Skip to content

TileScale v0.0.2

Latest

Choose a tag to compare

@Rachmanino Rachmanino released this 29 Jul 12:56
4e42b83

TileScale v0.0.2 Release

TileScale 0.0.2 establishes a new release baseline on upstream TileLang commit
550e25d,
with a hardened single-node multi-GPU programming model, CUDA IPC and VMM-backed
peer memory, multicast and multimem primitives, and release-ready packaging.

New Release Baseline

  • The Python distribution is now tilescale==0.0.2 while retaining the
    compatible tilelang import namespace.
  • The upstream TileLang compiler, JIT, cache, autotuning, CUDA, CuTeDSL,
    Blackwell/TCGEN05, ROCm, CPU, Metal, and WebGPU source surfaces are carried
    from the pinned upstream baseline.
  • Native dependencies are pinned to matching source and runtime ABIs, including
    apache-tvm-ffi==0.1.11 and z3-solver==4.15.4.0.

Single-Node Multi-GPU Features

  • Rank-aware remote load/store, peer put/get and TMA operations, system-scope
    signals, waits, atomics, barriers, and compile-once multi-process JIT
    coordination.
  • Zero-copy peer tensor views with collective allocation and deterministic
    ownership and teardown.
  • Automatic selection between CUDA VMM fabric handles and CUDA IPC, with
    runtime capability probes and explicit TILESCALE_USE_VMM=0/1 diagnostics.
  • Experimental multicast allocation and multimem load, store, reduction, and
    TMA operations on compatible NVSwitch/IMEX systems.
  • Executable all-gather GEMM, GEMM all-reduce, reduce-scatter, multicast, and
    multimem examples.

CUDA / Blackwell Correctness

  • CUDA 13 and Blackwell fixes cover metadata symbols, TMA alignment and rank-1
    descriptors, vector code generation, TCGEN05 fences, CuTeDSL/PDL behavior,
    multicast CLC cancellation, and reduction tolerances.
  • Remote TMA peer descriptors are range-checked instead of allowing invalid
    peers to fall through to another GPU.
  • Multimem data paths use consistent system-scope ordering; packed FP16/BF16
    layouts and numerical paths receive dedicated multi-GPU coverage.
  • Scalar remote load/store handles FP16 correctly and rejects misleading
    non-coherent load scope/semantic combinations.
  • Device IDs, constant remote-copy sizes, VMM environment values, optional peer
    arguments, and distributed signal diagnostics are validated explicitly.

JIT / Runtime / Build

  • Distributed modules and CUDA host bindings are imported lazily, preserving
    normal compiler and CPU-only imports when optional distributed dependencies
    are absent.
  • Allocator construction and teardown are collective and fail-closed, with
    staged error propagation, rollback, retryable close, and explicit ownership
    of imported mappings.
  • Release archives include license and notice material and are checked against
    required and restricted content before publication.
  • CPython 3.10+ is supported through CPython stable-ABI NVIDIA/CUDA wheels for
    Linux x86_64, Linux aarch64, and Windows x86_64.
  • This release publishes NVIDIA/CUDA wheel assets only; no ROCm, Metal, or
    source-distribution artifact is attached.

Validation

Validation ran on 8 NVIDIA B200 GPUs with CUDA 13.2, driver 590.48.01, full
NVSwitch topology, Fabric Manager, and an accessible IMEX channel. Independent
validation slices included:

  • 1,500+ upstream-compatible test cases across three broad groups, with
    hardware/backend capability skips reported separately.
  • 97 two-GPU distributed tests with forced CUDA IPC, plus three eight-GPU IPC
    examples.
  • 103 two-GPU distributed tests with automatic VMM selection and four
    eight-GPU VMM/multicast examples.
  • 58 focused VMM, multicast, and multimem primitive tests, native SM100 multimem
    TMA broadcast/reduction checks, and rank-1 TMA descriptor regressions.
  • A final four-GPU distributed suite and dedicated packed FP16/BF16 multimem
    numerical coverage.

The release wheels are additionally checked for archive contents, stable ABI
compliance, clean-environment installation, import behavior, and metadata.

Scope and Compatibility

  • TileScale 0.0.2 supports one host with one process per local GPU.
    Multi-node execution, NVSHMEM, DeepEP, and a general distributed runtime are
    outside this release scope.
  • VMM, multicast, and multimem remain capability-gated. Systems without the
    required GPUs, fabric, driver support, or accessible IMEX channel fall back
    to CUDA IPC where supported.
  • On Linux, distributed allocator imports require libcudart.so to be
    discoverable by the system dynamic loader. A CUDA runtime installed only in
    a Python package may require an explicit library search path.
  • On Windows, the distributed allocator currently attempts to load the POSIX
    libcudart.so name and is not functional; the base compiler/JIT wheel still
    installs and imports normally.
  • Upstream tilelang and tilescale must not be installed together because
    both distributions provide the tilelang import namespace.

Installation

Use a fresh environment, or remove both namespace providers before installing:

python -m pip uninstall -y tilescale tilelang
python -m pip install "tilescale[distributed]==0.0.2"
python -c "import tilelang; import importlib.metadata as m; print(m.version('tilescale'))"

Overall

This release centers on a hardened single-node multi-GPU TileLang extension,
automatic CUDA IPC/VMM peer-memory management, Blackwell multicast and
multimem support
, and portable, validated release packaging.

What's Changed

Full Changelog: 4704282...v0.0.2

What's Changed

  • [Feat] Implement get_dispatch_layout v1 for intranode DeepEP by @Rachmanino in #1
  • [Feat] Add copy_unrolled operation for optimized memory copying by @chengyupku in #2
  • [Feat]support example allgather gemm by @JiaqingFu in #10
  • [Feature] Support UVA copy with tile primitive by @chengyupku in #11
  • Merge upstream/main into main by @chengyupku in #12
  • [Bugfix] Fix the ipc bug caused by incorrect tensor management and update example_remote_copy by @Rachmanino in #13
  • [Feat] Add support for IPC-based warp-level pull from remote tensor and refactor relevent tests by @Rachmanino in #14
  • [Feat] Refactor UVA mechanism and introduce metadata table to support remote device-to-device copy by @chengyupku in #15
  • [Feat]support sp all2all without transpose example by @JiaqingFu in #16
  • [Feat][Doc] Enhance TileLang with NVSHMEM support and build system improvements by @chengyupku in #18
  • [BugFix] Robust import for different versions of cuda-python package by @tzj-fxz in #17
  • [Feat] Implement device and system level barrier for blocks based on atomic operations by @Rachmanino in #9
  • [Doc] Update README.md to include a link to the project roadmap for contributors by @chengyupku in #20
  • [Feat]support sequence parallel all2all with transpose example by @JiaqingFu in #21
  • [Benchmark] Add benchmark for IPC communication and support for fence operators by @Rachmanino in #22
  • [Feat] Add put_block/get_block implementation & examples and rename primitives by @Rachmanino in #23
  • [Feat]generalize example_all_to_all and use absolute import for benchmark_al… by @JiaqingFu in #24
  • [Feature][Example] Support return_peers in _allocate_tensor and add ag_gemm_ipc example by @chengyupku in #25
  • [Refactor][Example] Add AG-GEMM overlapping example by @chengyupku in #28
  • [Example] Update AG-GEMM example by @chengyupku in #30
  • Merge remote-tracking branch 'upstream/main' into main by @chengyupku in #31
  • [README] Update installation doc by @chengyupku in #32
  • [README] Update dependency in installation doc by @chengyupku in #34
  • [Feat]support internode copy with intranode copy by @JiaqingFu in #26
  • [Example] Update intra-node GEMM-RS example by @chengyupku in #35
  • [Example] Add CP example by @chengyupku in #37
  • [Feat]support container install docs by @JiaqingFu in #38
  • [BugFix] Fix port conflict in ci.yml pytest by @RMLYC in #29
  • [Feat] Support stride in T.serial by @Rachmanino in #41
  • [DeepEP] Intra-node DeepEP by @Rachmanino in #40
  • [Sync] Merge mainstream TileLang TVM-FFI features into TileScale by @chengyupku in #50
  • [BugFix] Fix multiple bugs by @Rachmanino in #53
  • [Doc] Add merge upstream TileLang skill doc by @Rachmanino in #54
  • [Release] TileScale 0.0.2 on upstream TileLang 550e25d by @Rachmanino in #61

New Contributors

Full Changelog: https://github.com/tile-ai/tilescale/commits/v0.0.2