[package]
name = "jouleclaw-loader-gguf"
version.workspace = true
edition.workspace = true
authors.workspace = true
license.workspace = true
repository.workspace = true
description = "GGUF model loader for JouleClaw. F32, F16, Q4_0, Q5_K, Q8_0 quants. mmap-zero-copy load path. Includes per-model decode (gemma, gemma4, llama). Ported from pattern-lang/joule-loader-gguf."

[dependencies]
jouleclaw-fabric = { path = "../jouleclaw-fabric", optional = true }
jouleclaw-backend-apple = { path = "../jouleclaw-backend-apple", optional = true }
jouleclaw-energy = { path = "../jouleclaw-energy", optional = true, features = ["ioreport"] }
jouleclaw-core = { path = "../jouleclaw-core" }
regex = "1"
fancy-regex = "0.20"   # HF tokenizer split patterns use lookahead (\s+(?!\S))
serde = { workspace = true }
serde_json = { workspace = true }
quick-xml = { workspace = true }
memmap2 = "0.9"
thiserror = { workspace = true }
metal = { version = "0.33", optional = true }
rayon = { version = "1", optional = true }
# Seal the energy-autotuner's tuning profile as an autotune/1 JCR-1 receipt
# (the `attest` feature) — reuses jouleclaw's own COSE_Sign1 envelope.
jouleclaw-jcr1 = { path = "../jouleclaw-jcr1", optional = true }
ed25519-dalek = { workspace = true, optional = true }

[features]
# Heterogeneous-compute placement: route batch-shaped matmuls
# through jouleclaw-fabric (AMX/Accelerate on Apple Silicon).
fabric = ["dep:jouleclaw-fabric", "dep:jouleclaw-backend-apple"]
# Fabric + the IOReport energy meter: enables the
# fabric_calibrate example that turns seeded placement
# tables into THIS machine's measured figures.
fabric-meter = ["fabric", "dep:jouleclaw-energy"]
# Seal a tuning profile as an autotune/1 JCR-1 receipt (COSE_Sign1 / Ed25519).
attest = ["dep:jouleclaw-jcr1", "dep:ed25519-dalek"]
default = []
# Direct Metal-buffer overlay over mmap'd GGUF regions. Only
# meaningful on Apple Silicon (UMA — the page cache IS the
# GPU's working set; zero CPU→GPU copy). Pulls in metal = 0.33
# transitively. Build flag, not a runtime switch.
metal = ["dep:metal"]
# io_uring async prefetch on Linux 5.1+. Replaces best-effort
# madvise WILLNEED with explicit async READ submissions +
# completion queue polling. Deterministic completion;
# measurable latency. Linux-only; no-ops on other targets.
io_uring = ["dep:io-uring"]
# CUDA UVA overlay — registers mmap'd pages with the CUDA
# driver via cudaHostRegister so the device can read host
# pages over PCIe without an explicit cudaMemcpy. Discrete-
# GPU equivalent of the Metal feature. Runtime-loaded via
# libloading; no CUDA toolkit needed at build time. Linux +
# Windows where the CUDA driver is installed; no-ops
# otherwise.
cuda = ["dep:libloading"]
# NUMA-aware allocation on Linux. Runtime-loaded libnuma.so.1
# via libloading; multi-socket inference servers can place
# weights on the right node + bind compute threads.
numa = ["dep:libloading"]
# Apple Accelerate / cblas_sgemm — uses AMX (Apple's matrix
# accelerator) on M-series chips. Routes f32 matmul through
# Accelerate.framework instead of the scalar / NEON path.
# macOS aarch64 only; no-op cfg-gated stub elsewhere.
accelerate = []
# Multi-thread fan-out via rayon. Adds parallel matmul
# variants that split work across CPU cores. Useful for
# batched inference (m ≥ 2) + n-wide single-token decode.
rayon = ["dep:rayon"]

[target.'cfg(any(target_os = "linux", target_os = "windows"))'.dependencies]
libloading = { version = "0.9", optional = true }

[target.'cfg(target_os = "linux")'.dependencies]
io-uring = { version = "0.7", optional = true }

[dev-dependencies]
criterion = { version = "0.8", default-features = false, features = ["plotters", "cargo_bench_support"] }

[[bench]]
name = "matmul_q8_0"
harness = false

[[bench]]
name = "matmul_q4_k"
harness = false

[[bench]]
name = "matmul_q56_k"
harness = false

[[bench]]
name = "attention"
harness = false

[[bench]]
name = "mmap_strategy"
harness = false

[[bench]]
name = "tier_store"
harness = false

# Energy-metered examples link the optional jouleclaw-energy
# crate, which only exists under `fabric-meter`. Gate them so a
# plain `--features "rayon fabric"` (or default) build does not
# try to compile them and fail on the unlinked crate.
[[example]]
name = "fabric_calibrate"
required-features = ["fabric-meter"]

[[example]]
name = "mpp_energy"
required-features = ["fabric-meter"]

[[example]]
name = "energy_autotune"
required-features = ["fabric-meter"]

[target.'cfg(unix)'.dependencies]
libc = "0.2"

[target.'cfg(windows)'.dependencies]
windows-sys = { version = "0.61", features = [
    "Win32_System_Memory",
    "Win32_Foundation",
    "Win32_System_Threading",
] }