OpenGVLab
3* 4090GPU OOM
Why do 3* 4090GPUs still out of memory (24*3>52GB)
0 NVIDIA GeForce RTX 4090 Off | 00000000:31:00.0 Off | Off | | 66% 24C P8 22W / 450W | 42MiB / 24564MiB | 0% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+ | 1 NVIDIA GeForce RTX 4090 Off | 00000000:B1:00.0 Off | Off | | 64% 23C P8 28W / 450W | 11MiB / 24564MiB | 0% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+ | 2 NVIDIA GeForce RTX 4090 Off | 00000000:E3:00.0 Off | Off | | 69% 25C P8 11W / 450W | 11MiB / 24564MiB | 0% Default | | | | N/A
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 192.00 MiB. GPU 0 has a total capacity of 23.65 GiB of which 34.75 MiB is free. Including non-PyTorch memory, this process has 23.57 GiB memory in use. Of the allocated memory 22.89 GiB is allocated by PyTorch, and 307.82 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
The automatic allocation scheme of device_map='auto' of transformers may not be reasonable, in which case you can try manually allocating GPU memory to achieve maximum utilization, for example:
device_map = {
'vision_model': 0,
'mlp1': 0,
'language_model.model.tok_embeddings': 0,
'language_model.model.layers.0': 0,
'language_model.model.layers.1': 0,
'language_model.model.layers.2': 0,
'language_model.model.layers.3': 0,
'language_model.model.layers.4': 0,
'language_model.model.layers.5': 0,
'language_model.model.layers.6': 0,
'language_model.model.norm': 2,
'language_model.output.weight': 2
}
for i in range(7, 28):
device_map[f'language_model.model.layers.{i}'] = 1
for i in range(28, 48):
device_map[f'language_model.model.layers.{i}'] = 2
print(device_map)
model = AutoModel.from_pretrained(
path,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=True,
trust_remote_code=True,
device_map=device_map
).eval()
The automatic allocation scheme of device_map='auto' of transformers may not be reasonable, in which case you can try manually allocating GPU memory to achieve maximum utilization, for example:
device_map = { 'vision_model': 0, 'mlp1': 0, 'language_model.model.tok_embeddings': 0, 'language_model.model.layers.0': 0, 'language_model.model.layers.1': 0, 'language_model.model.layers.2': 0, 'language_model.model.layers.3': 0, 'language_model.model.layers.4': 0, 'language_model.model.layers.5': 0, 'language_model.model.layers.6': 0, 'language_model.model.norm': 2, 'language_model.output.weight': 2 } for i in range(7, 28): device_map[f'language_model.model.layers.{i}'] = 1 for i in range(28, 48): device_map[f'language_model.model.layers.{i}'] = 2 print(device_map) model = AutoModel.from_pretrained( path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True, device_map=device_map ).eval()
thank you very much
Would two 4090s be enough? I have a 3090, thinking on getting another one.
Would two 4090s be enough? I have a 3090, thinking on getting another one.
Hello, thank you for your attention. You can now deploy the InternVL2 model following this document: https://internvl.readthedocs.io/en/latest/internvl2.0/deployment.html
See memory usage at here: https://internvl.readthedocs.io/en/latest/internvl2.0/deployment.html#memory-usage-testing