llm_load_tensors: offloading 80 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 81/81 layers to GPU
llm_load_tensors:   CPU_Mapped model buffer size =   563.62 MiB
llm_load_tensors:        CUDA0 model buffer size =  5648.81 MiB
llm_load_tensors:        CUDA1 model buffer size =  4777.06 MiB
llm_load_tensors:        CUDA2 model buffer size =  4835.88 MiB
llm_load_tensors:        CUDA3 model buffer size =  4777.06 MiB
llm_load_tensors:        CUDA4 model buffer size =  4777.06 MiB
llm_load_tensors:        CUDA5 model buffer size =  4835.88 MiB
llm_load_tensors:        CUDA6 model buffer size =  4835.88 MiB
llm_load_tensors:        CUDA7 model buffer size =  5491.86 MiB
llama_new_context_with_model: n_seq_max     = 4
llama_new_context_with_model: n_ctx         = 8192
llama_new_context_with_model: n_ctx_per_seq = 2048
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_new_context_with_model: n_ctx_per_seq (2048) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init: kv_size = 8192, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 80, can_shift = 1
llama_kv_cache_init:      CUDA0 KV buffer size =   352.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =   320.00 MiB
llama_kv_cache_init:      CUDA2 KV buffer size =   320.00 MiB
llama_kv_cache_init:      CUDA3 KV buffer size =   320.00 MiB
llama_kv_cache_init:      CUDA4 KV buffer size =   320.00 MiB
llama_kv_cache_init:      CUDA5 KV buffer size =   320.00 MiB
llama_kv_cache_init:      CUDA6 KV buffer size =   320.00 MiB
llama_kv_cache_init:      CUDA7 KV buffer size =   288.00 MiB
llama_new_context_with_model: KV self size  = 2560.00 MiB, K (f16): 1280.00 MiB, V (f16): 1280.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     2.08 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model:      CUDA0 compute buffer size =  1216.01 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =  1216.01 MiB
llama_new_context_with_model:      CUDA2 compute buffer size =  1216.01 MiB
llama_new_context_with_model:      CUDA3 compute buffer size =  1216.01 MiB
llama_new_context_with_model:      CUDA4 compute buffer size =  1216.01 MiB
llama_new_context_with_model:      CUDA5 compute buffer size =  1216.01 MiB
llama_new_context_with_model:      CUDA6 compute buffer size =  1216.01 MiB
llama_new_context_with_model:      CUDA7 compute buffer size =  1216.02 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    80.02 MiB
llama_new_context_with_model: graph nodes  = 2566
llama_new_context_with_model: graph splits = 9
time=2025-02-11T11:34:37.966+08:00 level=INFO source=server.go:594 msg="llama runner started in 231.27 seconds"
llama_model_loader: loaded meta data with 30 key-value pairs and 724 tensors from /root/.ollama/models/blobs/sha256-4cd576d9aa16961244012223abf01445567b061f1814b57dfef699e4cf8df339 (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              = DeepSeek R1 Distill Llama 70B
llama_model_loader: - kv   3:                           general.basename str              = DeepSeek-R1-Distill-Llama
llama_model_loader: - kv   4:                         general.size_label str              = 70B
llama_model_loader: - kv   5:                          llama.block_count u32              = 80
llama_model_loader: - kv   6:                       llama.context_length u32              = 131072
llama_model_loader: - kv   7:                     llama.embedding_length u32              = 8192
llama_model_loader: - kv   8:                  llama.feed_forward_length u32              = 28672
llama_model_loader: - kv   9:                 llama.attention.head_count u32              = 64
llama_model_loader: - kv  10:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  11:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  12:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  13:                 llama.attention.key_length u32              = 128
llama_model_loader: - kv  14:               llama.attention.value_length u32              = 128
llama_model_loader: - kv  15:                          general.file_type u32              = 15
llama_model_loader: - kv  16:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  17:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  18:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  19:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  20:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  21:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  22:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  23:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  24:                tokenizer.ggml.eos_token_id u32              = 128001
llama_model_loader: - kv  25:            tokenizer.ggml.padding_token_id u32              = 128001
llama_model_loader: - kv  26:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  27:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  28:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  29:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  162 tensors
llama_model_loader: - type q4_K:  441 tensors
llama_model_loader: - type q5_K:   40 tensors
llama_model_loader: - type q6_K:   81 tensors
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       = 1
llm_load_print_meta: model type       = ?B
llm_load_print_meta: model ftype      = all F32
llm_load_print_meta: model params     = 70.55 B
llm_load_print_meta: model size       = 39.59 GiB (4.82 BPW) 
llm_load_print_meta: general.name     = DeepSeek R1 Distill Llama 70B
llm_load_print_meta: BOS token        = 128000 '<|begin▁of▁sentence|>'
llm_load_print_meta: EOS token        = 128001 '<|end▁of▁sentence|>'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: EOM token        = 128008 '<|eom_id|>'
llm_load_print_meta: PAD token        = 128001 '<|end▁of▁sentence|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOG token        = 128001 '<|end▁of▁sentence|>'
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
llama_model_load: vocab only - skipping tensors

从日志来看,模型正在进行 张量卸载(offloading)到多个 GPU,这意味着模型的不同层正在被分配到不同的 GPU 上。以下是对日志内容的解释和可能的改进建议:

解释日志内容

  1. llm_load_tensors: offloading 80 repeating layers to GPU
    表示模型的 80 个重复层正在被卸载到 GPU 中。这通常意味着模型很大,超出了单个 GPU 的内存容量,因此需要跨多个 GPU 分配。

  2. llm_load_tensors: offloading output layer to GPU
    模型的输出层正在被卸载到 GPU。这个输出层通常比其他层大,因此它需要分配更多的内存。

  3. llm_load_tensors: offloaded X/Y layers to GPU
    显示已经成功卸载到各个 GPU 上的层数和每个 GPU 上的内存使用情况。

    • 比如,CUDA0(GPU 0)分配了 5648.81 MiB 内存,CUDA1(GPU 1)分配了 4777.06 MiB 内存,以此类推。
  4. CUDA{n} model buffer size
    显示每个 GPU 上分配的内存大小。这里说明了每个 GPU 上的内存分配情况,可能根据 GPU 的显存大小进行优化。

  5. llama_new_context_with_model: n_seq_max = 4
    这表示模型上下文的最大序列长度(n_seq_max)为 4,可能用于控制模型能处理的最大输入序列的长度。


分析与建议

  • 多 GPU 分配:日志显示模型已经成功地被卸载到多个 GPU 上,每个 GPU 分配了不同的内存量。你有 8 个 GPU (CUDA0CUDA7),而且模型的总内存需求也跨越了这些设备。这种分配方式表明模型过大,无法在单一 GPU 上运行,因此采用了跨多个 GPU 的并行计算。每个 GPU 上的内存大小差异,可能是由于显卡性能和显存容量不同。

  • 内存使用:某些 GPU(例如 CUDA0 和 CUDA6)分配了更多的内存(5648.81 MiB 和 4835.88 MiB),而其他 GPU 则分配了较少的内存(4777.06 MiB)。这可能表示模型层被优化并根据每个 GPU 的负载进行分配。

  • 优化 GPU 使用

    • 如果你有多个 GPU,Ollama 会尽量优化资源分配。如果 GPU 内存不足,可以考虑减少模型加载的大小,或将一些层分配给 CPU(通过参数设置),但这会影响推理速度。
    • 如果你只使用特定的 GPU 或者需要控制模型如何分布在 GPU 上,你可以通过配置 CUDA_VISIBLE_DEVICES 环境变量来指定使用哪些 GPU。例如:
      export CUDA_VISIBLE_DEVICES=0,1  # 只使用 GPU 0 和 GPU 1
      
  • 最大序列长度 (n_seq_max)

    • n_seq_max 设置为 4 表示模型能处理的最大序列长度为 4。根据任务的不同,你可能希望调整这个参数来优化处理能力。增大 n_seq_max 会影响内存使用和计算速度,但有时也需要调整它以适应输入数据的特定要求。

总结

  • 如果你想优化 GPU 使用,确保模型层合理分配到多个 GPU 上,且显存足够。检查 CUDA_VISIBLE_DEVICES 和相关环境配置,确保不会出现资源瓶颈。
  • 通过调整 n_seq_max 和序列长度,可以优化模型的输入输出。
  • 根据 GPU 的内存和性能配置,合理选择每个设备的工作负载,以确保模型推理的高效性和稳定性。
Logo

尧米是由西云算力与CSDN联合运营的AI算力和模型开源社区品牌,为基于DaModel智算平台的AI应用企业和泛AI开发者提供技术交流与成果转化平台。

更多推荐