Repository navigation
Releases
Releases · davidgraymi/neural-cost
Release v0.12.0
What's Changed
- feat: device-aware hardware detection and device-mismatch warning (#51) by @davidgraymi in #60
- feat(jax): model gather/scatter/dynamic slices and while trip counts (#52) by @davidgraymi in #61
- feat: report modeled vs unmodeled coverage of traced functions (#53) by @davidgraymi in #62
- feat: pipeline profiling with call counts and unexplained remainder (#54) by @davidgraymi in #63
- feat: device-aware peak calibration for JAX backends (#55) by @davidgraymi in #64
- feat: per-op breakdown table in render(), with JAX name-scope attribution (#56) by @davidgraymi in #65
- feat: compare one JAX function across platforms in subprocesses (#57) by @davidgraymi in #66
- docs: worked JAX + Flax + Apple Silicon example and jax-mps notes (#58) by @davidgraymi in #67
- feat: ship an installable AI agent skill for neural-cost (#59) by @davidgraymi in #68
- fix: issues found testing on real MPS (jax-mps) and Metal (jax-metal) GPUs by @davidgraymi in #69
Full Changelog: v0.11.1...v0.12.0
Release v0.11.1
What's Changed
- fix: TP recompute traffic, pipeline boundaries, position-dependent speculative acceptance by @davidgraymi in #49
Full Changelog: v0.11.0...v0.11.1
Release v0.11.0
What's Changed
- feat: cross-framework consistency, gap attribution, and published validation by @davidgraymi in #48
Full Changelog: v0.10.0...v0.11.0
Release v0.10.0
What's Changed
- feat: add zero-code Hugging Face model ingestion and SVG roofline exporter by @davidgraymi in #47
Full Changelog: v0.9.0...v0.10.0
Release v0.9.0
What's Changed
- feat: add activation checkpointing, selective recomputation, and training memory models by @davidgraymi in #46
Full Changelog: v0.8.0...v0.9.0
Release v0.8.0
What's Changed
- feat: add sub-byte quantization, dequantization tax, and kernel launch floor models by @davidgraymi in #45
Full Changelog: v0.7.0...v0.8.0
Release v0.7.0
What's Changed
- fix(ci): replace paulhatch/semantic-version with roll_version.py for correct minor bumps by @davidgraymi in #37
- feat(benchmarks): run and record benchmark suite for PRs 28-33 on Apple M1 by @davidgraymi in #38
- feat: add Mixture-of-Experts (MoE) and speculative decoding cost models by @davidgraymi in #39
- feat: add PagedAttention and continuous batching cost models by @davidgraymi in #40
- feat: add unified neural-cost CLI suite with profile, benchmark, and audit subcommands by @davidgraymi in #41
- feat: add State Space Model (Mamba/S6/SSD) cost models and CLI tooling by @davidgraymi in #42
- feat: add distributed 3D parallelism, interconnect rooflines, and cluster MFU analyzer by @davidgraymi in #43
- fix(ci): fix tags trigger indentation and python matrix in publish workflow by @davidgraymi in #44
Full Changelog: v0.6.3...v0.7.0
Release v0.6.3
What's Changed
- docs: add PyPI badges, repository links, and promote PyPI package by @davidgraymi in #36
Full Changelog: v0.6.2...v0.6.3
Release v0.6.2
What's Changed
- feat: realistic model topologies and scale tiers (#19) by @davidgraymi in #28
- feat: modernize operator suite with SDPA, RMSNorm, SwiGLU, and legacy gating (#23) by @davidgraymi in #29
- feat: multi-precision benchmark sweeps (FP16, BF16, and INT8) (#20) by @davidgraymi in #30
- feat: benchmark training workload phases and optimizer memory traffic (#21) by @davidgraymi in #31
- feat: benchmark LLM prefill vs. decode discrepancy and KV cache sizing (#22) by @davidgraymi in #32
- feat: add datacenter NVIDIA CUDA hardware support and benchmark script (#25) by @davidgraymi in #33
Full Changelog: v0.6.1...v0.6.2
Release v0.6.1
What's Changed
- feat: standard automated semantic version and PyPI release pipeline by @davidgraymi in #26
- feat: integrate peak memory telemetry and allocator tracking (#24) by @davidgraymi in #27
Full Changelog: v0.6.0...v0.6.1