Cargo.toml

[workspace]
resolver = "2"
members = [
    # Core (no backend dependencies)
    "crates/core/rlx-ir",          # Tensor IR: types, shapes, ops, graph, verifier
    "crates/core/rlx-tensor",      # Symbolic tensor DSL (operator-overloaded graph building)
    "crates/core/rlx-flow",        # Block assembly-line API for model builders
    "crates/core/rlx-fusion",      # MIR fusion passes + fused-op unfuse
    "crates/core/rlx-unfuse",      # Shared IR unfuse/decompose pass for CUDA/ROCm/wgpu (policy-parameterized)
    "crates/core/rlx-autodiff",    # Reverse/forward-mode AD on MIR
    "crates/core/rlx-compile",     # HIR → MIR → LIR pipeline
    "crates/core/rlx-opt",         # Facade: re-exports fusion + autodiff + compile
    "crates/core/rlx-optim",       # Training-step optimizers (Adam, AdamW, Lion, SOAP, Muon, …)
    "crates/core/rlx-extend",      # Facade: stable extension surface (custom ops + LayerStage + DSL) for downstream

    # Driver layer (devices, arenas, buffers, streams)
    "crates/core/rlx-driver",

    # In-graph collective ops (tensor-parallel all-reduce)
    "crates/core/rlx-collectives",
    # Multi-node pipeline-parallel execution (graph partition + weight-sharded stages)
    "crates/core/rlx-distributed",

    # CPU backend
    "crates/backends/rlx-cpu",         # SIMD kernels, BLAS FFI, thread pool, arena executor

    # GPU backends
    "crates/backends/rlx-metal",       # Apple Metal native (MPS + custom MSL kernels)
    "crates/backends/rlx-mlx",         # Apple MLX (lazy graph + eager modes, hand-rolled C++ shim)
    "crates/backends/rlx-coreml",      # Apple CoreML / Neural Engine (ANE) — IR → MIL ML Program

    "crates/backends/rlx-wgpu",        # Cross-platform GPU via wgpu (Metal/Vulkan/DX12/WebGPU)
    "crates/backends/rlx-cuda",        # NVIDIA CUDA (cuBLAS + NVRTC kernels via cudarc)
    "crates/backends/rlx-rocm",        # AMD ROCm/HIP — same kernels as rlx-cuda
    "crates/backends/rlx-xdna",        # AMD XDNA / Ryzen AI NPU (amdxdna) — Device::Xdna detection + AIE seam
    "crates/backends/rlx-tpu",         # Google TPU via libtpu's PJRT plugin (no Python)
    "crates/backends/rlx-egpu",        # External GPU over USB4/TB on Apple Silicon — IOKit PCI discovery + PCIDriverKit transport

    # Microcontroller backend
    "crates/backends/rlx-cortexm",     # ARMv7E-M (Cortex-M4/M7) INT8 kernels — nRF52840 etc.
    "crates/backends/rlx-cortexm/trainer",  # Native fp32 trainer that emits the INT8 model_weights.rs

    # FPGA backend
    "crates/backends/rlx-fpga",        # IR → Verilog → bitstream (per-graph datapath, integer-only Q0.31 requant)

    # Cerebras Wafer-Scale Engine backend
    "crates/backends/rlx-cerebras",    # IR → CSL → fabric simulator (single-PE matmul; multi-PE tiling WIP)

    # Native Vulkan compute backend (raw `ash` + embedded SPIR-V kernels)
    "crates/backends/rlx-vulkan",

    # Qualcomm Hexagon NPU backend (AI Engine Direct / QNN)
    "crates/backends/rlx-qnn",        # IR → QNN model C++ → qnn-net-run on the x86 reference backend (Adreno is rlx-vulkan)

    # Intel oneAPI Level Zero backend (Arc / Data Center Max) — dynamic-loaded
    # `libze_loader` + OpenCL-C→SPIR-V compute kernels (peak Intel path).
    "crates/backends/rlx-oneapi",

    # GPU/accelerator backends (future, ranked by hardware-coverage gap)
    # "rlx-directml",  # Windows NPUs (Copilot+ / Hexagon / Intel-AMD NPU) via DirectML

    "crates/io/rlx-gguf",        # GGUF parser + dequant + quant encoders + writer
    "crates/io/rlx-gguf-convert", # safetensors / ONNX → GGUF conversion with per-tensor quant
    "crates/io/rlx-bake",        # Offline weight bake: graph + weights → optimized *.rlx
    "crates/io/rlx-pkg",         # RLX package format (.rlxp): ZIP64/dir, mmap weights, sidecars, dist
    "crates/io/rlx-hub",         # Shard-aware HuggingFace download + verification (layer→shard planning)
    "crates/io/rlx-mlx-io",      # MLX weight layouts: mlx-community dirs, .npz/.npy, affine/mxfp dequant
    "crates/io/rlx-dduf",        # HuggingFace DDUF (.dduf) ZIP of nested safetensors + JSON
    "crates/io/rlx-nemo",        # NVIDIA NeMo `.nemo` loader (tar + torch.save ckpt + YAML)
    "crates/backends/rlx-gpu-dispatch", # Shape-keyed kernel-variant dispatch table shared by ALL GPU backends
    "crates/backends/rlx-gpu-kernels", # Shared CUDA/HIP .cu sources for rlx-cuda + rlx-rocm
    "crates/backends/rlx-gpu-host",    # Shared host-fallback (D2H→CPU→H2D) staging for rlx-cuda + rlx-rocm
    "crates/backends/rlx-mlx-sys",     # Vendored MLX C++ + rlx_mlx_shim C ABI

    # Proc macros
    "crates/core/rlx-macros",      # #[rlx_model] AOT compilation macro

    # Runtime (user-facing, feature-gated)
    "crates/core/rlx-runtime",     # Session API, backend dispatch

    # Prelude crate — re-exports IR + optimizer + runtime + companion
    # crates behind feature gates. The `cargo add rlx` entry point.
    "crates/rlx",

    # Downstream packages (registered against the custom-op scaffold)
    "crates/numerics/rlx-sparse",      # Sparse linear algebra (CSR LU/MatVec/CG) — JAX-style downstream
    "crates/numerics/rlx-linalg",    # Dense linalg (eigh/svd/qr/cholesky/solve_triangular) via LAPACK
    "crates/numerics/rlx-peft",      # Parameter-efficient adaptation: LoRA/IA3/AdaLoRA/DoRA/OFT
    "crates/numerics/rlx-rng",       # Bit-exact numpy/torch random streams
    "crates/numerics/rlx-vq",          # Fused vector-quantization kernel (nearest-codebook cdist→argmin)
    "crates/numerics/rlx-fdm",         # Force density method — pin-jointed form-finding (jax_fdm port)
    "crates/numerics/rlx-lbm",         # Moment-encoded lattice Boltzmann (HOME-LBM) — D2Q9/D3Q27
    "crates/numerics/rlx-rl",          # Flow-map policies + FMQ / QGBS (offline-to-online RL)
    "crates/numerics/rlx-bbo",         # Black-box + trust-region Q-guidance / QGBS (domain-agnostic f64)
    "crates/numerics/rlx-umap",        # UMAP / fast-umap custom ops (k-NN from pairwise distances, …)
    "crates/numerics/rlx-geo",         # Exact integer Delaunay + discrete Voronoi custom ops (per-backend)
    "crates/numerics/rlx-fem",         # Nonlinear scalar P1 FEM on triangles (Poisson family)

    # Benchmark framework (PLAN L5)
    "crates/tooling/rlx-bench",       # BenchmarkPattern + per-backend timing harness
    "crates/tooling/rlx-opscope",     # Data-pattern recording harness: inject stat taps, sweep synthetic data, mine exploitable structure
    "crates/tooling/rlx-hwprofile",   # Host GPU/VRAM detection for device selection + topology planning
    "crates/tooling/rlx-check",       # `cargo rlx check` — device-free static graph diagnostics
    "crates/tooling/rlx-corpus",      # Named kernel corpus a compiler change must clear
    "crates/tooling/rlx-megakernel",  # Standalone study: MoE block composition toward one device program
    "crates/tooling/rlxsl",  # Scalar-expr manifest → WGSL/CUDA/MSL/GLSL warm-tier kernel codegen
    "crates/io/rlx-onnx",        # ONNX Runtime inference for .onnx files (PLAN B7)
    "crates/io/rlx-onnx-proto",  # Vendored pure-Rust ONNX protobuf types (protoc-free)
    "crates/io/rlx-onnx-import", # ONNX → RLX HIR import + op lowering
    "crates/io/rlx-onnx-conformance", # ONNX op-level ORT parity harness
    "crates/io/rlx-torch-import", # PyTorch (torch.export) → RLX HIR import + bundle/crate emit

    # Text: tokenizer + chat template + sampling (public surface for downstream LM apps)
    "crates/io/rlx-text",

    # Python bindings
    "crates/bindings/pyrlx",           # PyO3 + maturin: run any RLX backend from Python

    # C ABI for the distributed node — iOS (staticlib → xcframework) and
    # embedded hosts (cdylib), including the host side of an FPGA rank.
    "crates/bindings/rlx-ffi",

    # WebAssembly entry point — run models in the browser (CPU backend +
    # async WebGPU device bring-up via rlx-wgpu)
    "crates/bindings/rlx-web",

    # WebGL2 GPGPU backend — render-to-texture execution in the browser
    # (no compute shaders on WebGL2), forward + backward.
    "crates/backends/rlx-webgl",
]

# ── Sanitizer / debug profiles (plan #41) ────────────────────────
#
# Borrowed from MAX's --config=asan / --config=debug-modular pattern.
# Run sanitized builds with:
#
#   cargo +nightly build --profile=asan
#   cargo +nightly test  --profile=asan
#
# AddressSanitizer needs the nightly toolchain on most targets; on
# Apple Silicon it compiles fine via `RUSTFLAGS="-Zsanitizer=address"`
# (which we set per-profile via .cargo/config.toml when present).
# `debug-rlx` is a stable-toolchain debug profile with extra checks
# (overflow, debug-assertions, less inlining for tractable backtraces).

[profile.asan]
inherits = "dev"
debug = "full"
debug-assertions = true
overflow-checks = true
opt-level = 1
lto = false
incremental = false

# ── Workspace-wide lint posture (plan #17) ─────────────────────
#
# Tightens the bar without going to "deny all the things." Crates can
# locally `#[allow(...)]` for genuinely-needed exceptions. Pair with
# `cargo clippy --all-targets -- -D warnings` in CI (see Justfile).
#
# Not yet enabled (would generate ~1000 warnings to fix in one go):
#   missing_docs / unsafe_op_in_unsafe_fn / clippy::pedantic.
# Add them as crates reach individual readiness.

[workspace.package]
version       = "0.2.16"
edition       = "2024"
rust-version  = "1.89"  # edition=2024 needs ≥1.85; `is_multiple_of` ≥1.87; AVX-VNNI
                        # `_mm256_dpbusd_*` intrinsics (rlx-cpu x86 Q2_0 dot) ≥1.89
license       = "MIT OR Apache-2.0"
authors       = ["Eugene Hauptmann", "Nataliya Kosmyna"]
repository    = "https://github.com/MIT-RLX/rlx"
homepage      = "https://github.com/MIT-RLX/rlx"
documentation = "https://docs.rs/rlx"
keywords      = ["ml", "tensor", "compiler", "neural-network", "autodiff"]
categories    = ["science", "algorithms", "compilers"]

[workspace.dependencies]
# Internal crates — one source of truth for path + version. Members
# reference these via `{ workspace = true, .. }` (features/optional per-use).
pyrlx = { path = "crates/bindings/pyrlx", version = "0.2.16" }
rlx = { path = "crates/rlx", version = "0.2.16" }
rlx-autodiff = { path = "crates/core/rlx-autodiff", version = "0.2.16" }
rlx-bake = { path = "crates/io/rlx-bake", version = "0.2.16" }
rlx-bbo = { path = "crates/numerics/rlx-bbo", version = "0.2.16" }
rlx-bench = { path = "crates/tooling/rlx-bench", version = "0.2.16" }
rlx-cerebras = { path = "crates/backends/rlx-cerebras", version = "0.2.16" }
rlx-check = { path = "crates/tooling/rlx-check", version = "0.2.16" }
rlx-collectives = { path = "crates/core/rlx-collectives", version = "0.2.16" }
rlx-compile = { path = "crates/core/rlx-compile", version = "0.2.16" }
rlx-coreml = { path = "crates/backends/rlx-coreml", version = "0.2.16" }
rlx-cortexm = { path = "crates/backends/rlx-cortexm", version = "0.2.16" }
rlx-cortexm-trainer = { path = "crates/backends/rlx-cortexm/trainer", version = "0.2.16" }
rlx-cpu = { path = "crates/backends/rlx-cpu", version = "0.2.16" }
rlx-cuda = { path = "crates/backends/rlx-cuda", version = "0.2.16" }
rlx-dduf = { path = "crates/io/rlx-dduf", version = "0.2.16" }
rlx-distributed = { path = "crates/core/rlx-distributed", version = "0.2.16" }
rlx-driver = { path = "crates/core/rlx-driver", version = "0.2.16" }
rlx-extend = { path = "crates/core/rlx-extend", version = "0.2.16" }
rlx-fdm = { path = "crates/numerics/rlx-fdm", version = "0.2.16" }
rlx-lbm = { path = "crates/numerics/rlx-lbm", version = "0.2.16" }
rlx-flow = { path = "crates/core/rlx-flow", version = "0.2.16" }
rlx-fpga = { path = "crates/backends/rlx-fpga", version = "0.2.16" }
rlx-fusion = { path = "crates/core/rlx-fusion", version = "0.2.16" }
rlx-geo = { path = "crates/numerics/rlx-geo", version = "0.2.16" }
rlx-gguf = { path = "crates/io/rlx-gguf", version = "0.2.16" }
rlx-gguf-convert = { path = "crates/io/rlx-gguf-convert", version = "0.2.16" }
rlx-gpu-host = { path = "crates/backends/rlx-gpu-host", version = "0.2.16" }
rlx-gpu-dispatch = { path = "crates/backends/rlx-gpu-dispatch", version = "0.2.16" }
rlx-gpu-kernels = { path = "crates/backends/rlx-gpu-kernels", version = "0.2.16" }
rlx-hub = { path = "crates/io/rlx-hub", version = "0.2.16" }
rlx-hwprofile = { path = "crates/tooling/rlx-hwprofile", version = "0.2.16" }
rlx-ir = { path = "crates/core/rlx-ir", version = "0.2.16" }
rlx-linalg = { path = "crates/numerics/rlx-linalg", version = "0.2.16" }
rlx-peft = { path = "crates/numerics/rlx-peft", version = "0.2.16" }
rlx-rng = { path = "crates/numerics/rlx-rng", version = "0.2.16" }
rlx-macros = { path = "crates/core/rlx-macros", version = "0.2.16" }
rlx-metal = { path = "crates/backends/rlx-metal", version = "0.2.16" }
rlx-mlx = { path = "crates/backends/rlx-mlx", version = "0.2.16" }
rlx-mlx-io = { path = "crates/io/rlx-mlx-io", version = "0.2.16" }
rlx-mlx-sys = { path = "crates/backends/rlx-mlx-sys", version = "0.2.16" }
rlx-nemo = { path = "crates/io/rlx-nemo", version = "0.2.16" }
rlx-oneapi = { path = "crates/backends/rlx-oneapi", version = "0.2.16" }
rlx-onnx = { path = "crates/io/rlx-onnx", version = "0.2.16" }
rlx-onnx-conformance = { path = "crates/io/rlx-onnx-conformance", version = "0.2.16" }
rlx-onnx-import = { path = "crates/io/rlx-onnx-import", version = "0.2.16" }
rlx-onnx-proto = { path = "crates/io/rlx-onnx-proto", version = "0.2.16" }
rlx-opscope = { path = "crates/tooling/rlx-opscope", version = "0.2.16" }
rlx-opt = { path = "crates/core/rlx-opt", version = "0.2.16" }
rlx-optim = { path = "crates/core/rlx-optim", version = "0.2.16" }
rlx-pkg = { path = "crates/io/rlx-pkg", version = "0.2.16" }
rlx-qnn = { path = "crates/backends/rlx-qnn", version = "0.2.16" }
rlx-rl = { path = "crates/numerics/rlx-rl", version = "0.2.16" }
rlx-rocm = { path = "crates/backends/rlx-rocm", version = "0.2.16" }
rlx-runtime = { path = "crates/core/rlx-runtime", version = "0.2.16" }
rlx-sparse = { path = "crates/numerics/rlx-sparse", version = "0.2.16" }
rlx-tensor = { path = "crates/core/rlx-tensor", version = "0.2.16" }
rlx-text = { path = "crates/io/rlx-text", version = "0.2.16" }
rlx-torch-import = { path = "crates/io/rlx-torch-import", version = "0.2.16" }
rlx-tpu = { path = "crates/backends/rlx-tpu", version = "0.2.16" }
rlx-umap = { path = "crates/numerics/rlx-umap", version = "0.2.16" }
rlx-unfuse = { path = "crates/core/rlx-unfuse", version = "0.2.16" }
rlx-vq = { path = "crates/numerics/rlx-vq", version = "0.2.16" }
rlx-vulkan = { path = "crates/backends/rlx-vulkan", version = "0.2.16" }
rlx-ffi = { path = "crates/bindings/rlx-ffi", version = "0.2.16" }
rlx-web = { path = "crates/bindings/rlx-web", version = "0.2.16" }
rlx-webgl = { path = "crates/backends/rlx-webgl", version = "0.2.16" }
rlx-wgpu = { path = "crates/backends/rlx-wgpu", version = "0.2.16" }
rlx-xdna = { path = "crates/backends/rlx-xdna", version = "0.2.16" }
rlx-egpu = { path = "crates/backends/rlx-egpu", version = "0.2.16" }
rlxsl = { path = "crates/tooling/rlxsl", version = "0.2.16" }

[workspace.lints.rust]
unused = { level = "warn" }

[workspace.lints.clippy]
all = { level = "warn", priority = -1 }
# Common false positives in numeric / SIMD code.
manual_range_contains = "allow"
needless_range_loop   = "allow"
too_many_arguments    = "allow"
upper_case_acronyms   = "allow"
# Macro / FFI-driven noise we don't want to silence per call-site.
useless_conversion          = "allow"  # pyo3 ?-into-PyErr boilerplate
missing_safety_doc          = "allow"  # cover incrementally; not 1.0.0-blocking
doc_overindented_list_items = "allow"  # rustfmt-vs-clippy disagreement
doc_lazy_continuation       = "allow"  # same family — multi-line list continuations
unnecessary_wraps           = "allow"  # API symmetry > "this could be infallible"
type_complexity             = "allow"  # IR types are intentionally explicit
large_enum_variant          = "allow"  # Op enum carries Box<Graph> in CustomFn / Scan
single_range_in_vec_init    = "allow"  # `vec![0..1]` is fine in tests / scaffolding
collapsible_if              = "allow"  # nested-if is often more readable in numeric code
should_implement_trait      = "allow"  # builder methods named like trait methods

[profile.debug-rlx]
inherits = "dev"
debug = "full"
debug-assertions = true
overflow-checks = true
opt-level = 1
lto = false
incremental = true