search

abetlen / llama-cpp-python

Python bindings for llama.cpp
https://llama-cpp-python.readthedocs.io
MIT License
7.65k stars 919 forks source link

Using the GPU is slower than the CPU #1123

Open AreckOVO opened 7 months ago

AreckOVO commented 7 months ago

I installed llamacpp using the instructions below: CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python the speed: llama_print_timings: eval time = 81.91 ms / 2 runs ( 40.95 ms per token, 1.02 tokens per second)

I installed llamacpp using the instructions below: pip install llama-cpp-python the speed: llama_print_timings: eval time = 81.91 ms / 2 runs ( 40.95 ms per token, 30.01 tokens per second)

My code is as follows: result = self.model( prompt, # Prompt

max_tokens=nt, # Generate up to 32 tokens

            # stop=["Q:", "\n"], # Stop generating just before the model would generate a new question
            # echo=True # Echo the prompt back in the output
        )

So how can i use GPU to speed up?

victordonat0 commented 7 months ago

you need set n_gpu_layers= your layers try with 5~10 make test increase it example code 


llm = Llama(model_path="model.gguf", n_gpu_layers=30, n_ctx=3584, n_batch=521, verbose=True)
# adjust n_gpu_layers as per your GPU and model
output = llm("Q: Name the planets in the solar system? A: ", max_tokens=32, stop=["Q:", "\n"], echo=True)
print(output)```
AreckOVO commented 7 months ago

you need set n_gpu_layers= your layers try with 5~10 make test increase it example code

llm = Llama(model_path="model.gguf", n_gpu_layers=30, n_ctx=3584, n_batch=521, verbose=True)
# adjust n_gpu_layers as per your GPU and model
output = llm("Q: Name the planets in the solar system? A: ", max_tokens=32, stop=["Q:", "\n"], echo=True)
print(output)```

I set n_gpu_layer=33, but it is still slow. model = Llama(model_path=model_path, n_gpu_layers=33,n_ctx=8192,n_batch=512,verbose=True) and the output is as follows: llama_model_loader: loaded meta data with 20 key-value pairs and 325 tensors from /models/phi-2/ggml-model-f32_q4_0.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 = phi2 llama_model_loader: - kv 1: general.name str = Phi2 llama_model_loader: - kv 2: phi2.context_length u32 = 2048 llama_model_loader: - kv 3: phi2.embedding_length u32 = 2560 llama_model_loader: - kv 4: phi2.feed_forward_length u32 = 10240 llama_model_loader: - kv 5: phi2.block_count u32 = 32 llama_model_loader: - kv 6: phi2.attention.head_count u32 = 32 llama_model_loader: - kv 7: phi2.attention.head_count_kv u32 = 32 llama_model_loader: - kv 8: phi2.attention.layer_norm_epsilon f32 = 0.000010 llama_model_loader: - kv 9: phi2.rope.dimension_count u32 = 32 llama_model_loader: - kv 10: general.file_type u32 = 2 llama_model_loader: - kv 11: tokenizer.ggml.add_bos_token bool = false llama_model_loader: - kv 12: tokenizer.ggml.model str = gpt2 llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,51200] = ["!", "\"", "#", "$", "%", "&", "'", ... llama_model_loader: - kv 14: tokenizer.ggml.token_type arr[i32,51200] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 15: tokenizer.ggml.merges arr[str,50000] = ["Ġ t", "Ġ a", "h e", "i n", "r e",... llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 50256 llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 50256 llama_model_loader: - kv 18: tokenizer.ggml.unknown_token_id u32 = 50256 llama_model_loader: - kv 19: general.quantization_version u32 = 2 llama_model_loader: - type f32: 195 tensors llama_model_loader: - type q4_0: 129 tensors llama_model_loader: - type q6_K: 1 tensors llm_load_vocab: mismatch in special tokens definition ( 910/51200 vs 944/51200 ). llm_load_print_meta: format = GGUF V3 (latest) llm_load_print_meta: arch = phi2 llm_load_print_meta: vocab type = BPE llm_load_print_meta: n_vocab = 51200 llm_load_print_meta: n_merges = 50000 llm_load_print_meta: n_ctx_train = 2048 llm_load_print_meta: n_embd = 2560 llm_load_print_meta: n_head = 32 llm_load_print_meta: n_head_kv = 32 llm_load_print_meta: n_layer = 32 llm_load_print_meta: n_rot = 32 llm_load_print_meta: n_embd_head_k = 80 llm_load_print_meta: n_embd_head_v = 80 llm_load_print_meta: n_gqa = 1 llm_load_print_meta: n_embd_k_gqa = 2560 llm_load_print_meta: n_embd_v_gqa = 2560 llm_load_print_meta: f_norm_eps = 1.0e-05 llm_load_print_meta: f_norm_rms_eps = 0.0e+00 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: n_ff = 10240 llm_load_print_meta: n_expert = 0 llm_load_print_meta: n_expert_used = 0 llm_load_print_meta: rope scaling = linear llm_load_print_meta: freq_base_train = 10000.0 llm_load_print_meta: freq_scale_train = 1 llm_load_print_meta: n_yarn_orig_ctx = 2048 llm_load_print_meta: rope_finetuned = unknown llm_load_print_meta: model type = 3B llm_load_print_meta: model ftype = Q4_0 llm_load_print_meta: model params = 2.78 B llm_load_print_meta: model size = 1.49 GiB (4.61 BPW) llm_load_print_meta: general.name = Phi2 llm_load_print_meta: BOS token = 50256 '<|endoftext|>' llm_load_print_meta: EOS token = 50256 '<|endoftext|>' llm_load_print_meta: UNK token = 50256 '<|endoftext|>' llm_load_print_meta: LF token = 128 'Ä' llm_load_tensors: ggml ctx size = 0.12 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: CPU buffer size = 1526.50 MiB ........................................................................................... llama_new_context_with_model: n_ctx = 8192 llama_new_context_with_model: freq_base = 10000.0 llama_new_context_with_model: freq_scale = 1 llama_kv_cache_init: CPU KV buffer size = 2560.00 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: graph splits (measure): 1 llama_new_context_with_model: CPU compute buffer size = 568.00 MiB AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | Model metadata: {'tokenizer.ggml.unknown_token_id': '50256', 'tokenizer.ggml.eos_token_id': '50256', 'tokenizer.ggml.bos_token_id': '50256', 'general.architecture': 'phi2', 'general.name': 'Phi2', 'phi2.context_length': '2048', 'general.quantization_version': '2', 'tokenizer.ggml.model': 'gpt2', 'tokenizer.ggml.add_bos_token': 'false', 'phi2.embedding_length': '2560', 'phi2.attention.head_count': '32', 'phi2.attention.head_count_kv': '32', 'phi2.feed_forward_length': '10240', 'phi2.attention.layer_norm_epsilon': '0.000010', 'phi2.block_count': '32', 'phi2.rope.dimension_count': '32', 'general.file_type': '2'} llama_print_timings: load time = 7734.04 ms llama_print_timings: sample time = 1.31 ms / 3 runs ( 0.44 ms per token, 2295.33 tokens per second) llama_print_timings: prompt eval time = 14847.33 ms / 920 tokens ( 16.14 ms per token, 61.96 tokens per second) llama_print_timings: eval time = 145.37 ms / 2 runs ( 72.69 ms per token, 13.76 tokens per second)

victordonat0 commented 7 months ago

look if when generating response your gpu memory increase.

vvsotnikov commented 7 months ago

Check pip install logs, there are probably some CUDA-related errors. Assuming you have an Nvidia GPU, your output should start like this:

ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
llama_model_loader: loaded meta data with 25 key-value pairs and 611 tensors from cognitivecomputations_laserxtral-Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.

Notice that in your example the the first 4 lines (CuBLAS-related) are missing.

AreckOVO commented 7 months ago

Check pip install logs, there are probably some CUDA-related errors. Assuming you have an Nvidia GPU, your output should start like this:

ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
llama_model_loader: loaded meta data with 25 key-value pairs and 611 tensors from cognitivecomputations_laserxtral-Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.

Notice that in your example the the first 4 lines (CuBLAS-related) are missing.

Thank you! I successfully run the program with GPU, but each card occupies about 1000MB, how can I make full use of each card? and the program runs with all GPUs by default, how do I specify which GPU to use for inference?

isaac-chung commented 7 months ago

Could you see if the suggestion in https://github.com/abetlen/llama-cpp-python/issues/794 works? specifically the following:

Use parameters main_gpu (e.g. 0) and tensor_split, which expects a list of floats, e.g. [0.5, 0.5].

Seems like if you might need to use the main GPU as part of the split.

kk2491 commented 6 months ago

Hello,

I am facing a similar issue, details given below:

Model used : phi-2.Q8_0.gguf
RAM : 7.4 GB
CPU : 2 vCPU Command used to run the server : python3 -m llama_cpp.server --model phi-2.Q8_0.gguf --port 13000 --n_gpu_layers -1
Output of the command : 

ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 4 CUDA devices:
  Device 0: Tesla T4, compute capability 7.5, VMM: yes
  Device 1: Tesla T4, compute capability 7.5, VMM: yes
  Device 2: Tesla T4, compute capability 7.5, VMM: yes
  Device 3: Tesla T4, compute capability 7.5, VMM: yes
llama_model_loader: loaded meta data with 20 key-value pairs and 325 tensors from phi-2.Q8_0.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              = phi2
llama_model_loader: - kv   1:                               general.name str              = Phi2
llama_model_loader: - kv   2:                        phi2.context_length u32              = 2048
llama_model_loader: - kv   3:                      phi2.embedding_length u32              = 2560
llama_model_loader: - kv   4:                   phi2.feed_forward_length u32              = 10240
llama_model_loader: - kv   5:                           phi2.block_count u32              = 32
llama_model_loader: - kv   6:                  phi2.attention.head_count u32              = 32
llama_model_loader: - kv   7:               phi2.attention.head_count_kv u32              = 32
llama_model_loader: - kv   8:          phi2.attention.layer_norm_epsilon f32              = 0.000010
llama_model_loader: - kv   9:                  phi2.rope.dimension_count u32              = 32
llama_model_loader: - kv  10:                          general.file_type u32              = 7
llama_model_loader: - kv  11:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,51200]   = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr[i32,51200]   = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  15:                      tokenizer.ggml.merges arr[str,50000]   = ["Ġ t", "Ġ a", "h e", "i n", "r e",...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 50256
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 50256
llama_model_loader: - kv  18:            tokenizer.ggml.unknown_token_id u32              = 50256
llama_model_loader: - kv  19:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:  195 tensors
llama_model_loader: - type q8_0:  130 tensors
llm_load_vocab: mismatch in special tokens definition ( 910/51200 vs 944/51200 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = phi2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 51200
llm_load_print_meta: n_merges         = 50000
llm_load_print_meta: n_ctx_train      = 2048
llm_load_print_meta: n_embd           = 2560
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 32
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 32
llm_load_print_meta: n_embd_head_k    = 80
llm_load_print_meta: n_embd_head_v    = 80
llm_load_print_meta: n_gqa            = 1
llm_load_print_meta: n_embd_k_gqa     = 2560
llm_load_print_meta: n_embd_v_gqa     = 2560
llm_load_print_meta: f_norm_eps       = 1.0e-05
llm_load_print_meta: f_norm_rms_eps   = 0.0e+00
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: n_ff             = 10240
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 2
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 2048
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: model type       = 3B
llm_load_print_meta: model ftype      = Q8_0
llm_load_print_meta: model params     = 2.78 B
llm_load_print_meta: model size       = 2.75 GiB (8.51 BPW) 
llm_load_print_meta: general.name     = Phi2
llm_load_print_meta: BOS token        = 50256 '<|endoftext|>'
llm_load_print_meta: EOS token        = 50256 '<|endoftext|>'
llm_load_print_meta: UNK token        = 50256 '<|endoftext|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_tensors: ggml ctx size =    0.62 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:        CPU buffer size =   132.81 MiB
llm_load_tensors:      CUDA0 buffer size =   718.15 MiB
llm_load_tensors:      CUDA1 buffer size =   638.36 MiB
llm_load_tensors:      CUDA2 buffer size =   638.36 MiB
llm_load_tensors:      CUDA3 buffer size =   691.59 MiB
.............................................................................................
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      CUDA0 KV buffer size =   180.00 MiB
llama_kv_cache_init:      CUDA1 KV buffer size =   160.00 MiB
llama_kv_cache_init:      CUDA2 KV buffer size =   160.00 MiB
llama_kv_cache_init:      CUDA3 KV buffer size =   140.00 MiB
llama_new_context_with_model: KV self size  =  640.00 MiB, K (f16):  320.00 MiB, V (f16):  320.00 MiB
llama_new_context_with_model:  CUDA_Host input buffer size   =    10.02 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =   167.00 MiB
llama_new_context_with_model:      CUDA1 compute buffer size =   167.00 MiB
llama_new_context_with_model:      CUDA2 compute buffer size =   167.00 MiB
llama_new_context_with_model:      CUDA3 compute buffer size =   167.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =     5.00 MiB
llama_new_context_with_model: graph splits (measure): 5
AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | 
Model metadata: {'tokenizer.ggml.unknown_token_id': '50256', 'tokenizer.ggml.eos_token_id': '50256', 'tokenizer.ggml.bos_token_id': '50256', 'general.architecture': 'phi2', 'general.name': 'Phi2', 'phi2.context_length': '2048', 'general.quantization_version': '2', 'tokenizer.ggml.model': 'gpt2', 'tokenizer.ggml.add_bos_token': 'false', 'phi2.embedding_length': '2560', 'phi2.attention.head_count': '32', 'phi2.attention.head_count_kv': '32', 'phi2.feed_forward_length': '10240', 'phi2.attention.layer_norm_epsilon': '0.000010', 'phi2.block_count': '32', 'phi2.rope.dimension_count': '32', 'general.file_type': '7'}
INFO:     Started server process [30532]
INFO:     Waiting for application startup.
INFO:     Application startup complete.

I have also set the environment variable TENSOR_SPLIT 

kishor@gpt-gpu:/mnt/data/gguf-models$ echo $TENSOR_SPLIT 
[0.25,0.25,0.25,0.25]

image

When the query is being executed, I could see that the usage for each GPU would be around 20%.

Is there a way I can maximize the utilization and get the response quicker?

Thank you,
KK

LiberiFatali commented 3 months ago

When the query is being executed, I could see that the usage for each GPU would be around 20%.

Do you mean the memory usage? If that's the case, you can tune the tensor_split parameter.

freckletonj commented 3 months ago

I'm facing a similar issue, where the latest version on pypi is relying on my CPUs mostly, and my 2 gpus barely. Llama 70B 5bit gguf occupies <8GB of VRAM, and 32 cores of CPU are pinned at 100% while GPU utilization is around 2%.

freckletonj commented 3 months ago

Finally solved it!

Reinstall:

CUDACXX=/usr/local/cuda-12/bin/nvcc CMAKE_ARGS="-DLLAMA_CUBLAS=on -DCMAKE_CUDA_ARCHITECTURES=all-major" FORCE_CMAKE=1 pip install llama-cpp-python --no-cache-dir --force-reinstall --upgrade

Inference like this:

        model = Llama(
            model_path=name_or_path,
            verbose=False,
            n_gpu_layers=-1,
            tensor_split=[1.0, 1.0],
            n_ctx=2048,
        )

that was... so painful. Great tok/s now.