Bug: RPC inference is drastically slower even on localhost

[hafezmg48]

What happened?

I am trying to run inference on RPC example. When running the llama-cli with rpc feature over a single rpc-server on localhost, the inference throughput is only 1.9 tok/sec for llama3.1-8B on CUDA, while the same llama-cli on local cuda build without rpc generates 25 tok/sec.

So about 13x slower even thought the server is in localhost, basically using same GPU locally but through rpc.

Name and Version

followed exact steps in https://github.com/ggerganov/llama.cpp/tree/master/examples/rpc

running cli with command: bin/llama-cli -m ./llama3.1-8B-F16.gguf -p "Hello, my name is" -n 64 --rpc localhost:50052 -ngl 99

running rpc-server: bin/rpc-server -p 50052 create_backend: using CUDA backend 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: Tesla V100-PCIE-32GB, compute capability 7.0, VMM: yes Starting RPC server on 0.0.0.0:50052, backend memory: 28170 MB Accepted client connection, free_mem=29538713600, total_mem=34079899648 Client connection closed

What operating system are you seeing the problem on?

Linux

Relevant log output

Log start
main: build = 3501 (b7a08fd5)
main: built with cc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 for x86_64-linux-gnu
main: seed  = 1722612132
llama_model_loader: loaded meta data with 27 key-value pairs and 291 tensors from ./llama3.1-8B-F16.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              = Llama3.1 8b
llama_model_loader: - kv   3:                           general.basename str              = llama3.1
llama_model_loader: - kv   4:                         general.size_label str              = 8B
llama_model_loader: - kv   5:                            general.license str              = llama3
llama_model_loader: - kv   6:                               general.tags arr[str,6]       = ["facebook", "meta", "pytorch", "llam...
llama_model_loader: - kv   7:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv   8:                          llama.block_count u32              = 32
llama_model_loader: - kv   9:                       llama.context_length u32              = 8192
llama_model_loader: - kv  10:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv  11:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv  12:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv  13:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  14:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv  15:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  16:                          general.file_type u32              = 1
llama_model_loader: - kv  17:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  18:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  19:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  20:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  21:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  22:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  23:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  24:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  25:                tokenizer.ggml.eos_token_id u32              = 128001
llama_model_loader: - kv  26:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type  f16:  226 tensors
llm_load_vocab: special tokens cache size = 256
llm_load_vocab: token to piece cache size = 0.8000 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      = 8192
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_head           = 32
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            = 4
llm_load_print_meta: n_embd_k_gqa     = 1024
llm_load_print_meta: n_embd_v_gqa     = 1024
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 14336
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 500000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 8192
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: model type       = 8B
llm_load_print_meta: model ftype      = F16
llm_load_print_meta: model params     = 8.03 B
llm_load_print_meta: model size       = 14.96 GiB (16.00 BPW)
llm_load_print_meta: general.name     = Llama3.1 8b
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128001 '<|end_of_text|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOT token        = 128009 '<|eot_id|>'
llm_load_print_meta: max token length = 256
llm_load_tensors: ggml ctx size =    0.27 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors: RPC[localhost:50052] buffer size = 13898.98 MiB
llm_load_tensors:        CPU buffer size =  1418.03 MiB
.........................................................................................
llama_new_context_with_model: n_ctx      = 8192
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: RPC[localhost:50052] KV buffer size =   992.00 MiB
llama_kv_cache_init:        CPU KV buffer size =    32.00 MiB
llama_new_context_with_model: KV self size  = 1024.00 MiB, K (f16):  512.00 MiB, V (f16):  512.00 MiB
llama_new_context_with_model:        CPU  output buffer size =     0.49 MiB
llama_new_context_with_model: RPC[localhost:50052] compute buffer size =   560.00 MiB
llama_new_context_with_model:        CPU compute buffer size =   560.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2

system_info: n_threads = 44 / 88 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |
sampling:
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order:
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temperature
generate: n_ctx = 8192, n_batch = 2048, n_predict = 64, n_keep = 1

Hello, my name is Paul. I am a graduate from the University of Michigan with a B.S. in Computer Science and a minor in Mathematics. I love math, and I know that it can be intimidating to some people. That is why I enjoy tutoring because it allows me to help those students who have trouble grasping certain concepts in math
llama_print_timings:        load time =   32468.00 ms
llama_print_timings:      sample time =      10.54 ms /    64 runs   (    0.16 ms per token,  6072.68 tokens per second)
llama_print_timings: prompt eval time =     617.41 ms /     6 tokens (  102.90 ms per token,     9.72 tokens per second)
llama_print_timings:        eval time =   32962.85 ms /    63 runs   (  523.22 ms per token,     1.91 tokens per second)
llama_print_timings:       total time =   33624.89 ms /    69 tokens
Log end

[ski422]

I'm not sure if it's the same issue (I am not using RPC), but the inference speed has dramatically slowed compared to the older version of llama.cpp.

Despite performing CPU-based inference, it's about 4-5 times slower than the version that ran "main" (2-3 tokens/sec -> 0.4 tokens/sec). For reference, the performance of the old and current versions of llama.cpp was compared on the same server.

Is there a solution to this performance issue?

[rgerganov]

@hafezmg48 I am not able to reproduce such regression with the latest code (commit 6e02327e8b783):

Results without RPC:

➜  build-rpc-cuda git:(master) ✗ bin/llama-bench -m ../models/tinyllama-1.1b-f16.gguf -ngl 99 -fa 1                      
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 GTX 1660, compute capability 7.5, VMM: yes
| model                          |       size |     params | backend    | ngl | fa |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ------------: | ---------------: |
| llama 1B F16                   |   2.05 GiB |     1.10 B | CUDA       |  99 |  1 |         pp512 |    315.02 ± 0.34 |
| llama 1B F16                   |   2.05 GiB |     1.10 B | CUDA       |  99 |  1 |         tg128 |     74.83 ± 0.05 |

build: 6e02327e (3565)

Results with RPC:

➜  build-rpc-cuda git:(master) ✗ bin/llama-bench -m ../models/tinyllama-1.1b-f16.gguf -ngl 99 -fa 1 --rpc localhost:50052
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 GTX 1660, compute capability 7.5, VMM: yes
| model                          |       size |     params | backend    | ngl | fa |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ------------: | ---------------: |
| llama 1B F16                   |   2.05 GiB |     1.10 B | CUDA+RPC   |  99 |  1 |         pp512 |    314.06 ± 0.30 |
| llama 1B F16                   |   2.05 GiB |     1.10 B | CUDA+RPC   |  99 |  1 |         tg128 |     68.74 ± 0.01 |

build: 6e02327e (3565)

Can you try llama-bench and post the results with different models?