Is there a way to terminate vllm.LLM and release the GPU memory

[sfc-gh-zhwang]

After below code, is there an api(maybe like llm.terminate) to kill llm and release the GPU memory?

from vllm import LLM, SamplingParams

prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
outputs = llm.generate(prompts, sampling_params)

[SuperBruceJia]

After below code, is there an api(maybe like llm.terminate) to kill llm and release the GPU memory?

from vllm import LLM, SamplingParams

prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
outputs = llm.generate(prompts, sampling_params)

Please check the codes below. It works.

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
gc.collect()
torch.cuda.empty_cache()
torch.distributed.destroy_process_group()
print("Successfully delete the llm pipeline and free the GPU memory!")

Best regards,

Shuyue
Dec. 3rd, 2023

[hijkzzz]

mark

[deepbrain]

Even after executing the code above, the GPU memory is not freed with the latest vllm built from source. Any recommendations?

[huylenguyen]

Are there any updates on this? the above code does not work for me either

[puddingfjz]

+1

[puddingfjz]

I find that we need to explicitly run "del llm.llm_engine.driver_worker" to release in when using a single worker.
Can anybody explain why this is the case?

[shyringo]

+1

[shyringo]

I find that we need to explicitly run "del llm.llm_engine.driver_worker" to release in when using a single worker. Can anybody explain why this is the case?

I tried the above code block and also this line "del llm.llm_engine.driver_worker". Both failed for me.

But I managed, with the following code, to terminate the vllm.LLM(), release the GPU memory, and shut down ray in convenience for using vllm.LLM() for the next model. After this, I succeeded in using vllm.LLM() again for the next model.

        #llm is a vllm.LLM object
        import gc
        import torch
        from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel
        
        destroy_model_parallel()
        #del a vllm.executor.ray_gpu_executor.RayGPUExecutor object
        del llm.llm_engine.model_executor
        del llm
        gc.collect()
        torch.cuda.empty_cache()
        import ray
        ray.shutdown()

Anyway, even if it works, it is just a temporary solution and this issue still needs fixing.

[shyringo]

I tried the above code block and also this line "del llm.llm_engine.driver_worker". Both failed for me.

But I managed, with the following code, to terminate the vllm.LLM(), release the GPU memory, and shut down ray in convenience for using vllm.LLM() for the next model. After this, I succeeded in using vllm.LLM() again for the next model.

        #llm is a vllm.LLM object
        import gc
        import torch
        from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel
        
        destroy_model_parallel()
        #del a vllm.executor.ray_gpu_executor.RayGPUExecutor object
        del llm.llm_engine.model_executor
        del llm
        gc.collect()
        torch.cuda.empty_cache()
        import ray
        ray.shutdown()

Anyway, even if it works, it is just a temporary solution and this issue still needs fixing.

update:
the following code would work better, without the possible dead lock warning.

        #llm is a vllm.LLM object
        import gc
        import torch
        from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel
        import os

        #avoid huggingface/tokenizers process dead lock
        os.environ["TOKENIZERS_PARALLELISM"] = "false"
        destroy_model_parallel()
        #del a vllm.executor.ray_gpu_executor.RayGPUExecutor object
        del llm.llm_engine.model_executor
        del llm
        gc.collect()
        torch.cuda.empty_cache()
        import ray
        ray.shutdown()

[ticoneva]

In the latest version of vLLM destroy_model_parallel has moved to vllm.distributed.parallel_state. The objects you have to delete have also changed:

from vllm.distributed.parallel_state import destroy_model_parallel
...
destroy_model_parallel()
del llm.llm_engine.model_executor.driver_worker
del llm # Isn't necessary for releasing memory, but why not
gc.collect()
torch.cuda.empty_cache()

[rbao2018]

In the latest version of vLLM destroy_model_parallel has moved to vllm.distributed.parallel_state. The objects you have to delete have also changed:

from vllm.distributed.parallel_state import destroy_model_parallel
...
destroy_model_parallel()
del llm.llm_engine.model_executor.driver_worker
del llm # Isn't necessary for releasing memory, but why not
gc.collect()
torch.cuda.empty_cache()

thx a lot

[mmoskal]

vLLM seems to hang to the first allocated LLM() instance. It does not hang to later instances. Maybe that helps with diagnosing the issue?

from vllm import LLM


def show_memory_usage():
    import torch.cuda
    import torch.distributed
    import gc

    print(f"cuda memory: {torch.cuda.memory_allocated()//1024//1024}MB")
    gc.collect()
    # torch.distributed.destroy_process_group()
    torch.cuda.empty_cache()
    print(f"  --> after gc: {torch.cuda.memory_allocated()//1024//1024}MB")


def gc_problem():
    show_memory_usage()
    print("loading llm0")
    llm0 = LLM(model="facebook/opt-125m", num_gpu_blocks_override=180)
    del llm0
    show_memory_usage()

    print("loading llm1")
    llm1 = LLM(model="facebook/opt-125m", num_gpu_blocks_override=500)
    del llm1
    show_memory_usage()

    print("loading llm2")
    llm2 = LLM(model="facebook/opt-125m", num_gpu_blocks_override=600)
    del llm2
    show_memory_usage()

gc_problem()

root@c09a058c2d5b:/workspaces/aici/py/vllm# python tests/core/block/e2e/gc_problem.py |grep -v INFO
cuda memory: 0MB
  --> after gc: 0MB
loading llm0
cuda memory: 368MB
  --> after gc: 368MB
loading llm1
cuda memory: 912MB
  --> after gc: 368MB
loading llm2
[rank0]:[W CUDAGraph.cpp:145] Warning: Waiting for pending NCCL work to finish before starting graph capture. (function operator())
cuda memory: 961MB
  --> after gc: 368MB
root@c09a058c2d5b:/workspaces/aici/py/vllm# 

The llm1 consumes more than llm0 but you can see that the allocated memory stays at llm0 level.

[yudataguy]

In the latest version of vLLM destroy_model_parallel has moved to vllm.distributed.parallel_state. The objects you have to delete have also changed:

from vllm.distributed.parallel_state import destroy_model_parallel
...
destroy_model_parallel()
del llm.llm_engine.model_executor.driver_worker
del llm # Isn't necessary for releasing memory, but why not
gc.collect()
torch.cuda.empty_cache()

Tried this including ray.shutdown() but the memory is not released on my end, any other suggestion?

[shyringo]

Tried this including ray.shutdown() but the memory is not released on my end, any other suggestion?

could try the "del llm.llm_engine.model_executor" in the following code instead:

update: the following code would work better, without the possible dead lock warning.

        #llm is a vllm.LLM object
        import gc
        import torch
        from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel
        import os

        #avoid huggingface/tokenizers process dead lock
        os.environ["TOKENIZERS_PARALLELISM"] = "false"
        destroy_model_parallel()
        #del a vllm.executor.ray_gpu_executor.RayGPUExecutor object
        del llm.llm_engine.model_executor
        del llm
        gc.collect()
        torch.cuda.empty_cache()
        import ray
        ray.shutdown()

[yudataguy]

Tried this including ray.shutdown() but the memory is not released on my end, any other suggestion?

could try the "del llm.llm_engine.model_executor" in the following code instead:

update: the following code would work better, without the possible dead lock warning.

        #llm is a vllm.LLM object
        import gc
        import torch
        from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel
        import os

        #avoid huggingface/tokenizers process dead lock
        os.environ["TOKENIZERS_PARALLELISM"] = "false"
        destroy_model_parallel()
        #del a vllm.executor.ray_gpu_executor.RayGPUExecutor object
        del llm.llm_engine.model_executor
        del llm
        gc.collect()
        torch.cuda.empty_cache()
        import ray
        ray.shutdown()

did that as well, still no change in gpu memory allocation. Not sure how to go further

[zheyang0825]

In the latest version of vLLM destroy_model_parallel has moved to vllm.distributed.parallel_state. The objects you have to delete have also changed:

from vllm.distributed.parallel_state import destroy_model_parallel
...
destroy_model_parallel()
del llm.llm_engine.model_executor.driver_worker
del llm # Isn't necessary for releasing memory, but why not
gc.collect()
torch.cuda.empty_cache()

We tried this in version 0.4.2, but GPU memory did not released.

[shyringo]

did that as well, still no change in gpu memory allocation. Not sure how to go further

Then I do not have a clue either. Meanwhile, I should add an information: the vllm version with which I succeeded with the above code was 0.4.0.post1

[mnoukhov]

@zheyang0825 does adding this lines at the end make it work?

torch.distributed.destroy_process_group()         

[yudataguy]

did that as well, still no change in gpu memory allocation. Not sure how to go further

Then I do not have a clue either. Meanwhile, I should add an information: the vllm version with which I succeeded with the above code was 0.4.0.post1

tried on 0.4.0.post1 and method worked, not sure what changed in the latest version that's not releasing the memory, possible bug?

[GurvanR]

Hello ! so if I'm not wrong, no one achieved to release memory on vllm 0.4.2 yet ?

[njhill]

A new bug was introduced in 0.4.2, but fixed in #4737. Please try with that PR or as a workaround you can also install tensorizer.

This should resolve such errors at least for TP=1. For TP > 1, there may be other issues with creating a new LLM instance after deleting one in the same process.

[GurvanR]

A new bug was introduced in 0.4.2, but fixed in #4737. Please try with that PR or as a workaround you can also install tensorizer.

This should resolve such errors at least for TP=1. For TP > 1, there may be other issues with creating a new LLM instance after deleting one in the same process.

I updated vllm yesterday and still have the problem, I'm using those lines :

destroy_model_parallel()
del llm.llm_engine.model_executor.driver_worker
del llm
gc.collect()
torch.cuda.empty_cache()

[Misterrendal]

This code is worked for me

vllm==0.4.0.post1

        import gc
        import ray
        from vllm.model_executor.parallel_utils.parallel_state import destroy_model_parallel

        print('service stopping ..')
        print(f"cuda memory: {torch.cuda.memory_allocated() // 1024 // 1024}MB")

        destroy_model_parallel()

        del model.llm_engine.model_executor.driver_worker
        del model

        gc.collect()
        torch.cuda.empty_cache()
        ray.shutdown()

        print(f"cuda memory: {torch.cuda.memory_allocated() // 1024 // 1024}MB")

        print("service stopped")

[cassanof]

There should be a built-in way! We cannot keep writing code that breaks on the next minor release :(

[youkaichao]

In general it is very difficult to clean up all resources correctly, especially when we use multiple GPUs, and might be prone to deadlocks .

I would say, the most stable way to terminate vLLM is to shut down the process.

[Vincent-Li-9701]

A new bug was introduced in 0.4.2, but fixed in #4737. Please try with that PR or as a workaround you can also install tensorizer.

This should resolve such errors at least for TP=1. For TP > 1, there may be other issues with creating a new LLM instance after deleting one in the same process.

I encountered this issue when TP = 8. I'm doing this in a iterative manner since I need to run the embedding model after the generative model so there are so loading / offloading. The first iteration is fine but the second iteration the instantiation of vllm ray server hangs.

[cassanof]

In general it is very difficult to clean up all resources correctly, especially when we use multiple GPUs, and might be prone to deadlocks .

I would say, the most stable way to terminate vLLM is to shut down the process.

I understand your point. However, this feature is extremely useful for situations where you need to switch between models. For instance, reinforcement learning loops. I am writing an off-policy RL loop, requiring me to train one model (target policy) while its previous version performs inference (behavior policy). As a result, I frequently load and unload models. While I know vLLM is not intended for training, using transformers would be too slow, making my technique unviable.

Let me know if this is a feature that's wanted and the team would be interested in maintaining it. I can open a separate issue and start working on it.

[DuZKai]

I don't know if anyone can currently clear memory correctly, but in version 0.4.2, I applied the above code that failed to clear memory correctly. I can only use a slightly extreme method of creating a new process before calling and closing the process after calling to roughly solve the problem:

from multiprocessing import Process, set_start_method
set_start_method('spawn', force=True)
def vllm_texts(model_path):
    prompts=""
    sampling_params = SamplingParams(max_tokens=512)
    llm = LLM(model=model_path)
    outputs = llm.generate(prompts, sampling_params)

...
print(torch.cuda.memory_summary())
p = Process(target=vllm_texts, args=(model_path))
p.start()
p.join()
if p.is_alive():
    p.terminate()
p.close()
print(torch.cuda.memory_summary())
...

I still hope there is a way in the future to correctly and perfectly clear memory

[SuperBruceJia]

While I am using multiple GPUs to serve a LLM (tensor_parallel_size > 1), the GPUs' memory is not released, except the first GPU (cuda:0).

[ywang96]

In general it is very difficult to clean up all resources correctly, especially when we use multiple GPUs, and might be prone to deadlocks .
I would say, the most stable way to terminate vLLM is to shut down the process.

I understand your point. However, this feature is extremely useful for situations where you need to switch between models. For instance, reinforcement learning loops. I am writing an off-policy RL loop, requiring me to train one model (target policy) while its previous version performs inference (behavior policy). As a result, I frequently load and unload models. While I know vLLM is not intended for training, using transformers would be too slow, making my technique unviable.

Let me know if this is a feature that's wanted and the team would be interested in maintaining it. I can open a separate issue and start working on it.

Glad to see you here @cassanof and to hear that you have been using vLLM in this kind of workflow!

Given how much wanted this feature seems to be, I will bring this back to the team to discuss! If multi-gpu instance is prone to deadlocks, then perhaps we can at least start with single-gpu instances. Everyone on the maintainer team does have limited bandwidth and we have a lot of things to work on, so contributions are very welcomed as always!

[kota-iizuka]

I tried inferring multiple models consecutively with vLLM v0.5.2. As mentioned above, the behavior differs depending on the value of TP.

• If you want to unload a model with TP=1 and then load a model with TP=1, the process described in Is there a way to terminate vllm.LLM and release the GPU memory #1908 (comment) will work without any problems.
• If you want to unload a model with TP=1 and then load a model with TP=2, or vice versa, in addition to the code above, you need to do export VLLM_WORKER_MULTIPROC_METHOD=spawn in [Bug]: When tensor_parallel_size>1, RuntimeError: Cannot re-initialize CUDA in forked subprocess. #6152 (comment), and run vllm.distributed.parallel_state.destroy_distributed_environment().
  • There is an issue where the warning message [Bug]: Producer process has been terminated before all shared CUDA tensors released (v 0.5.0 post1, v 0.4.3) #6025 appears after execution.
• The above does not work to unload and reload the model with TP=2. I tried ray.shutdown() but it did not work.

I use this function in a pipeline where I explain images with VLM and then summarize them with LLM. I hope that this kind of processing will be officially provided and become common.

from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import destroy_model_parallel, destroy_distributed_environment
import torch
import gc
import os

os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

def main():
    prompts = ["Hello, my name is"]
    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=1)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=2)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

if __name__ == "__main__":
    main()

[xansar]

I tried inferring multiple models consecutively with vLLM v0.5.2. As mentioned above, the behavior differs depending on the value of TP.

• If you want to unload a model with TP=1 and then load a model with TP=1, the process described in Is there a way to terminate vllm.LLM and release the GPU memory #1908 (comment) will work without any problems.

• If you want to unload a model with TP=1 and then load a model with TP=2, or vice versa, in addition to the code above, you need to do export VLLM_WORKER_MULTIPROC_METHOD=spawn in [Bug]: When tensor_parallel_size>1, RuntimeError: Cannot re-initialize CUDA in forked subprocess. #6152 (comment), and run vllm.distributed.parallel_state.destroy_distributed_environment().

  • There is an issue where the warning message [Bug]: Producer process has been terminated before all shared CUDA tensors released (v 0.5.0 post1, v 0.4.3) #6025 appears after execution.

• The above does not work to unload and reload the model with TP=2. I tried ray.shutdown() but it did not work.

I use this function in a pipeline where I explain images with VLM and then summarize them with LLM. I hope that this kind of processing will be officially provided and become common.

from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import destroy_model_parallel, destroy_distributed_environment
import torch
import gc
import os

os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

def main():
    prompts = ["Hello, my name is"]
    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=1)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=2)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

if __name__ == "__main__":
    main()

It works for me.

[bhavnicksm]

I tried inferring multiple models consecutively with vLLM v0.5.2. As mentioned above, the behavior differs depending on the value of TP.

• If you want to unload a model with TP=1 and then load a model with TP=1, the process described in Is there a way to terminate vllm.LLM and release the GPU memory #1908 (comment) will work without any problems.

• If you want to unload a model with TP=1 and then load a model with TP=2, or vice versa, in addition to the code above, you need to do export VLLM_WORKER_MULTIPROC_METHOD=spawn in [Bug]: When tensor_parallel_size>1, RuntimeError: Cannot re-initialize CUDA in forked subprocess. #6152 (comment), and run vllm.distributed.parallel_state.destroy_distributed_environment().

  • There is an issue where the warning message [Bug]: Producer process has been terminated before all shared CUDA tensors released (v 0.5.0 post1, v 0.4.3) #6025 appears after execution.

• The above does not work to unload and reload the model with TP=2. I tried ray.shutdown() but it did not work.

I use this function in a pipeline where I explain images with VLM and then summarize them with LLM. I hope that this kind of processing will be officially provided and become common.

from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import destroy_model_parallel, destroy_distributed_environment
import torch
import gc
import os

os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

def main():
    prompts = ["Hello, my name is"]
    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=1)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=2)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

if __name__ == "__main__":
    main()

With TP=1, I am able to unload the model without difficulty with the method described above, but the re-loading fails with a esoteric error, like:

File "/home/bhavnick/fd/workspace/vllm-api/modules/llm/generator.py", line 139, in load
    engine = AsyncLLMEngine.from_engine_args(args)
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 466, in from_engine_args
    engine = cls(
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 380, in __init__
    self.engine = self._init_engine(*args, **kwargs)
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 547, in _init_engine
    return engine_class(*args, **kwargs)
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/llm_engine.py", line 251, in __init__
    self.model_executor = executor_class(
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/executor/executor_base.py", line 47, in __init__
    self._init_executor()
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/executor/gpu_executor.py", line 35, in _init_executor
    self.driver_worker.init_device()
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/worker/worker.py", line 132, in init_device
    init_worker_distributed_environment(self.parallel_config, self.rank,
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/worker/worker.py", line 346, in init_worker_distributed_environment
    ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/distributed/parallel_state.py", line 920, in ensure_model_parallel_initialized
    backend = backend or torch.distributed.get_backend(
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1074, in get_backend
    return Backend(not_none(pg_store)[0])
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/torch/utils/_typing_utils.py", line 12, in not_none
    raise TypeError("Invariant encountered: value was None when it should not be")
TypeError: Invariant encountered: value was None when it should not be

Some differences I can think of is I am using the AysncLLMEngine instead of the synchronous version, but the unload works just the same.

[bhavnicksm]

I tried inferring multiple models consecutively with vLLM v0.5.2. As mentioned above, the behavior differs depending on the value of TP.

• If you want to unload a model with TP=1 and then load a model with TP=1, the process described in Is there a way to terminate vllm.LLM and release the GPU memory #1908 (comment) will work without any problems.

• If you want to unload a model with TP=1 and then load a model with TP=2, or vice versa, in addition to the code above, you need to do export VLLM_WORKER_MULTIPROC_METHOD=spawn in [Bug]: When tensor_parallel_size>1, RuntimeError: Cannot re-initialize CUDA in forked subprocess. #6152 (comment), and run vllm.distributed.parallel_state.destroy_distributed_environment().

  • There is an issue where the warning message [Bug]: Producer process has been terminated before all shared CUDA tensors released (v 0.5.0 post1, v 0.4.3) #6025 appears after execution.

• The above does not work to unload and reload the model with TP=2. I tried ray.shutdown() but it did not work.

I use this function in a pipeline where I explain images with VLM and then summarize them with LLM. I hope that this kind of processing will be officially provided and become common.

from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import destroy_model_parallel, destroy_distributed_environment
import torch
import gc
import os

os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

def main():
    prompts = ["Hello, my name is"]
    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=1)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

    llm = LLM(model="facebook/opt-125m", tensor_parallel_size=2)
    outputs = llm.generate(prompts, sampling_params)
    print(outputs)

    destroy_model_parallel()
    destroy_distributed_environment()
    del llm.llm_engine.model_executor
    del llm
    gc.collect()
    torch.cuda.empty_cache()

if __name__ == "__main__":
    main()

With TP=1, I am able to unload the model without difficulty with the method described above, but the re-loading fails with a esoteric error, like:

File "/home/bhavnick/fd/workspace/vllm-api/modules/llm/generator.py", line 139, in load
    engine = AsyncLLMEngine.from_engine_args(args)
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 466, in from_engine_args
    engine = cls(
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 380, in __init__
    self.engine = self._init_engine(*args, **kwargs)
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/async_llm_engine.py", line 547, in _init_engine
    return engine_class(*args, **kwargs)
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/engine/llm_engine.py", line 251, in __init__
    self.model_executor = executor_class(
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/executor/executor_base.py", line 47, in __init__
    self._init_executor()
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/executor/gpu_executor.py", line 35, in _init_executor
    self.driver_worker.init_device()
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/worker/worker.py", line 132, in init_device
    init_worker_distributed_environment(self.parallel_config, self.rank,
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/worker/worker.py", line 346, in init_worker_distributed_environment
    ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/vllm/distributed/parallel_state.py", line 920, in ensure_model_parallel_initialized
    backend = backend or torch.distributed.get_backend(
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 1074, in get_backend
    return Backend(not_none(pg_store)[0])
  File "/home/bhavnick/fd/workspace/pyenv/vllm-api/lib/python3.10/site-packages/torch/utils/_typing_utils.py", line 12, in not_none
    raise TypeError("Invariant encountered: value was None when it should not be")
TypeError: Invariant encountered: value was None when it should not be

Some differences I can think of is I am using the AysncLLMEngine instead of the synchronous version, but the unload works just the same.

UPDATE: Works with ray at TP=2 but for some reason not with multiprocessing with either fork or spawn. Also, only unload works with ray but not reload right now.

[jvlinsta]

@bhavnicksm I can reproduce the same error you have.

[jvlinsta]

When debugging, I found that when using 'spawn' the main GPU used (if using PCI_BUS) would still keep some small amount of memory allocated, indicating that the clean up is unsuccessful.

When I then check all available GPUs that have zero memory allocated and export CUDA_VISIBLE_DEVICES, I am able to progress into changing into another model until another issue hits that seemingly indicates the workers distributed state is still somewhere in memory.

    self.engine = self._init_engine(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/engine/async_llm_engine.py", line 551, in _init_engine
    return engine_class(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/engine/llm_engine.py", line 250, in __init__
    self.model_executor = executor_class(
  File "/home/ubuntu/vllm/vllm/executor/multiproc_gpu_executor.py", line 215, in __init__
    super().__init__(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/executor/distributed_gpu_executor.py", line 25, in __init__
    super().__init__(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/executor/executor_base.py", line 47, in __init__
    self._init_executor()
  File "/home/ubuntu/vllm/vllm/executor/multiproc_gpu_executor.py", line 137, in _init_executor
    self._run_workers("init_device")
  File "/home/ubuntu/vllm/vllm/executor/multiproc_gpu_executor.py", line 192, in _run_workers
    driver_worker_output = driver_worker_method(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/worker/worker.py", line 132, in init_device
    init_worker_distributed_environment(self.parallel_config, self.rank,
  File "/home/ubuntu/vllm/vllm/worker/worker.py", line 343, in init_worker_distributed_environment
    init_distributed_environment(parallel_config.world_size, rank,
  File "/home/ubuntu/vllm/vllm/distributed/parallel_state.py", line 874, in init_distributed_environment
    assert _WORLD.world_size == torch.distributed.get_world_size(), (
AssertionError: world group already initialized with a different world **size```

[jvlinsta]

Another update:

The problem is the global _WORLD variable that stays None. When overriding the backend to be 'nccl' hardcoded at vllm/distributed/parallel_state.py", line 920, in ensure_model_parallel_initialized, the whole state goes in deadlock:

    engine = AsyncLLMEngine.from_engine_args(engine_args)
  File "/home/ubuntu/vllm/vllm/engine/async_llm_engine.py", line 470, in from_engine_args
    engine = cls(
  File "/home/ubuntu/vllm/vllm/engine/async_llm_engine.py", line 380, in __init__
    self.engine = self._init_engine(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/engine/async_llm_engine.py", line 551, in _init_engine
    return engine_class(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/engine/llm_engine.py", line 250, in __init__
    self.model_executor = executor_class(
  File "/home/ubuntu/vllm/vllm/executor/multiproc_gpu_executor.py", line 215, in __init__
    super().__init__(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/executor/distributed_gpu_executor.py", line 25, in __init__
    super().__init__(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/executor/executor_base.py", line 47, in __init__
    self._init_executor()
  File "/home/ubuntu/vllm/vllm/executor/multiproc_gpu_executor.py", line 137, in _init_executor
    self._run_workers("init_device")
  File "/home/ubuntu/vllm/vllm/executor/multiproc_gpu_executor.py", line 192, in _run_workers
    driver_worker_output = driver_worker_method(*args, **kwargs)
  File "/home/ubuntu/vllm/vllm/worker/worker.py", line 132, in init_device
    init_worker_distributed_environment(self.parallel_config, self.rank,
  File "/home/ubuntu/vllm/vllm/worker/worker.py", line 346, in init_worker_distributed_environment
    ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
  File "/home/ubuntu/vllm/vllm/distributed/parallel_state.py", line 969, in ensure_model_parallel_initialized
    initialize_model_parallel(tensor_model_parallel_size,
  File "/home/ubuntu/vllm/vllm/distributed/parallel_state.py", line 931, in initialize_model_parallel
    _TP = init_model_parallel_group(group_ranks,
  File "/home/ubuntu/vllm/vllm/distributed/parallel_state.py", line 773, in init_model_parallel_group
    return GroupCoordinator(
  File "/home/ubuntu/vllm/vllm/distributed/parallel_state.py", line 154, in __init__
    self.pynccl_comm = PyNcclCommunicator(
  File "/home/ubuntu/vllm/vllm/distributed/device_communicators/pynccl.py", line 74, in __init__
    dist.broadcast(tensor, src=ranks[0], group=group)
  File "/opt/conda/envs/llm2/lib/python3.10/site-packages/torch/distributed/c10d_logger.py", line 79, in wrapper
    return func(*args, **kwargs)
  File "/opt/conda/envs/llm2/lib/python3.10/site-packages/torch/distributed/distributed_c10d.py", line 2213, in broadcast

[hammer-wang]

When using TP>1 for the first model, it seems there's no working method that can successfully release the GPU memory. I've tried all scripts in this thread and none worked.

[njhill]

@hammer-wang what version of vLLM are you using, which distributed backend (ray or multiprocessing) and how are you running the server?