What it is
Folder: sam2.1_segmentation_min_6gb_vram. Promptable segmentation with SAM 2.1 hiera-large (Apache-2.0): send an image and a point or box, get pixel-accurate masks back, ranked by confidence. A general tool — the base for interactive labeling, QC tooling, or any click-to-select feature.
It runs the standard predictor contract: one class, Sam2Segmenter, whose predict() accepts JSON with prompts or bare image bytes (which default to a center-point prompt). Weights (~3.5 GB) pull from HuggingFace Hub inside __init__; there is no weights upload.
inference.yaml (core)
model:
module: "sam2_serve"
class_name: "Sam2Segmenter"
init_args:
model_id: "facebook/sam2.1-hiera-large"
max_masks: 3 # top-K masks per prompt
dtype: "auto" # fp16 on GPU, fp32 on CPU
cluster:
num_workers: 1
gpus_per_worker: 1
cpus_per_worker: 8
min_vram_gb_per_worker: 6 # ~3.5 GB weights fp16 + activations
max_ongoing_requests: 8
device: "auto"
requirements:
- "torch"
- "transformers>=4.56,<5" # the Sam2 classes landed in 4.56
- "Pillow>=10.0.0"
- "numpy"Request and response
Prompts ride in a JSON body sent as octet-stream (the predict_json pattern; plain application/json would drop the prompt fields, see serve.py):
request
{
"image_b64": "<base64 png/jpeg>", // or send raw image bytes instead
"points": [[x, y]], // prompt points, in pixels
"labels": [1], // 1 = foreground, 0 = background
"boxes": [[x1, y1, x2, y2]], // optional box prompts
"multimask": true // return top-K candidates
}response
{
"masks": [
{"score": 0.97, "area": 12345, "bbox": [x1, y1, x2, y2],
"png_b64": "<base64 1-bit PNG, white = mask>"}
],
"image_size": {"width": W, "height": H},
"device": "cuda", "elapsed_ms": 42.0
}Submit and call
Shell
cd sam2.1_segmentation_min_6gb_vram
cp .env.example .env # RESON_API_KEY + S3 keys
python submit.py # prints endpoint + predict key
python predict.py # segments the sample image
python predict.py ./photo.jpg 420 310 # your image, point at (420, 310)From your own code, via the SDK client:
client.py
import base64
from resontech import InferenceClient
client = InferenceClient(url=ENDPOINT, predict_api_key=KEY, timeout=300.0)
result = client.predict_json({
"image_b64": base64.b64encode(open("photo.jpg", "rb").read()).decode(),
"points": [[420, 310]],
"labels": [1],
"multimask": True,
})
for mask in result["masks"]:
print(mask["score"], mask["bbox"])Hardware
note
Any 12 GB card runs it comfortably (RTX 5070 / 4070 / 3060-12G). For shared-worker deployments, switching
model_id to sam2.1-hiera-small (~1 GB) halves the footprint at some quality cost.