# "embeddings.position\_ids - UNEXPECTED" warning started showing

**URL:** <https://discuss.huggingface.co/t/embeddings-position-ids-unexpected-warning-started-showing/173102>\
**Category:** 🤗Hub\
**Created:** [February 5, 2026, 5:10am UTC](https://discuss.huggingface.co/t/embeddings-position-ids-unexpected-warning-started-showing/173102 "2026-02-05T05:10:33Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![lucasm55](https://avatars.discourse-cdn.com/v4/letter/l/dc4da7/32.png) [@lucasm55](https://discuss.huggingface.co/u/lucasm55)\
**Post date:** [February 5, 2026, 5:10am UTC](https://discuss.huggingface.co/t/embeddings-position-ids-unexpected-warning-started-showing/173102/1 "2026-02-05T05:10:34Z")

</div>

Hello all, I’m fairly new to HF and this type of package, so please keep that in mind.

About 2 months ago I started a project that uses all-MiniLM-L6-v2 and cross-encoder/ms-marco-MiniLM-L-6-v2

I finished the project and but recently came back to it to work on a new feature, however I’m getting lots of new verbose warnings/logging when downloading and using those models. These new warnings include not being authenticated while making requests to HF hub, and something to do with embeddings.position\_ids being unexpected. Here’s an example:

```shell
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
WARNING: Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Loading weights: 100%|███████████████████████| 103/103 [00:00<00:00, 5300.26it/s, Materializing param=pooler.dense.weight]
BertModel LOAD REPORT from: sentence-transformers/all-MiniLM-L6-v2
Key | Status | |
------------------------±-----------±-±
embeddings.position_ids | UNEXPECTED | |

Notes:

UNEXPECTED :can be ignored when loading from different task/architecture; not ok if you expect identical arch.
Loading weights: 100%|█████████████████████████| 105/105 [00:00<00:00, 5050.25it/s, Materializing param=classifier.weight]
BertForSequenceClassification LOAD REPORT from: cross-encoder/ms-marco-MiniLM-L-6-v2
Key | Status | |
-----------------------------±-----------±-±
bert.embeddings.position_ids | UNEXPECTED | |

Notes:

UNEXPECTED :can be ignored when loading from different task/architecture; not ok if you expect identical arch.

```

I’m not really sure what to make of these new outputs. The authentication seems straightforward except for the fact that I don’t understand why I wasn’t encountering this issue before, and I also didn’t realize you authentication to access HF was a thing. As for the unexpected status of the embeddings.position\_ids I really have no idea.

Hoping someone with more knowledge can help me understand why these appeared with no changes and how I can start to solve them. Functionality of my project seems the same/unimpacted FYI.

---

<div class="post-metadata">

**Author:** ![John6666](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/john6666/32/27664_2.png) [@John6666](https://discuss.huggingface.co/u/John6666)\
**Post date:** [February 5, 2026, 6:54am UTC](https://discuss.huggingface.co/t/embeddings-position-ids-unexpected-warning-started-showing/173102/2 "2026-02-05T06:54:42Z")

</div>

The message about that token basically appears to anyone who isn’t logged in and doesn’t pass a token.

The `position_ids` one is probably harmless too. This might be affected by [_the major update from Transformers 4 to 5_](https://huggingface.co/blog/transformers-v5).

* * *

## What changed (even if you didn’t change your code)

These messages are almost always caused by a **dependency update** (e.g., `huggingface_hub`, `transformers`, `accelerate`, `sentence-transformers`) or by your environment/cache changing (new venv/container, cleared `~/.cache`, etc.). Newer versions have become more explicit about:

- **Hub authentication / rate limits**
- **“state\_dict” key mismatches** when loading weights (the “LOAD REPORT” table)

* * *

## 1) “You are sending unauthenticated requests to the HF Hub … set a HF\_TOKEN”

### Background

Most Hugging Face models (including the two you use) are **public**. You do **not** need to be logged in to download them.

However, Hugging Face applies **rate limits** to Hub requests, and limits depend on whether requests are associated with a user account / plan. ([Hugging Face](https://huggingface.co/docs/hub/en/rate-limits "Hub Rate limits"))

Newer `huggingface_hub` versions may emit a warning when the Hub responds with a warning header (formatted like `X-HF-Warning: ...`). ([GitHub](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/utils/_http.py "huggingface\_hub/src/huggingface\_hub/utils/\_http.py at main"))  
So you can be totally “allowed” to download the model, but still get warned that **unauthenticated** downloads may be slower / more rate-limited.

### Why you didn’t see it before

Common reasons:

- You were previously loading from the **local cache** , so no Hub requests were made.
- You had previously logged in (token stored locally), but now you’re in a different environment or cleared the cache/token.
- Your dependencies updated and now **surface** this warning.

### How to fix (if you want)

**Option A — log in once (recommended):**

- `hf auth login` stores a token in `HF_HOME` (defaults to `~/.cache/huggingface/token`). ([Hugging Face](https://huggingface.co/docs/huggingface_hub/en/quick-start "Quickstart"))

**Option B — set an env var (good for servers/CI):**

- Set `HF_TOKEN` (a read token is enough for public downloads). ([Hugging Face](https://huggingface.co/docs/huggingface_hub/en/quick-start "Quickstart"))

**Option C — don’t authenticate, just reduce noise:**

- Set `HF_HUB_VERBOSITY=error` to quiet hub logging. ([Hugging Face](https://huggingface.co/docs/huggingface_hub/en/package_reference/environment_variables "Environment variables"))

* * *

## 2) `embeddings.position_ids | UNEXPECTED` in the “LOAD REPORT”

### What “UNEXPECTED” means

When loading weights, the loader compares:

- keys present in the checkpoint (`state_dict`)
- keys expected by the model class you instantiated

If a key exists in the checkpoint but not in the model, it’s **unexpected**. PyTorch tracks this as `unexpected_keys`; depending on `strict=True/False` it can be an error or just a report. ([PyTorch Docs](https://docs.pytorch.org/docs/stable/generated/torch.nn.Module.html "Module — PyTorch 2.10 documentation"))

### Why it’s happening specifically for `position_ids`

`position_ids` is typically a **buffer** used to build positional embeddings. Across Transformers versions, some models changed whether this buffer is **saved** in checkpoints (persistent vs non-persistent). Old checkpoints may contain `…position_ids`, while newer model code may not expect it.

This exact symptom is widely reported in Transformers issues, e.g.:

- `unexpected_keys=['bert.embeddings.position_ids']` appearing “now but not yesterday” ([GitHub](https://github.com/huggingface/transformers/issues/24921 "error when load model · Issue #24921 · huggingface/transformers · GitHub"))
- `Unexpected in state_dict: embeddings.position_ids` after a “recent update” ([GitHub](https://github.com/huggingface/transformers/issues/25330 "Issue from recent versions: Unexpected in state\_dict: embeddings.position\_ids · Issue #25330 · huggingface/transformers · GitHub"))

### Is it dangerous?

In your log, the only mismatch shown is `position_ids`, and your project behaves the same. In that case it’s **almost always safe to ignore**.

**When you should worry:** if you see many `MISSING` / `UNEXPECTED` entries for real weights (e.g., attention/MLP layers), because that can mean you loaded the wrong architecture or incompatible checkpoint.

* * *

## 3) Why you’re seeing duplicate warnings / lots of verbosity

Transformers uses both Python `warnings` and its own logging system, and it can route warnings through logging, which can create duplicate-looking output in some setups. ([Hugging Face](https://huggingface.co/docs/transformers/v4.37.0/main_classes/logging "Logging"))

* * *

## Practical “do this” checklist

### A) Record the versions that changed

Run:

```bash
python -c "import transformers, huggingface_hub, sentence_transformers, accelerate; \
print('transformers', transformers. __version__ ); \
print('huggingface_hub', huggingface_hub. __version__ ); \
print('sentence-transformers', sentence_transformers. __version__ ); \
print('accelerate', accelerate. __version__ )"

```

If these drifted since 2 months ago, that explains the new output.

### B) Fix authentication warning (optional, but recommended)

- `hf auth login` (stores token; simplest) ([Hugging Face](https://huggingface.co/docs/huggingface_hub/en/quick-start "Quickstart"))  
or set:
- `HF_TOKEN=...` ([Hugging Face](https://huggingface.co/docs/huggingface_hub/en/package_reference/environment_variables "Environment variables"))

### C) Quiet the logs (if you want)

Environment variables (best set _before_ importing libraries):

```bash
export HF_HUB_VERBOSITY=error
export TRANSFORMERS_VERBOSITY=error
export TRANSFORMERS_NO_ADVISORY_WARNINGS=1

```

`HF_HUB_VERBOSITY` is defined by `huggingface_hub`. ([Hugging Face](https://huggingface.co/docs/huggingface_hub/en/package_reference/environment_variables "Environment variables"))  
`TRANSFORMERS_VERBOSITY` / `TRANSFORMERS_NO_ADVISORY_WARNINGS` are documented by Transformers. ([Hugging Face](https://huggingface.co/docs/transformers/v4.37.0/main_classes/logging "Logging"))

* * *

## Bottom line

- The **HF\_TOKEN** message is about **rate limits / download experience** , not about needing permission for those public models. ([Hugging Face](https://huggingface.co/docs/hub/en/rate-limits "Hub Rate limits"))
- `…position_ids | UNEXPECTED` is a **known benign mismatch** that commonly appears after Transformers-version changes. ([GitHub](https://github.com/huggingface/transformers/issues/24921 "error when load model · Issue #24921 · huggingface/transformers · GitHub"))
- If you want this to stop happening unexpectedly in the future, pin versions (`pip-tools`, `poetry.lock`, `requirements.txt` with exact versions) so upgrades don’t silently change runtime behavior.