DEVELOPER CENTER

Build the open future.

Tools, documentation, and community for AI developers.

DEVELOPER START

Bring the model. Prove the path.

Start from familiar PyTorch, ONNX, and Python workflows, then import, compile, run, profile, and deploy. Every stage has a defined input, output, and validation gate.

01PREPARE

Install HOLYFLOW

Configure Python, the compiler toolchain, and the target runtime on Linux. Keep one project structure across simulation, development systems, and production.

02IMPORT

Enter through the graph

Import a PyTorch exported graph or ONNX model, then define shapes, precision policy, and dynamic dimensions for a repeatable baseline build.

03VERIFY

Correctness before speed

Compare outputs with the reference framework, inspect operator coverage and numerical error, then identify hotspots that need fusion or custom code.

04PRODUCTION

Package and deploy

Deliver a validated build to SPARK, NOVA, or AETHER and integrate batching, serving, telemetry, release control, and rollback.

HOLYFLOW / DEVELOPER PREVIEWMODEL → BUILD → RUN

# 1. Create a reproducible project environment
python -m venv .venv && source .venv/bin/activate
pip install holyflow-preview

# 2. Compile an ONNX model for a target
holyflow compile model.onnx --target spark-s20 --precision bf16

# 3. Run, verify, and produce a profile
holyflow run build/model.hfb --input sample.npy --verify
holyflow profile build/model.hfb --report profile.html

These commands describe the intended HOLYFLOW Developer Preview workflow. Packages, compatibility, and version requirements will be published with the preview.

CHOOSE YOUR PATH

Enter by the work, not by the product label.

A single model can require framework adaptation, kernel optimization, systems integration, and production operations. Start where your current constraint lives.

01 / MODEL

Model developer

Import PyTorch / ONNX models, resolve dynamic shapes, mixed precision, and quantization, then establish accuracy and performance baselines.

Export validation
Operator coverage report
End-to-end accuracy checks

02 / KERNEL

Operator & kernel developer

Extend operators through HOLYFLOW NN and Kernel APIs while controlling layout, on-chip memory, parallel tiling, and fusion.

Custom operator templates
Data-movement pipelines
Cycle-level profiling

03 / SYSTEM

Systems developer

Manage devices, queues, memory, and multi-card communication to extend execution from a model to a service, node, or cluster.

Runtime C++ / Python APIs
Asynchronous streams
Multi-card fabric topology

04 / PRODUCTION

Production engineer

Integrate build artifacts into online inference or batch systems with capacity, observability, staged rollout, and recovery policies.

Serving and concurrency
Latency, throughput, efficiency
Version pinning and rollback

MODEL ONBOARDING

Build compatibility evidence before promising production.

A model name is only the start. Task, parameter count, sequence length, resolution, precision targets, and uncovered operators determine the real porting work.

Reference workloadTaskFirst validation focusSuggested target

Llama 3.1 8B / Qwen 8B classText generationKV cache, long context, time to first tokenNOVA / AETHER

Qwen2.5-VL 7B classMultimodalVision encoder, dynamic resolution, cross-modal opsNOVA

BGE Large classEmbedding / RAGBatch size, vector precision, stable throughputSPARK / NOVA

Whisper Large classSpeech recognitionVariable audio, decode loop, real-time factorNOVA

Stable Diffusion classImage generationUNet / DiT coverage, peak memory, step latencyNOVA / AETHER

ResNet / ViT classVision inferencePreprocessing, batching, INT8 / BF16 accuracySPARK / NOVA

These are priority evaluation templates, not certified compatibility claims. Formal support will be defined in a versioned compatibility matrix.

OPEN TOOLCHAIN

See from framework graph to device execution.

Keep high-level productivity while exposing compiler, kernel, and runtime boundaries for teams that need deeper control.

FRAMEWORKSPyTorch ExportONNXPython / C++

Graph Import

Read graphs, shapes, and constants, then produce operator coverage and fallback reports.

HOLYFLOW Compiler

Apply graph optimization, operator fusion, layout selection, precision propagation, and target compilation.

NN & Kernel API

Provide composable operators and low-level extensions for uncovered or performance-critical paths.

Runtime

Manage devices, memory, command queues, events, streams, and multi-device execution.

Profiler

Connect graph, operator, kernel, and timeline views to expose wait, bandwidth, and utilization bottlenecks.

ONE SOFTWARE BASESPARK / EDGENOVA / DATACENTERAETHER / CLOUD

VALIDATION TO PRODUCTION

Scale only when the evidence is ready.

Do not jump from a one-off demo to fleet deployment. Progress through correctness, performance, stability, and operations gates.

GATE 01

Correctness

Pin data and framework baselines; record tolerances, failures, accuracy metrics, and a reproducible environment.

OUTPUT: VALIDATION REPORT

GATE 02

Performance

Measure warmup, P50/P95 latency, throughput, peak memory, and efficiency while separating host and device time.

OUTPUT: PERFORMANCE BASELINE

GATE 03

Stability

Test soak time, concurrency, thermals, malformed inputs, and device recovery to establish capacity boundaries.

OUTPUT: RELEASE GATE

GATE 04

Operations

Integrate logs, metrics, traces, version control, and rollback, then define failure domains and upgrade windows.

OUTPUT: RUNBOOK

RESOURCE STATUS

Know what is available, and what comes next.

Developer resources will ship as versioned releases. Preview content supports joint validation; production support will name exact platform, software, and model versions.

AVAILABLE NOW

Architecture & evaluation

Platform architecture, product intent, workload intake template, and a joint-validation contact path.

Submit a workload →

DEVELOPER PREVIEW

SDK & examples

HOLYFLOW compiler, runtime, reference models, profiler, and custom operator examples.

STAGED ACCESS

PLANNED

Compatibility & Model Zoo

Validated results and known limits by model, task, precision, hardware, and software version.

VERSIONED RELEASE

PLANNED

Docs, community & contribution

API reference, tutorials, issue tracking, technical discussion, and reproducible performance submissions.

OPEN ECOSYSTEM PATH

FAQ

Make the boundaries clear before you begin.

Can I download the HOLYFLOW SDK today?+

This page does not claim a public download yet. Submit a non-confidential workload, current baseline, and target device through the contact page to discuss Developer Preview validation.

What is needed to port a model?+

Provide model source and version, inputs and outputs, representative data, accuracy requirements, concurrency and latency goals, current hardware baseline, and required deployment environment.

When is a custom operator necessary?+

Start with the compiler coverage report and profiler. Move to NN / Kernel extensions only for unsupported behavior, special numerical requirements, or hotspots graph optimization cannot resolve.

How does a workload move from SPARK to NOVA or AETHER?+

Keep the model project, validation set, and metric definitions fixed, retarget the build, and pass the new performance and stability gates. Scale-out also requires data, model, and fabric topology validation.

READY TO BUILD?

Bring your workload. We’ll bring the platform.

hello@holycores.com →