# Distributed inference: how to store results in a global variable

**URL:** <https://discuss.huggingface.co/t/distributed-inference-how-to-store-results-in-a-global-variable/112246>\
**Category:** 🤗Accelerate\
**Created:** [October 15, 2024, 4:08pm UTC](https://discuss.huggingface.co/t/distributed-inference-how-to-store-results-in-a-global-variable/112246 "2024-10-15T16:08:08Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![silviuvo](https://avatars.discourse-cdn.com/v4/letter/s/c6cbf5/32.png) [@silviuvo](https://discuss.huggingface.co/u/silviuvo)\
**Post date:** [October 15, 2024, 4:08pm UTC](https://discuss.huggingface.co/t/distributed-inference-how-to-store-results-in-a-global-variable/112246/1 "2024-10-15T16:08:08Z")

</div>

Hi team! I am trying to speed up inference using a diffusion model.

Consider this (slightly modified) example from [https://huggingface.co/docs/diffusers/en/training/distributed\_inference:](https://huggingface.co/docs/diffusers/en/training/distributed_inference:)

```auto
import torch
import datasets
from accelerate import PartialState
from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True
)
distributed_state = PartialState()
pipeline.to(distributed_state.device)
ds = datasets.load_from_disk("/path/to/my/dataset")

with distributed_state.split_between_processes(ds) as proc_ds:
    proc_ds = proc_ds.map(lambda e: {
        "image": pipeline(e["prompt"]).images[0]
    })

```

How can I now combine the `proc_ds` datasets from across processes into one global dataset again that I continue to work on from the main process?

Can I just declare a list `results` outside of the `with` block, and then, after the `map` within the `with block`, do `results.append(proc_ds)`? Is that thread-safe? How does this actually work, how will that global list end up in the main process?

Sorry if this is obvious. Thank you for your help!

---

<div class="post-metadata">

**Author:** ![silviuvo](https://avatars.discourse-cdn.com/v4/letter/s/c6cbf5/32.png) [@silviuvo](https://discuss.huggingface.co/u/silviuvo)\
**Post date:** [October 16, 2024, 11:40am UTC](https://discuss.huggingface.co/t/distributed-inference-how-to-store-results-in-a-global-variable/112246/2 "2024-10-16T11:40:15Z")

</div>

Update: from withing each process, I am adding results to a list `proc_ds_lst`, then do `proc_ds_lst = accelerate.utils.gather_object(proc_ds_lst)`. Seems to work.

---

<div class="post-metadata">

**Author:** ![silviuvo](https://avatars.discourse-cdn.com/v4/letter/s/c6cbf5/32.png) [@silviuvo](https://discuss.huggingface.co/u/silviuvo)\
**Post date:** [October 16, 2024, 11:40am UTC](https://discuss.huggingface.co/t/distributed-inference-how-to-store-results-in-a-global-variable/112246/3 "2024-10-16T11:40:48Z")

</div>

This post helped: [How to properly gather results of PartialState for inference on 4xGPUs · Issue #2440 · huggingface/accelerate · GitHub](https://github.com/huggingface/accelerate/issues/2440)