v1.2.4

# 💪 Added

## 🧑‍🎨 RLE Mask Representation for Instance Segmentation

Instance segmentation models served by `inference`, `inference-models` and `inference` server can now return masks in **COCO RLE (Run-Length Encoding)** format as a memory-efficient alternative to polygon contours. This applies to all instance segmentation models in the library and:
* significantly reduces memory consumption both server-side and on the wire 
* ensures correct representation of non-continuous shapes (which old default, `polygon` format could not support)

To use the feature, pass `response_mask_format=rle` to your request as query parameter. Using `inference-sdk` client, you can just use the following snippet:
```python
from inference_sdk import InferenceHTTPClient, InferenceConfiguration

CLIENT = InferenceHTTPClient(
    api_url="https://serverless.roboflow.com",
    api_key="XXX",
).configure(InferenceConfiguration(response_mask_format="rle"))

with CLIENT.use_api_v0():
    results = CLIENT.infer(
        "/path/to/image.jpg",
        model_id="asl-instance-seg/155",
    )
```

If you choose to use RLE format, your prediction needs to be parsed differently compared to the old `polygon` responses, as each detection is be delivered with the`rle` key instead of `points` and new field (added to aid building client-side parsers) `mask_format` set appropriately:
```json
{
    "x": 58.5, "y": 72.5, "width": 71.0, "height": 95.0,
    "confidence": 0.94,
    "class": "N", "class_id": 13,
    "detection_id": "...",
    "rle": {
        "size": [200, 150],         # [height, width]
        "counts": "Ub46h5<L5K4M3M..." # ASCII-encoded compressed RLE
    },
    "mask_format": "rle"
}
```
<details>
<summary><b>👉 What is RLE?</b></summary>

**Run-Length Encoding** is a lossless compression scheme for binary data. Instead of storing every pixel of a mask, you store the *lengths of consecutive runs* of identical values.

Imagine a single row of a mask: `0 0 0 0 1 1 1 0 0 1 1 1 1 1`. Pixel-by-pixel that's 14 values. As RLE: `[4, 3, 2, 5]` — meaning "4 zeros, 3 ones, 2 zeros, 5 ones". The format always starts with a run of zeros (use `0` as the first count if the mask starts with 1), and alternates from there.

**COCO RLE** (the variant used here, via `pycocotools`) adds two specifics:

1. **Column-major traversal.** The image is flattened in Fortran order — top-to-bottom, then left-to-right — not row-major like you might expect. This is a historical quirk of the COCO format that you generally don't need to think about as long as you decode through `pycocotools`.

2. **ASCII-encoded counts.** Rather than transmitting the count integers as JSON numbers, COCO packs them with a custom variable-length encoding (similar in spirit to LEB128) and then base64-style-encodes the bytes into a printable ASCII string. This is what makes the `counts` field look like garbage characters — `Ub46h5<L5K4M3M202N...`. The encoding is much more compact than a JSON integer array, especially for masks with many short runs.

#### Why use it?

For a 1080p mask with one object, a dense binary representation is ~2 MB. Polygons help, but for masks with holes, ragged edges, or many small components, the polygon list grows quickly. RLE typically lands in the **single kilobytes** for a single object, regardless of shape complexity. For an inference server returning dozens of predictions per request, this compounds dramatically.
</details>

> [!IMPORTANT]  
> This feature works only with new `inference-models` backend - make sure that env variable `USE_INFERENCE_MODELS=True` (currently, that's the default value)

> [!NOTE]  
> Since the new format, if just replaced old `polygon` format, would be considered **a breaking change**, we added RLE as an alternative, while still believing this format is a better choice for our clients.

## 🧠 Gemma-4 added to `inference-model`

Thanks to @dkosowski87, we have [**Gemma-4 model**](https://huggingface.co/google/gemma-4-26B-A4B-it) added to `inference-models` (no suport in old `inference` yet).

```python
import cv2
from inference_models import AutoModel

model = AutoModel.from_pretrained("gemma-4-e2b-it")
image = cv2.imread("path/to/image.jpg")

answers = model.prompt(
    images=image,
    prompt="What is the main subject in this image?",
    max_new_tokens=256,
    do_sample=False,
)
print(answers[0])
```

PR:  https://github.com/roboflow/inference/pull/2227
See also [`inference-models` documentation](https://inference-models.roboflow.com/models/gemma4/) for more details.

# 🚧 Maintenance
* Add OTel spans to model artifact loading by @ecarrara in https://github.com/roboflow/inference/pull/2244
* Fix docs build: skip empty block sections in SUMMARY.md by @Erol444 in https://github.com/roboflow/inference/pull/2243
* Fix download page to resolve latest release assets dynamically by @Erol444 in https://github.com/roboflow/inference/pull/2250
* Try to fix owlv2 model in new transformers by @PawelPeczek-Roboflow in https://github.com/roboflow/inference/pull/2257
* Support optimal confidence from model-eval by @leeclemnet in https://github.com/roboflow/inference/pull/2242
* New roboflow model block versions with confidence dropdown by @leeclemnet in https://github.com/roboflow/inference/pull/2246
* Route all API calls through Secure Gateway proxy and rename LICENSE_SERVER by @alexnorell in https://github.com/roboflow/inference/pull/2254
* Fix classification confidence shape handling by @hansent in https://github.com/roboflow/inference/pull/2262
* Add add-inference-model Claude skill by @hansent in https://github.com/roboflow/inference/pull/2247
* Add SAM2 streaming video tracker (inference_models + workflow block) by @hansent in https://github.com/roboflow/inference/pull/2245
* Re-land remote cold start workflow headers by @hansent in https://github.com/roboflow/inference/pull/2224
* feat(cicd): switch to per-repo SA by @bigbitbus in https://github.com/roboflow/inference/pull/2259
* Add structured logging to inference compiler by @alexnorell in https://github.com/roboflow/inference/pull/2248
* fix(dynamic-blocks): tee stdout/stderr so print() still reaches Docke… by @rafel-roboflow in https://github.com/roboflow/inference/pull/2258
* Update GHA service account for docker build workflows by @ecarrara in https://github.com/roboflow/inference/pull/2264
* chore: update GHA service account for docker build workflows by @ecarrara in https://github.com/roboflow/inference/pull/2265
* Fix flash-attention compatibility issue for GLM-OCR model by @PawelPeczek-Roboflow in https://github.com/roboflow/inference/pull/2266


**Full Changelog**: https://github.com/roboflow/inference/compare/v1.2.3...v1.2.4