Distributed inference on Hugging Face Spaces

[rgerganov]

I have been playing with HF spaces and decided to try running rpc-server there. I had some partial success so I will share my experience here.

Their Docker space allows you to expose a TCP port to the outside world, so my first thought was to simply run rpc-server in a Docker container and expoose its port. Unfortunately this didn't work as they support only HTTP(2) based traffic and in rpc-server we use custom binary protocol. I have tried to workaround with some tunneling solutions like wstunnel with no success.

Then I found they have SSH support for PRO users and decided to try it out. The idea is to setup an SSH tunnel between two spaces and forward traffic to rpc-server:

flowchart LR
    sshc---|SSH tunnel|sshs
    subgraph hfs2[HF Space 2]
    sshs[SSH server]-->rpcs["rpc-server:50052"]
    end
    subgraph hfs1[HF Space 1]
    llamacli[llama-cli]-->sshc[SSH client]
    end

Loading

The steps to do this are:

1. SSH into Space 1 and create a tunnel to Space 2 with the following command (id_ed25519 is a disposable SSH key):

$ ssh -i ../id_ed25519 -L localhost:50052:localhost:50052 [email protected]

2. Start rpc-server in Space 2 on its default port (50052) which is now forwarded to Space 1

3. Run llama-cli in Space 1, offloading to localhost:50052

I gave this a try with Meta-Llama-3-70B-Instruct-IQ2_M.gguf (22.46 GiB) and two spaces each running on NVidia T4 small (16GB VRAM) and here is the result:

user@r-rgerganov-llama-rpc-server-p526g9pp-a5786-f2als:~/app$ ./llama-cli-cuda --rpc localhost:50052 -m models/Meta-Llama-3-70B-Instruct-IQ2_M.gguf -p "hello, my name is" -ngl 99 -n 256
Log start
main: build = 3615 (1731d423)
main: built with cc (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0 for x86_64-linux-gnu
main: seed  = 1724323519
llama_model_loader: loaded meta data with 26 key-value pairs and 723 tensors from models/Meta-Llama-3-70B-Instruct-IQ2_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.name str              = Meta-Llama-3-70B-Instruct
llama_model_loader: - kv   2:                          llama.block_count u32              = 80
llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 8192
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 28672
llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 64
llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 29
llama_model_loader: - kv  11:                           llama.vocab_size u32              = 128256
llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = llama-bpe
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,128256]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,128256]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  17:                      tokenizer.ggml.merges arr[str,280147]  = ["Ġ Ġ", "Ġ ĠĠĠ", "ĠĠ ĠĠ", "...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 128009
llama_model_loader: - kv  20:                    tokenizer.chat_template str              = {% set loop_messages = messages %}{% ...
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - kv  22:                      quantize.imatrix.file str              = /models/Meta-Llama-3-70B-Instruct-GGU...
llama_model_loader: - kv  23:                   quantize.imatrix.dataset str              = /training_data/calibration_datav3.txt
llama_model_loader: - kv  24:             quantize.imatrix.entries_count i32              = 560
llama_model_loader: - kv  25:              quantize.imatrix.chunks_count i32              = 125
llama_model_loader: - type  f32:  161 tensors
llama_model_loader: - type q4_K:   80 tensors
llama_model_loader: - type q5_K:    1 tensors
llama_model_loader: - type iq3_s:   91 tensors
llama_model_loader: - type iq2_s:  390 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           = 8192
llm_load_print_meta: n_layer          = 80
llm_load_print_meta: n_head           = 64
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            = 8
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             = 28672
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: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = 70B
llm_load_print_meta: model ftype      = IQ2_M - 2.7 bpw
llm_load_print_meta: model params     = 70.55 B
llm_load_print_meta: model size       = 22.46 GiB (2.73 BPW) 
llm_load_print_meta: general.name     = Meta-Llama-3-70B-Instruct
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128009 '<|eot_id|>'
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
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 T4, compute capability 7.5, VMM: yes
llm_load_tensors: ggml ctx size =    1.01 MiB
llm_load_tensors: offloading 80 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 81/81 layers to GPU
llm_load_tensors: RPC[localhost:50052] buffer size = 11233.53 MiB
llm_load_tensors:        CPU buffer size =   430.55 MiB
llm_load_tensors:      CUDA0 buffer size = 11330.38 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 =  1248.00 MiB
llama_kv_cache_init:      CUDA0 KV buffer size =  1312.00 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 =     0.49 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: RPC[localhost:50052] 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 = 3

system_info: n_threads = 2 / 4 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | 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 = 256, n_keep = 1


hello, my name is rahul and i am 18 years old. i am a student of class 12th. i have a passion for coding and i want to become a software engineer. i am new to this field and i don't know how to start. can you help me out? hello, my name is rahul and i am 18 years old. i am a student of class 12th. i have a passion for coding and i want to become a software engineer. i am new to this field and i don't know how to start. can you help me out? hello, my name is rahul and i am 18 years old. i am a student of class 12th. i have a passion for coding and i want to become a software engineer. i am new to this field and i don't know how to start. can you help me out?

Hello Rahul!

Congratulations on taking the first step towards your passion for coding! It's great to see enthusiasm and dedication at a young age.

Don't worry, I'm more than happy to help you get started. Here's a step-by-step guide to help you begin your coding journey:

**Step 1: Choose a Programming Language**
You'll need to decide which programming
llama_print_timings:        load time =  211714.39 ms
llama_print_timings:      sample time =      32.00 ms /   256 runs   (    0.12 ms per token,  8001.25 tokens per second)
llama_print_timings: prompt eval time =     561.64 ms /     6 tokens (   93.61 ms per token,    10.68 tokens per second)
llama_print_timings:        eval time =   84147.43 ms /   255 runs   (  329.99 ms per token,     3.03 tokens per second)
llama_print_timings:       total time =   84823.67 ms /   261 tokens
Log end

I believe the performance would be much better if there was a direct network connection between the two spaces but there is no way to prove this. It'd be great if HuggingFace add support for private networks between their Docker spaces.

If anyone wants to play with this, here is the space that I've been using.