v0.3.0

What's New

Scaling Infrastructure (M32-M34)

• Mixture of Experts — @moe decorator with top-k gating, capacity-based routing, and load-balancing aux loss
• Speculative Decoding — @speculative with tree attention, rejection sampling, and @medusa multi-head prediction
• Ring Attention — @context_parallel(ring_size=N) for cross-GPU sequence parallelism with causal masking

Quantization (M35)

• FP8 Compute — @fp8_compute decorator with E4M3/E5M2 scale management and automatic Tensor Core dispatch
• AWQ 4-bit — quant { dtype: awq4 } with in-register dequantize-in-GEMM (zero memory round-trip)
• GPTQ 4-bit/8-bit — quant { dtype: gptq4 } with Hessian-based optimal quantization

Compiler Intelligence (M36-M37)

• Memory Planning — compile-time tensor liveness analysis, interference graph, first-fit-decreasing slab assignment with 256-byte GPU alignment
• Roofline Cost Model — per-operation FLOP/byte/arithmetic-intensity analysis against a built-in GPU database (A100, H100, RTX-4090, RTX-3090, L40S); table, JSON, and Chrome tracing output formats

Language Features (M38-M40)

• Linear Types — ownership checker with use-after-move detection, branch consumption symmetry, loop consumption prevention, and @shared escape hatch
• Autodiff Safety — BackwardAccess classification for all 36 TapeOp variants (ShapeOnly/DataRequired/AuxDataRequired)
• vmap — @vmap(batch_dim=0) automatic batching with batch-variant/invariant tracking, dimension shifting, and matmul rewrite classification
• Source-to-Source AD — Wengert list extraction, 18 reverse-mode adjoint rules (reviewer-verified correct), dead gradient elimination, saved tensor analysis

Stats

• 408 unit tests passing across all crates
• 4,738 new lines of code across 42 files
• 18 new source modules + 7 implementation plans
• Clippy clean, release build verified

Breaking Changes

None. All new features are additive. Existing code compiles unchanged.

Known Limitations

• --linear-types, --vram-budget, --perf CLI flags are parsed but not yet wired through compile_entry() (same status as --fusion-report)
• Source AD and vmap are infrastructure-only (analysis libraries complete, codegen integration in progress)
• E2E tests that invoke nsl run require a C compiler (gcc/clang/MSVC) in PATH for linking

Full Changelog: v0.2.0...v0.3.0