Repository navigation
Releases
v0.1.10
Pins the simulator at fba948da (main, 2026-11-16). v0.1.9 pinned c7af104d; this release adds the fixes merged since, below.
NVIDIA runtime
- A stream's wait on an event keeps working after the event is destroyed.
cudaStreamWaitEventfollowed straight away bycudaEventDestroyused to release the waiting stream early and the next copy read stale data (5 of 30 rounds in a reproducer). tacuda's BatchAPI tests found it; they pass 319 of 319 now (#667). cudaGetDeviceFlagsandcudaSetDeviceFlagsanswer as a card does. A blocking stream that captures keeps the legacy stream out of the capture.cudaMemcpyAsyncwith an unknown copy kind answerscudaErrorInvalidMemcpyDirection(#655).- Driver: a fatbin with nothing the GPU can run is
CUDA_ERROR_NO_BINARY_FOR_GPU. cuBLAS: the Ex routines check their types in the card's order. cuSPARSE: SpMV takes the card's fifteen type combinations and no others (#655). - NPP's Gaussian filter, and more RTX 3060 differential fixes (#644).
The approximate math instructions give a real card's bits (#669)
MUFUin SASS and the PTX.approxforms were the function at host precision. A card computes them in a small fixed-point unit and differs in the last bit on 13 to 39 % of arguments. The unit was measured over all 2^32 operands of each of the ten operations on an RTX 3060 and modelled exactly. An A10G and an H100 give the same results on every operand; a T4 and an L4 through the math library.- This changes the results of existing programs that call
expf,powf,tanhf,rsqrtfand the like, or are built with-use_fast_math: they now print the card's last bits.VGPU_MUFU_EXACT=0gives the earlier results. Blackwell was not measured and gets the same model. - A program run from its PTX gives the same bits in the math library as from its SASS.
Tensor cores and matrix cores (#668)
- One rule with named parameters. Turing, Ada and GA100 devices take the measured Ampere sum, an H100 has its own sum, and the subnormal corners follow the FTTN paper's findings. An opt-in MI250X rule is behind a switch. The per-profile table, measured against published, is in docs/numerics.
Host-GPU mode (faster)
- Turbo mode's first-launch rules, and the hotspot case back to 1.85x (#645, #669).
- AMD tier: row handlers for the 64-bit forms (64-bit comparisons, double min/max, reciprocal and square root, 64-bit shifts, SDWA on whole rows), same bits, inq steepest_descent 18 % less CPU (#663).
- hiprtc keeps a compile cache.
AMD
- gfx1030 NaNs in the SDWA, DPP, fmamk/fmaak, packed, double and dot forms follow the card (#655).
- The
VGPU_DISPATCH_STATSline carries the kernel's name, its instruction mix, the matrix work by type, and the bytes each memory instruction moves; LDS bank conflicts are counted (#652).
Use it as before: uses: pantheongpu/setup-pantheonsim@v0.
v0.1.9
Pins the simulator at c7af104d (main, 2026-11-14). v0.1.8 pinned 26bb19f2; this release adds the fixes merged since, below.
SASS fix
- A lane group switched away from in the middle of a loop gets its uniform registers and predicates back when it runs again. pravirkr/loki's sm_90 EP scorer stopped after 64 trips with "a shared address ... naming block 141 of a cluster of 1"; it now runs as on the card (#642).
Host-GPU mode (faster, #641)
- Turbo mode runs more kernels on the card: short launches whose data is already on the card stay there (LU 11.6 s to 1.5 s). It now admits modules with
__device__/__constant__variables,-rdccalls, graph launches, and float and CAS atomics. Over 45 changed programs the turbo geomean went from 1.20x to 1.60x. - Verify mode keeps chunks whose card and simulator bytes match on the card (3DCONV verify 46.5 s to 8.5 s) (#569). An opt-in exact-mode sqrt idiom is behind
VGPU_HOST_GPU_SQRT=1(#570).
NVIDIA: answers as an RTX 3060 gives them
These were found by running the same edge cases on the simulator and on an RTX 3060 (CUDA 13.0) and comparing.
- Runtime: every failing call sets the thread's last error. An unknown copy kind is
cudaErrorInvalidMemcpyDirection. 2D and 3D pitches,ysizeand empty extents are checked as the card checks them. Flags the card refuses are refused bycudaEventCreateWithFlags,cudaHostAllocandcudaMallocManaged. An invalidated stream capture refuses later work untilcudaStreamEndCapture.cudaPointerGetAttributesof an unknown device address answers Unregistered. - Driver: dead contexts (destroyed, released or reset) answer
CUDA_ERROR_CONTEXT_IS_DESTROYED.cudaDeviceResetresets the primary context. The runtime's next call that uses the device makes the primary context current again. ThecuProfiler*andcudaProfiler*calls answer as the card does. - cuBLAS:
?gemvtakes negative sizes and zero increments as the card does, andCUBLAS_OP_CONJGis no longer taken as a conjugate transpose. Bad pointer and math modes are refused. The status texts are the card's. The level-2 routines and the GEMMs print the reference BLAS "On entry" line. - cuSOLVER: argument checks with LAPACK's info for getrs, potrf, potrs, syevd and gesvd. gesvd works with jobs 'O'. potrf and unpivoted getrf leave a failing matrix as the card does.
- cuSPARSE: descriptor and SpMV argument checks, and the card's status texts.
cusparseSpMV_bufferSizeasks for the card's workspace size and a null buffer is refused, so call it beforecusparseSpMV.cusparseCsr2cscEx2checks its arguments. - Under
vgpu run --preload, CUPTI and NVSHMEM find the privately loaded driver.
AMD fixes
- The SDWA, DPP and VOP3P modifiers, and the denormal and IEEE modes around them, were audited against AMD's guides and an RX 6800. The output modifier is ignored with IEEE mode or output denormals on, on every target.
v_div_scaleresults are clamped and scaled. neg_lo and neg_hi apply to packed integer instructions. gfx90a and gfx1030 follow the guides'v_minandv_maxrules. gfx1030 SDWA comparisons are implemented. - The clamp bit saturates integer results on every target, including the carry-out adds and subtracts and
v_mad_u64_u32andv_mad_i64_i32.v_mad_i64_i32's scalar bit is the sign of the exact result.
Use it as before: uses: pantheongpu/setup-pantheonsim@v0.
v0.1.8: the v0.1.7 AMD regression fixed, and 25 more fixes
Pins the simulator at 26bb19f2 (main, 2026-11-13). v0.1.7 pinned 1dc43311; this release adds the fixes merged since, below.
Regression fixed
- AMD: a whole-wave load into accumulation registers the wave had not used yet lost all but the last of them, so Kokkos HIP kernels read small invalid addresses (#621). This came with v0.1.7; exadis passes 25/25 again.
Faster
- cuBLAS band and packed routines (tbmv, tpmv, gbmv, sbmv/hbmv, spmv/hpmv) work on the band instead of building the full matrix on every call, and glibc's heap thresholds are set at start-up. SCS's max_ent on a kernel 7.0 runner went from 574 s to 5.5 s with the same iterations (#551). The action also sets transparent huge pages to
alwayson GitHub-hosted runners (new inputhuge-pages, on by default). - AMD code objects load with fewer copies and page faults: rocBLAS start-up about 1.7x faster, same bits (#565).
NVIDIA fixes
- Kernels no longer die with SIGFPE when the host program enabled floating-point traps, as gfortran's
-ffpe-trapdoes (#605). - Driver functions from
cudaGetDriverEntryPointwork in programs that do not link libcuda (#604).cuGetProcAddressresolves names by CUDA version, as the driver does, which cuda-core 1.3 and so numba-cuda need (#626).cuProfilerInitializeis exported for cudarc 0.19 (#627). vgpu run --preloadloads the driver privately, so programs that export variables namedcuInitand so on (ThrustRTC) no longer crash (#597). NVRTC's fallback compiles the way NVRTC does, without the host's headers (#619).- Peer access: after
cudaDeviceEnablePeerAccess, kernels read and write the peer device's memory (#598). - Copies, fills and launches queued behind a pending host function no longer block the caller (#634).
cublasDdotand the other dot products store a device result in host pointer mode as the card does (#638).- cuRAND exports the Poisson API (#629); NPP's stream-context functions for CUDA 12 (#617, partly).
ld.param::entry/::funcin PTX (jax 0.12) (#589); destroying texture or surface object 0 succeeds (#618);nvidia-smi nvlink -c/-cBridge(#640).
AMD fixes
v_div_fmas/v_div_fixuphonour clamp and the output modifier (#613); atomics of overlapping dispatches are no longer lost (#610); dynamic-stack kernels get the stack size of scratch (#585).hipGraphInstantiateWithParamstakes the Upload flag (#628); a failedhipModuleLoadis the thread's last error (#635); Tensile's kernel lookup no longer prints spurious "no kernel named" lines (#636).
Downstream checks on the pinned commit: exadis 25/25, colvars 216/216, OpenMM HIP Double 21/21, celeritas 37/37, GauXC including its xc-integrator device case.
Use it as before: uses: pantheongpu/setup-pantheonsim@v0.
v0.1.7: a newer simulator
Pins the simulator at 1dc43311 (main, 2026-11-13). v0.1.6 pinned 735c2626, so this release brings the 26 simulator pull requests merged since then.
Fixes
cudaLaunchHostFuncruns its callback on a thread of its own instead of inline (#487), so programs that wait on a host function from the stream no longer hang (zulianp/sfem'sCudaExecutionContextTest).- PTX:
vote.syncballots at two sites now rendezvous, so abar.syncafter them is reachable (#577, #574). - SASS and NVRTC:
NANOSLEEPspin loops,ld.weak/st.weak, NVRTC float options, the JIT error log, large.regranges (#562). - Results that differed from a real card: min/max forms, real cuFFT transforms, GEMM input rounding, the WMMA layout, packed f16 SASS (#556).
- Kernel thread limits are counted in whole warps per sub-partition, as an RTX 5080 reports them (#586, from @PhilipDeegan).
- AMD: a scalar load into VCC, a negative scratch offset sum, packed zero constants (#584); f32 and f64 division and
v_ldexp_f32(#511). libcudartexports thecudaGL*interop functions (#540).- Host-mapped (zero-copy) memory is faster: zerocopy-cuda 1.8x, AMD zerocopy 3x (#507).
The action itself
- The CUDA install from NVIDIA's repository has a time limit and is retried three times (#596).
- An apt mirror fallback for when the runner's Ubuntu mirror does not resolve (#594).
- The build cache is saved right after the build, so a failing test later in the job no longer throws it away (#583).
Opt-in features (off unless you set them)
VGPU_EXACT_THREADS, parallel blocks with the result of blocks in order (#572).- JIT work: region linking, fused runs, and a cheaper wait on the plain loop (#573, #590, #592, #546).
- Host-GPU mode on a machine with the real card (#529, #557, #560).
Use it as before: uses: pantheongpu/setup-pantheonsim@v0.
v0.1.6: no Node 20 warning
Pins the simulator at ae270779: 735c2626 with actions/cache v6.1.0, so the Node.js 20 deprecation warning is gone. Nothing else changes.
v0.1.5: a newer simulator
Pins the simulator at 735c2626. Adds #477, #484, #485, #494 (abs.s16), #500 (indexed ULDC) and the SASS and speed work since v0.2.0.
v0.1.4: simulator 0.2.0
This release moves the simulator from pantheongpu/pantheonsim@0591eb7b to pantheongpu/pantheonsim@840bbf6c, the commit tagged v0.2.0, the first official PantheonSim release. The inputs are unchanged, so jobs on @v0 need no edits.
What changed
- Simulator 0.2.0: 369 commits since the previous pin. 48 GPU profiles (38 NVIDIA, 10 AMD; 14 NVIDIA and 3 AMD read from physical cards), and every document checked against the code. Release notes and full list.
- Other CI systems: the simulator's
scripts/ci-setup.shdoes for GitLab CI, CircleCI, Forgejo, Jenkins and Buildkite what this action does for GitHub. - The simulator checks behaviour and has no timing model, so keep a run on a physical GPU before you ship.
v0.1.3: the cuDNN headers are fetched at build time
This release moves the simulator from pantheongpu/pantheonsim@407a112e to pantheongpu/pantheonsim@0591eb7b. The inputs are unchanged, so jobs on @v0 need no edits.
What changed
- Licensing: the simulator's source tree no longer contains NVIDIA's cuDNN headers. They are downloaded from the
nvidia-cudnn-cu12wheel when the simulator is built (pantheonsim#338), so the build the action does has no redistributed NVIDIA headers in it. The first build, or any build whose cache misses, downloads that wheel (several hundred MB); runs that restore the cache are unaffected. - Newer simulator: about 90 commits since the previous pin: more GPU profiles, more AMD and NVIDIA instruction coverage, and library fixes. Full list.
v0.1.2: a newer simulator, cuStateVec, and the fixes outside projects needed
This release moves the simulator from pantheongpu/pantheonsim@4f1f39f to pantheongpu/pantheonsim@407a112e. The inputs are unchanged, so jobs on @v0 need no edits.
What changed for CUDA code
- PTX: device functions that return vectors through call slots, and signed
ld.param. Device calls with barriers inside them, and indirect calls with a different target per lane. - Driver API:
- managed memory and
cuLaunchHostFunc; - CUDA arrays,
cuMemcpy2D/3Dand stream-ordered allocation; - pitched allocation, module globals and stream callbacks;
cuEventElapsedTime's refusals.
These are what bindings that loadlibcudaalone (rust-cuda's cust, cudarc) reach for.
- managed memory and
- Runtime: the per-thread default stream's entry points (
--default-stream per-thread), managed variables, device-sideprintf, and more of the API. - Libraries: cuBLAS Set/Get Vector and Matrix, Dgemv, asum and grouped batched GEMM; cuRAND's log-normal doubles; cuBLASLt's logger.
- cuStateVec (new): a cuStateVec of the simulator's own, written from NVIDIA's documentation, covering what QuEST's cuQuantum backend calls. NVIDIA's own libcustatevec carries a static CUDA runtime that can't reach a simulated driver; programs linked against it pick up the simulator's under
vgpu run. - Toolkit install: the CUDA profiler headers (
cuda_profiler_api.h,cudaProfiler.h) now install with the toolkit.
What changed for HIP code
More of HIP's API as ROCm answers it: every function ROCm 7.1's libamdhip64 exports, memory copies and pools, graphs and stream capture, virtual memory and IPC between processes, textures and images on RDNA, and HSA SVM.
Things to know
- If your job links against the toolkit's own cuBLAS, cuSPARSE and so on, set
library-path: falseand run the tests undervgpu run. - It checks behaviour, not performance. Keep a run on physical GPUs before a release.
@v0 moves to this release. To stay on one version, pin @v0.1.2, or pin this release's commit SHA.
v0.1.1: a newer simulator and six more GPUs
This release moves the simulator from pantheongpu/pantheonsim@04dde05 to pantheongpu/pantheonsim@4f1f39f, which is 66 merged pull requests later. The inputs are unchanged, so jobs on @v0 need no edits.
New GPUs
- NVIDIA: GeForce RTX 5090 and B300 (Blackwell Ultra), 16 profiles in all.
- AMD: MI250X, Radeon RX 6900 XT, RX 7900 XTX and RX 9070 XT, 7 profiles in all.
What changed for CUDA code
- Runtime: copy directions follow the hardware's rules, and managed and pinned memory count as device memory for copies and memsets. Zero-byte copies succeed.
-rdcprograms link every relocatable piece.cudaLimitStackSizenow reaches the launch. When libcuda and libcudart are loaded separately, they share one simulated machine. - PTX:
atom.cashas the right operand order. Other additions: device-sidealloca, named-barrierbar.red, the video instructions, and surface.clamp/.zero. Allmma.sync/mma.spforms before sm_120's are checked against an RTX 3060. Blackwell's block-scaled and sparsetcgen05.mmaand fp4/fp6 are in, and so are the B300's instructions. - Libraries: more of cuBLAS, cuSPARSE, cuSOLVER, cuFFT and NVRTC, which is enough for PyTorch's
torch.fft,torch.linalg,torch.sparseand complex tensors. Some examples: the cuBLAS_64forms,Sdgmm/Ddgmm, cuSPARSE CSR to CSC, and cuSPARSE calls made during graph capture. - Toolkit install: nvJitLink is now installed alongside the toolkit, because linking against its cuSPARSE needs it.
- Speed: tensor-core MMAs are faster, and a grid spreads over the CPUs the job may use.
What changed for HIP code
- ROCm's own
rocminfo,rocm-smiandamd-smirun unchanged, and so do rocBLAS, hipSPARSELt, RCCL and Ollama's HIP backend. - PyTorch runs on every simulated AMD GPU, and vLLM runs on the MI300X.
torch.compileworks on the MI350X.
Things to know
- If your job links against the toolkit's own cuBLAS, cuSPARSE and so on, set
library-path: falseand run the tests undervgpu run. The input's description explains why. - It checks behaviour, not performance. Keep a run on physical GPUs before a release.
@v0 moves to this release. To stay on one version, pin @v0.1.1, or pin this release's commit SHA.