unload the model

[osafaimal]

Hi,
i m sorry, i don't find how unload model. like i load a model, i delete the object and i call the garbage collector but it does nothing.
How we are suppose to unload model?
I want to load a model do a batch, load an other do a batch, like that for multiple models for comparing them. But for now i must stop python each time.

[hmellor]

Try calling torch.cuda.empty_cache() after you delete the LLM object

[chenxu2048]

You can also use gc.collect() to remove *garbage* objects immediately, after you delete them.

[osafaimal]

both doesn't work.

[chenxu2048]

You should also clean Notebook output: https://stackoverflow.com/questions/24816237/ipython-notebook-clear-cell-output-in-code

[osafaimal]

i always do (In the GUI not in my cells)

[mnoukhov]

this seems mostly solved by #1908 with

import gc

import torch
from vllm import LLM, SamplingParams
from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel

# Load the model via vLLM
llm = LLM(model=model_name, download_dir=saver_dir, tensor_parallel_size=num_gpus, gpu_memory_utilization=0.70)

# Delete the llm object and free the memory
destroy_model_parallel()
del llm.llm_engine.driver_worker
del llm
gc.collect()
torch.cuda.empty_cache()
torch.distributed.destroy_process_group()
print("Successfully delete the llm pipeline and free the GPU memory!")

[osafaimal]

this seems mostly solved by #1908 with

import gc

import torch
from vllm import LLM, SamplingParams
from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel

# Load the model via vLLM
llm = LLM(model=model_name, download_dir=saver_dir, tensor_parallel_size=num_gpus, gpu_memory_utilization=0.70)

# Delete the llm object and free the memory
destroy_model_parallel()
del llm.llm_engine.driver_worker
del llm
gc.collect()
torch.cuda.empty_cache()
torch.distributed.destroy_process_group()
print("Successfully delete the llm pipeline and free the GPU memory!")

i had already read that. My problem stay unsolved when i use the Vllm from llamaindex otherwise it almost works. I've a little of memory that stay used (~1GB) but at least i can load and unload the models. the problem is that i don't find how access to the member llm_engine of Vllm.LLM

[vvolhejn]

@chenxu2048 the notebook output is just computed data shown to the user, the Python kernel computes it but it's a one-way communication - the output doesn't affect the kernel at all. Therefore clearing the output will have no effect on GPU memory or any other state of the kernel.

[david-koleckar]

No resulute answer given. Can be model unload from gpu ram with vllm? Yes or no