TornadoVM 6.0.0 #1062

Pinned

stratika announced in announcements

stratika
Sep 2, 2026
Maintainer

02/09/26

Improvements

Expand unit-test coverage for Java language corner cases and compiler restrictions #1018: Expand unit-test coverage for Java language corner cases and compiler restrictions
Add TornadoVM contributor Claude skill (OpenCL/CUDA/Metal) #980: Add TornadoVM contributor Claude skill (OpenCL/CUDA/Metal)
[feat] CUDA backend: native FP8 conversion + FP8/BF16 tensor-core MMA and cp.async copies #952: [feat] CUDA backend: native FP8 conversion + FP8/BF16 tensor-core MMA and cp.async copies
Packed half2 support for the CUDA backend #962: Packed half2 support for the CUDA backend
Add FP8 (E4M3/E5M2) support for the CUDA backend + fix NaN/Infinity float constants #928: Add FP8 (E4M3/E5M2) support for the CUDA backend + fix NaN/Infinity float constants
Auto-disable intra-plan concurrency on serial task graphs #934: Auto-disable intra-plan concurrency on serial task graphs
Stage large H2D transfers through a pinned host buffer ring #936: Stage large H2D transfers through a pinned host buffer ring
[cuda] NVTX instrumentation for library tasks + cuSPARSE hybrid provider (SpMV/SpMM) #919: [cuda] NVTX instrumentation for library tasks + cuSPARSE hybrid provider (SpMV/SpMM)
[cuda] CUTLASS hybrid-API provider: FP32/FP16 GEMM + fused bias/ReLU/GELU epilogues as library tasks #912: [cuda] CUTLASS hybrid-API provider: FP32/FP16 GEMM + fused bias/ReLU/GELU epilogues as library tasks
[hybrid] Hybrid API: native library tasks with cuDNN provider #888: [hybrid] Hybrid API: native library tasks with cuDNN provider
[hybrid] Hybrid API: native library tasks with cuFFT provider #887: [hybrid] Hybrid API: native library tasks with cuFFT provider
[Metal] Add hardware matrix units (simdgroup_float8x8) for fast gemms to M-silicon #850: [Metal] Add hardware matrix units (`simdgroup_float8x8`) for fast gemms to M-silicon

Compatibility

Remove dependency to JVMCI for JDK27+ backwards compatiblity for JDKs 21-26 #894: Remove dependency to JVMCI for JDK27+ backwards compatiblity for JDKs 21-26
Correct the hybrid-API documentation against the code on develop #1017: Correct the hybrid-API documentation against the code on develop
[docs] Revision of the documentation page for v5.0.0 #938: [docs] Revision of the documentation page for v5.0.0
[ci] Fix release pipeline triggers, rename pipeline stages, skip redundant CI #925: [ci] Fix release pipeline triggers, rename pipeline stages, skip redundant CI

Bug Fixes

[hotfix] Emit a switch statement for the table and hash switch lowerings instead of nothing #1054: [hotfix] Emit a switch statement for the table and hash switch lowerings instead of nothing
[runtime] Allocate every output of a batched graph instead of reusing another object's buffer #1035: [runtime] Allocate every output of a batched graph instead of reusing another object's buffer
[bugfix] cuSPARSE native library is never packaged, so the provider is dead on Linux/macOS #1012: [bugfix] cuSPARSE native library is never packaged, so the provider is dead on Linux/macOS
Fix develop build: TestAtomicRmw references the removed SPIRV and PTX backend types #1007: Fix develop build: TestAtomicRmw references the removed SPIRV and PTX backend types
[fix] Restore cutlass-jni build on GCC 15+ hosts (-Wtemplate-body) #945: [fix] Restore cutlass-jni build on GCC 15+ hosts (-Wtemplate-body)
Fix vmDeps path to dispatch KernelContext/WorkerGrid kernels #932: Fix vmDeps path to dispatch KernelContext/WorkerGrid kernels
[cuda] Fix misleading NVRTC compilation errors: memoized FP16 header resolution + explanatory driver/toolkit diagnostics #910: [cuda] Fix misleading NVRTC compilation errors: memoized FP16 header resolution + explanatory driver/toolkit diagnostics
[hotfix] Resolving unit-tests failure in CUDA backend #877: [hotfix] Resolving unit-tests failure in CUDA backend
Fix for stride-16 swizzled FP16 shared-memory load and store that caused segfault in tests #903: Fix for stride-16 swizzled FP16 shared-memory load and store that caused segfault in tests
[hotfix] Multiple fixes after enabling CUDA backend into develop #874: [hotfix] Multiple fixes after enabling CUDA backend into develop
Fix batch processing with a remainder chunk under device buffer reuse #886: Fix batch processing with a remainder chunk under device buffer reuse
Fix repeated execution of batched plans #882: Fix repeated execution of batched plans

Other Changes

docs: add class-level Javadoc to examples #35 #1060: docs: add class-level Javadoc to examples Add Javadoc and document the examples in TornadoVM #35
Replace JNI with the Java FFM API in the backends and library tasks #1058: Replace JNI with the Java FFM API in the backends and library tasks
[docs] Bring the hybrid API guide up to date with the providers and execution-plan options on master #1056: [docs] Bring the hybrid API guide up to date with the providers and execution-plan options on master
Pass CUDA compiler flags to NVRTC, and add named compile profiles (default|fast|debug|repro) #1010: Pass CUDA compiler flags to NVRTC, and add named compile profiles (default|fast|debug|repro)
[profiler] Count copy-out bytes even when the transfer produced no event to time #1048: [profiler] Count copy-out bytes even when the transfer produced no event to time
[runtime] Fix silent zero output when a prebuiltTask follows a JIT task in the same graph #1049: [runtime] Fix silent zero output when a prebuiltTask follows a JIT task in the same graph
[docs] Remove stale planning notes left at the repository root #1052: [docs] Remove stale planning notes left at the repository root
[release] Fix Windows release-SDK build/validate failures: argnd MAX_PATH, and jdk22plus JVMCI vendoring #1050: [release] Fix Windows release-SDK build/validate failures: argnd MAX_PATH, and jdk22plus JVMCI vendoring
[api] TornadoExecutionPlan.transferToDevice(): ad-hoc copy-in without running a kernel to force it #1036: [api] TornadoExecutionPlan.transferToDevice(): ad-hoc copy-in without running a kernel to force it
[docs] Add a tornadovm-perf-campaign skill: profiling recipes, probe harness and optimisation-pattern catalogue #1038: [docs] Add a tornadovm-perf-campaign skill: profiling recipes, probe harness and optimisation-pattern catalogue
[runtime] Clear only the written prefix of each wait-list row between executions (1.57x on dependency-tracked decode) #1027: [runtime] Clear only the written prefix of each wait-list row between executions (1.57x on dependency-tracked decode)
[runtime] Dispatch interpreter bytecodes through a switch over named handlers #1040: [runtime] Dispatch interpreter bytecodes through a switch over named handlers
[runtime] Replace the per-backend instanceof chains with one native-array layout table #1039: [runtime] Replace the per-backend instanceof chains with one native-array layout table
[runtime] Read non-terminal copy-outs asynchronously instead of blocking on each one (1.27x on an 8-output graph) #1029: [runtime] Read non-terminal copy-outs asynchronously instead of blocking on each one (1.27x on an 8-output graph)
[cuda] Propagate CUDA driver failures to Java instead of only logging them #1025: [cuda] Propagate CUDA driver failures to Java instead of only logging them
Add CUDA-samples-derived compute examples with measured results (GEMM, N-body, binomial options, FDTD stencil) #1020: Add CUDA-samples-derived compute examples with measured results (GEMM, N-body, binomial options, FDTD stencil)
[build] Validate the BACKEND variable in the Makefile against the supported backends #1045: [build] Validate the BACKEND variable in the Makefile against the supported backends
Fix emitter miscompile of nested short-circuit booleans #1043: Fix emitter miscompile of nested short-circuit booleans
Add warp-vote intrinsics (simdAny, simdAll, simdBallot) to KernelContext on the CUDA backend #984: Add warp-vote intrinsics (simdAny, simdAll, simdBallot) to KernelContext on the CUDA backend
Add KernelContext coverage for globalBarrier, cp.async byte copies and flash attention #1016: Add KernelContext coverage for globalBarrier, cp.async byte copies and flash attention
[bugfix] KernelContext operations a backend cannot intrinsify silently compute wrong results #1013: [bugfix] KernelContext operations a backend cannot intrinsify silently compute wrong results
[bugfix] CUDA: getFloat32() on a swizzle-loaded half float fails to compile #1014: [bugfix] CUDA: getFloat32() on a swizzle-loaded half float fails to compile
[bugfix] Library tasks can allocate device memory inside a CUDA-graph capture region #1015: [bugfix] Library tasks can allocate device memory inside a CUDA-graph capture region
Fix atomicAdd on local-memory arrays with a runtime-computed index (CUDA backend) #1019: Fix atomicAdd on local-memory arrays with a runtime-computed index (CUDA backend)
Allocate wait-event lists on demand #1002: Allocate wait-event lists on demand
Cache compiled CUDA modules on disk (re-land of #1000, which never reached develop) #1008: Cache compiled CUDA modules on disk (re-land of Cache compiled CUDA modules on disk: 2.2x faster start-up for kernel-heavy applications #1000, which never reached develop)
[cuda] Add atomicCAS, atomicExchange, atomicMin and atomicMax on local arrays to KernelContext #991: [cuda] Add atomicCAS, atomicExchange, atomicMin and atomicMax on local arrays to KernelContext
Size the bytecode buffer for the graph: many-task graphs silently produce wrong results #1004: Size the bytecode buffer for the graph: many-task graphs silently produce wrong results
Take fewer map and monitor operations on the interpreter's per-object hot path #997: Take fewer map and monitor operations on the interpreter's per-object hot path
Emit variable declarations in a deterministic order #999: Emit variable declarations in a deterministic order
Fix consumeFromDevice silently reading a stale buffer when the named producer is not the previously executed graph #996: Fix consumeFromDevice silently reading a stale buffer when the named producer is not the previously executed graph
Remove FPGA support (build infra, codegen, API, docs) #978: Remove FPGA support (build infra, codegen, API, docs)
Remove the SPIR-V backend #951: Remove the SPIR-V backend
Remove PTX backend in favour of the newly added CUDA backend #872: Remove PTX backend in favour of the newly added CUDA backend
[metal][fix] Fix inflated Metal backend profiling metrics (#905) #908: [metal][fix] Fix inflated Metal backend profiling metrics (Metal backend reports grossly inflated GPU/dispatch timing metrics (TOTAL_KERNEL_TIME, TOTAL_DISPATCH_KERNEL_TIME, COPY_IN_TIME, COPY_OUT_TIME) despite comparable or better end-to-end execution time #905)
Fix packed FP16 (Half2) support on the OpenCL backend and add cross-backend test coverage #994: Fix packed FP16 (Half2) support on the OpenCL backend and add cross-backend test coverage
Extend TaskGraph task/libraryTask arity to 20 for hybrid API #958: Extend TaskGraph task/libraryTask arity to 20 for hybrid API
Add tornado-fuzz: differential + sanitizer fuzz tester for the CUDA backend #947: Add tornado-fuzz: differential + sanitizer fuzz tester for the CUDA backend
[cuda] BFloat16Array: first-class bfloat16 native array + JIT conversions + cuBLAS/CUTLASS bf16 GEMM #967: [cuda] BFloat16Array: first-class bfloat16 native array + JIT conversions + cuBLAS/CUTLASS bf16 GEMM
[cuda] Saturate CUTLASS provider: SiLU/Sigmoid/Tanh/HardSwish epilogues + batched FP16 GEMM #963: [cuda] Saturate CUTLASS provider: SiLU/Sigmoid/Tanh/HardSwish epilogues + batched FP16 GEMM
[cuda] Fix HalfFloat codegen for inline new HalfFloat(computedExpr) writes #969: [cuda] Fix HalfFloat codegen for inline new HalfFloat(computedExpr) writes
[fix] Windows support for the CUDA backend (build, runtime DLL resolution, correctness) #921: [fix] Windows support for the CUDA backend (build, runtime DLL resolution, correctness)
[cleanup] remove deprecated files #923: [cleanup] remove deprecated files
[docs] Complete hybrid API guide (HYBRID_API_GUIDE.md): all providers, examples, flags, custom-provider walkthrough #915: [docs] Complete hybrid API guide (HYBRID_API_GUIDE.md): all providers, examples, flags, custom-provider walkthrough
Update level-zero library version to v1.18.1 #864: Update level-zero library version to v1.18.1
Fix CUDA-backend build on develop (cuFFT broken merge + cudnn SDPA) #893: Fix CUDA-backend build on develop (cuFFT broken merge + cudnn SDPA)
Feat/update logo #901: Feat/update logo
[JDK21] Support for CUDA streams for overlapping data transfers with TaskGraph execution #800: [JDK21] Support for CUDA streams for overlapping data transfers with TaskGraph execution
Revise README to include latest TornadoVM updates #852: Revise README to include latest TornadoVM updates
[hybrid] Hybrid API: native library tasks with cuBLAS/cuBLASLt provider #879: [hybrid] Hybrid API: native library tasks with cuBLAS/cuBLASLt provider
[cuda] Add CUDA C backend via NVRTC to complement the PTX backend #861: [cuda] Add CUDA C backend via NVRTC to complement the PTX backend
[hotfix][PTX] Align compileTask exception handling with other backends to preserve original exception on bailout-disabled path #890: [hotfix][PTX] Align compileTask exception handling with other backends to preserve original exception on bailout-disabled path
[OpenCL] Fix pointer-to-ulong conversions rejected by strict OpenCL compilers (Intel iGPUs) #884: [OpenCL] Fix pointer-to-ulong conversions rejected by strict OpenCL compilers (Intel iGPUs)
Replace heuristic brace-placement with structured control-flow recovery (OpenCL + Metal) #848: Replace heuristic brace-placement with structured control-flow recovery (OpenCL + Metal)
Fix nested persisted-object list on task-graph reuse and empty consume #870: Fix nested persisted-object list on task-graph reuse and empty consume

How to use TornadoVM SDK

Select the archive matching your JDK (21, 25, 26, or 27) and target accelerator backend (opencl, cuda, metal, or full) from this Release's assets. Ensure JAVA_HOME points to a matching JDK installation before running. The example below uses jdk21 + opencl for Linux/macOS/Windows — swap jdk21-opencl for e.g. jdk26-cuda or jdk27-metal as needed.

Linux Systems

wget https://github.com/beehive-lab/TornadoVM/releases/download/v6.0.0/tornadovm-6.0.0-jdk21-opencl-linux-amd64.zip
unzip tornadovm-6.0.0-jdk21-opencl-linux-amd64.zip
export TORNADOVM_HOME="$(pwd)/tornadovm-6.0.0-jdk21-opencl"
export PATH=$TORNADOVM_HOME/bin:$PATH
tornado --devices
tornado --version

macOS Systems

wget https://github.com/beehive-lab/TornadoVM/releases/download/v6.0.0/tornadovm-6.0.0-jdk21-opencl-mac-aarch64.zip
unzip tornadovm-6.0.0-jdk21-opencl-mac-aarch64.zip
export TORNADOVM_HOME="$(pwd)/tornadovm-6.0.0-jdk21-opencl"
export PATH=$TORNADOVM_HOME/bin:$PATH
tornado --devices
tornado --version

Windows (10+) Systems

curl -L -o tornadovm-6.0.0-jdk21-opencl-windows-amd64.zip https://github.com/beehive-lab/TornadoVM/releases/download/v6.0.0/tornadovm-6.0.0-jdk21-opencl-windows-amd64.zip
tar -xf tornadovm-6.0.0-jdk21-opencl-windows-amd64.zip
set TORNADOVM_HOME=%cd%\tornadovm-6.0.0-jdk21-opencl
set PATH=%TORNADOVM_HOME%\bin;%PATH%
tornado --devices
tornado --version

This discussion was created from the release TornadoVM 6.0.0.

Replies: 0 comments

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment