ollama 在ROCm上卸载层时出现内存不足

3gtaxfhh 于 2个月前发布在其他

关注(0)|答案(5)|浏览(25)

问题是什么？

我注意到最近还有一些关于卸载的问题，但是由于我使用的是不同的设置，所以我认为打开一个单独的问题是有意义的。我既不使用Docker,也不使用Nvidia。但是在Debian稳定版上运行的是ROCm 6.1.3。

在每次测试之前，我都停止了ollama(因为我希望这会释放任何剩余的GPU内存声明)。

在使用ollama运行时，我得到了：

$ free -m
               total        used        free      shared  buff/cache   available
Mem:           63508        2349         753           8       61127       61158
Swap:              0           0           0

此外，没有任何东西声称拥有GPU(我使用的是桌面，但在内部GPU(设备1)上):

$ rocm-smi  --device=0  --showmemuse  --showpids

======================= ROCm System Management Interface =======================
============================== Current Memory Use ==============================
GPU[0]          : GPU memory use (%): 0
GPU[0]          : Memory Activity: N/A
================================================================================
================================ KFD Processes =================================
No KFD PIDs currently running
================================================================================
============================= End of ROCm SMI Log ==============================

当我在openwebui中按下回车键运行相同的命令时，我得到了：

$ rocm-smi  --device=0  --showmemuse  --showpids

======================= ROCm System Management Interface =======================
============================== Current Memory Use ==============================
GPU[0]          : GPU memory use (%): 2
GPU[0]          : Memory Activity: N/A
================================================================================
================================ KFD Processes =================================
KFD process information:
PID     PROCESS NAME    GPU(s)  VRAM USED       SDMA USED       CU OCCUPANCY
115056  ollama_llama_se 1       25240158208     0               0
================================================================================
============================= End of ROCm SMI Log ==============================

错误输出示例

[Ollama服务的输出，点击展开]

Jul 24 16:07:04 vega systemd[1]: Stopping ollama.service - Ollama Service...
Jul 24 16:07:04 vega systemd[1]: ollama.service: Deactivated successfully.
Jul 24 16:07:04 vega systemd[1]: Stopped ollama.service - Ollama Service.
Jul 24 16:07:04 vega systemd[1]: Started ollama.service - Ollama Service.
Jul 24 16:07:04 vega ollama[114917]: 2024/07/24 16:07:04 routes.go:1100: INFO server config env="map[CUDA_VISIBLE_DEVICES: GPU_DEVICE_ORDINAL: HIP_VISIBLE_DEVICES: HSA_OVERRIDE_GFX_VERSION: OLLAMA_DEBUG:false OLLAMA_FLASH_ATTENTION:false OLLAMA_HOST:http://0.0.0.0:11434 OLLAMA_INTEL_GPU:false OLLAMA_KEEP_ALIVE:5m0s OLLAMA_LLM_LIBRARY: OLLAMA_MAX_LOADED_MODELS:0 OLLAMA_MAX_QUEUE:512 OLLAMA_MODELS:/usr/share/ollama/.ollama/models OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:0 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://*] OLLAMA_RUNNERS_DIR: OLLAMA_SCHED_SPREAD:false OLLAMA_TMPDIR: ROCR_VISIBLE_DEVICES:]"
Jul 24 16:07:04 vega ollama[114917]: time=2024-07-24T16:07:04.312+02:00 level=INFO source=images.go:784 msg="total blobs: 47"
Jul 24 16:07:04 vega ollama[114917]: time=2024-07-24T16:07:04.312+02:00 level=INFO source=images.go:791 msg="total unused blobs removed: 0"
Jul 24 16:07:04 vega ollama[114917]: time=2024-07-24T16:07:04.312+02:00 level=INFO source=routes.go:1147 msg="Listening on [::]:11434 (version 0.2.8)"
Jul 24 16:07:04 vega ollama[114917]: time=2024-07-24T16:07:04.313+02:00 level=INFO source=payload.go:30 msg="extracting embedded files" dir=/tmp/ollama3590369896/runners
Jul 24 16:07:06 vega ollama[114917]: time=2024-07-24T16:07:06.996+02:00 level=INFO source=payload.go:44 msg="Dynamic LLM libraries [cpu_avx cpu_avx2 cuda_v11 rocm_v60102 cpu]"
Jul 24 16:07:06 vega ollama[114917]: time=2024-07-24T16:07:06.996+02:00 level=INFO source=gpu.go:205 msg="looking for compatible GPUs"
Jul 24 16:07:07 vega ollama[114917]: time=2024-07-24T16:07:07.004+02:00 level=INFO source=amd_linux.go:330 msg="amdgpu is supported" gpu=0 gpu_type=gfx1100
Jul 24 16:07:07 vega ollama[114917]: time=2024-07-24T16:07:07.004+02:00 level=INFO source=amd_linux.go:259 msg="unsupported Radeon iGPU detected skipping" id=1 total="512.0 MiB"
Jul 24 16:07:07 vega ollama[114917]: time=2024-07-24T16:07:07.004+02:00 level=INFO source=types.go:105 msg="inference compute" id=0 library=rocm compute=gfx1100 driver=6.7 name=1002:744c total="24.0 GiB" available="24.0 GiB"
Jul 24 16:07:51 vega ollama[114917]: time=2024-07-24T16:07:51.812+02:00 level=INFO source=memory.go:309 msg="offload to rocm" layers.requested=-1 layers.model=81 layers.offload=48 layers.split="" memory.available="[24.0 GiB]" memory.required.full="39.3 GiB" memory.required.partial="23.9 GiB" memory.required.kv="640.0 MiB" memory.required.allocations="[23.9 GiB]" memory.weights.total="36.5 GiB" memory.weights.repeating="35.7 GiB" memory.weights.nonrepeating="822.0 MiB" memory.graph.full="324.0 MiB" memory.graph.partial="1.1 GiB"
Jul 24 16:07:51 vega ollama[114917]: time=2024-07-24T16:07:51.813+02:00 level=INFO source=server.go:383 msg="starting llama server" cmd="/tmp/ollama3590369896/runners/rocm_v60102/ollama_llama_server --model /usr/share/ollama/.ollama/models/blobs/sha256-0bd51f8f0c975ce910ed067dcb962a9af05b77bafcdc595ef02178387f10e51d --ctx-size 2048 --batch-size 512 --embedding --log-disable --n-gpu-layers 48 --parallel 1 --port 36639"
Jul 24 16:07:51 vega ollama[114917]: time=2024-07-24T16:07:51.813+02:00 level=INFO source=sched.go:437 msg="loaded runners" count=1
Jul 24 16:07:51 vega ollama[114917]: time=2024-07-24T16:07:51.813+02:00 level=INFO source=server.go:583 msg="waiting for llama runner to start responding"
Jul 24 16:07:51 vega ollama[114917]: time=2024-07-24T16:07:51.813+02:00 level=INFO source=server.go:617 msg="waiting for server to become available" status="llm server error"
Jul 24 16:07:51 vega ollama[115056]: INFO [main] build info | build=1 commit="d94c6e0" tid="140302582510400" timestamp=1721830071
Jul 24 16:07:51 vega ollama[115056]: INFO [main] system info | n_threads=8 n_threads_batch=-1 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 0 | " tid="140302582510400" timestamp=1721830071 total_threads=16
Jul 24 16:07:51 vega ollama[115056]: INFO [main] HTTP server listening | hostname="127.0.0.1" n_threads_http="15" port="36639" tid="140302582510400" timestamp=1721830071
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: loaded meta data with 22 key-value pairs and 723 tensors from /usr/share/ollama/.ollama/models/blobs/sha256-0bd51f8f0c975ce910ed067dcb962a9af05b77bafcdc595ef02178387f10e51d (version GGUF V3 (latest))
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   0:                       general.architecture str              = llama
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   1:                               general.name str              = Meta-Llama-3-70B-Instruct
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   2:                          llama.block_count u32              = 80
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   4:                     llama.embedding_length u32              = 8192
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 28672
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 64
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 500000.000000
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  10:                          general.file_type u32              = 2
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  11:                           llama.vocab_size u32              = 128256
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = llama-bpe
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 128000
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 128009
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  20:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - kv  21:               general.quantization_version u32              = 2
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - type  f32:  161 tensors
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - type q4_0:  561 tensors
Jul 24 16:07:51 vega ollama[114917]: llama_model_loader: - type q6_K:    1 tensors
Jul 24 16:07:52 vega ollama[114917]: time=2024-07-24T16:07:52.065+02:00 level=INFO source=server.go:617 msg="waiting for server to become available" status="llm server loading model"
Jul 24 16:07:52 vega ollama[114917]: llm_load_vocab: special tokens cache size = 256
Jul 24 16:07:52 vega ollama[114917]: llm_load_vocab: token to piece cache size = 0.8000 MB
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: format           = GGUF V3 (latest)
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: arch             = llama
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: vocab type       = BPE
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_vocab          = 128256
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_merges         = 280147
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: vocab_only       = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_ctx_train      = 8192
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_embd           = 8192
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_layer          = 80
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_head           = 64
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_head_kv        = 8
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_rot            = 128
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_swa            = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_embd_head_k    = 128
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_embd_head_v    = 128
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_gqa            = 8
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_embd_k_gqa     = 1024
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_embd_v_gqa     = 1024
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: f_norm_eps       = 0.0e+00
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: f_clamp_kqv      = 0.0e+00
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: f_max_alibi_bias = 0.0e+00
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: f_logit_scale    = 0.0e+00
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_ff             = 28672
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_expert         = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_expert_used    = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: causal attn      = 1
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: pooling type     = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: rope type        = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: rope scaling     = linear
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: freq_base_train  = 500000.0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: freq_scale_train = 1
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: n_ctx_orig_yarn  = 8192
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: rope_finetuned   = unknown
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: ssm_d_conv       = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: ssm_d_inner      = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: ssm_d_state      = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: ssm_dt_rank      = 0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: model type       = 70B
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: model ftype      = Q4_0
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: model params     = 70.55 B
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: model size       = 37.22 GiB (4.53 BPW)
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: general.name     = Meta-Llama-3-70B-Instruct
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: EOS token        = 128009 '<|eot_id|>'
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: LF token         = 128 'Ä'
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
Jul 24 16:07:52 vega ollama[114917]: llm_load_print_meta: max token length = 256
Jul 24 16:07:53 vega ollama[114917]: ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
Jul 24 16:07:53 vega ollama[114917]: ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
Jul 24 16:07:53 vega ollama[114917]: ggml_cuda_init: found 1 ROCm devices:
Jul 24 16:07:53 vega ollama[114917]:   Device 0: Radeon RX 7900 XTX, compute capability 11.0, VMM: no
Jul 24 16:07:53 vega ollama[114917]: llm_load_tensors: ggml ctx size =    0.68 MiB
Jul 24 16:07:55 vega ollama[114917]: llm_load_tensors: offloading 48 repeating layers to GPU
Jul 24 16:07:55 vega ollama[114917]: llm_load_tensors: offloaded 48/81 layers to GPU
Jul 24 16:07:55 vega ollama[114917]: llm_load_tensors:      ROCm0 buffer size = 22035.00 MiB
Jul 24 16:07:55 vega ollama[114917]: llm_load_tensors:        CPU buffer size = 38110.61 MiB
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: n_ctx      = 2048
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: n_batch    = 512
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: n_ubatch   = 512
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: flash_attn = 0
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: freq_base  = 500000.0
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: freq_scale = 1
Jul 24 16:07:57 vega ollama[114917]: llama_kv_cache_init:      ROCm0 KV buffer size =   384.00 MiB
Jul 24 16:07:57 vega ollama[114917]: llama_kv_cache_init:  ROCm_Host KV buffer size =   256.00 MiB
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: KV self size  =  640.00 MiB, K (f16):  320.00 MiB, V (f16):  320.00 MiB
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model:  ROCm_Host  output buffer size =     0.52 MiB
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model:      ROCm0 compute buffer size =  1104.45 MiB
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model:  ROCm_Host compute buffer size =    20.01 MiB
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: graph nodes  = 2566
Jul 24 16:07:57 vega ollama[114917]: llama_new_context_with_model: graph splits = 356
Jul 24 16:07:58 vega ollama[115056]: INFO [main] model loaded | tid="140302582510400" timestamp=1721830078
Jul 24 16:07:58 vega ollama[114917]: time=2024-07-24T16:07:58.350+02:00 level=INFO source=server.go:622 msg="llama runner started in 6.54 seconds"
Jul 24 16:07:58 vega ollama[114917]: CUDA error: out of memory
Jul 24 16:07:58 vega ollama[114917]:   current device: 0, in function alloc at /go/src/github.com/ollama/ollama/llm/llama.cpp/ggml/src/ggml-cuda.cu:291
Jul 24 16:07:58 vega ollama[114917]:   ggml_cuda_device_malloc(&ptr, look_ahead_size, device)
Jul 24 16:07:58 vega ollama[114917]: GGML_ASSERT: /go/src/github.com/ollama/ollama/llm/llama.cpp/ggml/src/ggml-cuda.cu:101: !"CUDA error"
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115057]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115058]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115059]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115060]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115061]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115062]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115063]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115064]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115065]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115066]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115067]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115068]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115069]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115070]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115071]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115072]
Jul 24 16:07:58 vega ollama[115138]: [New LWP 115083]
Jul 24 16:07:58 vega ollama[115138]: [Thread debugging using libthread_db enabled]
Jul 24 16:07:58 vega ollama[115138]: Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
Jul 24 16:07:58 vega ollama[115138]: 0x00007f9abdcf2b57 in __GI___wait4 (pid=115138, stat_loc=0x0, options=0, usage=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
Jul 24 16:07:58 vega ollama[114917]: 30        ../sysdeps/unix/sysv/linux/wait4.c: No such file or directory.
Jul 24 16:07:58 vega ollama[115138]: #0  0x00007f9abdcf2b57 in __GI___wait4 (pid=115138, stat_loc=0x0, options=0, usage=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
Jul 24 16:07:58 vega ollama[115138]: 30        in ../sysdeps/unix/sysv/linux/wait4.c
Jul 24 16:07:58 vega ollama[115138]: #1  0x00000000178d6c87 in ggml_cuda_error(char const*, char const*, char const*, int, char const*) ()
Jul 24 16:07:58 vega ollama[115138]: #2  0x00000000178ec50e in ggml_cuda_pool_leg::alloc(unsigned long, unsigned long*) ()
Jul 24 16:07:58 vega ollama[115138]: #3  0x00000000178ecc60 in ggml_cuda_pool_alloc<__half>::alloc(unsigned long) ()
Jul 24 16:07:58 vega ollama[115138]: #4  0x00000000178e2afb in ggml_cuda_mul_mat(ggml_backend_cuda_context&, ggml_tensor const*, ggml_tensor const*, ggml_tensor*) ()
Jul 24 16:07:58 vega ollama[115138]: #5  0x00000000178da448 in ggml_backend_cuda_graph_compute(ggml_backend*, ggml_cgraph*) ()
Jul 24 16:07:58 vega ollama[115138]: #6  0x000000001789eac8 in ggml_backend_sched_graph_compute_async ()
Jul 24 16:07:58 vega ollama[115138]: #7  0x0000000017a85789 in llama_decode ()
Jul 24 16:07:58 vega ollama[115138]: #8  0x00000000177aac61 in llama_server_context::update_slots() ()
Jul 24 16:07:58 vega ollama[115138]: #9  0x00000000177ace2a in llama_server_queue::start_loop() ()
Jul 24 16:07:58 vega ollama[115138]: #10 0x000000001779073f in main ()
Jul 24 16:07:58 vega ollama[115138]: [Inferior 1 (process 115056) detached]
Jul 24 16:07:59 vega ollama[114917]: [GIN] 2024/07/24 - 16:07:59 | 200 |  7.978114451s | 2a02:b30:fac:1f00:7285:c2ff:fe0f:69f3 | POST     "/api/chat"

有任何关于这里可能出错的地方的建议吗？
运行较小的模型运行得很好。

OS

Linux 6.1.0-22-amd64(debian稳定版)

GPU

AMD Radeon RX 7900 XTX(24 GiB VRAM)

CPU

AMD Ryzen 7 7700X

Ollama版本

0.2.8

型号

llama3:70b-instruct-q4_0

ollama

来源：https://github.com/ollama/ollama/issues/5913

5条答案

按热度按时间

zvms9eto1#

没有答案，只有观察。您的卡有24G,llama.cpp正在分配23.9 GiB。有两种方法可以缓解这个问题。第一种方法是减少llama.cpp卸载到卡上的层数，可以通过在API调用中添加"options": {"num_gpu": 46}来实现，其中46是要卸载的层数，请参阅日志中的offloaded -较低的数字将减轻内存压力。由于您正在使用UI,可能无法微调API调用，因此可以创建一个具有内置层数的新模型：

$ ollama show --modelfile llama3.1:70b-instruct-q4_0 > Modelfile
# edit Modelfile and add "PARAMETER num_gpu 46"
$ ollama create llama3.1:70b-instruct-n46-q4_0 -f Modelfile

您还可以尝试通过在服务器环境中添加OLLAMA_FLASH_ATTENTION=1来打开闪存注意力。闪存注意力更好地利用KV缓存，因此也可能减轻内存压力。
话虽如此，我也偶尔遇到OOM问题，所以内存计算需要一些细心呵护。

赞(0）回复(0）举报 1个月前

bihw5rsg2#

感谢rick-github提供的提示！
尽管在这种情况下，OLLAMA_FLASH_ATTENTION=1似乎不会对任何事情产生影响，但我可以通过手动使用46层来确认这一点确实可以解决这个问题。

赞(0）回复(0）举报 1个月前

tkqqtvp13#

你正在尝试加载哪个模型？

赞(0）回复(0）举报 1个月前

67up9zun4#

llama3:70b-instruct-q4_0

赞(0）回复(0）举报 1个月前

zpf6vheq5#

@dhiltgen llama3:70b
FWIW: 我刚刚尝试了 CognitiveComputations/dolphin-qwen2:72b-v2.9.2-q4_k_s ,发现这里的卸载似乎可以正常工作。对于 LLAMA3,它是 memory.required.partial="23.9 GiB"
[点击展开 ollama service 的输出]

Jul 24 21:50:02 vega ollama[132945]: time=2024-07-24T21:50:02.743+02:00 level=INFO source=memory.go:309 msg="offload to rocm" layers.requested=-1 layers.model=81 layers.offload=43 layers.split="" memory.available="[24.0 GiB]" memory.required.full="43.1 GiB" memory.required.partial="23.7 GiB" memory.required.kv="640.0 MiB" memory.required.allocations="[23.7 GiB]" memory.weights.total="39.9 GiB" memory.weights.repeating="38.9 GiB" memory.weights.nonrepeating="974.6 MiB" memory.graph.full="313.0 MiB" memory.graph.partial="1.3 GiB"
Jul 24 21:50:02 vega ollama[132945]: time=2024-07-24T21:50:02.743+02:00 level=INFO source=server.go:383 msg="starting llama server" cmd="/tmp/ollama3633236186/runners/rocm_v60102/ollama_llama_server --model /usr/share/ollama/.ollama/models/blobs/sha256-11fe3cb8956204489b5d545e5c8bae6cf888d110f0fd39bc2afdb6a5644d76be --ctx-size 2048 --batch-size 512 --embedding --log-disable --n-gpu-layers 43 --parallel 1 --port 45365"
Jul 24 21:50:02 vega ollama[132945]: time=2024-07-24T21:50:02.743+02:00 level=INFO source=sched.go:437 msg="loaded runners" count=1
Jul 24 21:50:02 vega ollama[132945]: time=2024-07-24T21:50:02.743+02:00 level=INFO source=server.go:583 msg="waiting for llama runner to start responding"
Jul 24 21:50:02 vega ollama[132945]: time=2024-07-24T21:50:02.743+02:00 level=INFO source=server.go:617 msg="waiting for server to become available" status="llm server error"
Jul 24 21:50:02 vega ollama[174233]: INFO [main] build info | build=1 commit="d94c6e0" tid="139910788932416" timestamp=1721850602
Jul 24 21:50:02 vega ollama[174233]: INFO [main] system info | n_threads=8 n_threads_batch=-1 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 0 | " tid="139910788932416" timestamp=1721850602 total_threads=16
Jul 24 21:50:02 vega ollama[174233]: INFO [main] HTTP server listening | hostname="127.0.0.1" n_threads_http="15" port="45365" tid="139910788932416" timestamp=1721850602
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: loaded meta data with 24 key-value pairs and 963 tensors from /usr/share/ollama/.ollama/models/blobs/sha256-11fe3cb8956204489b5d545e5c8bae6cf888d110f0fd39bc2afdb6a5644d76be (version GGUF V3 (latest))
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   0:                       general.architecture str              = qwen2
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   1:                               general.name str              = qwen2
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   2:                          qwen2.block_count u32              = 80
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   3:                       qwen2.context_length u32              = 131072
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   4:                     qwen2.embedding_length u32              = 8192
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   5:                  qwen2.feed_forward_length u32              = 29568
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   6:                 qwen2.attention.head_count u32              = 64
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   7:              qwen2.attention.head_count_kv u32              = 8
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   8:                       qwen2.rope.freq_base f32              = 1000000.000000
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv   9:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000010
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  10:                          general.file_type u32              = 14
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = gpt2
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  12:                         tokenizer.ggml.pre str              = qwen2
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  15:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32              = 151645
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  17:            tokenizer.ggml.padding_token_id u32              = 151643
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  18:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  19:               general.quantization_version u32              = 2
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  20:                      quantize.imatrix.file str              = /models/qwen2-GGUF/qwen2.imatrix
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  21:                   quantize.imatrix.dataset str              = /training_data/calibration_datav3.txt
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  22:             quantize.imatrix.entries_count i32              = 560
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - kv  23:              quantize.imatrix.chunks_count i32              = 128
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - type  f32:  401 tensors
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - type q5_0:   70 tensors
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - type q5_1:   10 tensors
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - type q4_K:  401 tensors
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - type q5_K:   80 tensors
Jul 24 21:50:02 vega ollama[132945]: llama_model_loader: - type q6_K:    1 tensors
Jul 24 21:50:02 vega ollama[132945]: llm_load_vocab: special tokens cache size = 421
Jul 24 21:50:02 vega ollama[132945]: time=2024-07-24T21:50:02.994+02:00 level=INFO source=server.go:617 msg="waiting for server to become available" status="llm server loading model"
Jul 24 21:50:03 vega ollama[132945]: llm_load_vocab: token to piece cache size = 0.9352 MB
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: format           = GGUF V3 (latest)
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: arch             = qwen2
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: vocab type       = BPE
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_vocab          = 152064
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_merges         = 151387
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: vocab_only       = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_ctx_train      = 131072
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_embd           = 8192
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_layer          = 80
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_head           = 64
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_head_kv        = 8
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_rot            = 128
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_swa            = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_embd_head_k    = 128
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_embd_head_v    = 128
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_gqa            = 8
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_embd_k_gqa     = 1024
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_embd_v_gqa     = 1024
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: f_norm_eps       = 0.0e+00
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: f_clamp_kqv      = 0.0e+00
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: f_max_alibi_bias = 0.0e+00
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: f_logit_scale    = 0.0e+00
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_ff             = 29568
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_expert         = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_expert_used    = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: causal attn      = 1
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: pooling type     = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: rope type        = 2
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: rope scaling     = linear
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: freq_base_train  = 1000000.0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: freq_scale_train = 1
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: n_ctx_orig_yarn  = 131072
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: rope_finetuned   = unknown
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: ssm_d_conv       = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: ssm_d_inner      = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: ssm_d_state      = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: ssm_dt_rank      = 0
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: model type       = 70B
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: model ftype      = Q4_K - Small
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: model params     = 72.71 B
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: model size       = 40.87 GiB (4.83 BPW)
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: general.name     = qwen2
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: BOS token        = 11 ','
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: LF token         = 148848 'ÄĬ'
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
Jul 24 21:50:03 vega ollama[132945]: llm_load_print_meta: max token length = 256
Jul 24 21:50:04 vega ollama[132945]: ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
Jul 24 21:50:04 vega ollama[132945]: ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
Jul 24 21:50:04 vega ollama[132945]: ggml_cuda_init: found 1 ROCm devices:
Jul 24 21:50:04 vega ollama[132945]:   Device 0: Radeon RX 7900 XTX, compute capability 11.0, VMM: no
Jul 24 21:50:04 vega ollama[132945]: llm_load_tensors: ggml ctx size =    0.85 MiB
Jul 24 21:50:05 vega ollama[132945]: time=2024-07-24T21:50:05.702+02:00 level=INFO source=server.go:617 msg="waiting for server to become available" status="llm server not responding"
Jul 24 21:50:06 vega ollama[132945]: time=2024-07-24T21:50:06.542+02:00 level=INFO source=server.go:617 msg="waiting for server to become available" status="llm server loading model"
Jul 24 21:50:06 vega ollama[132945]: llm_load_tensors: offloading 43 repeating layers to GPU
Jul 24 21:50:06 vega ollama[132945]: llm_load_tensors: offloaded 43/81 layers to GPU
Jul 24 21:50:06 vega ollama[132945]: llm_load_tensors:      ROCm0 buffer size = 21533.93 MiB
Jul 24 21:50:06 vega ollama[132945]: llm_load_tensors:        CPU buffer size = 41850.31 MiB
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: n_ctx      = 2048
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: n_batch    = 512
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: n_ubatch   = 512
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: flash_attn = 0
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: freq_base  = 1000000.0
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: freq_scale = 1
Jul 24 21:50:08 vega ollama[132945]: llama_kv_cache_init:      ROCm0 KV buffer size =   344.00 MiB
Jul 24 21:50:08 vega ollama[132945]: llama_kv_cache_init:  ROCm_Host KV buffer size =   296.00 MiB
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: KV self size  =  640.00 MiB, K (f16):  320.00 MiB, V (f16):  320.00 MiB
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model:  ROCm_Host  output buffer size =     0.61 MiB
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model:      ROCm0 compute buffer size =  1287.53 MiB
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model:  ROCm_Host compute buffer size =    20.01 MiB
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: graph nodes  = 2806
Jul 24 21:50:08 vega ollama[132945]: llama_new_context_with_model: graph splits = 522
Jul 24 21:50:09 vega ollama[174233]: INFO [main] model loaded | tid="139910788932416" timestamp=1721850609
Jul 24 21:50:09 vega ollama[132945]: time=2024-07-24T21:50:09.556+02:00 level=INFO source=server.go:622 msg="llama runner started in 6.81 seconds"
Jul 24 21:51:18 vega ollama[132945]: [GIN] 2024/07/24 - 21:51:18 | 200 |         1m15s |       127.0.0.1 | POST     "/api/generate"
Jul 24 21:53:00 vega ollama[132945]: [GIN] 2024/07/24 - 21:53:00 | 200 |      14.768µs |       127.0.0.1 | HEAD     "/"
Jul 24 21:53:00 vega ollama[132945]: [GIN] 2024/07/24 - 21:53:00 | 200 |    9.223694ms |       127.0.0.1 | POST     "/api/show"
Jul 24 21:53:19 vega ollama[132945]: [GIN] 2024/07/24 - 21:53:19 | 200 | 18.637106356s |       127.0.0.1 | POST     "/api/generate"
Jul 24 21:53:37 vega ollama[132945]: [GIN] 2024/07/24 - 21:53:37 | 200 |    1.310385ms | 2a02:b30:fac:1f00:7285:c2ff:fe0f:69f3 | GET      "/api/tags"
Jul 24 21:53:37 vega ollama[132945]: [GIN] 2024/07/24 - 21:53:37 | 200 |      34.534µs | 2a02:b30:fac:1f00:7285:c2ff:fe0f:69f3 | GET      "/api/version"
Jul 24 21:54:48 vega ollama[132945]: [GIN] 2024/07/24 - 21:54:48 | 200 | 53.933274734s | 2a02:b30:fac:1f00:7285:c2ff:fe0f:69f3 | POST     "/api/chat"
Jul 24 21:54:54 vega ollama[132945]: [GIN] 2024/07/24 - 21:54:54 | 200 |  5.307397265s | 2a02:b30:fac:1f00:7285:c2ff:fe0f:69f3 | POST     "/v1/chat/completions"
Jul 24 21:55:40 vega ollama[132945]: [GIN] 2024/07/24 - 21:55:40 | 200 |      30.527µs | 2a02:b30:fac:1f00:7285:c2ff:fe0f:69f3 | GET      "/api/version"
Jul 24 21:56:42 vega ollama[132945]: [GIN] 2024/07/24 - 21:56:42 | 200 |      13.635µs |       127.0.0.1 | HEAD     "/"
Jul 24 21:56:42 vega ollama[132945]: [GIN] 2024/07/24 - 21:56:42 | 200 |    9.354188ms |       127.0.0.1 | POST     "/api/show"
Jul 24 21:57:43 vega ollama[132945]: [GIN] 2024/07/24 - 21:57:43 | 200 |          1m1s |       127.0.0.1 | POST     "/api/generate"

赞(0）回复(0）举报 1个月前

我来回答

ollama 在ROCm上卸载层时出现内存不足

问题是什么？

错误输出示例

OS

GPU

CPU

Ollama版本

型号

5条答案

相关问题

热门标签

最新问答