v0.1.15 #3311

LeiWang1999 · 2026-09-30T00:29:56Z

LeiWang1999
Sep 30, 2026
Maintainer

Highlights

Native Huawei Ascend 950 support ([Public Release 9/30] Introduce Ascend 950 backend #3308): an end-to-end NPU backend with native code generation, automatic Cube/Vector scheduling and synchronization, and mixed SIMD/SIMT programming.
Automatic CUDA warp specialization ([Feature] [CUDA] Role-based automatic warp specialization #3059, [Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization #3185): an opt-in role-based scheduler that assigns TMA loads, MMA computation, TMA stores, and worker operations to specialized warp groups.
Unified block-scaled GEMM ([Do not review][Op][Language][CUDA] Add common block-scaled GEMM semantics and backend dispatch #3237, [CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids #3257, [CUDA] Reject conflicting SM120 scale fragment layouts #3284): common T.gemm_blockscaled semantics with dedicated backend dispatch, improved SM100 instruction selection, and expanded SM120 fragment support.
More expressive Python frontend ([Frontend] Support Python iterables and comprehensions #3230): compile-time iteration over Python iterables, enumerate, zip, comprehensions, and generator expressions.

Ascend 950

Add tilelang.ascend.language and target="ascend" for Huawei Ascend 950 (dav-3510).
Combine Cube GEMM and Vector computation in one kernel, with T.SimdVF and T.SimtVF regions.
Support explicit UB/L1/L0 storage, tiled copies, cross-core transfers, and MXFP8/MXFP4 block-scaled GEMM.
Add automatic scheduling, pipelining, multi-buffering, layout inference, and synchronization insertion.
Integrate Bisheng compilation, tvm_ffi and Cython execution, PyTorch NPU tensors and streams, and NPU profiling.
Include GEMM, DeepGEMM-style kernels, FlashAttention forward/backward, RMSNorm, and FP8 quantization examples.

See the Ascend 950 guide for installation and usage. Ascend A2/A3 support remains in the community-maintained TileLang-Ascend projects.

CUDA

Enable role-based automatic warp specialization with TL_ENABLE_AUTO_WARP_SPECIALIZATION: "role_based"; add GEMM and FlashAttention examples ([Feature] [CUDA] Role-based automatic warp specialization #3059, [Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization #3185).
Extend SM120 block-scaled GEMM to fragment-resident A operands, row-major scale fragments, and odd per-warp atom grids; diagnose conflicting scale-fragment layouts ([CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids #3257, [CUDA] Reject conflicting SM120 scale fragment layouts #3284).
Add round-to-nearest T.fma and T.fmul, and complete additional FP16/BF16 math bridges ([Language][CUDA] Add T.fma and T.fmul round-to-nearest intrinsics #3134, [CUDA] Complete 16-bit bridges for CUDA-lowered unary math #3132, [CUDA] Add 16-bit hpow and hfmod bridges #3163).
Preserve packed FP8 vector copies and fuse exact FP4-to-FP8 conversions through FP32 ([CUDA] Keep FP8 vector copies packed #3276, [CUDA] Fuse exact FP4 to FP8 conversion through FP32 #3204).
Fix WGMMA/UMMA K-panel strides for operand layouts and insert async-proxy fences for sparse MMA ([BugFix][Hopper][Blackwell] Read the WGMMA/UMMA K-panel stride from the layout instead of the operand extent #2965, [CUDA] Add async-proxy fences for sparse MMA intrinsics #3139).
Correct atomic vectorization for invariant or non-contiguous destinations; keep shared FP32 atomics scalar on SM90 ([BugFix][Transform] Keep shared FP32 atomics scalar on SM90 #3129, [BugFix][Vectorize] Keep atomic_add scalar for invariant/non-contiguous destinations #3219, [Transform][CUDA] Plan atomic vector widths from destination addresses #3238).
Fix NVRTC warp reductions, kernel-body assertions, boolean bitwise negation, and non-constant int4/uint4 broadcasts; restrict 256-bit global loads/stores to SM100+ ([NVRTC] Fix warp reduction compilation #3260, [BugFix][CUDA] Lower a kernel-body assert to a device-legal check #3206, [CUDA] Fix boolean bitwise negation codegen #3228, [BugFix][CUDA] Support non-constant int4/uint4 broadcast #3114, [BugFix][CUDA] Only emit 256-bit global load/store on SM100+ targets #3248).
Add an FP8 sparse MLA forward example for DeepSeek V3.2 on Hopper ([Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper #3224).

Language and Compiler

Make kernel launch encoding target-neutral, with launch options and operation hints owned by their backend dialects ([Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations #3186, [Refactor][Language] Move backend-specific op hints into their owning dialects #3203).
Support Python compile-time iteration and comprehensions while preserving device-loop semantics for for ... in range(...) ([Frontend] Support Python iterables and comprehensions #3230).
Support unary plus on symbolic expressions, fix loop-variable binding, remove warnings for immutable rebinding, and reject unsupported loop else clauses ([Bugfix][Language] Support unary plus on PrimExpr in the eager frontend #3141, [BugFix] Bind loop targets independently of mutable scalar variables #3232, [Frontend] Remove warnings for immutable variable rebinding #3262, [Bugfix][Language] Reject loop else clauses in the eager frontend #3142).
Treat T.copy / T.async_copy coalesced_width as a hint and clamp it to the achievable vector width ([Fix]Clamp T.copy/T.async_copy coalesced_width to achievable vector size instead of LOG(FATAL) #3246).
Preserve dynamic reduction tail guards, fix parallel-loop lowering with let inlining disabled, and restore async-copy lowering with partitioned layouts ([Fix][TVM] Update TVM to preserve dynamic reduction tail guards #3294, [BugFix] Fix parallel loop lowering with let inlining disabled #3269, [Transform][CUDA] Fix async copy lowering with partitioned layouts #3278).
Improve reducer layout planning and symbolic layout validation; reject unsupported AllReduce thread strides and reduction NaN-propagation dtypes ([Layout] Consider scalar reducer plans in register-count search #3171, [BugFix] Restore symbolic loop-layout injectivity proof; reject layouts on symbolic shared tiles #3233, [BugFix] Reject non-power-of-two AllReduce thread strides #3266, [Fix] Reject unsupported reduction NaN propagation dtypes #3273).
Fix NaN-propagating clamp lowering and eliminate unused bindings during simplification ([BugFix][Language] Lower NaN-propagating clamp through device templates #3205, [BugFix] Eliminate unused bindings in Simplify #3293).

ROCm, CPU, and Metal

ROCm: support the DeepSeek V3.2 Top-K selector, wave64 Hadamard transforms, and GLM-5.3 k-pool examples, including Top-K index transformation ([ROCm] Support DeepSeek-V3.2 Top-K selector #3147, [ROCm] Support wave64 Hadamard transforms #3154, [ROCm] Add GLM-5.3 k-pool Top-K transform #3254).
Expand ROCm validation for attention kernels and portable examples, and document pip installation ([ROCm] Validate block-causal attention kernels #3146, [ROCm] Validate attention sink kernels #3148, [ROCm] Run portable example validation in CI #3165, [Doc] Document pip installation on AMD GPUs (ROCm) #3168).
CPU: compute scalar GEMM products in the accumulator dtype and legalize BF16 arithmetic ([BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype #2917, [BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline #3201).
Metal: support 32-bit integer atomic add and respect GEMM buffer-region offsets ([Metal] Support 32-bit integer atomic add #3211, [BugFix][Metal] Respect GEMM buffer region offsets #3209).

Runtime, Build, and Tooling

Store cached kernel parameters as JSON instead of cloudpickle ([Cache] Store cached kernel params as JSON instead of cloudpickle #3143).
Publish CUDA binaries and metadata atomically in immutable cache directories, preventing readers from observing partially published entries ([CUDA][Cache] Publish immutable binary cache directories #3177).
Fix dynamic-output allocation when the sizing input follows the output, and add missing uint64 argument mappings ([BugFix][JIT] Allocate a dynamic-shape output that precedes its sizing input #3207, [JIT] Add missing uint64 argument type mappings #3229).
Add wall-clock benchmarking and MPS-compatible timing helpers ([Profiler] Split timing helpers and add wall-clock benchmarking #3234).
Improve CUDA/ROCm compiler discovery and library loading for symlinked installations ([CUDA] Resolve symlinked nvcc before deriving CUDA_HOME #2839, [BugFix][ROCm] Resolve hipcc robustly instead of trusting bare PATH lookup #3166, [Runtime] Fix library loading for symlink installs #3227).
Enable optimization for default single-config native builds and restore effective Windows wheel-build caching ([Build] Optimize default single-config builds #3191, [CI] Fix slow Windows wheel builds by making ccache effective #3133, [Build][Windows] Restore ccache hits for clang-cl wheel builds #3305).
Validate built wheels on GPU runners and make performance regressions fail CI ([CI] Validate built wheels on GPU runners in the Dist workflow #3167, [CI] Fail the perf bot on a regression instead of only reporting it #3182).

Compatibility Notes

Remove the unused sync and group parameters from T.Pipelined, and k_pack from T.gemm_sp ([Cleanup][Language] Remove dead T.Pipelined(sync/group) and T.gemm_sp(k_pack) parameters #3202).
ROCm kernels using T.gemm(k_pack=...) should import tilelang.rocm.language ([Refactor][Language] Move backend-specific op hints into their owning dialects #3203).
Kernels querying thread extents during tracing must specify threads= explicitly in T.Kernel ([Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations #3186).
T.symbolic remains available as a deprecated alias; use T.dynamic for new code ([Language] Restore deprecated T.symbolic alias #3216).
Remove TILELANG_CACHE_VERIFY_HASH; binary artifact hash verification is now mandatory. Legacy cache formats are rebuilt automatically ([Cache] Store cached kernel params as JSON instead of cloudpickle #3143, [CUDA][Cache] Publish immutable binary cache directories #3177).

Full Changelog: v0.1.14...v0.1.15

What's Changed

[CI] Cap torch<2.14 for CUDA tests until flash-attn supports torch 2.14 by @LeiWang1999 in [CI] Cap torch<2.14 for CUDA tests until flash-attn supports torch 2.14 #3136
Bump transformers from 5.5.0 to 5.10.1 in /examples/bitnet-1.58b by @dependabot[bot] in Bump transformers from 5.5.0 to 5.10.1 in /examples/bitnet-1.58b #3131
[CI] Fix slow Windows wheel builds by making ccache effective by @LeiWang1999 in [CI] Fix slow Windows wheel builds by making ccache effective #3133
[Language][CUDA] Add T.fma and T.fmul round-to-nearest intrinsics by @LeiWang1999 in [Language][CUDA] Add T.fma and T.fmul round-to-nearest intrinsics #3134
[CUDA] Complete 16-bit bridges for CUDA-lowered unary math by @Chennesxu in [CUDA] Complete 16-bit bridges for CUDA-lowered unary math #3132
[Refactor] Restore LoopUnswitching test altered by [BugFix] Always vectorize T.Parallel loops #3121 by @Yongqi-Zhuo in [Refactor] Restore LoopUnswitching test altered by #3121 #3124
[BugFix][CUDA] Support non-constant int4/uint4 broadcast by @jjppp in [BugFix][CUDA] Support non-constant int4/uint4 broadcast #3114
[BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype by @Dino1844 in [BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype #2917
[BugFix][Hopper][Blackwell] Read the WGMMA/UMMA K-panel stride from the layout instead of the operand extent by @bigSheep123 in [BugFix][Hopper][Blackwell] Read the WGMMA/UMMA K-panel stride from the layout instead of the operand extent #2965
[BugFix][Carver] Gate _legalize_info fallback on sm_version by @mocusez in [BugFix][Carver] Gate _legalize_info fallback on sm_version #3145
[ROCm] Validate block-causal attention kernels by @andyluo7 in [ROCm] Validate block-causal attention kernels #3146
[ROCm] Validate attention sink kernels by @andyluo7 in [ROCm] Validate attention sink kernels #3148
[ROCm] Support DeepSeek-V3.2 Top-K selector by @andyluo7 in [ROCm] Support DeepSeek-V3.2 Top-K selector #3147
[Bugfix][Language] Reject loop else clauses in the eager frontend by @rishabhsinha17 in [Bugfix][Language] Reject loop else clauses in the eager frontend #3142
[Example] Fix hadamard reference precision on TF32 and add pytest entry for CI by @Guan-jeans in [Example] Fix hadamard reference precision on TF32 and add pytest entry for CI #3126
[BugFix][Carver] Compare sm_version numerically in plan_rasterization by @mocusez in [BugFix][Carver] Compare sm_version numerically in plan_rasterization #3144
[Misc] Expose .agents skills to Claude Code via .claude/skills symlinks by @LeiWang1999 in [Misc] Expose .agents skills to Claude Code via .claude/skills symlinks #3161
[Language] Cache dtype.as_torch and demote storage-dtype fallback logs to debug by @LeiWang1999 in [Language] Cache dtype.as_torch and demote storage-dtype fallback logs to debug #3160
[Refactor][CUDA] Replace per-lane-count vector constructors with variadic packers by @LeiWang1999 in [Refactor][CUDA] Replace per-lane-count vector constructors with variadic packers #3158
[CUDA] Add 16-bit hpow and hfmod bridges by @Chennesxu in [CUDA] Add 16-bit hpow and hfmod bridges #3163
[BugFix][ROCm] Resolve hipcc robustly instead of trusting bare PATH lookup by @LeiWang1999 in [BugFix][ROCm] Resolve hipcc robustly instead of trusting bare PATH lookup #3166
[CI] Validate built wheels on GPU runners in the Dist workflow by @LeiWang1999 in [CI] Validate built wheels on GPU runners in the Dist workflow #3167
[Doc] Document pip installation on AMD GPUs (ROCm) by @LeiWang1999 in [Doc] Document pip installation on AMD GPUs (ROCm) #3168
[Bugfix][Language] Support unary plus on PrimExpr in the eager frontend by @rishabhsinha17 in [Bugfix][Language] Support unary plus on PrimExpr in the eager frontend #3141
[CUDA] Add async-proxy fences for sparse MMA intrinsics by @ZenAlexa in [CUDA] Add async-proxy fences for sparse MMA intrinsics #3139
[Feature] [CUDA] Role-based automatic warp specialization by @Yongqi-Zhuo in [Feature] [CUDA] Role-based automatic warp specialization #3059
[CUDA] Resolve symlinked nvcc before deriving CUDA_HOME by @morluto in [CUDA] Resolve symlinked nvcc before deriving CUDA_HOME #2839
[Language] Remove deprecated T.symbolic alias by @SiriusNEO in [Language] Remove deprecated T.symbolic alias #3122
[ROCm] Support wave64 Hadamard transforms by @andyluo7 in [ROCm] Support wave64 Hadamard transforms #3154
[Cache] Store cached kernel params as JSON instead of cloudpickle by @rishabhsinha17 in [Cache] Store cached kernel params as JSON instead of cloudpickle #3143
[BugFix][Transform] Keep shared FP32 atomics scalar on SM90 by @ZenAlexa in [BugFix][Transform] Keep shared FP32 atomics scalar on SM90 #3129
[Layout] Consider scalar reducer plans in register-count search by @LeiWang1999 in [Layout] Consider scalar reducer plans in register-count search #3171
[Transform] Add snapshot mode to FreshenMutableReads by @LeiWang1999 in [Transform] Add snapshot mode to FreshenMutableReads #3174
[CUDA][Cache] Publish immutable binary cache directories by @LeiWang1999 in [CUDA][Cache] Publish immutable binary cache directories #3177
[CI] [pre-commit.ci] autoupdate by @pre-commit-ci[bot] in [CI] [pre-commit.ci] autoupdate #3181
[CI] Fail the perf bot on a regression instead of only reporting it by @cklxx in [CI] Fail the perf bot on a regression instead of only reporting it #3182
[Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization by @Yongqi-Zhuo in [Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization #3185
[Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations by @LeiWang1999 in [Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations #3186
[Cleanup][Language] Remove dead T.Pipelined(sync/group) and T.gemm_sp(k_pack) parameters by @LeiWang1999 in [Cleanup][Language] Remove dead T.Pipelined(sync/group) and T.gemm_sp(k_pack) parameters #3202
[Refactor][Language] Move backend-specific op hints into their owning dialects by @LeiWang1999 in [Refactor][Language] Move backend-specific op hints into their owning dialects #3203
[BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline by @anerli in [BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline #3201
[Language] Restore deprecated T.symbolic alias by @LeiWang1999 in [Language] Restore deprecated T.symbolic alias #3216
[Refactor][JIT] Replace callee-allocated-output target sniffing with a backend capability flag by @LeiWang1999 in [Refactor][JIT] Replace callee-allocated-output target sniffing with a backend capability flag #3218
[Transform] Restore bind-before-use order when merging constraint sets by @LeiWang1999 in [Transform] Restore bind-before-use order when merging constraint sets #3221
[BugFix][Metal] Respect GEMM buffer region offsets by @anerli in [BugFix][Metal] Respect GEMM buffer region offsets #3209
[Runtime] Fix library loading for symlink installs by @sepcnt in [Runtime] Fix library loading for symlink installs #3227
[CUDA] Fix boolean bitwise negation codegen by @sepcnt in [CUDA] Fix boolean bitwise negation codegen #3228
[Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper by @xuebozhang525-alt in [Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper #3224
[Profiler] Split timing helpers and add wall-clock benchmarking by @SiriusNEO in [Profiler] Split timing helpers and add wall-clock benchmarking #3234
[CPU] Import the CPU dialect in CPU tests by @penguin-wwy in [CPU] Import the CPU dialect in CPU tests #3236
[BugFix][Vectorize] Keep atomic_add scalar for invariant/non-contiguous destinations by @Dino1844 in [BugFix][Vectorize] Keep atomic_add scalar for invariant/non-contiguous destinations #3219
[Transform][CUDA] Plan atomic vector widths from destination addresses by @LeiWang1999 in [Transform][CUDA] Plan atomic vector widths from destination addresses #3238
[Do not review][Op][Language][CUDA] Add common block-scaled GEMM semantics and backend dispatch by @LeiWang1999 in [Do not review][Op][Language][CUDA] Add common block-scaled GEMM semantics and backend dispatch #3237
[JIT] Add missing uint64 argument type mappings by @sepcnt in [JIT] Add missing uint64 argument type mappings #3229
[BugFix] Restore symbolic loop-layout injectivity proof; reject layouts on symbolic shared tiles by @sepcnt in [BugFix] Restore symbolic loop-layout injectivity proof; reject layouts on symbolic shared tiles #3233
[ROCm] Run portable example validation in CI by @andyluo7 in [ROCm] Run portable example validation in CI #3165
[BugFix][CUDA] Only emit 256-bit global load/store on SM100+ targets by @penguin-wwy in [BugFix][CUDA] Only emit 256-bit global load/store on SM100+ targets #3248
[Metal] Support 32-bit integer atomic add by @anerli in [Metal] Support 32-bit integer atomic add #3211
[Testing] Drop duplicated codegen-smoke tests; fix fastmath self-comparison assertions by @penguin-wwy in [Testing] Drop duplicated codegen-smoke tests; fix fastmath self-comparison assertions #3252
[ROCm] Add GLM-5.3 k-pool Top-K transform by @andyluo7 in [ROCm] Add GLM-5.3 k-pool Top-K transform #3254
[Frontend] Support Python iterables and comprehensions by @sepcnt in [Frontend] Support Python iterables and comprehensions #3230
[Frontend] Remove warnings for immutable variable rebinding by @LJC00118 in [Frontend] Remove warnings for immutable variable rebinding #3262
[BugFix] Reject non-power-of-two AllReduce thread strides by @LeiWang1999 in [BugFix] Reject non-power-of-two AllReduce thread strides #3266
[NVRTC] Fix warp reduction compilation by @sepcnt in [NVRTC] Fix warp reduction compilation #3260
[BugFix][CUDA] Lower a kernel-body assert to a device-legal check by @xy200303 in [BugFix][CUDA] Lower a kernel-body assert to a device-legal check #3206
[BugFix][JIT] Allocate a dynamic-shape output that precedes its sizing input by @xy200303 in [BugFix][JIT] Allocate a dynamic-shape output that precedes its sizing input #3207
[CUDA] Fuse exact FP4 to FP8 conversion through FP32 by @ZenAlexa in [CUDA] Fuse exact FP4 to FP8 conversion through FP32 #3204
[Fix]Clamp T.copy/T.async_copy coalesced_width to achievable vector size instead of LOG(FATAL) by @edragain2nd in [Fix]Clamp T.copy/T.async_copy coalesced_width to achievable vector size instead of LOG(FATAL) #3246
[BugFix] Bind loop targets independently of mutable scalar variables by @sepcnt in [BugFix] Bind loop targets independently of mutable scalar variables #3232
[BugFix][Language] Lower NaN-propagating clamp through device templates by @xy200303 in [BugFix][Language] Lower NaN-propagating clamp through device templates #3205
[Fix] Reject unsupported reduction NaN propagation dtypes by @ZenAlexa in [Fix] Reject unsupported reduction NaN propagation dtypes #3273
[BugFix] Fix parallel loop lowering with let inlining disabled by @penguin-wwy in [BugFix] Fix parallel loop lowering with let inlining disabled #3269
[Transform][CUDA] Fix async copy lowering with partitioned layouts by @LeiWang1999 in [Transform][CUDA] Fix async copy lowering with partitioned layouts #3278
[CUDA] Keep FP8 vector copies packed by @LeiWang1999 in [CUDA] Keep FP8 vector copies packed #3276
[CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids by @sepcnt in [CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids #3257
[CUDA] Reject conflicting SM120 scale fragment layouts by @LeiWang1999 in [CUDA] Reject conflicting SM120 scale fragment layouts #3284
[Build] Optimize default single-config builds by @KellyFrog in [Build] Optimize default single-config builds #3191
[Fix][TVM] Update TVM to preserve dynamic reduction tail guards by @LeiWang1999 in [Fix][TVM] Update TVM to preserve dynamic reduction tail guards #3294
[BugFix] Eliminate unused bindings in Simplify by @LJC00118 in [BugFix] Eliminate unused bindings in Simplify #3293
[Build][Windows] Restore ccache hits for clang-cl wheel builds by @LeiWang1999 in [Build][Windows] Restore ccache hits for clang-cl wheel builds #3305
[Public Release 9/30] Introduce Ascend 950 backend by @SiriusNEO in [Public Release 9/30] Introduce Ascend 950 backend #3308
[Release] Bump version into 0.1.15 by @LeiWang1999 in [Release] Bump version into 0.1.15 #3309

New Contributors

@Dino1844 made their first contribution in [BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype #2917
@mocusez made their first contribution in [BugFix][Carver] Gate _legalize_info fallback on sm_version #3145
@rishabhsinha17 made their first contribution in [Bugfix][Language] Reject loop else clauses in the eager frontend #3142
@Guan-jeans made their first contribution in [Example] Fix hadamard reference precision on TF32 and add pytest entry for CI #3126
@anerli made their first contribution in [BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline #3201
@xuebozhang525-alt made their first contribution in [Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper #3224
@xy200303 made their first contribution in [BugFix][CUDA] Lower a kernel-body assert to a device-legal check #3206

Full Changelog: v0.1.14...v0.1.15

This discussion was created from the release v0.1.15.

melitahambasic43-afk · 2026-10-02T09:36:41Z

melitahambasic43-afk
Oct 2, 2026

Highlights

Native Huawei Ascend 950 support ([Public Release 9/30] Introduce Ascend 950 backend #3308): an end-to-end NPU backend with native code generation, automatic Cube/Vector scheduling and synchronization, and mixed SIMD/SIMT programming.

Automatic CUDA warp specialization ([Feature] [CUDA] Role-based automatic warp specialization #3059, [Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization #3185): an opt-in role-based scheduler that assigns TMA loads, MMA computation, TMA stores, and worker operations to specialized warp groups.

Unified block-scaled GEMM ([Do not review][Op][Language][CUDA] Add common block-scaled GEMM semantics and backend dispatch #3237, [CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids #3257, [CUDA] Reject conflicting SM120 scale fragment layouts #3284): common T.gemm_blockscaled semantics with dedicated backend dispatch, improved SM100 instruction selection, and expanded SM120 fragment support.

More expressive Python frontend ([Frontend] Support Python iterables and comprehensions #3230): compile-time iteration over Python iterables, enumerate, zip, comprehensions, and generator expressions.

Ascend 950

Add tilelang.ascend.language and target="ascend" for Huawei Ascend 950 (dav-3510).

Combine Cube GEMM and Vector computation in one kernel, with T.SimdVF and T.SimtVF regions.

Support explicit UB/L1/L0 storage, tiled copies, cross-core transfers, and MXFP8/MXFP4 block-scaled GEMM.

Add automatic scheduling, pipelining, multi-buffering, layout inference, and synchronization insertion.

Integrate Bisheng compilation, tvm_ffi and Cython execution, PyTorch NPU tensors and streams, and NPU profiling.

Include GEMM, DeepGEMM-style kernels, FlashAttention forward/backward, RMSNorm, and FP8 quantization examples.

See the Ascend 950 guide for installation and usage. Ascend A2/A3 support remains in the community-maintained TileLang-Ascend projects.

CUDA

Enable role-based automatic warp specialization with TL_ENABLE_AUTO_WARP_SPECIALIZATION: "role_based"; add GEMM and FlashAttention examples ([Feature] [CUDA] Role-based automatic warp specialization #3059, [Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization #3185).

Extend SM120 block-scaled GEMM to fragment-resident A operands, row-major scale fragments, and odd per-warp atom grids; diagnose conflicting scale-fragment layouts ([CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids #3257, [CUDA] Reject conflicting SM120 scale fragment layouts #3284).

Add round-to-nearest T.fma and T.fmul, and complete additional FP16/BF16 math bridges ([Language][CUDA] Add T.fma and T.fmul round-to-nearest intrinsics #3134, [CUDA] Complete 16-bit bridges for CUDA-lowered unary math #3132, [CUDA] Add 16-bit hpow and hfmod bridges #3163).

Preserve packed FP8 vector copies and fuse exact FP4-to-FP8 conversions through FP32 ([CUDA] Keep FP8 vector copies packed #3276, [CUDA] Fuse exact FP4 to FP8 conversion through FP32 #3204).

Fix WGMMA/UMMA K-panel strides for operand layouts and insert async-proxy fences for sparse MMA ([BugFix][Hopper][Blackwell] Read the WGMMA/UMMA K-panel stride from the layout instead of the operand extent #2965, [CUDA] Add async-proxy fences for sparse MMA intrinsics #3139).

Correct atomic vectorization for invariant or non-contiguous destinations; keep shared FP32 atomics scalar on SM90 ([BugFix][Transform] Keep shared FP32 atomics scalar on SM90 #3129, [BugFix][Vectorize] Keep atomic_add scalar for invariant/non-contiguous destinations #3219, [Transform][CUDA] Plan atomic vector widths from destination addresses #3238).

Fix NVRTC warp reductions, kernel-body assertions, boolean bitwise negation, and non-constant int4/uint4 broadcasts; restrict 256-bit global loads/stores to SM100+ ([NVRTC] Fix warp reduction compilation #3260, [BugFix][CUDA] Lower a kernel-body assert to a device-legal check #3206, [CUDA] Fix boolean bitwise negation codegen #3228, [BugFix][CUDA] Support non-constant int4/uint4 broadcast #3114, [BugFix][CUDA] Only emit 256-bit global load/store on SM100+ targets #3248).

Add an FP8 sparse MLA forward example for DeepSeek V3.2 on Hopper ([Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper #3224).

Language and Compiler

Make kernel launch encoding target-neutral, with launch options and operation hints owned by their backend dialects ([Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations #3186, [Refactor][Language] Move backend-specific op hints into their owning dialects #3203).

Support Python compile-time iteration and comprehensions while preserving device-loop semantics for for ... in range(...) ([Frontend] Support Python iterables and comprehensions #3230).

Support unary plus on symbolic expressions, fix loop-variable binding, remove warnings for immutable rebinding, and reject unsupported loop else clauses ([Bugfix][Language] Support unary plus on PrimExpr in the eager frontend #3141, [BugFix] Bind loop targets independently of mutable scalar variables #3232, [Frontend] Remove warnings for immutable variable rebinding #3262, [Bugfix][Language] Reject loop else clauses in the eager frontend #3142).

Treat T.copy / T.async_copy coalesced_width as a hint and clamp it to the achievable vector width ([Fix]Clamp T.copy/T.async_copy coalesced_width to achievable vector size instead of LOG(FATAL) #3246).

Preserve dynamic reduction tail guards, fix parallel-loop lowering with let inlining disabled, and restore async-copy lowering with partitioned layouts ([Fix][TVM] Update TVM to preserve dynamic reduction tail guards #3294, [BugFix] Fix parallel loop lowering with let inlining disabled #3269, [Transform][CUDA] Fix async copy lowering with partitioned layouts #3278).

Improve reducer layout planning and symbolic layout validation; reject unsupported AllReduce thread strides and reduction NaN-propagation dtypes ([Layout] Consider scalar reducer plans in register-count search #3171, [BugFix] Restore symbolic loop-layout injectivity proof; reject layouts on symbolic shared tiles #3233, [BugFix] Reject non-power-of-two AllReduce thread strides #3266, [Fix] Reject unsupported reduction NaN propagation dtypes #3273).

Fix NaN-propagating clamp lowering and eliminate unused bindings during simplification ([BugFix][Language] Lower NaN-propagating clamp through device templates #3205, [BugFix] Eliminate unused bindings in Simplify #3293).

ROCm, CPU, and Metal

ROCm: support the DeepSeek V3.2 Top-K selector, wave64 Hadamard transforms, and GLM-5.3 k-pool examples, including Top-K index transformation ([ROCm] Support DeepSeek-V3.2 Top-K selector #3147, [ROCm] Support wave64 Hadamard transforms #3154, [ROCm] Add GLM-5.3 k-pool Top-K transform #3254).

Expand ROCm validation for attention kernels and portable examples, and document pip installation ([ROCm] Validate block-causal attention kernels #3146, [ROCm] Validate attention sink kernels #3148, [ROCm] Run portable example validation in CI #3165, [Doc] Document pip installation on AMD GPUs (ROCm) #3168).

CPU: compute scalar GEMM products in the accumulator dtype and legalize BF16 arithmetic ([BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype #2917, [BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline #3201).

Metal: support 32-bit integer atomic add and respect GEMM buffer-region offsets ([Metal] Support 32-bit integer atomic add #3211, [BugFix][Metal] Respect GEMM buffer region offsets #3209).

Runtime, Build, and Tooling

Store cached kernel parameters as JSON instead of cloudpickle ([Cache] Store cached kernel params as JSON instead of cloudpickle #3143).

Publish CUDA binaries and metadata atomically in immutable cache directories, preventing readers from observing partially published entries ([CUDA][Cache] Publish immutable binary cache directories #3177).

Fix dynamic-output allocation when the sizing input follows the output, and add missing uint64 argument mappings ([BugFix][JIT] Allocate a dynamic-shape output that precedes its sizing input #3207, [JIT] Add missing uint64 argument type mappings #3229).

Add wall-clock benchmarking and MPS-compatible timing helpers ([Profiler] Split timing helpers and add wall-clock benchmarking #3234).

Improve CUDA/ROCm compiler discovery and library loading for symlinked installations ([CUDA] Resolve symlinked nvcc before deriving CUDA_HOME #2839, [BugFix][ROCm] Resolve hipcc robustly instead of trusting bare PATH lookup #3166, [Runtime] Fix library loading for symlink installs #3227).

Enable optimization for default single-config native builds and restore effective Windows wheel-build caching ([Build] Optimize default single-config builds #3191, [CI] Fix slow Windows wheel builds by making ccache effective #3133, [Build][Windows] Restore ccache hits for clang-cl wheel builds #3305).

Validate built wheels on GPU runners and make performance regressions fail CI ([CI] Validate built wheels on GPU runners in the Dist workflow #3167, [CI] Fail the perf bot on a regression instead of only reporting it #3182).

Compatibility Notes

Remove the unused sync and group parameters from T.Pipelined, and k_pack from T.gemm_sp ([Cleanup][Language] Remove dead T.Pipelined(sync/group) and T.gemm_sp(k_pack) parameters #3202).

ROCm kernels using T.gemm(k_pack=...) should import tilelang.rocm.language ([Refactor][Language] Move backend-specific op hints into their owning dialects #3203).

Kernels querying thread extents during tracing must specify threads= explicitly in T.Kernel ([Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations #3186).

T.symbolic remains available as a deprecated alias; use T.dynamic for new code ([Language] Restore deprecated T.symbolic alias #3216).

Remove TILELANG_CACHE_VERIFY_HASH; binary artifact hash verification is now mandatory. Legacy cache formats are rebuilt automatically ([Cache] Store cached kernel params as JSON instead of cloudpickle #3143, [CUDA][Cache] Publish immutable binary cache directories #3177).

Full Changelog: v0.1.14...v0.1.15

What's Changed

[CI] Cap torch<2.14 for CUDA tests until flash-attn supports torch 2.14 by @LeiWang1999 in [CI] Cap torch<2.14 for CUDA tests until flash-attn supports torch 2.14 #3136

Bump transformers from 5.5.0 to 5.10.1 in /examples/bitnet-1.58b by @dependabot[bot] in Bump transformers from 5.5.0 to 5.10.1 in /examples/bitnet-1.58b #3131

[CI] Fix slow Windows wheel builds by making ccache effective by @LeiWang1999 in [CI] Fix slow Windows wheel builds by making ccache effective #3133

[Language][CUDA] Add T.fma and T.fmul round-to-nearest intrinsics by @LeiWang1999 in [Language][CUDA] Add T.fma and T.fmul round-to-nearest intrinsics #3134

[CUDA] Complete 16-bit bridges for CUDA-lowered unary math by @Chennesxu in [CUDA] Complete 16-bit bridges for CUDA-lowered unary math #3132

[Refactor] Restore LoopUnswitching test altered by [BugFix] Always vectorize T.Parallel loops #3121 by @Yongqi-Zhuo in [Refactor] Restore LoopUnswitching test altered by #3121 #3124

[BugFix][CUDA] Support non-constant int4/uint4 broadcast by @jjppp in [BugFix][CUDA] Support non-constant int4/uint4 broadcast #3114

[BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype by @Dino1844 in [BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype #2917

[BugFix][Hopper][Blackwell] Read the WGMMA/UMMA K-panel stride from the layout instead of the operand extent by @bigSheep123 in [BugFix][Hopper][Blackwell] Read the WGMMA/UMMA K-panel stride from the layout instead of the operand extent #2965

[BugFix][Carver] Gate _legalize_info fallback on sm_version by @mocusez in [BugFix][Carver] Gate _legalize_info fallback on sm_version #3145

[ROCm] Validate block-causal attention kernels by @andyluo7 in [ROCm] Validate block-causal attention kernels #3146

[ROCm] Validate attention sink kernels by @andyluo7 in [ROCm] Validate attention sink kernels #3148

[ROCm] Support DeepSeek-V3.2 Top-K selector by @andyluo7 in [ROCm] Support DeepSeek-V3.2 Top-K selector #3147

[Bugfix][Language] Reject loop else clauses in the eager frontend by @rishabhsinha17 in [Bugfix][Language] Reject loop else clauses in the eager frontend #3142

[Example] Fix hadamard reference precision on TF32 and add pytest entry for CI by @Guan-jeans in [Example] Fix hadamard reference precision on TF32 and add pytest entry for CI #3126

[BugFix][Carver] Compare sm_version numerically in plan_rasterization by @mocusez in [BugFix][Carver] Compare sm_version numerically in plan_rasterization #3144

[Misc] Expose .agents skills to Claude Code via .claude/skills symlinks by @LeiWang1999 in [Misc] Expose .agents skills to Claude Code via .claude/skills symlinks #3161

[Language] Cache dtype.as_torch and demote storage-dtype fallback logs to debug by @LeiWang1999 in [Language] Cache dtype.as_torch and demote storage-dtype fallback logs to debug #3160

[Refactor][CUDA] Replace per-lane-count vector constructors with variadic packers by @LeiWang1999 in [Refactor][CUDA] Replace per-lane-count vector constructors with variadic packers #3158

[CUDA] Add 16-bit hpow and hfmod bridges by @Chennesxu in [CUDA] Add 16-bit hpow and hfmod bridges #3163

[BugFix][ROCm] Resolve hipcc robustly instead of trusting bare PATH lookup by @LeiWang1999 in [BugFix][ROCm] Resolve hipcc robustly instead of trusting bare PATH lookup #3166

[CI] Validate built wheels on GPU runners in the Dist workflow by @LeiWang1999 in [CI] Validate built wheels on GPU runners in the Dist workflow #3167

[Doc] Document pip installation on AMD GPUs (ROCm) by @LeiWang1999 in [Doc] Document pip installation on AMD GPUs (ROCm) #3168

[Bugfix][Language] Support unary plus on PrimExpr in the eager frontend by @rishabhsinha17 in [Bugfix][Language] Support unary plus on PrimExpr in the eager frontend #3141

[CUDA] Add async-proxy fences for sparse MMA intrinsics by @ZenAlexa in [CUDA] Add async-proxy fences for sparse MMA intrinsics #3139

[Feature] [CUDA] Role-based automatic warp specialization by @Yongqi-Zhuo in [Feature] [CUDA] Role-based automatic warp specialization #3059

[CUDA] Resolve symlinked nvcc before deriving CUDA_HOME by @morluto in [CUDA] Resolve symlinked nvcc before deriving CUDA_HOME #2839

[Language] Remove deprecated T.symbolic alias by @SiriusNEO in [Language] Remove deprecated T.symbolic alias #3122

[ROCm] Support wave64 Hadamard transforms by @andyluo7 in [ROCm] Support wave64 Hadamard transforms #3154

[Cache] Store cached kernel params as JSON instead of cloudpickle by @rishabhsinha17 in [Cache] Store cached kernel params as JSON instead of cloudpickle #3143

[BugFix][Transform] Keep shared FP32 atomics scalar on SM90 by @ZenAlexa in [BugFix][Transform] Keep shared FP32 atomics scalar on SM90 #3129

[Layout] Consider scalar reducer plans in register-count search by @LeiWang1999 in [Layout] Consider scalar reducer plans in register-count search #3171

[Transform] Add snapshot mode to FreshenMutableReads by @LeiWang1999 in [Transform] Add snapshot mode to FreshenMutableReads #3174

[CUDA][Cache] Publish immutable binary cache directories by @LeiWang1999 in [CUDA][Cache] Publish immutable binary cache directories #3177

[CI] [pre-commit.ci] autoupdate by @pre-commit-ci[bot] in [CI] [pre-commit.ci] autoupdate #3181

[CI] Fail the perf bot on a regression instead of only reporting it by @cklxx in [CI] Fail the perf bot on a regression instead of only reporting it #3182

[Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization by @Yongqi-Zhuo in [Refactor] [CUDA] Rename AutoSchedule to AutoWarpSpecialization #3185

[Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations by @LeiWang1999 in [Refactor][Language] Make T.Kernel target-neutral with dialect-owned launch annotations #3186

[Cleanup][Language] Remove dead T.Pipelined(sync/group) and T.gemm_sp(k_pack) parameters by @LeiWang1999 in [Cleanup][Language] Remove dead T.Pipelined(sync/group) and T.gemm_sp(k_pack) parameters #3202

[Refactor][Language] Move backend-specific op hints into their owning dialects by @LeiWang1999 in [Refactor][Language] Move backend-specific op hints into their owning dialects #3203

[BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline by @anerli in [BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline #3201

[Language] Restore deprecated T.symbolic alias by @LeiWang1999 in [Language] Restore deprecated T.symbolic alias #3216

[Refactor][JIT] Replace callee-allocated-output target sniffing with a backend capability flag by @LeiWang1999 in [Refactor][JIT] Replace callee-allocated-output target sniffing with a backend capability flag #3218

[Transform] Restore bind-before-use order when merging constraint sets by @LeiWang1999 in [Transform] Restore bind-before-use order when merging constraint sets #3221

[BugFix][Metal] Respect GEMM buffer region offsets by @anerli in [BugFix][Metal] Respect GEMM buffer region offsets #3209

[Runtime] Fix library loading for symlink installs by @sepcnt in [Runtime] Fix library loading for symlink installs #3227

[CUDA] Fix boolean bitwise negation codegen by @sepcnt in [CUDA] Fix boolean bitwise negation codegen #3228

[Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper by @xuebozhang525-alt in [Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper #3224

[Profiler] Split timing helpers and add wall-clock benchmarking by @SiriusNEO in [Profiler] Split timing helpers and add wall-clock benchmarking #3234

[CPU] Import the CPU dialect in CPU tests by @penguin-wwy in [CPU] Import the CPU dialect in CPU tests #3236

[BugFix][Vectorize] Keep atomic_add scalar for invariant/non-contiguous destinations by @Dino1844 in [BugFix][Vectorize] Keep atomic_add scalar for invariant/non-contiguous destinations #3219

[Transform][CUDA] Plan atomic vector widths from destination addresses by @LeiWang1999 in [Transform][CUDA] Plan atomic vector widths from destination addresses #3238

[Do not review][Op][Language][CUDA] Add common block-scaled GEMM semantics and backend dispatch by @LeiWang1999 in [Do not review][Op][Language][CUDA] Add common block-scaled GEMM semantics and backend dispatch #3237

[JIT] Add missing uint64 argument type mappings by @sepcnt in [JIT] Add missing uint64 argument type mappings #3229

[BugFix] Restore symbolic loop-layout injectivity proof; reject layouts on symbolic shared tiles by @sepcnt in [BugFix] Restore symbolic loop-layout injectivity proof; reject layouts on symbolic shared tiles #3233

[ROCm] Run portable example validation in CI by @andyluo7 in [ROCm] Run portable example validation in CI #3165

[BugFix][CUDA] Only emit 256-bit global load/store on SM100+ targets by @penguin-wwy in [BugFix][CUDA] Only emit 256-bit global load/store on SM100+ targets #3248

[Metal] Support 32-bit integer atomic add by @anerli in [Metal] Support 32-bit integer atomic add #3211

[Testing] Drop duplicated codegen-smoke tests; fix fastmath self-comparison assertions by @penguin-wwy in [Testing] Drop duplicated codegen-smoke tests; fix fastmath self-comparison assertions #3252

[ROCm] Add GLM-5.3 k-pool Top-K transform by @andyluo7 in [ROCm] Add GLM-5.3 k-pool Top-K transform #3254

[Frontend] Support Python iterables and comprehensions by @sepcnt in [Frontend] Support Python iterables and comprehensions #3230

[Frontend] Remove warnings for immutable variable rebinding by @LJC00118 in [Frontend] Remove warnings for immutable variable rebinding #3262

[BugFix] Reject non-power-of-two AllReduce thread strides by @LeiWang1999 in [BugFix] Reject non-power-of-two AllReduce thread strides #3266

[NVRTC] Fix warp reduction compilation by @sepcnt in [NVRTC] Fix warp reduction compilation #3260

[BugFix][CUDA] Lower a kernel-body assert to a device-legal check by @xy200303 in [BugFix][CUDA] Lower a kernel-body assert to a device-legal check #3206

[BugFix][JIT] Allocate a dynamic-shape output that precedes its sizing input by @xy200303 in [BugFix][JIT] Allocate a dynamic-shape output that precedes its sizing input #3207

[CUDA] Fuse exact FP4 to FP8 conversion through FP32 by @ZenAlexa in [CUDA] Fuse exact FP4 to FP8 conversion through FP32 #3204

[Fix]Clamp T.copy/T.async_copy coalesced_width to achievable vector size instead of LOG(FATAL) by @edragain2nd in [Fix]Clamp T.copy/T.async_copy coalesced_width to achievable vector size instead of LOG(FATAL) #3246

[BugFix] Bind loop targets independently of mutable scalar variables by @sepcnt in [BugFix] Bind loop targets independently of mutable scalar variables #3232

[BugFix][Language] Lower NaN-propagating clamp through device templates by @xy200303 in [BugFix][Language] Lower NaN-propagating clamp through device templates #3205

[Fix] Reject unsupported reduction NaN propagation dtypes by @ZenAlexa in [Fix] Reject unsupported reduction NaN propagation dtypes #3273

[BugFix] Fix parallel loop lowering with let inlining disabled by @penguin-wwy in [BugFix] Fix parallel loop lowering with let inlining disabled #3269

[Transform][CUDA] Fix async copy lowering with partitioned layouts by @LeiWang1999 in [Transform][CUDA] Fix async copy lowering with partitioned layouts #3278

[CUDA] Keep FP8 vector copies packed by @LeiWang1999 in [CUDA] Keep FP8 vector copies packed #3276

[CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids by @sepcnt in [CUDA] Support SM120 block-scaled GEMM fragments and odd warp atom grids #3257

[CUDA] Reject conflicting SM120 scale fragment layouts by @LeiWang1999 in [CUDA] Reject conflicting SM120 scale fragment layouts #3284

[Build] Optimize default single-config builds by @KellyFrog in [Build] Optimize default single-config builds #3191

[Fix][TVM] Update TVM to preserve dynamic reduction tail guards by @LeiWang1999 in [Fix][TVM] Update TVM to preserve dynamic reduction tail guards #3294

[BugFix] Eliminate unused bindings in Simplify by @LJC00118 in [BugFix] Eliminate unused bindings in Simplify #3293

[Build][Windows] Restore ccache hits for clang-cl wheel builds by @LeiWang1999 in [Build][Windows] Restore ccache hits for clang-cl wheel builds #3305

[Public Release 9/30] Introduce Ascend 950 backend by @SiriusNEO in [Public Release 9/30] Introduce Ascend 950 backend #3308

[Release] Bump version into 0.1.15 by @LeiWang1999 in [Release] Bump version into 0.1.15 #3309

New Contributors

@Dino1844 made their first contribution in [BugFix][CPU] Compute scalar GEMM products in accum dtype instead of input dtype #2917

@mocusez made their first contribution in [BugFix][Carver] Gate _legalize_info fallback on sm_version #3145

@rishabhsinha17 made their first contribution in [Bugfix][Language] Reject loop else clauses in the eager frontend #3142

@Guan-jeans made their first contribution in [Example] Fix hadamard reference precision on TF32 and add pytest entry for CI #3126

@anerli made their first contribution in [BugFix][CPU] Legalize BF16 arithmetic in the CPU pipeline #3201

@xuebozhang525-alt made their first contribution in [Example] FP8 sparse MLA forward for DeepSeek V3.2 on Hopper #3224

@xy200303 made their first contribution in [BugFix][CUDA] Lower a kernel-body assert to a device-legal check #3206

Full Changelog: v0.1.14...v0.1.15

This discussion was created from the release v0.1.15.

0 replies

Navigation Menu

Repository navigation

v0.1.15 #3311

Uh oh!

{{title}}

Uh oh!

Replies: 1 comment

Uh oh!

{{title}}

Uh oh!

Highlights

Ascend 950

CUDA

Language and Compiler

ROCm, CPU, and Metal

Runtime, Build, and Tooling

Compatibility Notes

What's Changed

New Contributors

Select a reply

Uh oh!

v0.1.15 #3311

Uh oh!

LeiWang1999 Sep 30, 2026 Maintainer

Highlights

Ascend 950

CUDA

Language and Compiler

ROCm, CPU, and Metal

Runtime, Build, and Tooling

Compatibility Notes

What's Changed

New Contributors

Replies: 1 comment

Uh oh!

melitahambasic43-afk Oct 2, 2026

Highlights

Ascend 950

CUDA

Language and Compiler

ROCm, CPU, and Metal

Runtime, Build, and Tooling

Compatibility Notes

What's Changed

New Contributors

LeiWang1999
Sep 30, 2026
Maintainer

melitahambasic43-afk
Oct 2, 2026