Native Rust runtime for actor-based concurrent and distributed systems. Umbrella crate — pull in subsystems via feature flags.
GPU observability for atomr-accel: NVTX ranges, NVML metrics actor, and CUPTI activity tracing. Implements the KernelTrace hooks exported by atomr-accel-cuda and registers atomr-telemetry probes.
CUB-backed device-wide reductions, scans, sorts, histograms, and selects, surfaced as an atomr actor compiled per-(op, dtype) via NVRTC against the atomr-accel-cuda Phase 0.6 disk cache.
CUTLASS kernel-template instantiation via NVRTC for atomr-accel. Provides GEMM, grouped GEMM, implicit-GEMM convolution, and EVT (epilogue visitor tree) actors that JIT CUTLASS C++ templates against the per-arch toolchain pinned by atomr-accel-cuda's NvrtcActor.
Backend-agnostic compute-acceleration core. Defines the AccelBackend trait, AccelRef<T> typed pointers, AccelError enum, and CompletionStrategy — the abstraction layer that lets atomr-accel-cuda (NVIDIA), and future ROCm / Metal / oneAPI / Vulkan backends plug into the same actor surface.
GPU acceleration via the actor model. Wraps NVIDIA CUDA libraries (cuBLAS, cuDNN, cuFFT, cuRAND, cuSOLVER, cuSPARSE, cuTENSOR, cuBLASLt, NVRTC, NCCL) as supervised atomr actors with generation-validated buffers and a uniform async surface.
FlashAttention v2 + v3 kernel templates for atomr-accel — fp16/bf16/fp8, causal, varlen, ALiBi, sliding window, sink tokens, MQA/GQA, paged KV-cache, and chunked prefill, dispatched through NVRTC + Phase 0.6 cubin cache.
TensorRT engine builder + runtime as supervised atomr actors. Wraps libnvinfer / libnvonnxparser at runtime (proprietary library, not vendored). ONNX import, INT8/FP8 PTQ calibration, IPluginV3 Rust trampolines, dynamic shapes, refit.
Composable agentic framework on top of atomr actors and atomr-infer.
atomr-agents host runtime — long-lived agent identity, memory, skills, hooks, schedules, channels, atop atomr actors.
Avatar orchestrator: composes STT (perception), atomr-infer (cognition), TTS (synthesis), viseme→blendshape mapping, and the Sync-Manager into a supervised actor pipeline. Drives an AvatarSink (Live Link UDP to UE5 MetaHuman) with timecode-aligned audio + ARKit blendshape frames.
Avatar audio→ARKit-blendshape provider backed by NVIDIA Audio2Face-3D. Currently a documented stub: the real runtime lives upstream in atomr-infer once FR-A2F-001 (audio2face modality) lands. x86_64 + NVIDIA-GPU-only.