myelion-csf 1.0.0

MYELION CSF Core RT — inference and training runtime for .csf model files

CSF Core RT — Python SDK

v1.0 · Stable

Inference runtime for .csf model files produced by the CSF Core training framework.
Device-locked · Hardware-optimised · Precision-configurable.

---

Requirements

Requirement	Minimum version
Python	3.9
Linux	x86-64 or ARM64
CUDA (GPU)	11.8
ROCm (AMD GPU)	5.6
numpy	1.22

GPU inference additionally requires torch>=2.0.

---

Installation

CSF Core RT is distributed via the MYELION private package registry.
A valid licence key is required to authenticate.

# Configure the MYELION registry
pip config set global.index-url https://pkg.myelion.com/simple/
pip config set global.extra-index-url https://pypi.org/simple/

# Install the SDK
pip install csfrt csfrt_ops

# Verify
python -c "import csfrt; print(csfrt.__version__)"

---

Licence Activation

Each device must be activated before the SDK will load models.

# Online activation
python -m csfrt.activate MYELION-XXXX-XXXX-XXXX

# Check status (no network required)
python -m csfrt.status

Air-gap environments:

# 1. Extract fingerprint on the isolated device
python -m csfrt.fingerprint
# → Copy the base64 string and send to [e-mail]

# 2. After receiving license.mkey, deploy it:
sudo cp license.mkey /etc/csfrt/license.mkey
sudo chmod 644 /etc/csfrt/license.mkey
python -m csfrt.status

---

Quick Start

import csfrt
import numpy as np

# Load model — licence verified automatically
model = csfrt.load("model.csf")

# Synchronous inference
input_data = np.random.rand(1, 3, 512, 512).astype(np.float32)
result = model.infer(input_data)
print(result.shape)   # → (1, num_classes)

---

API Reference

csfrt.load()

csfrt.load(
    path: str,
    device: str = "cuda",
    precision: str = "fp16",
    license: str | None = None,
    num_threads: int = 4,
) -> CSFModel

Parameter	Type	Description
path	str	Path to the .csf model file.
device	str	"cuda", "cpu", "cuda:N", or "rocm".
precision	str	"fp16" (default) or "int8".
license	str | None	Custom path to .mkey file.
num_threads	int	CPU threads for preprocessing. Default 4.

---

CSFModel.infer()

model.infer(
    inputs: np.ndarray | torch.Tensor,
    return_numpy: bool = True,
) -> np.ndarray | torch.Tensor

Synchronous inference. Blocks until the result is available.

---

CSFModel.infer_async()

future = model.infer_async(input_data)
do_other_work()
result = future.get(timeout=5.0)

Non-blocking inference. Returns a CSFInferenceFuture.

---

CSFStream

Real-time streaming with automatic micro-batching and callbacks.

from csfrt import CSFStream

def on_result(result, metadata):
    print(f"Frame {metadata['frame_id']}: {result.shape}")

stream = CSFStream(
    model_path="model.csf",
    callback=on_result,
    batch_size=4,
    precision="fp16",
)

stream.start()
for frame in data_source:
    stream.push(frame, metadata={"frame_id": frame.id})
stream.stop()

---

Precision Modes

Mode	Flag	Notes
FP16	"fp16"	Default. Half-precision float. Recommended for GPU.
INT8	"int8"	~2× throughput. Requires calibration data in .csf file.

Note: INT8 mode requires quantisation calibration data bundled into the .csf file at training time. Models compiled without it will raise CSFPrecisionError.

---

Exceptions

Exception	When raised
CSFLicenceError	No valid token, or fingerprint mismatch.
CSFLicenceExpiredError	Token validity period elapsed.
CSFModelError	Corrupt/incompatible .csf file.
CSFPrecisionError	INT8 requested but model has no calibration data.
CSFDeviceError	Requested device not available.
CSFShapeError	Input shape does not match model specification.

---

Docker

# CUDA (NVIDIA)
docker pull pkg.myelion.com/csfrt:1.0-cuda11.8

# ROCm (AMD)
docker pull pkg.myelion.com/csfrt:1.0-rocm5.6

# CPU only
docker pull pkg.myelion.com/csfrt:1.0-cpu

docker run --gpus all \
  -v /etc/csfrt:/etc/csfrt:ro \
  -v /path/to/model.csf:/model.csf:ro \
  pkg.myelion.com/csfrt:1.0-cuda11.8 \
  python inference.py

---

Benchmarking

python -m csfrt.bench model.csf --iterations 100 --batch 1
python -m csfrt.bench model.csf --precision fp16 int8 --compare
python -m csfrt.bench model.csf --output benchmark.json

---

Support

Documentation: myelion.com/docs
Licensing: [e-mail]