feat: Unified HPC Toolchain — merge Nexa_Vortex runtime + Nexa_Inference apps + CUTLASS kernels into PyC

[DarkStarStrix]

Summary

This PR evolves PyC into a vertically integrated HPC toolchain by merging the best components from Nexa_Vortex and Nexa_Inference directly into this repository, and adding a new CUTLASS-backed kernel library.

---

Architecture

[apps/inference_api/]   ← Nexa_Inference: FastAPI SciML inference server
        ↓
[compiler/]             ← PyC: IR, pass manager, optimizer policy
[compiler/cutlass_kernels/] ← NEW: CUTLASS GEMM / Conv2d / Attention kernels
        ↓
[runtime/vortex_core/]  ← Nexa_Vortex: async Rust execution engine
  - Async conveyor belt pipeline (eliminates GPU starvation)
  - Lock-free crossbeam dispatcher (up to 15x lower tail latency)
  - NUMA-aware pinned allocator (~2x H2D bandwidth)
  - Hardware topology profiler
  - C-ABI FFI bridge to PyC compiler (auto-generated via bindgen)
        ↓
[Hardware]              ← GPU (CUDA/CUTLASS Tensor Cores), CPU, NVLink

---

Files Added / Changed

Runtime (from Nexa_Vortex)

File	Description
runtime/vortex_core/src/pipeline.rs	Async conveyor belt — overlaps CPU prep, H2D transfer, GPU compute
runtime/vortex_core/src/cpu_dispatch.rs	Lock-free crossbeam worker pool
runtime/vortex_core/src/allocator.rs	NUMA-local pinned memory allocator
runtime/vortex_core/src/hw_profile.rs	Hardware topology detection
runtime/vortex_core/src/ffi/mod.rs	Safe Rust wrappers for PyC C-ABI
runtime/vortex_core/build.rs	bindgen: auto-generates FFI from include/pyc/ headers
runtime/vortex_core/src/integrations/	Mesocarp lock-free primitive integrations
python/pyc/runtime/control_plane.py	Python control plane (from Nexa_Vortex)
python/pyc/runtime/telemetry_manager.py	Python telemetry manager (from Nexa_Vortex)

CUTLASS Kernels (new)

File	Description
compiler/cutlass_kernels/cutlass_gemm.cu	FP16/BF16 Tensor Core + FP32 SIMT GEMM
compiler/cutlass_kernels/cutlass_conv2d.cu	FP16/BF16 Tensor Core Conv2d
compiler/cutlass_kernels/cutlass_attention.cu	FP16/BF16 fused QKV attention
compiler/cutlass_kernels/cutlass_registry_init.cu	Auto-registers all kernels at .so load time

Application Layer (from Nexa_Inference)

File	Description
apps/inference_api/src/main.py	FastAPI server entry point
apps/inference_api/src/inference.py	Core inference logic
apps/inference_api/src/engines.py	Engine dispatch
apps/inference_api/src/pipelines.py	SciML pipeline definitions
apps/inference_api/models/schemas.py	Pydantic request/response schemas
apps/inference_api/src/auth.py	Authentication
apps/inference_api/src/config.py	Configuration

Python SDK

File	Description
python/pyc/__init__.py	Top-level API: pyc.init(), pyc.detect_hardware()
python/pyc/compiler/__init__.py	ctypes wrapper for compiler C-ABI + kernel policy
python/pyc/runtime/hw_profile.py	Hardware detection (native + pure-Python fallback)
python/pyc/runtime/pipeline.py	Pipeline wrapper (PyO3 + pure-Python stub)

Build System

File	Description
pyproject.toml	Maturin-based Python package definition
runtime/CMakeLists.txt	ExternalProject Cargo build integration
compiler/CMakeLists.txt	Updated with CUTLASS kernel target
include/pyc/cuda_backend.h	New public header for CUDA dispatch C-ABI
scripts/migrate_sources.sh	Helper to pull remaining C sources from original repos

---

How to Build

# 1. Configure
cmake -B build -DPYC_BUILD_CUDA=ON

# 2. Build C/CUDA/Rust
cmake --build build -j$(nproc)

# 3. Install Python package
export PYC_COMPILER_LIB_DIR=$(pwd)/build/compiler
maturin develop --features python_ext

# 4. Test
pytest tests/

---

Checklist

• Runtime sources stripped from Nexa_Vortex and integrated
• Inference app sources stripped from Nexa_Inference and integrated
• CUTLASS GEMM / Conv2d / Attention kernels added
• C-ABI FFI bridge implemented (Rust ↔ C compiler)
• Python SDK updated with unified pyc package
• Build system updated (CMake + Cargo + Maturin)
• Full CI/CD pipeline update (follow-up PR)
• Integration tests for CUTLASS kernels (follow-up PR)