ENH: multiple GPU support for llama.cpp engine #1202
Description
Is your feature request related to a problem? Please describe
启动GGUF模型时，总是只能使用一颗GPU
xinference | 2024-03-28 01:34:02,909 xinference.core.worker 202 DEBUG Enter launch_builtin_model, args: (<xinference.core.worker.WorkerActor object at 0x7f3831ffda30>,), kwargs: {'model_uid': 'qwen1.5-72b-q3-1-0', 'model_name': 'qwen1.5-chat-offline', 'model_size_in_billions': 72, 'model_format': 'ggufv2', 'quantization': 'q3_k_m', 'model_type': 'LLM', 'n_gpu': 'auto', 'request_limits': None, 'peft_model_path': None, 'image_lora_load_kwargs': None, 'image_lora_fuse_kwargs': None}
xinference | 2024-03-28 01:34:02,910 xinference.core.worker 202 DEBUG GPU selected: [0] for model qwen1.5-72b-q3-1-0
xinference | 2024-03-28 01:34:17,402 xinference.model.llm.llm_family 202 INFO Caching from URI: file:///opt/models/llm-gguf/Qwen/Qwen1.5-72B-Chat-GGUF
xinference | 2024-03-28 01:34:17,411 xinference.model.llm.llm_family 202 INFO Cache /opt/models/llm-gguf/Qwen/Qwen1.5-72B-Chat-GGUF exists
xinference | 2024-03-28 01:34:17,412 xinference.model.llm.core 202 DEBUG Launching qwen1.5-72b-q3-1-0 with LlamaCppChatModel
xinference | ggml_init_cublas: GGML_CUDA_FORCE_MMQ: no
xinference | ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
xinference | ggml_init_cublas: found 1 CUDA devices:
xinference | Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
xinference | llama_model_loader: loaded meta data with 21 key-value pairs and 963 tensors from /opt/models/llm-gguf/Qwen/Qwen1.5-72B-Chat-GGUF/qwen1_5-72b-chat-q3_k_m.gguf (version GGUF V3 (latest))
xinference | llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
xinference | llama_model_loader: - kv 0: general.architecture str = qwen2
xinference | llama_model_loader: - kv 1: general.name str = Qwen1.5-72B-Chat-AWQ-fp16
xinference | llama_model_loader: - kv 2: qwen2.block_count u32 = 80
xinference | llama_model_loader: - kv 3: qwen2.context_length u32 = 32768
xinference | llama_model_loader: - kv 4: qwen2.embedding_length u32 = 8192
xinference | llama_model_loader: - kv 5: qwen2.feed_forward_length u32 = 24576
xinference | llama_model_loader: - kv 6: qwen2.attention.head_count u32 = 64
xinference | llama_model_loader: - kv 7: qwen2.attention.head_count_kv u32 = 64
xinference | llama_model_loader: - kv 8: qwen2.attention.layer_norm_rms_epsilon f32 = 0.000001
xinference | llama_model_loader: - kv 9: qwen2.rope.freq_base f32 = 1000000.000000
xinference | llama_model_loader: - kv 10: qwen2.use_parallel_residual bool = true
xinference | llama_model_loader: - kv 11: tokenizer.ggml.model str = gpt2
xinference | llama_model_loader: - kv 12: tokenizer.ggml.tokens arr[str,152064] = ["!", "\"", "#", "$", "%", "&", "'", ...
xinference | llama_model_loader: - kv 13: tokenizer.ggml.token_type arr[i32,152064] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
xinference | llama_model_loader: - kv 14: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
xinference | llama_model_loader: - kv 15: tokenizer.ggml.eos_token_id u32 = 151645
xinference | llama_model_loader: - kv 16: tokenizer.ggml.padding_token_id u32 = 151643
xinference | llama_model_loader: - kv 17: tokenizer.ggml.bos_token_id u32 = 151643
xinference | llama_model_loader: - kv 18: tokenizer.chat_template str = {% for message in messages %}{{'<|im_...
xinference | llama_model_loader: - kv 19: general.quantization_version u32 = 2
xinference | llama_model_loader: - kv 20: general.file_type u32 = 12
xinference | llama_model_loader: - type f32: 401 tensors
xinference | llama_model_loader: - type q3_K: 321 tensors
xinference | llama_model_loader: - type q4_K: 155 tensors
xinference | llama_model_loader: - type q5_K: 85 tensors
xinference | llama_model_loader: - type q6_K: 1 tensors
xinference | llm_load_vocab: special tokens definition check successful ( 421/152064 ).
xinference | llm_load_print_meta: format = GGUF V3 (latest)
xinference | llm_load_print_meta: arch = qwen2
xinference | llm_load_print_meta: vocab type = BPE
xinference | llm_load_print_meta: n_vocab = 152064
xinference | llm_load_print_meta: n_merges = 151387
xinference | llm_load_print_meta: n_ctx_train = 32768
xinference | llm_load_print_meta: n_embd = 8192
xinference | llm_load_print_meta: n_head = 64
xinference | llm_load_print_meta: n_head_kv = 64
xinference | llm_load_print_meta: n_layer = 80
xinference | llm_load_print_meta: n_rot = 128
xinference | llm_load_print_meta: n_embd_head_k = 128
xinference | llm_load_print_meta: n_embd_head_v = 128
xinference | llm_load_print_meta: n_gqa = 1
xinference | llm_load_print_meta: n_embd_k_gqa = 8192
xinference | llm_load_print_meta: n_embd_v_gqa = 8192
xinference | llm_load_print_meta: f_norm_eps = 0.0e+00
xinference | llm_load_print_meta: f_norm_rms_eps = 1.0e-06
xinference | llm_load_print_meta: f_clamp_kqv = 0.0e+00
xinference | llm_load_print_meta: f_max_alibi_bias = 0.0e+00
xinference | llm_load_print_meta: n_ff = 24576
xinference | llm_load_print_meta: n_expert = 0
xinference | llm_load_print_meta: n_expert_used = 0
xinference | llm_load_print_meta: pooling type = 0
xinference | llm_load_print_meta: rope type = 2
xinference | llm_load_print_meta: rope scaling = linear
xinference | llm_load_print_meta: freq_base_train = 1000000.0
xinference | llm_load_print_meta: freq_scale_train = 1
xinference | llm_load_print_meta: n_yarn_orig_ctx = 32768
xinference | llm_load_print_meta: rope_finetuned = unknown
xinference | llm_load_print_meta: ssm_d_conv = 0
xinference | llm_load_print_meta: ssm_d_inner = 0
xinference | llm_load_print_meta: ssm_d_state = 0
xinference | llm_load_print_meta: ssm_dt_rank = 0
xinference | llm_load_print_meta: model type = 70B
xinference | llm_load_print_meta: model ftype = Q3_K - Medium
xinference | llm_load_print_meta: model params = 72.29 B
xinference | llm_load_print_meta: model size = 33.45 GiB (3.98 BPW)
xinference | llm_load_print_meta: general.name = Qwen1.5-72B-Chat-AWQ-fp16
xinference | llm_load_print_meta: BOS token = 151643 '<|endoftext|>'
xinference | llm_load_print_meta: EOS token = 151645 '<|im_end|>'
xinference | llm_load_print_meta: PAD token = 151643 '<|endoftext|>'
xinference | llm_load_print_meta: LF token = 148848 'ÄĬ'
xinference | llm_load_tensors: ggml ctx size = 0.74 MiB
xinference | llm_load_tensors: offloading 80 repeating layers to GPU
xinference | llm_load_tensors: offloading non-repeating layers to GPU
xinference | llm_load_tensors: offloaded 81/81 layers to GPU
xinference | llm_load_tensors: CUDA_Host buffer size = 510.47 MiB
xinference | llm_load_tensors: CUDA0 buffer size = 33747.06 MiB
xinference | ....................................2024-03-28 01:35:30,477 xinference.core.supervisor 202 DEBUG Enter launch_builtin_model, model_uid: qwen1.5-72b-q3, model_name: qwen1.5-chat-offline, model_size: 72, model_format: ggufv2, quantization: q3_k_m, replica: 1
xinference | 2024-03-28 01:35:30,480 xinference.api.restful_api 1 ERROR [address=0.0.0.0:53955, pid=202] Model is already in the model list, uid: qwen1.5-72b-q3
xinference | Traceback (most recent call last):
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/api/restful_api.py", line 722, in launch_model
xinference | model_uid = await (await self._get_supervisor_ref()).launch_builtin_model(
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 227, in send
xinference | return self._process_result_message(result)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 102, in _process_result_message
xinference | raise message.as_instanceof_cause()
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 659, in send
xinference | result = await self._run_coro(message.message_id, coro)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 370, in _run_coro
xinference | return await coro
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/api.py", line 384, in __on_receive__
xinference | return await super().__on_receive__(message) # type: ignore
xinference | File "xoscar/core.pyx", line 558, in __on_receive__
xinference | raise ex
xinference | File "xoscar/core.pyx", line 520, in xoscar.core._BaseActor.__on_receive__
xinference | async with self._lock:
xinference | File "xoscar/core.pyx", line 521, in xoscar.core._BaseActor.__on_receive__
xinference | with debug_async_timeout('actor_lock_timeout',
xinference | File "xoscar/core.pyx", line 526, in xoscar.core._BaseActor.__on_receive__
xinference | result = await result
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/supervisor.py", line 780, in launch_builtin_model
xinference | raise ValueError(f"Model is already in the model list, uid: {model_uid}")
xinference | ValueError: [address=0.0.0.0:53955, pid=202] Model is already in the model list, uid: qwen1.5-72b-q3
xinference | ..............................................................
xinference | llama_new_context_with_model: n_ctx = 32768
xinference | llama_new_context_with_model: freq_base = 1000000.0
xinference | llama_new_context_with_model: freq_scale = 1
xinference | 2024-03-28 01:37:54,732 xinference.core.supervisor 202 DEBUG Enter list_model_registrations, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'LLM'), kwargs: {'detailed': True}
xinference | 2024-03-28 01:37:56,346 xinference.core.supervisor 202 DEBUG Leave list_model_registrations, elapsed time: 1 s
xinference | 2024-03-28 01:37:58,053 xinference.core.supervisor 202 DEBUG Enter list_model_registrations, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'rerank'), kwargs: {'detailed': False}
xinference | 2024-03-28 01:37:58,055 xinference.core.supervisor 202 DEBUG Leave list_model_registrations, elapsed time: 0 s
xinference | 2024-03-28 01:37:58,109 xinference.core.supervisor 202 DEBUG Enter list_model_registrations, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'embedding'), kwargs: {'detailed': False}
xinference | 2024-03-28 01:37:58,111 xinference.core.supervisor 202 DEBUG Leave list_model_registrations, elapsed time: 0 s
xinference | 2024-03-28 01:37:58,130 xinference.core.supervisor 202 DEBUG Enter list_model_registrations, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'LLM'), kwargs: {'detailed': False}
xinference | 2024-03-28 01:37:58,132 xinference.core.supervisor 202 DEBUG Leave list_model_registrations, elapsed time: 0 s
xinference | 2024-03-28 01:37:58,150 xinference.core.supervisor 202 DEBUG Enter get_model_registration, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'LLM', 'chatglm2-offline'), kwargs: {}
xinference | 2024-03-28 01:37:58,151 xinference.core.supervisor 202 DEBUG Leave get_model_registration, elapsed time: 0 s
xinference | 2024-03-28 01:37:58,155 xinference.core.supervisor 202 DEBUG Enter get_model_registration, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'LLM', 'chatglm3-offline'), kwargs: {}
xinference | 2024-03-28 01:37:58,157 xinference.core.supervisor 202 DEBUG Leave get_model_registration, elapsed time: 0 s
xinference | 2024-03-28 01:37:58,160 xinference.core.supervisor 202 DEBUG Enter get_model_registration, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'LLM', 'qwen1.5-chat-offline'), kwargs: {}
xinference | 2024-03-28 01:37:58,162 xinference.core.supervisor 202 DEBUG Leave get_model_registration, elapsed time: 0 s
xinference | 2024-03-28 01:38:01,173 xinference.core.supervisor 202 DEBUG Enter unregister_model, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'LLM', 'qwen1.5-chat-offline'), kwargs: {}
xinference | 2024-03-28 01:38:01,186 xinference.core.supervisor 202 DEBUG Leave unregister_model, elapsed time: 0 s
xinference | ggml_backend_cuda_buffer_type_alloc_buffer: allocating 81920.00 MiB on device 0: cudaMalloc failed: out of memory
xinference | llama_kv_cache_init: failed to allocate buffer for kv cache
xinference | llama_new_context_with_model: llama_kv_cache_init() failed for self-attention cache
xinference | 2024-03-28 01:38:02,330 xinference.core.worker 202 ERROR Failed to load model qwen1.5-72b-q3-1-0
xinference | Traceback (most recent call last):
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/worker.py", line 569, in launch_builtin_model
xinference | await model_ref.load()
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 227, in send
xinference | return self._process_result_message(result)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 102, in _process_result_message
xinference | raise message.as_instanceof_cause()
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 659, in send
xinference | result = await self._run_coro(message.message_id, coro)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 370, in _run_coro
xinference | return await coro
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/api.py", line 384, in __on_receive__
xinference | return await super().__on_receive__(message) # type: ignore
xinference | File "xoscar/core.pyx", line 558, in __on_receive__
xinference | raise ex
xinference | File "xoscar/core.pyx", line 520, in xoscar.core._BaseActor.__on_receive__
xinference | async with self._lock:
xinference | File "xoscar/core.pyx", line 521, in xoscar.core._BaseActor.__on_receive__
xinference | with debug_async_timeout('actor_lock_timeout',
xinference | File "xoscar/core.pyx", line 524, in xoscar.core._BaseActor.__on_receive__
xinference | result = func(*args, **kwargs)
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/model.py", line 239, in load
xinference | self._model.load()
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/model/llm/ggml/llamacpp.py", line 171, in load
xinference | self._llm = Llama(
xinference | File "/opt/conda/lib/python3.10/site-packages/llama_cpp/llama.py", line 325, in __init__
xinference | self._ctx = _LlamaContext(
xinference | File "/opt/conda/lib/python3.10/site-packages/llama_cpp/_internals.py", line 265, in __init__
xinference | raise ValueError("Failed to create llama_context")
xinference | ValueError: [address=0.0.0.0:38499, pid=258] Failed to create llama_context
xinference | 2024-03-28 01:38:05,054 xinference.core.supervisor 202 DEBUG Enter terminate_model, args: (<xinference.core.supervisor.SupervisorActor object at 0x7f3831ffda80>, 'qwen1.5-72b-q3'), kwargs: {'suppress_exception': True}
xinference | 2024-03-28 01:38:05,056 xinference.core.supervisor 202 DEBUG Leave terminate_model, elapsed time: 0 s
xinference | 2024-03-28 01:38:05,066 xinference.api.restful_api 1 ERROR [address=0.0.0.0:38499, pid=258] Failed to create llama_context
xinference | Traceback (most recent call last):
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/api/restful_api.py", line 722, in launch_model
xinference | model_uid = await (await self._get_supervisor_ref()).launch_builtin_model(
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 227, in send
xinference | return self._process_result_message(result)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 102, in _process_result_message
xinference | raise message.as_instanceof_cause()
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 659, in send
xinference | result = await self._run_coro(message.message_id, coro)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 370, in _run_coro
xinference | return await coro
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/api.py", line 384, in __on_receive__
xinference | return await super().__on_receive__(message) # type: ignore
xinference | File "xoscar/core.pyx", line 558, in __on_receive__
xinference | raise ex
xinference | File "xoscar/core.pyx", line 520, in xoscar.core._BaseActor.__on_receive__
xinference | async with self._lock:
xinference | File "xoscar/core.pyx", line 521, in xoscar.core._BaseActor.__on_receive__
xinference | with debug_async_timeout('actor_lock_timeout',
xinference | File "xoscar/core.pyx", line 526, in xoscar.core._BaseActor.__on_receive__
xinference | result = await result
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/supervisor.py", line 796, in launch_builtin_model
xinference | await _launch_model()
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/supervisor.py", line 760, in _launch_model
xinference | await _launch_one_model(rep_model_uid)
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/supervisor.py", line 741, in _launch_one_model
xinference | await worker_ref.launch_builtin_model(
xinference | File "xoscar/core.pyx", line 284, in __pyx_actor_method_wrapper
xinference | async with lock:
xinference | File "xoscar/core.pyx", line 287, in xoscar.core.__pyx_actor_method_wrapper
xinference | result = await result
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/utils.py", line 45, in wrapped
xinference | ret = await func(*args, **kwargs)
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/worker.py", line 569, in launch_builtin_model
xinference | await model_ref.load()
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 227, in send
xinference | return self._process_result_message(result)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/context.py", line 102, in _process_result_message
xinference | raise message.as_instanceof_cause()
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 659, in send
xinference | result = await self._run_coro(message.message_id, coro)
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/backends/pool.py", line 370, in _run_coro
xinference | return await coro
xinference | File "/opt/conda/lib/python3.10/site-packages/xoscar/api.py", line 384, in __on_receive__
xinference | return await super().__on_receive__(message) # type: ignore
xinference | File "xoscar/core.pyx", line 558, in __on_receive__
xinference | raise ex
xinference | File "xoscar/core.pyx", line 520, in xoscar.core._BaseActor.__on_receive__
xinference | async with self._lock:
xinference | File "xoscar/core.pyx", line 521, in xoscar.core._BaseActor.__on_receive__
xinference | with debug_async_timeout('actor_lock_timeout',
xinference | File "xoscar/core.pyx", line 524, in xoscar.core._BaseActor.__on_receive__
xinference | result = func(*args, **kwargs)
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/core/model.py", line 239, in load
xinference | self._model.load()
xinference | File "/opt/conda/lib/python3.10/site-packages/xinference/model/llm/ggml/llamacpp.py", line 171, in load
xinference | self._llm = Llama(
xinference | File "/opt/conda/lib/python3.10/site-packages/llama_cpp/llama.py", line 325, in __init__
xinference | self._ctx = _LlamaContext(
xinference | File "/opt/conda/lib/python3.10/site-packages/llama_cpp/_internals.py", line 265, in __init__
xinference | raise ValueError("Failed to create llama_context")
xinference | ValueError: [address=0.0.0.0:38499, pid=258] Failed to create llama_context
尝试修改参数
curl 'http://localhost:9997/v1/models' \
-H 'Content-Type: application/json' \
-H 'User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/122.0.0.0 Safari/537.36 Edg/122.0.0.0' \
--data-raw '{"model_uid":"qwen1.5-72b-q4","model_name":"qwen1.5-chat-offline","model_format":"ggufv2","model_size_in_billions":72,"quantization":"q4_k_m","n_gpu":"2","replica":1}'
报错，不支持设置n_gpu
{"detail":"[address=0.0.0.0:53955, pid=202] Currently `n_gpu` only supports `auto`."}
>>> import torch
>>> torch.cuda.is_available()
True
>>> torch.cuda.device_count()
2
Describe the solution you'd like
- llama.cpp 已支持多GPU Multi GPU support, CUDA refactor, CUDA scratch buffer ggml-org/llama.cpp#1703
- 使用LM Studio测试GTX3090*2加载qwen-72b q4可以正常加载