwapFree: 4194300 kB Zswap: 0 kB Zswapped: 0 kB Dirty: 96 kB Writeback: 0 kB AnonPages: 3386220 kB Mapped: 1475556 kB Shmem: 436000 kB KReclaimable: 159404 kB Slab: 341540 kB SReclaimable: 159404 kB SUnreclaim: 182136 kB KernelStack: 21568 kB PageTables: 45092 kB SecPageTables: 1052 kB NFS_Unstable: 0 kB Bounce: 0 kB WritebackTmp: 0 kB CommitLimit: 36984276 kB Committed_AS: 9430624 kB VmallocTotal: 34359738367 kB VmallocUsed: 139168 kB VmallocChunk: 0 kB Percpu: 14208 kB HardwareCorrupted: 0 kB AnonHugePages: 499712 kB ShmemHugePages: 0 kB ShmemPmdMapped: 0 kB FileHugePages: 2768896 kB FilePmdMapped: 354304 kB CmaTotal: 0 kB CmaFree: 0 kB Unaccepted: 0 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 2048 kB Hugetlb: 0 kB DirectMap4k: 1957432 kB DirectMap2M: 39720960 kB DirectMap1G: 25165824 kB ```

`cat /proc/meminfo` after `llama-server`

```log MemTotal: 65579952 kB MemFree: 17215096 kB MemAvailable: 22615424 kB Buffers: 3228 kB Cached: 6354176 kB SwapCached: 0 kB Active: 43082444 kB Inactive: 4406484 kB Active(anon): 41571876 kB Inactive(anon): 0 kB Active(file): 1510568 kB Inactive(file): 4406484 kB Unevictable: 17308 kB Mlocked: 17308 kB SwapTotal: 4194300 kB SwapFree: 4194300 kB Zswap: 0 kB Zswapped: 0 kB Dirty: 1104 kB Writeback: 0 kB AnonPages: 3400176 kB Mapped: 1475480 kB Shmem: 435924 kB KReclaimable: 159484 kB Slab: 332816 kB SReclaimable: 159484 kB SUnreclaim: 173332 kB KernelStack: 21632 kB PageTables: 45088 kB SecPageTables: 1052 kB NFS_Unstable: 0 kB Bounce: 0 kB WritebackTmp: 0 kB CommitLimit: 36984276 kB Committed_AS: 9429912 kB VmallocTotal: 34359738367 kB VmallocUsed: 139200 kB VmallocChunk: 0 kB Percpu: 14208 kB HardwareCorrupted: 0 kB AnonHugePages: 499712 kB ShmemHugePages: 0 kB ShmemPmdMapped: 0 kB FileHugePages: 2768896 kB FilePmdMapped: 354304 kB CmaTotal: 0 kB CmaFree: 0 kB Unaccepted: 0 kB HugePages_Total: 0 HugePages_Free: 0 HugePages_Rsvd: 0 HugePages_Surp: 0 Hugepagesize: 2048 kB Hugetlb: 0 kB DirectMap4k: 1973816 kB DirectMap2M: 40753152 kB DirectMap1G: 24117248 kB ```

It seem like the memory leak is related to "Active(anon)", but IIUC Anonymouse memory are only used for heap, stack and anonymous mmap, which will be reclaimed after program exit.

So I don't know how it is possible to leak.

Name and Version

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 4060 Ti, compute capability 8.9, VMM: yes version: 3933 (f010b77a) built with cc (GCC) 14.2.1 20240910 for x86_64-pc-linux-gnu

What operating system are you seeing the problem on?

Linux

Relevant log output

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 4060 Ti, compute capability 8.9, VMM: yes
build: 3933 (f010b77a) with cc (GCC) 14.2.1 20240910 for x86_64-pc-linux-gnu
system info: n_threads = 8, n_threads_batch = 8, total_threads = 24

system_info: n_threads = 8 (n_threads_batch = 8) / 24 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | 

main: HTTP server is listening, hostname: 127.0.0.1, port: 8080, http threads: 23
main: loading model
llama_load_model_from_file: using device CUDA0 (NVIDIA GeForce RTX 4060 Ti) - 15264 MiB free
llama_model_loader: loaded meta data with 28 key-value pairs and 255 tensors from models/Llama-3.2-3B.Q4_K_M.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.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Models
llama_model_loader: - kv   3:                         general.size_label str              = 3.2B
llama_model_loader: - kv   4:                            general.license str              = llama3.2
llama_model_loader: - kv   5:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv   6:                          general.languages arr[str,8]       = ["en", "de", "fr", "it", "pt", "hi", ...
llama_model_loader: - kv   7:                          llama.block_count u32              = 28
llama_model_loader: - kv   8:                       llama.context_length u32              = 131072
llama_model_loader: - kv   9:                     llama.embedding_length u32              = 3072
llama_model_loader: - kv  10:                  llama.feed_forward_length u32              = 8192
llama_model_loader: - kv  11:                 llama.attention.head_count u32              = 24
llama_model_loader: - kv  12:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  13:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  14:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  15:                 llama.attention.key_length u32              = 128
llama_model_loader: - kv  16:               llama.attention.value_length u32              = 128
llama_model_loader: - kv  17:                          general.file_type u32              = 15
llama_model_loader: - kv  18:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  19:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  20:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  21:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  22:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  23:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  24:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  25:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  26:                tokenizer.ggml.eos_token_id u32              = 128001
llama_model_loader: - kv  27:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   58 tensors
llama_model_loader: - type q4_K:  168 tensors
llama_model_loader: - type q6_K:   29 tensors
llm_load_vocab: control token: 128254 '<|reserved_special_token_246|>' is not marked as EOG
llm_load_vocab: control token: 128249 '<|reserved_special_token_241|>' is not marked as EOG
llm_load_vocab: control token: 128246 '<|reserved_special_token_238|>' is not marked as EOG
llm_load_vocab: control token: 128243 '<|reserved_special_token_235|>' is not marked as EOG
llm_load_vocab: control token: 128242 '<|reserved_special_token_234|>' is not marked as EOG
llm_load_vocab: control token: 128241 '<|reserved_special_token_233|>' is not marked as EOG
llm_load_vocab: control token: 128240 '<|reserved_special_token_232|>' is not marked as EOG
llm_load_vocab: control token: 128235 '<|reserved_special_token_227|>' is not marked as EOG
llm_load_vocab: control token: 128231 '<|reserved_special_token_223|>' is not marked as EOG
llm_load_vocab: control token: 128230 '<|reserved_special_token_222|>' is not marked as EOG
llm_load_vocab: control token: 128228 '<|reserved_special_token_220|>' is not marked as EOG
llm_load_vocab: control token: 128225 '<|reserved_special_token_217|>' is not marked as EOG
llm_load_vocab: control token: 128218 '<|reserved_special_token_210|>' is not marked as EOG
llm_load_vocab: control token: 128214 '<|reserved_special_token_206|>' is not marked as EOG
llm_load_vocab: control token: 128213 '<|reserved_special_token_205|>' is not marked as EOG
llm_load_vocab: control token: 128207 '<|reserved_special_token_199|>' is not marked as EOG
llm_load_vocab: control token: 128206 '<|reserved_special_token_198|>' is not marked as EOG
llm_load_vocab: control token: 128204 '<|reserved_special_token_196|>' is not marked as EOG
llm_load_vocab: control token: 128200 '<|reserved_special_token_192|>' is not marked as EOG
llm_load_vocab: control token: 128199 '<|reserved_special_token_191|>' is not marked as EOG
llm_load_vocab: control token: 128198 '<|reserved_special_token_190|>' is not marked as EOG
llm_load_vocab: control token: 128196 '<|reserved_special_token_188|>' is not marked as EOG
llm_load_vocab: control token: 128194 '<|reserved_special_token_186|>' is not marked as EOG
llm_load_vocab: control token: 128193 '<|reserved_special_token_185|>' is not marked as EOG
llm_load_vocab: control token: 128188 '<|reserved_special_token_180|>' is not marked as EOG
llm_load_vocab: control token: 128187 '<|reserved_special_token_179|>' is not marked as EOG
llm_load_vocab: control token: 128185 '<|reserved_special_token_177|>' is not marked as EOG
llm_load_vocab: control token: 128184 '<|reserved_special_token_176|>' is not marked as EOG
llm_load_vocab: control token: 128180 '<|reserved_special_token_172|>' is not marked as EOG
llm_load_vocab: control token: 128179 '<|reserved_special_token_171|>' is not marked as EOG
llm_load_vocab: control token: 128178 '<|reserved_special_token_170|>' is not marked as EOG
llm_load_vocab: control token: 128177 '<|reserved_special_token_169|>' is not marked as EOG
llm_load_vocab: control token: 128176 '<|reserved_special_token_168|>' is not marked as EOG
llm_load_vocab: control token: 128175 '<|reserved_special_token_167|>' is not marked as EOG
llm_load_vocab: control token: 128171 '<|reserved_special_token_163|>' is not marked as EOG
llm_load_vocab: control token: 128170 '<|reserved_special_token_162|>' is not marked as EOG
llm_load_vocab: control token: 128169 '<|reserved_special_token_161|>' is not marked as EOG
llm_load_vocab: control token: 128168 '<|reserved_special_token_160|>' is not marked as EOG
llm_load_vocab: control token: 128165 '<|reserved_special_token_157|>' is not marked as EOG
llm_load_vocab: control token: 128162 '<|reserved_special_token_154|>' is not marked as EOG
llm_load_vocab: control token: 128158 '<|reserved_special_token_150|>' is not marked as EOG
llm_load_vocab: control token: 128156 '<|reserved_special_token_148|>' is not marked as EOG
llm_load_vocab: control token: 128155 '<|reserved_special_token_147|>' is not marked as EOG
llm_load_vocab: control token: 128154 '<|reserved_special_token_146|>' is not marked as EOG
llm_load_vocab: control token: 128151 '<|reserved_special_token_143|>' is not marked as EOG
llm_load_vocab: control token: 128149 '<|reserved_special_token_141|>' is not marked as EOG
llm_load_vocab: control token: 128147 '<|reserved_special_token_139|>' is not marked as EOG
llm_load_vocab: control token: 128146 '<|reserved_special_token_138|>' is not marked as EOG
llm_load_vocab: control token: 128144 '<|reserved_special_token_136|>' is not marked as EOG
llm_load_vocab: control token: 128142 '<|reserved_special_token_134|>' is not marked as EOG
llm_load_vocab: control token: 128141 '<|reserved_special_token_133|>' is not marked as EOG
llm_load_vocab: control token: 128138 '<|reserved_special_token_130|>' is not marked as EOG
llm_load_vocab: control token: 128136 '<|reserved_special_token_128|>' is not marked as EOG
llm_load_vocab: control token: 128135 '<|reserved_special_token_127|>' is not marked as EOG
llm_load_vocab: control token: 128134 '<|reserved_special_token_126|>' is not marked as EOG
llm_load_vocab: control token: 128133 '<|reserved_special_token_125|>' is not marked as EOG
llm_load_vocab: control token: 128131 '<|reserved_special_token_123|>' is not marked as EOG
llm_load_vocab: control token: 128128 '<|reserved_special_token_120|>' is not marked as EOG
llm_load_vocab: control token: 128124 '<|reserved_special_token_116|>' is not marked as EOG
llm_load_vocab: control token: 128123 '<|reserved_special_token_115|>' is not marked as EOG
llm_load_vocab: control token: 128122 '<|reserved_special_token_114|>' is not marked as EOG
llm_load_vocab: control token: 128119 '<|reserved_special_token_111|>' is not marked as EOG
llm_load_vocab: control token: 128115 '<|reserved_special_token_107|>' is not marked as EOG
llm_load_vocab: control token: 128112 '<|reserved_special_token_104|>' is not marked as EOG
llm_load_vocab: control token: 128110 '<|reserved_special_token_102|>' is not marked as EOG
llm_load_vocab: control token: 128109 '<|reserved_special_token_101|>' is not marked as EOG
llm_load_vocab: control token: 128108 '<|reserved_special_token_100|>' is not marked as EOG
llm_load_vocab: control token: 128106 '<|reserved_special_token_98|>' is not marked as EOG
llm_load_vocab: control token: 128103 '<|reserved_special_token_95|>' is not marked as EOG
llm_load_vocab: control token: 128102 '<|reserved_special_token_94|>' is not marked as EOG
llm_load_vocab: control token: 128101 '<|reserved_special_token_93|>' is not marked as EOG
llm_load_vocab: control token: 128097 '<|reserved_special_token_89|>' is not marked as EOG
llm_load_vocab: control token: 128091 '<|reserved_special_token_83|>' is not marked as EOG
llm_load_vocab: control token: 128090 '<|reserved_special_token_82|>' is not marked as EOG
llm_load_vocab: control token: 128089 '<|reserved_special_token_81|>' is not marked as EOG
llm_load_vocab: control token: 128087 '<|reserved_special_token_79|>' is not marked as EOG
llm_load_vocab: control token: 128085 '<|reserved_special_token_77|>' is not marked as EOG
llm_load_vocab: control token: 128081 '<|reserved_special_token_73|>' is not marked as EOG
llm_load_vocab: control token: 128078 '<|reserved_special_token_70|>' is not marked as EOG
llm_load_vocab: control token: 128076 '<|reserved_special_token_68|>' is not marked as EOG
llm_load_vocab: control token: 128075 '<|reserved_special_token_67|>' is not marked as EOG
llm_load_vocab: control token: 128073 '<|reserved_special_token_65|>' is not marked as EOG
llm_load_vocab: control token: 128068 '<|reserved_special_token_60|>' is not marked as EOG
llm_load_vocab: control token: 128067 '<|reserved_special_token_59|>' is not marked as EOG
llm_load_vocab: control token: 128065 '<|reserved_special_token_57|>' is not marked as EOG
llm_load_vocab: control token: 128063 '<|reserved_special_token_55|>' is not marked as EOG
llm_load_vocab: control token: 128062 '<|reserved_special_token_54|>' is not marked as EOG
llm_load_vocab: control token: 128060 '<|reserved_special_token_52|>' is not marked as EOG
llm_load_vocab: control token: 128059 '<|reserved_special_token_51|>' is not marked as EOG
llm_load_vocab: control token: 128057 '<|reserved_special_token_49|>' is not marked as EOG
llm_load_vocab: control token: 128054 '<|reserved_special_token_46|>' is not marked as EOG
llm_load_vocab: control token: 128046 '<|reserved_special_token_38|>' is not marked as EOG
llm_load_vocab: control token: 128045 '<|reserved_special_token_37|>' is not marked as EOG
llm_load_vocab: control token: 128044 '<|reserved_special_token_36|>' is not marked as EOG
llm_load_vocab: control token: 128043 '<|reserved_special_token_35|>' is not marked as EOG
llm_load_vocab: control token: 128038 '<|reserved_special_token_30|>' is not marked as EOG
llm_load_vocab: control token: 128036 '<|reserved_special_token_28|>' is not marked as EOG
llm_load_vocab: control token: 128035 '<|reserved_special_token_27|>' is not marked as EOG
llm_load_vocab: control token: 128032 '<|reserved_special_token_24|>' is not marked as EOG
llm_load_vocab: control token: 128028 '<|reserved_special_token_20|>' is not marked as EOG
llm_load_vocab: control token: 128027 '<|reserved_special_token_19|>' is not marked as EOG
llm_load_vocab: control token: 128024 '<|reserved_special_token_16|>' is not marked as EOG
llm_load_vocab: control token: 128023 '<|reserved_special_token_15|>' is not marked as EOG
llm_load_vocab: control token: 128022 '<|reserved_special_token_14|>' is not marked as EOG
llm_load_vocab: control token: 128021 '<|reserved_special_token_13|>' is not marked as EOG
llm_load_vocab: control token: 128018 '<|reserved_special_token_10|>' is not marked as EOG
llm_load_vocab: control token: 128016 '<|reserved_special_token_8|>' is not marked as EOG
llm_load_vocab: control token: 128015 '<|reserved_special_token_7|>' is not marked as EOG
llm_load_vocab: control token: 128013 '<|reserved_special_token_5|>' is not marked as EOG
llm_load_vocab: control token: 128011 '<|reserved_special_token_3|>' is not marked as EOG
llm_load_vocab: control token: 128005 '<|reserved_special_token_2|>' is not marked as EOG
llm_load_vocab: control token: 128004 '<|finetune_right_pad_id|>' is not marked as EOG
llm_load_vocab: control token: 128002 '<|reserved_special_token_0|>' is not marked as EOG
llm_load_vocab: control token: 128252 '<|reserved_special_token_244|>' is not marked as EOG
llm_load_vocab: control token: 128190 '<|reserved_special_token_182|>' is not marked as EOG
llm_load_vocab: control token: 128183 '<|reserved_special_token_175|>' is not marked as EOG
llm_load_vocab: control token: 128137 '<|reserved_special_token_129|>' is not marked as EOG
llm_load_vocab: control token: 128182 '<|reserved_special_token_174|>' is not marked as EOG
llm_load_vocab: control token: 128040 '<|reserved_special_token_32|>' is not marked as EOG
llm_load_vocab: control token: 128048 '<|reserved_special_token_40|>' is not marked as EOG
llm_load_vocab: control token: 128092 '<|reserved_special_token_84|>' is not marked as EOG
llm_load_vocab: control token: 128215 '<|reserved_special_token_207|>' is not marked as EOG
llm_load_vocab: control token: 128107 '<|reserved_special_token_99|>' is not marked as EOG
llm_load_vocab: control token: 128208 '<|reserved_special_token_200|>' is not marked as EOG
llm_load_vocab: control token: 128145 '<|reserved_special_token_137|>' is not marked as EOG
llm_load_vocab: control token: 128031 '<|reserved_special_token_23|>' is not marked as EOG
llm_load_vocab: control token: 128129 '<|reserved_special_token_121|>' is not marked as EOG
llm_load_vocab: control token: 128201 '<|reserved_special_token_193|>' is not marked as EOG
llm_load_vocab: control token: 128074 '<|reserved_special_token_66|>' is not marked as EOG
llm_load_vocab: control token: 128095 '<|reserved_special_token_87|>' is not marked as EOG
llm_load_vocab: control token: 128186 '<|reserved_special_token_178|>' is not marked as EOG
llm_load_vocab: control token: 128143 '<|reserved_special_token_135|>' is not marked as EOG
llm_load_vocab: control token: 128229 '<|reserved_special_token_221|>' is not marked as EOG
llm_load_vocab: control token: 128007 '<|end_header_id|>' is not marked as EOG
llm_load_vocab: control token: 128055 '<|reserved_special_token_47|>' is not marked as EOG
llm_load_vocab: control token: 128056 '<|reserved_special_token_48|>' is not marked as EOG
llm_load_vocab: control token: 128061 '<|reserved_special_token_53|>' is not marked as EOG
llm_load_vocab: control token: 128153 '<|reserved_special_token_145|>' is not marked as EOG
llm_load_vocab: control token: 128152 '<|reserved_special_token_144|>' is not marked as EOG
llm_load_vocab: control token: 128212 '<|reserved_special_token_204|>' is not marked as EOG
llm_load_vocab: control token: 128172 '<|reserved_special_token_164|>' is not marked as EOG
llm_load_vocab: control token: 128160 '<|reserved_special_token_152|>' is not marked as EOG
llm_load_vocab: control token: 128041 '<|reserved_special_token_33|>' is not marked as EOG
llm_load_vocab: control token: 128181 '<|reserved_special_token_173|>' is not marked as EOG
llm_load_vocab: control token: 128094 '<|reserved_special_token_86|>' is not marked as EOG
llm_load_vocab: control token: 128118 '<|reserved_special_token_110|>' is not marked as EOG
llm_load_vocab: control token: 128236 '<|reserved_special_token_228|>' is not marked as EOG
llm_load_vocab: control token: 128148 '<|reserved_special_token_140|>' is not marked as EOG
llm_load_vocab: control token: 128042 '<|reserved_special_token_34|>' is not marked as EOG
llm_load_vocab: control token: 128139 '<|reserved_special_token_131|>' is not marked as EOG
llm_load_vocab: control token: 128173 '<|reserved_special_token_165|>' is not marked as EOG
llm_load_vocab: control token: 128239 '<|reserved_special_token_231|>' is not marked as EOG
llm_load_vocab: control token: 128157 '<|reserved_special_token_149|>' is not marked as EOG
llm_load_vocab: control token: 128052 '<|reserved_special_token_44|>' is not marked as EOG
llm_load_vocab: control token: 128026 '<|reserved_special_token_18|>' is not marked as EOG
llm_load_vocab: control token: 128003 '<|reserved_special_token_1|>' is not marked as EOG
llm_load_vocab: control token: 128019 '<|reserved_special_token_11|>' is not marked as EOG
llm_load_vocab: control token: 128116 '<|reserved_special_token_108|>' is not marked as EOG
llm_load_vocab: control token: 128161 '<|reserved_special_token_153|>' is not marked as EOG
llm_load_vocab: control token: 128226 '<|reserved_special_token_218|>' is not marked as EOG
llm_load_vocab: control token: 128159 '<|reserved_special_token_151|>' is not marked as EOG
llm_load_vocab: control token: 128012 '<|reserved_special_token_4|>' is not marked as EOG
llm_load_vocab: control token: 128088 '<|reserved_special_token_80|>' is not marked as EOG
llm_load_vocab: control token: 128163 '<|reserved_special_token_155|>' is not marked as EOG
llm_load_vocab: control token: 128001 '<|end_of_text|>' is not marked as EOG
llm_load_vocab: control token: 128113 '<|reserved_special_token_105|>' is not marked as EOG
llm_load_vocab: control token: 128250 '<|reserved_special_token_242|>' is not marked as EOG
llm_load_vocab: control token: 128125 '<|reserved_special_token_117|>' is not marked as EOG
llm_load_vocab: control token: 128053 '<|reserved_special_token_45|>' is not marked as EOG
llm_load_vocab: control token: 128224 '<|reserved_special_token_216|>' is not marked as EOG
llm_load_vocab: control token: 128247 '<|reserved_special_token_239|>' is not marked as EOG
llm_load_vocab: control token: 128251 '<|reserved_special_token_243|>' is not marked as EOG
llm_load_vocab: control token: 128216 '<|reserved_special_token_208|>' is not marked as EOG
llm_load_vocab: control token: 128006 '<|start_header_id|>' is not marked as EOG
llm_load_vocab: control token: 128211 '<|reserved_special_token_203|>' is not marked as EOG
llm_load_vocab: control token: 128077 '<|reserved_special_token_69|>' is not marked as EOG
llm_load_vocab: control token: 128237 '<|reserved_special_token_229|>' is not marked as EOG
llm_load_vocab: control token: 128086 '<|reserved_special_token_78|>' is not marked as EOG
llm_load_vocab: control token: 128227 '<|reserved_special_token_219|>' is not marked as EOG
llm_load_vocab: control token: 128058 '<|reserved_special_token_50|>' is not marked as EOG
llm_load_vocab: control token: 128100 '<|reserved_special_token_92|>' is not marked as EOG
llm_load_vocab: control token: 128209 '<|reserved_special_token_201|>' is not marked as EOG
llm_load_vocab: control token: 128084 '<|reserved_special_token_76|>' is not marked as EOG
llm_load_vocab: control token: 128071 '<|reserved_special_token_63|>' is not marked as EOG
llm_load_vocab: control token: 128070 '<|reserved_special_token_62|>' is not marked as EOG
llm_load_vocab: control token: 128049 '<|reserved_special_token_41|>' is not marked as EOG
llm_load_vocab: control token: 128197 '<|reserved_special_token_189|>' is not marked as EOG
llm_load_vocab: control token: 128072 '<|reserved_special_token_64|>' is not marked as EOG
llm_load_vocab: control token: 128000 '<|begin_of_text|>' is not marked as EOG
llm_load_vocab: control token: 128223 '<|reserved_special_token_215|>' is not marked as EOG
llm_load_vocab: control token: 128217 '<|reserved_special_token_209|>' is not marked as EOG
llm_load_vocab: control token: 128111 '<|reserved_special_token_103|>' is not marked as EOG
llm_load_vocab: control token: 128203 '<|reserved_special_token_195|>' is not marked as EOG
llm_load_vocab: control token: 128051 '<|reserved_special_token_43|>' is not marked as EOG
llm_load_vocab: control token: 128030 '<|reserved_special_token_22|>' is not marked as EOG
llm_load_vocab: control token: 128117 '<|reserved_special_token_109|>' is not marked as EOG
llm_load_vocab: control token: 128010 '<|python_tag|>' is not marked as EOG
llm_load_vocab: control token: 128238 '<|reserved_special_token_230|>' is not marked as EOG
llm_load_vocab: control token: 128255 '<|reserved_special_token_247|>' is not marked as EOG
llm_load_vocab: control token: 128202 '<|reserved_special_token_194|>' is not marked as EOG
llm_load_vocab: control token: 128132 '<|reserved_special_token_124|>' is not marked as EOG
llm_load_vocab: control token: 128248 '<|reserved_special_token_240|>' is not marked as EOG
llm_load_vocab: control token: 128167 '<|reserved_special_token_159|>' is not marked as EOG
llm_load_vocab: control token: 128127 '<|reserved_special_token_119|>' is not marked as EOG
llm_load_vocab: control token: 128105 '<|reserved_special_token_97|>' is not marked as EOG
llm_load_vocab: control token: 128039 '<|reserved_special_token_31|>' is not marked as EOG
llm_load_vocab: control token: 128232 '<|reserved_special_token_224|>' is not marked as EOG
llm_load_vocab: control token: 128166 '<|reserved_special_token_158|>' is not marked as EOG
llm_load_vocab: control token: 128130 '<|reserved_special_token_122|>' is not marked as EOG
llm_load_vocab: control token: 128114 '<|reserved_special_token_106|>' is not marked as EOG
llm_load_vocab: control token: 128234 '<|reserved_special_token_226|>' is not marked as EOG
llm_load_vocab: control token: 128191 '<|reserved_special_token_183|>' is not marked as EOG
llm_load_vocab: control token: 128064 '<|reserved_special_token_56|>' is not marked as EOG
llm_load_vocab: control token: 128140 '<|reserved_special_token_132|>' is not marked as EOG
llm_load_vocab: control token: 128096 '<|reserved_special_token_88|>' is not marked as EOG
llm_load_vocab: control token: 128098 '<|reserved_special_token_90|>' is not marked as EOG
llm_load_vocab: control token: 128192 '<|reserved_special_token_184|>' is not marked as EOG
llm_load_vocab: control token: 128093 '<|reserved_special_token_85|>' is not marked as EOG
llm_load_vocab: control token: 128150 '<|reserved_special_token_142|>' is not marked as EOG
llm_load_vocab: control token: 128222 '<|reserved_special_token_214|>' is not marked as EOG
llm_load_vocab: control token: 128233 '<|reserved_special_token_225|>' is not marked as EOG
llm_load_vocab: control token: 128220 '<|reserved_special_token_212|>' is not marked as EOG
llm_load_vocab: control token: 128034 '<|reserved_special_token_26|>' is not marked as EOG
llm_load_vocab: control token: 128033 '<|reserved_special_token_25|>' is not marked as EOG
llm_load_vocab: control token: 128253 '<|reserved_special_token_245|>' is not marked as EOG
llm_load_vocab: control token: 128195 '<|reserved_special_token_187|>' is not marked as EOG
llm_load_vocab: control token: 128099 '<|reserved_special_token_91|>' is not marked as EOG
llm_load_vocab: control token: 128189 '<|reserved_special_token_181|>' is not marked as EOG
llm_load_vocab: control token: 128210 '<|reserved_special_token_202|>' is not marked as EOG
llm_load_vocab: control token: 128174 '<|reserved_special_token_166|>' is not marked as EOG
llm_load_vocab: control token: 128083 '<|reserved_special_token_75|>' is not marked as EOG
llm_load_vocab: control token: 128080 '<|reserved_special_token_72|>' is not marked as EOG
llm_load_vocab: control token: 128104 '<|reserved_special_token_96|>' is not marked as EOG
llm_load_vocab: control token: 128082 '<|reserved_special_token_74|>' is not marked as EOG
llm_load_vocab: control token: 128219 '<|reserved_special_token_211|>' is not marked as EOG
llm_load_vocab: control token: 128017 '<|reserved_special_token_9|>' is not marked as EOG
llm_load_vocab: control token: 128050 '<|reserved_special_token_42|>' is not marked as EOG
llm_load_vocab: control token: 128205 '<|reserved_special_token_197|>' is not marked as EOG
llm_load_vocab: control token: 128047 '<|reserved_special_token_39|>' is not marked as EOG
llm_load_vocab: control token: 128164 '<|reserved_special_token_156|>' is not marked as EOG
llm_load_vocab: control token: 128020 '<|reserved_special_token_12|>' is not marked as EOG
llm_load_vocab: control token: 128069 '<|reserved_special_token_61|>' is not marked as EOG
llm_load_vocab: control token: 128245 '<|reserved_special_token_237|>' is not marked as EOG
llm_load_vocab: control token: 128121 '<|reserved_special_token_113|>' is not marked as EOG
llm_load_vocab: control token: 128079 '<|reserved_special_token_71|>' is not marked as EOG
llm_load_vocab: control token: 128037 '<|reserved_special_token_29|>' is not marked as EOG
llm_load_vocab: control token: 128244 '<|reserved_special_token_236|>' is not marked as EOG
llm_load_vocab: control token: 128029 '<|reserved_special_token_21|>' is not marked as EOG
llm_load_vocab: control token: 128221 '<|reserved_special_token_213|>' is not marked as EOG
llm_load_vocab: control token: 128066 '<|reserved_special_token_58|>' is not marked as EOG
llm_load_vocab: control token: 128120 '<|reserved_special_token_112|>' is not marked as EOG
llm_load_vocab: control token: 128014 '<|reserved_special_token_6|>' is not marked as EOG
llm_load_vocab: control token: 128025 '<|reserved_special_token_17|>' is not marked as EOG
llm_load_vocab: control token: 128126 '<|reserved_special_token_118|>' is not marked as EOG
llm_load_vocab: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.7999 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 128256
llm_load_print_meta: n_merges         = 280147
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 131072
llm_load_print_meta: n_embd           = 3072
llm_load_print_meta: n_layer          = 28
llm_load_print_meta: n_head           = 24
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 128
llm_load_print_meta: n_embd_head_v    = 128
llm_load_print_meta: n_gqa            = 3
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             = 8192
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  = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 131072
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: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = 3B
llm_load_print_meta: model ftype      = Q4_K - Medium
llm_load_print_meta: model params     = 3.21 B
llm_load_print_meta: model size       = 1.87 GiB (5.01 BPW) 
llm_load_print_meta: general.name     = Models
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128001 '<|end_of_text|>'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: EOM token        = 128008 '<|eom_id|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOG token        = 128001 '<|end_of_text|>'
llm_load_print_meta: EOG token        = 128008 '<|eom_id|>'
llm_load_print_meta: EOG token        = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.24 MiB
llm_load_tensors: offloading 10 repeating layers to GPU
llm_load_tensors: offloaded 10/29 layers to GPU
llm_load_tensors:        CPU buffer size =  1918.35 MiB
llm_load_tensors:      CUDA0 buffer size =   588.19 MiB
...........................................................................
llama_new_context_with_model: n_ctx      = 131072
llama_new_context_with_model: n_batch    = 2048
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base  = 500000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:        CPU KV buffer size =  9216.00 MiB
llama_kv_cache_init:      CUDA0 KV buffer size =  5120.00 MiB
llama_new_context_with_model: KV self size  = 14336.00 MiB, K (f16): 7168.00 MiB, V (f16): 7168.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.98 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =  7180.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =   266.01 MiB
llama_new_context_with_model: graph nodes  = 902
llama_new_context_with_model: graph splits = 238
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 131072
main: model loaded
main: chat template, built_in: 1, chat_example: '<|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 127.0.0.1:8080 - starting the main loop
srv  update_slots: all slots are idle

ngxson commented 3 weeks ago

It seem like the memory leak is related to "Active(anon)", but IIUC Anonymouse memory are only used for heap, stack and anonymous mmap, which will be reclaimed after program exit.

IIRC "Active(anon)" does not need to be freed immediately after the process exits, but will be cleaned by kernel after a while. Anw, I don't think this is a llama.cpp-related problem. You should probably review & understand how linux manages virtual memory: https://blogs.oracle.com/linux/post/understanding-linux-kernel-memory-statistics

LRU lists

Currently, the kernel maintains an active list and an inactive list for both page types.When the system’s memory is below a watermark threshold, the kernel starts to scan the tails of inactive lists to reclaim pages which are likely to be idle for a while. If an inactive list becomes short, the kernel scans its corresponding active list at the tail to deactivate pages and move them to the inactive list.

edwin0cheng commented 3 weeks ago

You can see MemFree is dropped from 26684720 to 17215096, which according your document:

MemFree is the amount of free, unused RAM

And the section you are mentioned, IIRC is for Used Memory, you can confirm by the following paragraph (which stated that the in used anonymous page will be swapped) :

Among the two types of pages, page cache is easier to be reclaimed. A page cache page can be directly freed if not dirty. Otherwise, a write-back operation is needed. However, reclaiming an anonymous page requires to save the page to swap space.

In fact, the first time I found this bug were system hang by no memory available. But maybe you are right, it is not related to llama.cpp. I guess maybe related to CUDA...

IlyaBizyaev commented 3 weeks ago

I think I observe the same problem with llama-cli on Linux. The consequence is that I can only rerun it 3–4 times before my system runs out of memory. System monitors fail to attribute this usage to any process.

I also don't know if the cause lies in llama.cpp or NVIDIA's code.

katyo commented 3 weeks ago

I have same issue too.

Before starting llama.cpp:

$ free -h
               total        used        free      shared  buff/cache   available
Mem:            62Gi       1,1Gi        60Gi        27Mi       1,1Gi        60Gi
Swap:           63Gi          0B        63Gi

After exiting:

$ free -h
               total        used        free      shared  buff/cache   available
Mem:            62Gi        41Gi       421Mi        18Mi        20Gi        20Gi
Swap:           63Gi       758Mi        63Gi

Smem totals:

  PID User     Command                         Swap      USS      PSS      RSS
-------------------------------------------------------------------------------
   67 18                                     462.9M   601.2M   661.8M   980.2M

My system:

$ uname -a
Linux nixie-farm 6.11.4 #1-NixOS SMP PREEMPT_DYNAMIC Thu Oct 17 13:27:02 UTC 2024 x86_64 GNU/Linux

$ nvidia-smi
Thu Oct 24 02:07:28 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 555.58.02              Driver Version: 555.58.02      CUDA Version: 12.5     |
|-----------------------------------------+------------------------+----------------------+
| 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  NVIDIA GeForce RTX 4060 Ti     Off |   00000000:05:00.0  On |                  N/A |
|  0%   44C    P8             18W /  165W |      41MiB /  16380MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A     12298      G   ...iz78r52md-xorg-server-21.1.13/bin/X         36MiB |
+-----------------------------------------------------------------------------------------+

ngxson commented 3 weeks ago

Could someone please try these and see what's the result?

Add --no-mmap argument
With -ngl 0
Maybe with a CPU-only build? (even with -ngl 0, a CUDA build will still load CUDA library when it runs)

edwin0cheng commented 3 weeks ago

I just tried it:

* Add `--no-mmap` argument
* With `-ngl 0`

Still memory leak.

* Maybe with a CPU-only build? (even with `-ngl 0`, a CUDA build will still load CUDA library when it runs)

./llama-server -m models/Llama-3.2-3B.Q4_K_M.gguf 
build: 3933 (f010b77a) with cc (GCC) 14.2.1 20240910 for x86_64-pc-linux-gnu
system info: n_threads = 8, n_threads_batch = 8, total_threads = 24

system_info: n_threads = 8 (n_threads_batch = 8) / 24 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |

No memory leak at all.

katyo commented 3 weeks ago

I also have memory leak with --no-mmap and -ngl 0

$ free -h
               total        used        free      shared  buff/cache   available
Mem:            62Gi        39Gi        21Gi        13Mi       1,8Gi        22Gi
Swap:           63Gi       531Mi        63Gi

ngxson commented 3 weeks ago

OK so based on these results, we can reduce the scope of the issue to CUDA-only (happens even with -ngl 0)

Since this is outside of my knowledge, I'd like to ask if you could help debugging this further? @slaren @JohannesGaessler

slaren commented 3 weeks ago

You could try setting the environment variable GGML_CUDA_NO_PINNED to see if it is related to using host memory, but ultimately, llama.cpp is a user mode application and it cannot keep resources beyond the lifetime of the process, so whatever is happening, it must be at the kernel level. If you think that this may be a CUDA driver bug, the best course of action would be to report this to NVIDIA. If they confirm the issue and suggest a workaround until it is fixed in the driver, we can try implementing it, but otherwise there is not much we can do about it.

edwin0cheng commented 3 weeks ago

`GGML_CUDA_NO_PINNED = 1 fixed the leakage, so it is related to using host memory i guess.

edwin0cheng commented 3 weeks ago

I found this post in nvidia developer forums, it seem like a bug related in the kernel introduced in 6.11 kernel.

aiseei commented 3 weeks ago

Have the same issue: SPECS: 4 * T4 with ~64GB VRAM, 440GB RAM, 64 CPU Running cpp-server built off release b3655 on docker with cuda 12.4. Running Mistral-Nemo-Instruct-2407-Q6_K_L.gguf with np 2, cont batching, ngl 99 and flash attention. Even after 30 mins post generation the memory isnt released. GGML_CUDA_NO_PINNED = 1 did not work me. Screenshot attached

Any inputs would be much appreciated !

JohannesGaessler commented 3 weeks ago

I am not seeing any discrepancies with Linux 6.6.54-2-MANJARO, drivers v550.120-1, CUDA v12.6.1-1, and the latest llama.cpp master commit.

edwin0cheng commented 3 weeks ago

FYI:

According to this linux kernel dev email thread, I tried the following commands to test whether it is related:

# before
> grep foll /proc/vmstat
    nr_foll_pin_acquired 5344
    nr_foll_pin_released 5344

# after
> grep foll /proc/vmstat
    nr_foll_pin_acquired 2173154
    nr_foll_pin_released 76002

Bingo! So I will close this issue when this patch lands in the future Linux kernel.

vlovich commented 3 weeks ago

Since the patch isn't applied in 6.12 nor mainline, I've filed a bug to make sure it's not lost: https://bugzilla.kernel.org/show_bug.cgi?id=219427

ggerganov / llama.cpp