话说能支持下苹果的MPS吗,现在mac m2上运行报错

[wacdev]

❯ python web_demo.py

[2023-05-21 21:29:01,122] [INFO] DeepSpeed/CUDA is not installed, fallback to Pytorch checkpointing.
[2023-05-21 21:29:01,599] [WARNING] Failed to load cpm_kernels:Unknown platform: darwin
[2023-05-21 21:29:01,601] [INFO] building VisualGLMModel model ...
59203
[2023-05-21 21:29:01,625] [INFO] [RANK 0] > initializing model parallel with size 1
[2023-05-21 21:29:01,627] [INFO] [RANK 0] You are using model-only mode.
For torch.distributed users or loading model parallel models, set environment variables RANK, WORLD_SIZE and LOCAL_RANK.


/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/nn/init.py:405: UserWarning: Initializing zero-element tensors is a no-op
  warnings.warn("Initializing zero-element tensors is a no-op")
[2023-05-21 21:29:13,787] [INFO] [RANK 0]  > number of parameters on model parallel rank 0: 7810582016
[2023-05-21 21:29:14,203] [INFO] [RANK 0] Torch not compiled with CUDA enabled
[2023-05-21 21:29:14,203] [INFO] [RANK 0] global rank 0 is loading checkpoint /Users/z/.sat_models/visualglm-6b/1/mp_rank_00_model_states.pt
[2023-05-21 21:29:28,809] [INFO] [RANK 0] > successfully loaded /Users/z/.sat_models/visualglm-6b/1/mp_rank_00_model_states.pt
Traceback (most recent call last):
  File "/Users/z/git/VisualGLM-6B/web_demo.py", line 128, in <module>
    main(args)
  File "/Users/z/git/VisualGLM-6B/web_demo.py", line 81, in main
    model, tokenizer = get_infer_setting(gpu_device=0, quant=args.quant)
  File "/Users/z/git/VisualGLM-6B/model/infer_util.py", line 27, in get_infer_setting
    model = model.cuda()
  File "/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/nn/modules/module.py", line 905, in cuda
    return self._apply(lambda t: t.cuda(device))
  File "/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/nn/modules/module.py", line 797, in _apply
    module._apply(fn)
  File "/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/nn/modules/module.py", line 797, in _apply
    module._apply(fn)
  File "/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/nn/modules/module.py", line 797, in _apply
    module._apply(fn)
  File "/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/nn/modules/module.py", line 820, in _apply
    param_applied = fn(param)
  File "/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/nn/modules/module.py", line 905, in <lambda>
    return self._apply(lambda t: t.cuda(device))
  File "/Users/z/git/VisualGLM-6B/.direnv/python-3.10.11/lib/python3.10/site-packages/torch/cuda/__init__.py", line 239, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

[jamiesun]

model = AutoModel.from_pretrained("THUDM/visualglm-6b", trust_remote_code=True).half().cuda() 改为

model = AutoModel.from_pretrained("THUDM/visualglm-6b", trust_remote_code=True).half().to(“mps”)

[longkeyy]

Traceback (most recent call last): File "/Users/longkeyy/PycharmProjects/hf_demo/llm.py", line 4, in model = AutoModel.from_pretrained("THUDM/visualglm-6b", trust_remote_code=True).half().to("mps") File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/transformers/models/auto/auto_factory.py", line 459, in from_pretrained model_class = get_class_from_dynamic_module( File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/transformers/dynamic_module_utils.py", line 425, in get_class_from_dynamic_module final_module = get_cached_module_file( File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/transformers/dynamic_module_utils.py", line 305, in get_cached_module_file get_cached_module_file( File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/transformers/dynamic_module_utils.py", line 267, in get_cached_module_file modules_needed = check_imports(resolved_module_file) File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/transformers/dynamic_module_utils.py", line 145, in check_imports importlib.import_module(imp) File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/importlib/init.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 1006, in _find_and_load_unlocked File "", line 688, in _load_unlocked File "", line 883, in exec_module File "", line 241, in _call_with_frames_removed File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/cpm_kernels/init.py", line 1, in from . import library File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/cpm_kernels/library/init.py", line 1, in from . import nvrtc File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/cpm_kernels/library/nvrtc.py", line 5, in nvrtc = Lib("nvrtc") File "/Users/longkeyy/miniconda3/envs/hf_demo/lib/python3.10/site-packages/cpm_kernels/library/base.py", line 59, in init raise RuntimeError("Unknown platform: %s" % sys.platform) RuntimeError: Unknown platform: darwin

[Sleepychord]

看上去你使用了量化，目前只支持cuda上的量化。

[xxaier]

能参考 stable-diffusion-webui 在官方代码调整下让在mps上可以跑吗？我不太懂机器学习，不知道怎么改

https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/devices.py

  if sys.platform == "darwin":
      from modules import mac_specific


  def has_mps() -> bool:
      if sys.platform != "darwin":
          return False
      else:
          return mac_specific.has_mps

  def extract_device_id(args, name):
      for x in range(len(args)):
          if name in args[x]:
              return args[x + 1]

      return None


  def get_cuda_device_string():
      from modules import shared

      if shared.cmd_opts.device_id is not None:
          return f"cuda:{shared.cmd_opts.device_id}"

      return "cuda"


  def get_optimal_device_name():
      if torch.cuda.is_available():
          return get_cuda_device_string()

      if has_mps():
          return "mps"

      return "cpu"


  def get_optimal_device():
      return torch.device(get_optimal_device_name())

https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/mac_specific.py

[see2023]

运行python web_demo.py 用cpu跑错误是 "slow_conv2d_cpu" not implemented for 'Half'

用mps跑是
loc("varianceEps"("(mpsFileLoc): /AppleInternal/Library/BuildRoots/97f6331a-ba75-11ed-a4bc-863efbbaf80d/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShadersGraph/mpsgraph/MetalPerformanceShadersGraph/Core/Files/MPSGraphUtilities.mm":228:0)): error: input types 'tensor<1x257x1xf16>' and 'tensor<1xf32>' are not broadcast compatible

[see2023]

更新torch到2.1后是能用fp16和mps跑的，但是貌似有内存泄露，问一个问题后内存就从18G涨到28G，swap一用就扛不住了。

pip list|grep torch
torch 2.1.0.dev20230606 torchaudio 2.1.0.dev20230606 torchvision 0.16.0.dev20230606