v0.12.0-pre.2预发布2026-11-09
What's Changed
- chore: bump version to 0.12.0-pre.2 (#1965) @laggui
- fix(cuda,hip): apply an allocation window to its stream at once (#1963) @SamuelBelanger
- Store: an eager store serves a value the storage refused to write (#1970) @nathanielsimard
- feat(tile): accumulator runs as one A operand; NVFP4 stages by TMA (#1971) @louisfd
- fix(tune): run a pending key's fallback in the plan's order (#1697) @vboussot
- Serialize DeviceProperties and pack a QuantScheme into one u64 (#1973) @nathanielsimard
- Close HIP devices before HIP tears itself down at exit (#1972) @nathanielsimard
- feat: Speculative PRE (#1924) @wingertge
- fix(cpp): use the target's native log1p (#1817) @shinaoka
- feat: native Metal on par with wgpu-msl; tile shared-memory reuse (#1975) @louisfd
- fix(runtime): a lazy read holds its source until its bytes are read (#1967) @ThierryCantin-Demers
- ci: run CI only on PRs to main (#1993) @louisfd
- feat(llvm): NVPTX lowers cp.async and block-scaled mma on e2m1x2 (#1995) @louisfd
- feat(common): spawn a device task and join it later (#1989) @ThierryCantin-Demers
- feat(tile): accumulator rows along either factor; NVPTX e2m1 decode (#2006) @louisfd
- fix(server): align with a buffer's queued writer, not only its owner (#2002) @ThierryCantin-Demers
- feat(llvm): no fp6 format; native e4m3/e5m2 decode from sm_89 (#2008) @louisfd
- feat(environment): create, fork, load and save environments by id (#2014) @nathanielsimard
- fix(hip): keep code objects as bytes, whatever the sign of c_char (#2018) @Marc-AnthonyG
- docs(readme): refresh CubeCL branding and badge layout (#2020) @laggui
- Implement native Rust arrays and fix some array semantics (#2021) @wingertge
- feat(tile): a masked operand binds at a width dividing its extent (#2024) @nathanielsimard
- Tile: a register prefetch falls back to the slots where it cannot fetch straight (#1982) @SamuelBelanger
- usize isize harness support (#1944) @Redhawk18
- feat!: cubecl's own Metal runtime; cubecl-tile API trimmed; exp2 (#2016) @louisfd
- RDNA4 fp8 and RDNA int8 WMMA in the LLVM backend (#2032) @nathanielsimard
- fix(tile): monoid Max and Min keep NaN; strided fragment loads (#2035) @louisfd
- fix(runtime): retain lazy read sources while host bytes alias mutable storage (#2036) @laggui
- feat(hip): collectives and transfers between AMD GPUs over RCCL (#2030) @ThierryCantin-Demers
- Decode quantized fp8 in software in cubecl-std (#2040) @SamuelBelanger
- feat(environment): merge an environment into another (#2046) @nathanielsimard
- feat: tile improvement (#2053) @marcantoinem
- feat(tile): Indexed and TopK monoids; attention backward fragments (#2054) @louisfd
- feat: kernels read and write another GPU's memory (peer access) (#2048) @nathanielsimard
- basic iterator functions (#2037) @Redhawk18
- fix(gelu): use host sqrt during comptime expansion (#2058) @istaari
- perf(tile): two B fragments of a grid load in one ldmatrix (#2056) @louisfd
- fix: handle negative exponents in integer powi (#2043) @b84330808
- fix(spirv): infer variable-pointer capabilities from usage (#2059) @laggui
- fix(server): compile a kernel batch on every core, not the server's cache (#2063) @nathanielsimard
- feat(tile): a zero tile the softmax takes as no bias (#2069) @louisfd
- Unroll a constant loop only where a local array's index changes each turn (#2070) @SamuelBelanger
- perf(tile): a short window's lane runs read a clamped row (#2068) @nathanielsimard
- fix(capture): record in thread-local mode, so several devices can capture at once (#2075) @ThierryCantin-Demers
- perf(tile): Coords::unravel skips the division for a one-row window (#2074) @louisfd
- fix(tile): wrap the units of a paired ldmatrix onto the pair's matrices (#2073) @SamuelBelanger
- fix(llvm): advertise copy_async on NVPTX (#2052) @Charles23R
- Fix/environment wal restarts (#2087) @Charles23R
- Contract a gathered convolution input in a dense tiled stage, forward and backward (#2071) @SamuelBelanger
- feat(runtime): time a benchmark on several devices at once (#2078) @ThierryCantin-Demers
- perf(tile): faster top-k; capture per thread for several devices (#2086) @louisfd
- fix(examples): the multi_device example turns on no runtime by default (#2089) @ThierryCantin-Demers
- feat(collective): members queue their own part; faster top-k for long lists (#2093) @louisfd
- refactor: warn on a needless comptime!; write-only tile ops are IR ops (#2097) @louisfd
- fix(tune): a settled winner that refuses a call falls back down the plan (#2098) @louisfd
- feat(tile): a storage partitioning can stagger its coarsest pieces (#2077) @nathanielsimard
- fix(ir): a fold by one is not the other operand, for remainder and mul_hi (#2105) @nathanielsimard
- chore: prepare 0.12.0-pre.2 release (#2109) @laggui
v0.12.0-pre.1预发布2026-10-28
What's Changed
- fix(opt): make simple CSE linear in the operations of a region (#1767) @ThierryCantin-Demers
- fix(runtime): leave a launch a dry run drops untimed (#1773) @ThierryCantin-Demers - also released in 0.11.1.
- chore: declare intra-workspace dependencies once in the root manifest (#1772) @ThierryCantin-Demers
- Perf/minifloat decode (#1777) @louisfd
- Feat/parallel compilation (#1770) @nathanielsimard
- update readme (#1771) @louisfd
- fix(runtime): queue both halves of a device transfer under one lock (#1760) @ThierryCantin-Demers
- perf(macros): share a kernel's define markers across kernels (#1781) @nathanielsimard
- fix(wgpu): create and drop Vulkan instances one thread at a time (#1783) @ThierryCantin-Demers - also released in 0.11.1.
- fix(wgpu): fail instead of waiting forever on a hung device (#1778) @ThierryCantin-Demers
- fix(runtime): stop a tune on a lost device instead of deciding a winner (#1779) @ThierryCantin-Demers
- test(cmma): run the fp4 scaled mma under NVFP4's e4m3 scales (#1784) @louisfd
- Feat/build progress (#1787) @nathanielsimard
- docs(fp4): no private or ambiguous links (#1795) @louisfd
- fix(metal): use bulk copies for contiguous tensors (#1718) @ArthurBrussee
- fix(cuda): stop a two-digit PTX placeholder being eaten by its prefix (#1691) @vboussot - also released in 0.11.1.
- feat(cuda): offer the k = 8 manual mma for f16 from Turing on, and for bf16 from Ampere on (#1802) @louisfd
- fix(cuda): give transfers their own communicator and stream (#1782) @ThierryCantin-Demers
- fix(throughput): fail a compute probe whose kernel did not run (#1793) @SamuelBelanger
- feat(quant)!: scheme axis updates, a float value check, and E2M1/E8M0 without std (#1798) @ThierryCantin-Demers
- fix: respect disabled default features in WGPU compiler dependencies (#1792) @laggui - also released in 0.11.1.
- fix(docs): skip LLVM build steps on docs.rs (#1794) @laggui - also released in 0.11.1.
- feat: Early return and continue (#1788) @wingertge - The loop-condition scope fix was also included in 0.11.1.
- Feat/wgmma (#1791) @nathanielsimard - The f16 stmatrix packing fix was also included in 0.11.2 via #1880.
- perf(channel): park the idle device server and wake it when a buffer fills (#1806) @louisfd - also released in 0.11.3.
- Fix/wgsl flat workgroup (#1809) @nathanielsimard
- Run the adaptive pool tests serial with the process mode overrides (#1814) @nathanielsimard
- fix: improve floating-point powi accuracy (#1803) @laggui
- fix(environment): let Mutex and RwLock hold unsized values (#1830) @ThierryCantin-Demers
- fix(autotune): key the tune cache on the adapter and API a device resolved to (#1811) @arcusbuilds
- Change Float::new to f64 (#1815) @wingertge
- Store the cmma throughput probe's tile from the whole first plane (#1828) @SamuelBelanger
- fix(deps): update rustls to 0.23.46 (#1832) @ThierryCantin-Demers - also released in 0.11.1.
- Feat/cubin compilation (#1831) @Charles23R
- docs(wasm): explain constructor overhead and explicit startup initialization (#1834) @laggui - also released in 0.11.1.
- fix(cpu): report f16 arithmetic as emulated on hosts without native f16 (#1807) @skewballfox
- feat: specify native cpu for llvm (#1833) @marcantoinem - also released in 0.11.1.
- perf(std): parallel perpendicular copy on GPU (#1813) @nathanielsimard
- fix(core,cpp,spirv): restore support for boolean comparisons and XOR (#1825) @egilll
- feat: downcast a cmma fragment or a pending wgmma accumulator, as a slice is; fix: a cmma cast between two names of one element is no copy (#1836) @louisfd
- feat(cpu): add persistent JIT compilation cache (#1527) (#1827) @Sadik00789
- fix(cuda): offer FP8 mma.sync from sm89 and FP6/FP4 on the sm120 family only (#1774) @Marc-AnthonyG
- feat(cuda): sm120 ldmatrix/stmatrix forms and scalevec::4X ue8m0 (#1775) @Marc-AnthonyG
- perf(opt): visit only an alloc's own ops in mem2reg (#1835) @ThierryCantin-Demers
- fix(llvm): keep the NVPTX options across LLD links (#1837) @skewballfox
- perf(opt): order mem2reg's same-block users with block_order (#1844) @ThierryCantin-Demers
- cube root and inverse cube root support (#1786) @Redhawk18
- feat(ir): report shared-memory atomics apart from global ones (#1845) @louisfd
- fix(core): cbrt's constants are the f64 that Float::new takes (#1847) @louisfd
- fix(core): the inverse cube root test's infinity is the f64 as_type! takes (#1849) @louisfd
- feat(macros): a define! lasts to the end of its block (#1846) @louisfd
- fix(environment): an import commits a transaction per 64 MiB, not only per 1024 entries (#1851) @louisfd
- feat(runtime): allgather, reducescatter, broadcast and alltoall between devices (#1843) @Thie
v0.11.32026-10-22
What's Changed
- fix(cpp): lower half tanh for older CUDA toolkits (#1856) @arcusbuilds
- perf(channel): park the idle device server and wake it when a buffer fills (#1806) @louisfd
v0.11.22026-10-21
What's Changed
- fix(llvm): give f16 its width when stmatrix packs registers into words (#1880) @nixvio64
- backported from #1791 by @nathanielsimard
- fix(cubecl-std): use fromint128 in 64-bit FastDivmod::register to prevent u64>=2^63 wrap panic (#1859) @ssoj13
- fix(cubecl-macros): AutotuneKey's new writes an unanchored field in shorthand (#1879) @louisfd
- fix(spirv): a plane barrier orders the workgroup's memory (#1871) @skewballfox
- fix(cpp): e8m0 converts through its source's converter (#1869) @louisfd
- fix(deps): update webpki dependencies to fix audit @laggui
- fix(macros): preserve expanded type equality in CubeTypeMut bounds (#1888) @laggui
v0.11.12026-10-20
What's Changed
- fix(cuda): stop a two-digit PTX placeholder being eaten by its prefix (#1691) @vboussot
- fix(wgpu): create and drop Vulkan instances one thread at a time (#1783) @ThierryCantin-Demers
- fix(runtime): leave a launch a dry run drops untimed (#1773) @ThierryCantin-Demers
- fix(core): read loop conditions in their own scope (extracted from #1788) @wingertge
- fix: respect disabled default features in WGPU compiler dependencies (#1792) @laggui
- fix(docs): skip LLVM build steps on docs.rs (#1794) @laggui
- fix(docs): enable branches/std for persistence build (#1864) @laggui
- docs(wasm): explain constructor overhead and explicit startup initialization (#1834) @laggui
- feat: specify native cpu for llvm (#1833) @marcantoinem
- fix(deps): update rustls to 0.23.46 (#1832) @ThierryCantin-Demers
- Fix: SPIR-V buffer tagging when read only (#1861) @ChillFish8
- chore(deps): bump tracel-llvm-bundler to 23.1.0-4 (#1866) @laggui
v0.11.02026-10-06
What's Changed
- Add tiled layout (#1329) @Sublime12
- Bump version to 0.11.0-pre.1 (#1326) @laggui
- feat: streaming: add stream priority hint, wire into CUDA backend (#1324) @lilith
- perf: lighter cpu runtime (#1330) @marcantoinem
- feat: Add support for VKEXTshaderlongvector (#1188) @wingertge
- mega-refactor: Totally change the frontend to enable references, among other things (#1322) @wingertge
- Add f64 support back for CUDA. (#1321) @vaijira
- fix: Ensure constant block args are generated in the correct block (#1332) @wingertge
- Add workgroupUniformLoad primitive (#1327) @ArthurBrussee
- fix: shared memories allocation calculation were wrong (#1337) @marcantoinem
- make renderdoc optional (does not compile on mac os) (#1339) @louisfd
- Metal: fix atomic syntax (#1340) @louisfd
- Metal: missing space caused compilation errors (#1342) @louisfd
- Feat/obfuscation (#1341) @nathanielsimard
- fix metal compilation (#1343) @louisfd
- fix: Fix more compilation issues on metal (#1347) @wingertge
- feat(runtime): add memoryusagetotal aggregating all streams (#1333) @ArthurBrussee
- Refactor/wgpu compilers (#1346) @nathanielsimard
- Improve CPU backend by removing need to flush, add atomics test and add atomic to cubecl-cpu (#1345) @marcantoinem
- fix: CubeOr expand methods computed AND instead of OR (#1350) @LucaCappelletti94
- fix: short-circuit || and && in #[cube] code (#1348) @LucaCappelletti94
- refactor: Add a lifetime to views (#1344) @wingertge
- fix: Assign/init_mut (#1355) @wingertge
- fix: cubecl-cpu synchronisation problem (#1356) @marcantoinem
- fix: skip kernel launch when cube count is zero (#1349) @LucaCappelletti94
- fix: Fix atomics on CUDA again (#1354) @wingertge
- Refactor/readme (#1358) @nathanielsimard
- fix: re-export wgpu::Backend from cubecl-wgpu (#1357) @zhan-wei-919
- refactor(runtime): aggregate memoryusage + memorycleanup in client (#1360) @ArthurBrussee
- Tiling as a composable view layout (#1362) @louisfd
- fix(wgpu): bound live timestamp query sets per device (Metal counter-sample-buffer exhaustion) (#1361) @AdrianEddy
- fix shared memory bytes (#1367) @louisfd
- fix: reject unsupported kernel argument types instead of panicking (#1373) @LucaCappelletti94
- fix: don't panic when calling a non-path expression in a kernel (#1372) @LucaCappelletti94
- fixed typo check and bitwise test errors (#1369) @Andy2887
- fix: don't panic when const-folding an unfoldable const expression (#1374) @LucaCappelletti94
- book: constants (plane) (#1320) @Redhawk18
- feat: add expm1 lowering across backends (#1301) @shinaoka
- refactor: Simplify Variable to align it with existing IRs (#1378) @wingertge
- Feat/bytes improvements (#1379) @nathanielsimard
- Fix imports (#1383) @nathanielsimard
- Fix handling one tuple (#1382) @akiradeveloper
- fixes return Self::Scalar in implunaryfuncscalarout (Fixes #1283) (#1338) @ethqnol
- Feat/fix abusive allocation in CPU backend (#1385) @marcantoinem
- feat: Coordinates CoordsDynI type (#1386) @SamuelBelanger
- Add read_lazy method for non-WASM targets in ComputeClient (#1392) @jwric
- add as read to view mut (#1393) @louisfd
- Cpu dump improvement (#1394) @marcantoinem
- add citation (#1335) @Redhawk18
- burn element trait related changes for e4m3, e5m2 (#1389) @skewballfox
- Fix: returning panic payload in device channel tasks to caller and re-raising panic for blocking APIs (#1376) @Andy2887
- feat: metal backend (#1175) @dcvz
- fix(metal): attribute launch errors to the issuing stream (#1403) @dcvz
- fix(metal): validate shared memory limit before pipeline creation (#1405) @dcvz
- fixed cargo xtask validate fails on Linux (#1397) @Andy2887
- refactor: replace dirs with etcetera (drops MPL-2.0 option-ext) (#1395) @shimwell
- Feat/new CPU runtime (#1400) @marcantoinem
- Peak device throughput (#1408) @ThierryCantin-Demers
- disaggregate-oob-consts (#1409) @nathanielsimard
- Fix/cpu index unit (#1411) @louisfd
- fix(cpu): reserve shared memory from a dedicated pool (#1412) @louisfd
- Feat/dynamic memory pool config (#1417) @nathanielsimard
- Feat/hip captured graph (#1415) @nathanielsimard
- update wgpu version (#1416) @Charles23R
- Fix/graph safety (#1419) @nathanielsimard
- Disable persistent tune cache option (#1423) @SamuelBelanger
- Feat/autotune throughput (#1422) @SamuelBelanger
- fix(cpu): parallelize sync_cube and clamp test cube dims to core count (#1424) @louisfd
- Cpu: Allow more threads than cores (#1426) @louisfd
- Wmma compile error (#1427) @nathanielsimard
- add cross-stream input bindings pinning wgpu (#1434) @Charles23R
- fix atomic import (#1441) @Charles23R
- Allow nested tuple destructure (#1420) @akiradeveloper
- feat: add AsIndex fallible dimension index conversion (#1445) @mattisonchao
- chore: use tracel-llvm version 22.1.4-5 (#1447) @laggui
- fix: add missing cubecl-metal publish (#1448) @laggui
- Feat/autotune observability (#1437) @SamuelBelanger
- CubeCL Environment: multiplatform HPC runtime API (#1435) @nathanielsimard
- Fix initialize memory failure with error handling + fallback (#1454) @nathanielsimard
- Invalidate the kernel cache when a kernel changes (#1455) @nathanielsimard
- fix(quant): map QuantParam::UE4M3 to FloatKind::E4M3 (#1450) @ThierryCantin-Demers
- Feat/dry run environment (#1457) @nathanielsimard
- fix(runtime): check PendingDropQueue during gpu writes to prevent u32 overflow (#1456) @for-lack-of-a-better-name-j
- Adaptive autotune scheduler with early elimination (#1449) @SamuelBelanger @nathanielsimard
- feat(runtime): DeviceProperties::identity (#1465) @nathanielsimard
- feat(tune): roofline bounds from a Work amount and per-resource thresholds (#1462) @SamuelBelanger
- fix(float): correct the e4m3 and e5m2 limits and exponent constants (#1472) @nathanielsimard
- feature(cubecl-cuda): add cuda graph capture and replay (#1469) @michaelvsinko
- fix(cuda): reclaim and retry a failed device reserve (#1468) @laurigates
- fix(common): give device runner threads a shutdown path (#1470) @nathanielsimard
- fix(runtime): classify driver out-of-memory apart from BufferTooBig (#1474) @ThierryCantin-Demers
- Add a two-level quantization level (#1453) @nathanielsimard
- fix(common): call ceil through numtraits so roundup builds on no_std (#1475) @ThierryCantin-Demers
- fix(cpp): materialize constant bitcasts as lvalues (#1477) @jcwal1516
- fix(common): release the caller's borrows before waking it in run_scoped (#1478) @ThierryCantin-Demers
- fix(spirv): name the FPEncoding variants added in tracel-rspirv 0.13.3 (#1484) @louisfd
- fix(cubecl-hip): implement sync_warp instead of emitting #error (#1483) @nathanielsimard
- feat(common): add ComptimeFloat and Ratio for hashable comptime kernel params (#1482) @SamuelBelanger
- chore: version Cargo.lock (#1485) @ThierryCantin-Demers
- chore: bump version to 0.11.0-pre.2 (#1489) @laggui
- fix(publish): add missing cubecl-environment (#1490) @laggui
- fix(publish): add missing cubecl-std requirement to cubecl-metal (#1492) @laggui
- Feat/throughput memory curve (#1493) @louisfd
- feat(common): support negative numerators in Ratio (#1491) @SamuelBelanger
- fix Mutex imports in cubecl-metal (#1481) @olukowski
- refactor: Make everything Pliron (#1402) @wingertge
- Fix memory read throughput runner (#1495) @wingertge
- Reorder passes to avoid CondBrOp panic (#1496) @wingertge
- Fix features for cubecl-spirv (#1497) @wingertge
- Suffixes the llvm.is.fpclass intrinsic with the type to fix crash in cubek-reduce (#1498) @wingertge
- Add missing dynamic extract lowering for CPU (#1499) @wingertge
- Clean up dead regions before lowering branches to avoid weird crash in DCE (#1500) @wingertge
- Fix ue8m0 display (#1501) @wingertge
- Apply the global scale in the quantized view for two level quant (#1479) @ThierryCantin-Demers
- Improve memory management and memory poo
v0.11.0-pre.4预发布2026-09-22
What's Changed
- fix(runtime): lazy error handling on the logical stream (#1548) @nathanielsimard
- Perf/add fma detection (#1550) @marcantoinem
- feat(cpu): report the device last level cache size (#1551) @ThierryCantin-Demers
- fix(throughput): probe a working set at its own footprint (#1552) @ThierryCantin-Demers
- fix(throughput): walk the last window of every cycle (#1554) @ThierryCantin-Demers
- Fix/e2m1 software codec (#1557) @nathanielsimard
- fix(llvm): name an overloaded intrinsic for the type it is overloaded on (#1561) @ThierryCantin-Demers
- feat(cpu): Vectorize exp,log,sin,cos,tanh with polynomial approximation (#1553) @ThierryCantin-Demers
- Feat/erased sink (#1563) @nathanielsimard
- Feat/execution observer (#1573) @nathanielsimard
- Refactor/runtimes (#1568) @nathanielsimard
- fix(ir): an scf.if reports its condition as its result to SCCP (#1574) @nathanielsimard
- fix(llvm): give ue8m0 a type, and report it on the CPU backend (#1565) @ThierryCantin-Demers
- rename the fetch_update deprecrated method to remove miri warnings (#1569) @ThierryCantin-Demers
- fix(cpp): an accumulator fragment names no layout (#1570) @ThierryCantin-Demers
- fix(plane): document the shuffle up and down boundary as undefined (#1564) @ThierryCantin-Demers
- fix(tests): stop f16 tolerating an error of eight in assertequalsapprox (#1567) @ThierryCantin-Demers
- Add data flow solver (#1572) @wingertge
- feat: add complex support and validation (#1300) @shinaoka
- Feat/add amdgpu backend (#1571) @marcantoinem
- fix(throughput): stop the memory curve measuring cache, and collapse its modes (#1559) @ThierryCantin-Demers
- Llvm amdgpu fixes (#1577) @nathanielsimard
- fix(core): pin a scalar's WithScalar to itself (#1578) @ThierryCantin-Demers
- Refactor/amdgpu llvm review (#1582) @nathanielsimard
- Revert "rename the fetch_update deprecrated method to remove miri warnings (#…" (#1583) @ThierryCantin-Demers
- feat(profiling): device timing for the HIP and CUDA backends (#1580) @nathanielsimard
- feat: add dp4a intrinsic for packed int8×4 DOT accumulate (#1471) @saivishwak @wingertge
- Chore/reduce dependencies (#1587) @nathanielsimard
- fix: stop advertising tensor cores on dies that have none (#1588) @ThierryCantin-Demers
- fix macos llvm (#1586) @louisfd
- Fix/throughput probe rotation (#1589) @nathanielsimard
- Add fine-grained memory clobbers for gpu_asm (#1576) @wingertge
- fix(cpp): emit casts using destination type (#1549) @y7nieSEl5
- feat: update tracel llvm (#1593) @marcantoinem
- test(runtime): coordinate the profile tests with the refusal toggle (#1579) @ThierryCantin-Demers
- fix(cuda): stop advertising MMA and tensor cores on dies that have none (#1591) @ThierryCantin-Demers
- fix(throughput): measure the peaks the hardware can carry, and say why a device has none (#1598) @ThierryCantin-Demers
- fix(wgpu): close a device profile on its last compute pass (#1592) @ThierryCantin-Demers
- fix(spirv, wgsl): emit real fmas and fuse a product into the subtraction consuming it (#1562) @ThierryCantin-Demers
- Refactor/runtime erasure (#1590) @nathanielsimard
- refactor: ComputeClient and ComputeServer are Client and Server (#1603) @nathanielsimard
- fix(topology): a cube's units are contiguous in ABSOLUTE_POS (#1595) @ThierryCantin-Demers
- feat(device): make Device a device in its own right (#1604) @nathanielsimard
- Feat/metadata tiling (#1605) @louisfd
- Fix/ordered tune groups (#1615) @nathanielsimard
- feat(tune): an eviction the tuner runs before every measured sample (#1621) @nathanielsimard
- Feat/metadata logical shape (#1622) @louisfd
- fix(cpu): report the host CPU as the device name (#1623) @ThierryCantin-Demers
- fix: add integer powi lowering (#1624) @laggui
- fix(throughput): stop reporting ceilings the hardware beats (#1607) @ThierryCantin-Demers
- feat(zspace): the stride and shape mutators refuse a storage-tiled metadata (#1627) @louisfd
- fix(ir): decline the self-operand folds when the result is a vector (#1626) @ThierryCantin-Demers
- feat: Uniformity analysis (#1620) @wingertge
- fix: don't compile AMDGPU for MacOS (#1597) @marcantoinem
- fix(cpp): lower CUDA complex casts with cuComplex helpers (#1581) @shinaoka
- fix(cuda): brace the ldmatrix destination when the vector is 1 wide (#1596) @michaelvsinko
- fix(llvm): gate the last AmdGpu match arm on the amdgpu feature (#1631) @louisfd
- fix(llvm): gate the AMDGPU arm of lang_tag (#1630) @ThierryCantin-Demers
- fix(wgpu): a window with no timing reports the absence, not a zero (#1632) @louisfd
- Fix/profile/absence is not zero (#1634) @louisfd
- fix format (#1637) @ThierryCantin-Demers
- fix(llvm): keep amdgpu out of cubecl-llvm's default features (#1638) @ThierryCantin-Demers
- fix: metal kernel corruption (#1636) @Marc-AnthonyG
- Refactor/crate layout (#1606) @nathanielsimard
- fix(runtime): disambiguate kernel entrypoint names and avoid launch-path allocation (#1625) (#1633) @Sadik00789
- fix audit (#1644) @ThierryCantin-Demers
- ci: pin the miri job to nightly-2026-09-09 (#1640) @ThierryCantin-Demers
- feat(cpu): evaluate f16 arithmetic in f32 where the host has none (#1629) @ThierryCantin-Demers
- chore: bump version to 0.11.0-pre.4 (#1648) @laggui
- fix(runtime): saturate FlushingPolicyState counters to prevent add-overflow panic (#1359) (#1413) @teddytennant
- device fence (#1654) @louisfd
- fix(server): skip copy-layout validation for empty transfers (#1660) @laggui
- feat: MemorySSA (#1656) @wingertge
- fix(throughput): probe memory outside the installed pools, and fail when it cannot be allocated (#1659) @SamuelBelanger
- feat(wgpu): write host data straight into mapped storage on integrated Vulkan GPUs (#1651) @SamuelBelanger
- chore: update to pliron 0.18 (#1662) @laggui
- fix launch overhead (#1664) @louisfd
- Feat/cuda llvm backend (#1609) @nathanielsimard
- Fix/hip runtime on no hip platform (#1649) @syl20bnr
- Feat/profiling (#1666) @nathanielsimard
- fix(runtime): refuse collectives on a runtime with no device transport (#1653) @ThierryCantin-Demers
- feat(ir): report the physical card behind a device (#1652) @ThierryCantin-Demers
- fix(llvm): build the bitcode linker shim for NVPTX too (#1667) @ThierryCantin-Demers
- feat: state the device's capacity in MemoryDeviceProperties (#1672) @nathanielsimard
- fix(opt): walk memory phis on an explicit stack (#1669) @SamuelBelanger
- chore: use tracel-rspirv and plrion-spirv 0.15.0 (#1681) @laggui
- refactor e2m1 (#1682) @louisfd
- Feat/environment records (#1680) @nathanielsimard
- feat: fix hip test on machine without ROCm support and bump llvm (#1679) @syl20bnr
- fix(ci): publish LLVM before CUDA (#1683) @laggui
- Refactor/turso persistence (#1677) @jwric
- fix(llvm): emit the PTX version the CUDA driver loads (#1684) @SamuelBelanger
- fix(throughput): release the probe pool lock before the device call (#1678) @ThierryCantin-Demers
v0.11.0-pre.3预发布2026-08-25
What's Changed
- fix(publish): add missing cubecl-std requirement to cubecl-metal (#1492) @laggui
- Feat/throughput memory curve (#1493) @louisfd
- feat(common): support negative numerators in Ratio (#1491) @SamuelBelanger
- fix Mutex imports in cubecl-metal (#1481) @olukowski
- refactor: Make everything Pliron (#1402) @wingertge
- Fix memory read throughput runner (#1495) @wingertge
- Reorder passes to avoid CondBrOp panic (#1496) @wingertge
- Fix features for cubecl-spirv (#1497) @wingertge
- Suffixes the llvm.is.fpclass intrinsic with the type to fix crash in cubek-reduce (#1498) @wingertge
- Add missing dynamic extract lowering for CPU (#1499) @wingertge
- Clean up dead regions before lowering branches to avoid weird crash in DCE (#1500) @wingertge
- Fix ue8m0 display (#1501) @wingertge
- Apply the global scale in the quantized view for two level quant (#1479) @ThierryCantin-Demers
- Improve memory management and memory pools (#1494) @nathanielsimard
- Add SROA pass @wingertge
- Cache metadata info uniforms in the wgpu runtime (#1504) @nathanielsimard
- Zero-init arrays (#1511) @wingertge
- Fix MSL codegen (#1507) @louisfd
- Fix/pliron codegen regressions (#1508) @nathanielsimard
- Add version check to ensure required features are supported in MSL (#1512) @wingertge
- Fix atomics and elect on CUDA and WGSL (#1517) @wingertge
- Fix Pliron timer panic on WebAssembly (#1519) @syl20bnr
- Refactor/comptime option launch derive (#1513) @ThierryCantin-Demers
- Fix/unroll initializer type (#1515) @louisfd
- Update to syn 3.0 (#1521) @wingertge
- Declare wasm-bindgen-futures for WASM runtime (#1518) @syl20bnr
- Fix cubecl-runtime build for wasm32 (#1524) @ArthurBrussee
- Sync before the load in workgroupuniformload (#1525) @ArthurBrussee
- fix broken merge (#1528) @wingertge
- Add two-level caching for CUDA and HIP (#1526) @wingertge
- Wgpu graph capture (#1505) @nathanielsimard
- Replace QuantLevel with per-tensor and per-block scale levels (#1514) @ThierryCantin-Demers
- Feat/quant lookup (#1532) @louisfd
- Fix/cpu shared memory (#1533) @nathanielsimard
- feat(cpu): implement thread affinity and topology on Windows (#1537) @ThierryCantin-Demers
- ci: pin the shared actions to v11 (#1535) @ThierryCantin-Demers
- Add non-pointer MMA dialect (#1530) @wingertge
- Fix/plane reductions and emitter errors (#1520) @nathanielsimard
- feat(quant): Software fp8 conversion on every backend, native where the hardware has it (#1534) @ThierryCantin-Demers
- chore: tracel github actions v11 (#1539) @syl20bnr
- chore: fix clippy lints (#1541) @laggui
- refactor: split cubecl-cpu in two to extract cubecl-llvm (#1542) @marcantoinem
- fix(metal): correct MSL capability detection (#1540) @laggui
- fix(wgpu): keep the query set current names out of cleanup (#1543) @nathanielsimard
- chore: update to xtask v5 (#1544) @syl20bnr
- fix(cpu): yield between idle polls instead of spinning (#1545) @ThierryCantin-Demers
- feat(bench): write-only memory probe and roofline scoring for benchmarks (#1546) @ThierryCantin-Demers
- fix(unroll): unroll dynamic vector extract and insert (#1547) @nathanielsimard
- chore: bump version to 0.11.0-pre.3 (#1555) @laggui
- fix(cubecl-llvm): add missing publish and use released tracel-llvm-bundler (#1556) @laggui
v0.11.0-pre.2预发布2026-08-10
What's Changed
- Feat/autotune observability (#1437) @SamuelBelanger
- CubeCL Environment: multiplatform HPC runtime API (#1435) @nathanielsimard
- Fix initialize memory failure with error handling + fallback (#1454) @nathanielsimard
- Invalidate the kernel cache when a kernel changes (#1455) @nathanielsimard
- fix(quant): map QuantParam::UE4M3 to FloatKind::E4M3 (#1450) @ThierryCantin-Demers
- Feat/dry run environment (#1457) @nathanielsimard
- fix(runtime): check PendingDropQueue during gpu writes to prevent u32 overflow (#1456) @for-lack-of-a-better-name-j
- Adaptive autotune scheduler with early elimination (#1449) @SamuelBelanger @nathanielsimard
- feat(runtime): DeviceProperties::identity (#1465) @nathanielsimard
- feat(tune): roofline bounds from a Work amount and per-resource thresholds (#1462) @SamuelBelanger
- fix(float): correct the e4m3 and e5m2 limits and exponent constants (#1472) @nathanielsimard
- feature(cubecl-cuda): add cuda graph capture and replay (#1469) @michaelvsinko
- fix(cuda): reclaim and retry a failed device reserve (#1468) @laurigates
- fix(common): give device runner threads a shutdown path (#1470) @nathanielsimard
- fix(runtime): classify driver out-of-memory apart from BufferTooBig (#1474) @ThierryCantin-Demers
- Add a two-level quantization level (#1453) @nathanielsimard
- fix(common): call ceil through numtraits so roundup builds on no_std (#1475) @ThierryCantin-Demers
- fix(cpp): materialize constant bitcasts as lvalues (#1477) @jcwal1516
- fix(common): release the caller's borrows before waking it in run_scoped (#1478) @ThierryCantin-Demers
- fix(spirv): name the FPEncoding variants added in tracel-rspirv 0.13.3 (#1484) @louisfd
- fix(cubecl-hip): implement sync_warp instead of emitting #error (#1483) @nathanielsimard
- feat(common): add ComptimeFloat and Ratio for hashable comptime kernel params (#1482) @SamuelBelanger
- chore: version Cargo.lock (#1485) @ThierryCantin-Demers
- chore: bump version to 0.11.0-pre.2 (#1489) @laggui
- fix(publish): add missing cubecl-environment (#1490) @laggui
v0.11.0-pre.1预发布2026-07-29
What's Changed
- Add tiled layout (#1329) @Sublime12
- Bump version to 0.11.0-pre.1 (#1326) @laggui
- feat: streaming: add stream priority hint, wire into CUDA backend (#1324) @lilith
- perf: lighter cpu runtime (#1330) @marcantoinem
- feat: Add support for VKEXTshaderlongvector (#1188) @wingertge
- mega-refactor: Totally change the frontend to enable references, among other things (#1322) @wingertge
- Add f64 support back for CUDA. (#1321) @vaijira
- fix: Ensure constant block args are generated in the correct block (#1332) @wingertge
- Add workgroupUniformLoad primitive (#1327) @ArthurBrussee
- fix: shared memories allocation calculation were wrong (#1337) @marcantoinem
- make renderdoc optional (does not compile on mac os) (#1339) @louisfd
- Metal: fix atomic syntax (#1340) @louisfd
- Metal: missing space caused compilation errors (#1342) @louisfd
- Feat/obfuscation (#1341) @nathanielsimard
- fix metal compilation (#1343) @louisfd
- fix: Fix more compilation issues on metal (#1347) @wingertge
- feat(runtime): add memoryusagetotal aggregating all streams (#1333) @ArthurBrussee
- Refactor/wgpu compilers (#1346) @nathanielsimard
- Improve CPU backend by removing need to flush, add atomics test and add atomic to cubecl-cpu (#1345) @marcantoinem
- fix: CubeOr expand methods computed AND instead of OR (#1350) @LucaCappelletti94
- fix: short-circuit || and && in #[cube] code (#1348) @LucaCappelletti94
- refactor: Add a lifetime to views (#1344) @wingertge
- fix: Assign/init_mut (#1355) @wingertge
- fix: cubecl-cpu synchronisation problem (#1356) @marcantoinem
- fix: skip kernel launch when cube count is zero (#1349) @LucaCappelletti94
- fix: Fix atomics on CUDA again (#1354) @wingertge
- Refactor/readme (#1358) @nathanielsimard
- fix: re-export wgpu::Backend from cubecl-wgpu (#1357) @zhan-wei-919
- refactor(runtime): aggregate memoryusage + memorycleanup in client (#1360) @ArthurBrussee
- Tiling as a composable view layout (#1362) @louisfd
- fix(wgpu): bound live timestamp query sets per device (Metal counter-sample-buffer exhaustion) (#1361) @AdrianEddy
- fix shared memory bytes (#1367) @louisfd
- fix: reject unsupported kernel argument types instead of panicking (#1373) @LucaCappelletti94
- fix: don't panic when calling a non-path expression in a kernel (#1372) @LucaCappelletti94
- fixed typo check and bitwise test errors (#1369) @Andy2887
- fix: don't panic when const-folding an unfoldable const expression (#1374) @LucaCappelletti94
- book: constants (plane) (#1320) @Redhawk18
- feat: add expm1 lowering across backends (#1301) @shinaoka
- refactor: Simplify Variable to align it with existing IRs (#1378) @wingertge
- Feat/bytes improvements (#1379) @nathanielsimard
- Fix imports (#1383) @nathanielsimard
- Fix handling one tuple (#1382) @akiradeveloper
- fixes return Self::Scalar in implunaryfuncscalarout (Fixes #1283) (#1338) @ethqnol
- Feat/fix abusive allocation in CPU backend (#1385) @marcantoinem
- feat: Coordinates CoordsDynI type (#1386) @SamuelBelanger
- Add read_lazy method for non-WASM targets in ComputeClient (#1392) @jwric
- add as read to view mut (#1393) @louisfd
- Cpu dump improvement (#1394) @marcantoinem
- add citation (#1335) @Redhawk18
- burn element trait related changes for e4m3, e5m2 (#1389) @skewballfox
- Fix: returning panic payload in device channel tasks to caller and re-raising panic for blocking APIs (#1376) @Andy2887
- feat: metal backend (#1175) @dcvz
- fix(metal): attribute launch errors to the issuing stream (#1403) @dcvz
- fix(metal): validate shared memory limit before pipeline creation (#1405) @dcvz
- fixed cargo xtask validate fails on Linux (#1397) @Andy2887
- refactor: replace dirs with etcetera (drops MPL-2.0 option-ext) (#1395) @shimwell
- Feat/new CPU runtime (#1400) @marcantoinem
- Peak device throughput (#1408) @ThierryCantin-Demers
- disaggregate-oob-consts (#1409) @nathanielsimard
- Fix/cpu index unit (#1411) @louisfd
- fix(cpu): reserve shared memory from a dedicated pool (#1412) @louisfd
- Feat/dynamic memory pool config (#1417) @nathanielsimard
- Feat/hip captured graph (#1415) @nathanielsimard
- update wgpu version (#1416) @Charles23R
- Fix/graph safety (#1419) @nathanielsimard
- Disable persistent tune cache option (#1423) @SamuelBelanger
- Feat/autotune throughput (#1422) @SamuelBelanger
- fix(cpu): parallelize sync_cube and clamp test cube dims to core count (#1424) @louisfd
- Cpu: Allow more threads than cores (#1426) @louisfd
- Wmma compile error (#1427) @nathanielsimard
- add cross-stream input bindings pinning wgpu (#1434) @Charles23R
- fix atomic import (#1441) @Charles23R
- Allow nested tuple destructure (#1420) @akiradeveloper
- feat: add AsIndex fallible dimension index conversion (#1445) @mattisonchao
- chore: use tracel-llvm version 22.1.4-5 (#1447) @laggui
- fix: add missing cubecl-metal publish (#1448) @laggui