Add sccache support to build scripts with ROCm compiler wrapping by subodh-dubey-amd · Pull Request #3482 · ROCm/TheRock

subodh-dubey-amd · 2026-02-18T09:53:21Z

Motivation

Add sccache support to PyTorch wheel builds for S3-backed distributed caching. Script placed in build_tools/ per reviewer feedback, modeled after build_tools/setup_ccache.py.

Part of sccache PR sequence: #3369 → #3389 → this → workflow wiring.

Technical Details

New: build_tools/setup_sccache_rocm.py — generic sccache ROCm helper (CLI + importable):
- find_sccache() — locate binary; hard fail if missing
- setup_rocm_sccache() — wrap clang/clang++ with sccache stubs (Linux only)
- restore_rocm_compilers() — undo wrapping
Modified: external-builds/pytorch/build_prod_wheels.py:
- --use-ccache / --use-sccache mutually exclusive args
- Both hard-fail with RuntimeError if the requested cache tool is not found (per review) — no silent fallback
- Added explicit ccache availability check (previously would fail with an unclear subprocess error)
- sccache: wrap compilers → set CMAKE launchers → try/finally around build for guaranteed compiler restore + stats
- Moved ccache stats reporting into finally block for consistent reporting on both success and failure

Test Result

No workflow changes — sccache wired but not yet invoked by CI (next PR adds cache_type input + AWS config).

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

… wrapping

… readability

…--start-server or --zero-stats fail, compilers are still restored

…ild_prod_wheels.py

ScottTodd · 2026-02-18T19:21:22Z

+        sccache_path = find_sccache()
+        if not sccache_path:
+            raise RuntimeError(
+                "sccache not found. Please install it:\n"
+                "  - Linux: sccache should be pre-installed in the Docker image\n"
+                "  - Windows: choco install sccache"
+            )


nit: These scripts are runnable by developers who may be building outside of "the Docker image" that our CI/CD systems use, so I would prefer to link to general documentation like https://github.com/mozilla/sccache or https://github.com/mozilla/sccache?tab=readme-ov-file#installation and/or be specific about which Docker image we suggest building within (https://github.com/ROCm/TheRock/tree/main/dockerfiles#build_manylinux_dockerfile)

Updated. Error messages now link to sccache installation docs and TheRock dockerfiles instead of vaguely referencing "the Docker image". Applied in both build_prod_wheels.py and setup_sccache_rocm.py.

ScottTodd · 2026-02-18T19:23:49Z

+        sccache_setup_attempted = True
+        setup_rocm_sccache(rocm_dir, sccache_path)
+
+    try:
+        if args.use_sccache and sccache_setup_attempted:


The sccache_setup_attempted variable doesn't seem useful, can it be removed? If setup fails, that should raise an exception and this code path would not run anyways.

Removed. Replaced with sccache_wrapped — it tracks whether compilers were actually wrapped so finally knows whether to restore. Failures in setup_rocm_sccache() raise before reaching the build.

ScottTodd · 2026-02-18T19:31:06Z

+This script wraps the ROCm LLVM compilers (clang, clang++) with sccache stubs
+to enable caching of HIP device code compilation. This is necessary because
+hipcc invokes clang via absolute paths, bypassing CMAKE_*_COMPILER_LAUNCHER.


We might be able to use this as part of regular source builds of TheRock with some effort (as we build clang during one CMake invocation then point other subproject builds at dist/ directories using that clang using toolchain files which we do control and could edit), but I think a more natural approach might be to change hipcc itself. Fine to first solve for pytorch using this method that has already been tested there.

We'd also need to be careful about making edits in the build directory that would feed into released packages (the clang that we ship should just be clang, not a symlink to a wrapper script that uses sccache)

I did some more research, seems like

[Feature]: Support executing Clang through a wrapper/launcher for caching llvm-project#72

Add HIP_CLANG_LAUNCHER for launching Clang through a wrapper executable llvm-project#14

[Feature/RFC]: Allow caching of HIP compilations ROCm#2817

Might let us drop the wrapper code 🤔

Agreed — added a TODO referencing the upstream issues (ROCm/llvm-project#72, ROCm/llvm-project#1490, ROCm/ROCm#2817). Once HIP_CLANG_LAUNCHER lands, the wrapper approach can be replaced with an env var.

Cool, that's making progress. Filed #3760 to follow up.

ScottTodd · 2026-02-18T19:33:38Z

+    compiler_dir = compiler_path.parent
+    original_dir = compiler_dir / "original"
+    original_dir.mkdir(exist_ok=True)
+
+    original_path_file = original_dir / f"{compiler_path.name}.path"
+
+    if original_path_file.exists():
+        print(f"  {compiler_path} already wrapped (path file exists)")
+        return


Please document the side effects of this script somewhere prominent (docstring for this function, docstring for the file itself).

I wonder if a directory name like backup_for_sccache/ might be better than just original/?

Done. Added a prominent WARNING — Side effects section to the module docstring and a Side effects on the filesystem section to the create_sccache_wrapper() docstring. Renamed original/ to backup_for_sccache/ throughout.

ScottTodd · 2026-02-18T19:37:57Z

+    On Windows, compiler wrapping is skipped because hipcc calls clang.exe
+    directly and shell script wrappers won't intercept these calls.
+    Windows builds rely on CMAKE_C/CXX_COMPILER_LAUNCHER for host code caching.
+    """
+    if is_windows:
+        print("Skipping ROCm compiler wrapping on Windows (using CMAKE launchers)")
+        return


We could also add a mode to this script for Linux that skips the wrapper scripts, as those have pretty disruptive side effects in their current form that developers may want to skip.

Added --sccache-no-wrap flag. With --use-sccache --sccache-no-wrap, only CMAKE launchers are set (caches host C/C++) — no compiler wrapping, no restore needed. Useful for developers who want basic sccache without touching compiler binaries.

GZGavinZhao · 2026-02-18T20:25:54Z

I'd suggest checking to make sure MIOpen builds with sccache. We had to turn off sccache caching for MIOpen specifically because somehow with sccache we produce different object files that contain the same __hip_cuid_XXXXX, causing a "duplicate symbol" linking error. I suspected that it's some function symbol visibility problem but didn't have time to look further.

subodh-dubey-amd · 2026-02-19T04:51:35Z

I'd suggest checking to make sure MIOpen builds with sccache. We had to turn off sccache caching for MIOpen specifically because somehow with sccache we produce different object files that contain the same __hip_cuid_XXXXX, causing a "duplicate symbol" linking error. I suspected that it's some function symbol visibility problem but didn't have time to look further.

@GZGavinZhao
Thanks for the heads up! This PR scopes sccache to PyTorch wheel builds only (via build_prod_wheels.py), so MIOpen is not directly affected here. For future work extending sccache to TheRock source builds, we'll need to investigate the __hip_cuid_ duplicate symbol issue separately.

GZGavinZhao · 2026-02-19T04:57:20Z

Ah so this is just for PyTorch, not for the entire ROCm stack. I agree with your point. Thanks for the clarification!

…ls.py` to include backup and restore functionality for ROCm compilers. Update usage instructions and add a flag to skip compiler wrapping.

…p` flags, clarifying their functionalities and mutual exclusivity with `--use-ccache`.

## Summary Adds `sccache` with S3 remote storage to all four PyTorch wheel build workflows, significantly reducing build times through distributed compiler caching. **PR sequence:** #3369 → #3306 → #3389 → #3482 → **this** → #3189 ([based on Reviewer's Feedback](#3171 (comment))) ## How It Works | | Linux | Windows | |---|---|---| | **Host C/C++** | CMake compiler launchers | CMake compiler launchers | | **HIP device code** | Wraps ROCm `clang`/`clang++` with sccache | Not supported | | **Cleanup** | Restores original compilers via try/finally | N/A | Cache is stored in the `therock-<workflow>-pytorch-sccache` S3 bucket, keyed by `<os>/<arch>/` prefix. ## S3 Cache Configuration Each workflow uses a dedicated S3 bucket and IAM role, keyed by `<os>/<arch>/` prefix: | Workflow | S3 Bucket | IAM Role | |----------|-----------|----------| | Linux CI | `therock-ci-pytorch-sccache` | `therock-ci` | | Windows CI | `therock-ci-pytorch-sccache` | `therock-ci` | | Linux Release | `therock-{release_type}-pytorch-sccache` | `therock-{release_type}` | | Windows Release | `therock-{release_type}-pytorch-sccache` | `therock-{release_type}` | Where `release_type` is one of: `dev`, `nightly`, `prerelease`. ## Impact | Platform | Cold → Warm | Improvement | |----------|------------|-------------| | Linux | ~70m → ~37m | **~49%** | | Windows | ~42m → ~26m | **~38%** | Windows is lower — sccache cannot wrap HIP device compilation on Windows, only host C/C++ via CMAKE launchers. ## Tests ### Linux: - [Linux (Cache Population)](https://github.com/ROCm/TheRock/actions/runs/22226347964/job/64293924748) - 70 mins - [Linux (Cache Hit)](https://github.com/ROCm/TheRock/actions/runs/22231743387/job/64312966557) - 37 mins ### Windows: - [Windows (Cache Population)](https://github.com/ROCm/TheRock/actions/runs/22219252671/job/64280583887) - 42 mins - [Windows (Cache Hit)](https://github.com/ROCm/TheRock/actions/runs/22223608689/job/64284721704) - 26 mins ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.' > Forks: S3 caching is only active for ROCm-owned runs. Fork users can set cache_type to ccache or none, or leave the default — sccache will work locally without S3 access. ---------

## Motivation Bump rocm-systems from 93bc019 to 093b66caa3 (includes fix for hip-tests issue and revert for mathlib hiprtc issues and revert for rccl-test, added revert for miopen failures due to PR 653): Commits: 093b66caa3 (HEAD, origin/develop, origin/HEAD) Revert "SWDEV-546177 - hipModuleGetLoadingMode API impl (#653)" (#3858) d8a0adbc9f [AMD-SMI] Hide libamd_smi.so internal symbols (#3777) d4da458f94 [rocprofiler-sdk] [Documentation ] Updating changelog (#3827) 19fadeb082 (origin/users/abchoudh/fix_dispatch_count) [RCCL][Tuner Plugin] Enable tuning of RCCL tuning constants (#3757) b4f5f8a6a8 rocr: Fix IPC dmabuf hang with large allocations (#3211) 64efea0435 RCCL: allow users to override max and per job memory & fix defaults. (#3797) 9b3dd101bb Removing ready_for_review (#3849) 7e43880a64 [rocprofiler-systems] Update ROCm version to 7.2.0 in CI workflows for Debian, RedHat, and Ubuntu (#3431) 1fdb6b9827 [rocshmem] add gda/topology unit tests (#3715) be1ea24a96 Move hipMipmappedArrayGetMemoryRequirements test to common tests e4513f04c8 Update amdgpu-windows-interop with latest changes, pal 58aa0bab2ced0cc9ebe8d2d0932db6774feb4e49 2026-03-04(#3773) b1f964d796 [rocprofiler-compute] Ensure long kernel name fully shows in compute analyze (#3665) 4dcf1e3ce0 SWDEV-567112 - Replace test names (#3787) 33f5f302e5 ROCM-2428 - fixes hipStreamBatchMemOp invalid operation checks (#3099) 139f4bfff8 [SWDEV-556456] Align HIP_UUID with rocminfo (#3614) 8e8928544c Reduce buffers alignment to 4 bytes (#3821) 51be29a647 AIRUNTIME-125: Consolidate Windows optimization and debug flags (#3825) 1407392240 [AMD-SMI] CI: Fix root workflow to use ASIC-specific test filters (#3807) 63f78a98d7 (origin/users/mcao/fix_rocpdsummary) [ROCM-SMI] Fix DRM include dirs leaking absolute build paths to consumers (#3808) caf2f7e1eb [ROCM-186] amd-smi: Add support for a VRAM and GTT tuning interface (#3636) a0712d4c2a [TheRock CI] Update projects_to_test lists (#3749) 02090c42c9 rocrtst: install gfx .hsaco files to share/rocrtst (#3744) 4a0a1cbfce Merge other simd table (#3696) 0d07657d78 Add missing kwargs from rocprofiler_add_integration_validate_test in .cmake-format.yaml (#2336) 3a3df301dc Optimize device counting service GPU interactions (#1583) 95d9da0098 Add SPM Enable flag in build infrastructure (#3677) 12bb9435b2 [rocprofiler-sdk] On-demand GPU profile queue creation/destruction (#3586) 941057c2c0 Navi4 tuning table iter 1 (#3052) dbf2b7369f [AMD-SMI] Display N/A for cu_occupancy when file is unavailable (#3589) b0efc7c639 [RCCL] [UT] Add ROCTX test (#3625) ba7a20ea18 Reducing the p2pnChannels for half-subscription A2A on multi-node MI350 (#3381) 75238c98a2 [clr] Fix memory leak in getOrCreateHostcallBuffer (#3699) af2ee0e8ad [hip-tests] ASAN Check for image support before we create context (#3834) ad4496678e Update windows ci subtree in include amdgpu-windows-interop (#3814) c8ad252208 [rocprofiler-register] Fix compilation with system fmt/glog (#1243) 781881544d Update README to include dbgapi and debug agent components (#3731) 88e4a7837e ROCProfiler and ROCTracer: Modifying deprecation note (#3831) b5918a5f35 [ROCM-3124-3125-3126] CUID file generation hangs on MI350 systems/CUID test failures/Segmentation fault in CUID example code (#3548) 97a5dd993c Update copyright to use SPDX IDs (#3805) 511730ab45 [rocshmem]: add flood-amo tester (#3653) 2d650a0065 [clr] Fix heap use after free error in device allocations (#3789) b6b179ad81 Disable hipHostRegister_Negative test for ASAN (#3832) 39ec318c8d [RCCL] Add GDA alltoallv via rocshmem integration (#3613) fb0f4d53b1 [RCCL] [CUMEM] Fix cuMem multi-process runs (#3811) c3de7d4bf6 SWDEV-526201 - Fix and enable disabled HIP tests from warp group (#3089) 8d9a8ca161 roofline: code cleanup and refactor vector types (#3813) 8957e49028 Don't wait on command completion if worker thread is destroyed (#3790) 9e7586a5fa [rocshmem] Add barrier APIs and expose `ROCSHMEM_TEAM_WORLD` on device (#3651) 91b09235b0 Revert "fix local gpu release static build failure (#3667)" (#3799) 0fda754b1b libhsakmt: Add secondary KFD context creation support ee43db95b0 Revert "Update TheRock reference to 20260303 commit (#3709)" (#3826) 86e28b9fae Added fix to update GL2C counters instance count for GFX11.5 (#3100) 93f69f7de4 Adjust includes to match use (#3742) e9fbc3f1a2 (develop) Update TheRock reference to 20260303 commit (#3709) be0675a1a6 (HEAD) Revert "Support fp8 types in hiprtc (#2605)" (#3792) 3e3a94a4ef [rocprofiler-systems] Add trace_cache support for std::optional<T> serialization (#3490) 0b42a7f472 clr: Eliminate unnecessary kernel name string copies (#3774) b6b0d77b29 rocr: Add hsa_amd_memory_async_batch_copy API for batched memory copies (#3259) 486e6d12d2 Resolve staircase RS regression with 48 max channels (#3684) eb59c85ac4 [gfx942][gfx950] Leverage new cache bypass builtins for simple protocol where available (#2847) 4d74d27f0e (origin/users/raramakr/rocm-smi-target) Revert "Auto Labeler: Add ci:regression-detection label to rccl PRs (#3543)" (#3769) 8f0795517c [AMD-SMI] CI: Use ASIC-specific test blacklists in workflows (#3775) 7cef5b64c1 Fix MFMA total FLOPS calculation (#3371) aea37512ba Remove duplicated tests (#3235) b6c656fdd4 Remove duplicated tests in memory module (#3087) ca3137d8f9 [rocprofiler-sdk] Install integration tests without building for therock & Misc. fixes (#3047) 0ab5c41f65 [rdc] Enable on-demand queue mode in rocprofiler-sdk to prevent inference degradation (#3629) a1eb2a1f7c rocr/wsl: a library should not output to std::out by default (#3718) b7da296cc8 Reenable flood_put/get testers on mlx5 since they should work after pr2732 (#3748) 000e24de2f [rocprofiler-sdk] Add automatic late-start support to rocprofiler_force_configure (#2168) 64ea87f592 [hip-tests] Fix memory leaks in hipMemPoolTrimTo tests (#3643) 543a7d765f rocr: Include code object allocs in lightweight coredump a58da378d4 [rocdecode] - update rocdecode ctest (#3768) f88e4ee44d [rocprofiler-systems] Make CDash submit non-fatal and add GitHub Actions logging (#3525) cb14debc3a [rocprofiler-systems] Update nlohmann-json submodule (#3391) 449253009a SWDEV-567112 - Introduce new mechanism for tagging and disabling tests - Part 2 (#3707) 8ca991393d disabling rccl from full build (linux), covered in RCCL CI (#3770) c4fdb20b74 [ATT] Re-enable tests. Add option to specify perf to target CU only (#2819) 615aab95ed ROCM-3816 Out of Memory fix (#3588) 8ffad41b24 Fix rocm_smi64 exporting invalid absolute paths to consumers (#3717) 042d76a626 rocr: Remove dependency on KFD in Runtime::VMemoryHandleMap (#2515) 555db59b2a [AMD-SMI] CPU: Added support for family 1A Models 50h-57h (#3206) 3affa2c7a3 [SWDEV-555935] Fix shared mutex and self-heal (#3729) ba0bf0f3db Replace hipMemGetInfo with ihipMemGetInfo and use it for internal calls. (#2845) c5cef9b18e Fix HIP_RETURN on all HIP API calls. (#2838) 241ce7ba83 Revert "memory: fix "contiguous_bytes" calculation in generic conversion (#3285)" (#3755) 8a690f482e [kpack/clr] Windows PE/COFF support for kpack artifact splitting and runtime loading (#3728) 863bdf8aa8 MFMA pre-processor guards for ipc.hip (#3724) 90bb9b1921 Release queue outside of vgpusAccess lock (#3705) de4523910c clr: Add build support of ROCR and PAL backends together (#3722) dfb7abc2d8 [rocprofiler-sdk] RCCL API changes for RCCL_API_TRACE_VERSION_PATCH = 3 (#3477) d69d4f23db [AICOMRCCL-633] - Fixed warnings in tests (#3402) 067d86dcaa rocr/wsl: Disable AQL Queue usage with flag ROCR_USE_PM4 (#3663) 594eb60d42 [TheRock CI] rocm-systems build full ROCm stack (#3182) 27d17e8ea0 [ROCProfiler-SDK] Fix SWDEV-556922: Handle comments before checking for pmc: (#1723) c80d90439d memory: fix "contiguous_bytes" calculation in generic conversion (#3285) 669987c83f [hip-tests] ASAN - add missing release handles (#3735) a24bbd75a4 fix local gpu release static build failure (#3667) 259b2ff913 Speed up DeviceId (#2803) 65d9264bf4 Simplify MPI trace merge logic and remove legacy guards (#3562) 1076c083cb use system to look for zcat path instead (#3720) 22f1d19db3 [AICOMRCCL-355] Enable threshold-based p2p-batching (#3000) a2e4c794d2 Partially flatten template tests cases (#2597) e242abe219 Pass space separated gfx target list to RCCL build command (#3701) 4f78aea66d SWDEV-570074 - Refactor Memset memory object handling. (#2228) b3ad12d834 Support Nvidia build on theRock for HIP-tests (#3335) a1cf15ea9a Support fp8 types in hiprtc (#2605) 8ef84b0a50 [rocprofiler-systems] Add HPC examples to automated testing (#3437) db3a70dfa0 Free memory which was allocated in tests (#3710) 27e6809c7e [rocprofiler-systems]: Fix rhel CI failure on for MPI and UCX tests (#3700) 0d9aaf59d8 rccl/topo_expl: fix build issue. (#3719) be04d75765 Fix zcat path used for checking kernel configs (#3423) cab60a7b27 rocr/thunk/win: Add CU mask support (#3518) 5b3d826c05 [CUMEM] Initial support for cuMem APIs (#2763) 0606ff491f [HIP] [PLAT-194496] Improve Stress_hipMalloc_HighSizeAlloc reliability (#3550) 05750a77cc fix hip-test name in config (#3716) 33f777f3e9 hsakmt: Remove --high functionality from run_kfdtest.sh (#2486) e4c46e3480 Hide the retain under direct dispatch check (#3698) bfe0ca0279 Add rocprof trace decoder to CI tests (#3690) a769b6f54e [rocSHMEM] Edgar/abstract allocator ipc part1 (#3411) 659fb52243 [AMD-SMI] Fix bugs, improve error handling, and clean up NIC/switch code (#3654) 0eb26ea571 hsakmt: Fix Import/Export of dmabuf_fd for WSL/Windows (#3348) a122936abb [SWDEV-567812] Add UBB power and power_limit fields to npm_info (#3262) c3bec090c5 [rocprofiler-sdk][rocprofv3][rocpd] Updates for KFD data (#340) 7c44d47740 SWDEV-547659 - Remove HIP_VERSION_GITHASH in logs (#448) 74b6487a6a SWDEV-547008 - Documentation fix for function return values (#463) af21cd44f1 SWDEV-545553 - Improve clarity and robustness of CALLBACK unit tests (#546) 180d639044 SWDEV-544900 - Change hip-test test case name (#547) feeca99950 Doc improvements (#3688) c1822b6336 ROCprofiler-SDK: deprecation of legacy tools (#3609) 5d7aff8462 Fix rocprof-compute-viewer link (#3459) 0b0b4846f0 AIRUNTIME-129 - Fix Ocl test failures of 2D image with pitches. (#3584) ac569b87e0 Fix memory tests config (#3687) 603fe7a5cf [hip-tests] Enable hipMipmappedArrayGetMemoryRequirements test via cmake 4fad4452d9 [hip] Docs: Updates to some memory management pages 8cc59559fe AICOMRCCL-656 fix memory leak in ncclCommInitRankFunc (#3628) 94a4595a5d Fix missing amd_comgr linkage in pc-sampling integration test (#3453) 2a68565dce rocrtst: CMAke file: strip xnack/feature suffixes from gfxNum in build_kernel (#3652) c3542bfb2b [rocprofv3] Deprecating input text files for counter collection (#1562) ff122e7ed7 SWDEV-573073 - Cleanup hipHostAlloc/Malloc/Register tests (#3017) 5b1deaf29d SWDEV-567112 - Introduce new mechanism for tagging and disabling tests - Part 1 - Core (#2351) 6e0cc309e1 rocrtst: MaxSingleAllocationTest: skip CPU NUMA nodes >0 (#3208) d65f601195 [AICOMRCCL-667] rccl: Change GDR selection logic. (#3607) f1c44ab200 Patch Back to Old Repo: fixes from manual runs (#3621) fe53bcd715 [AMD-SMI] Allow amdsmi init to succeed when no NIC hardware is present (#3403) b25600efdb [ROCM SMI] Fix fw pldm version not displayed in default amd-smi (#3594) 169d2ef763 root to module wiring, remove legacy source collection (#3482) 7469781988 [LRT][clr] SWDEV-512963-Fix CTS test failures for 1D buffer copy (#3520) c8f55d9b86 Adding rocprof trace decoder (#3576) 425e983502 Trace decoder codeowners (#3600) a176efd648 [hip-tests] Add return statements to HIP_SKIP_TEST (#3647) 32687cf183 rocrtst: CPUAccessToGPUMemoryTest: Cap host allocation to 512 MB under ASAN (#3407) 97c0206753 Update codeowners for thunk DXG (#3334) be44b28bb6 [rocdecode][rocjpeg] - ctest CMakeLists cleanup (#3632) 80ff0b8942 Various memory leak fixes in hip-tests (#3605) 0988f67a85 fix typo in help text (#3314) 9f823c53f1 Fix CUID file lookup by loading files before searching entries (#3436) 064c89261b SWDEV-546177 - hipModuleGetLoadingMode API impl (#653) 006213e112 ROCM-2696: Ignare size and base if null ptr (#3336) 6060b99d83 Improve atomic min max test perf (#2580) 3fbcc13602 Change printf capture impl (#1127) 93bc01937c (tag: hip-version_7.12.60610, origin/users/mradosav-amd/rocprofsys-selective-region) [ROCM-CORE] Update rdhc script to support rocm install prefix (ROCm/rocm-systems#3596) [AICOMRCCL-355]: https://amd-hub.atlassian.net/browse/AICOMRCCL-355?atlOrigin=eyJpIjoiNWRkNTljNzYxNjVmNDY3MDlhMDU5Y2ZhYzA5YTRkZjUiLCJwIjoiZ2l0aHViLWNvbS1KU1cifQ

Integrate sccache support in build_prod_wheels.py for ROCm compiler…

10c483d

… wrapping

github-project-automation Bot added this to TheRock Triage Feb 18, 2026

github-project-automation Bot moved this to TODO in TheRock Triage Feb 18, 2026

subodh-dubey-amd added 3 commits February 18, 2026 09:55

Refactor permission settings in create_sccache_wrapper for improved…

958fbf9

… readability

try/finally now starts immediately after setup_rocm_sccache(), so if …

3fc8939

…--start-server or --zero-stats fail, compilers are still restored

Check for ccache installation before building with --use-ccache in bu…

d7f3e69

…ild_prod_wheels.py

subodh-dubey-amd marked this pull request as ready for review February 18, 2026 10:30

subodh-dubey-amd requested a review from ScottTodd February 18, 2026 10:30

ScottTodd reviewed Feb 18, 2026

View reviewed changes

ScottTodd requested a review from amd-shiraz February 18, 2026 19:38

ScottTodd mentioned this pull request Feb 18, 2026

Add HIP_CLANG_LAUNCHER for launching Clang through a wrapper executable ROCm/llvm-project#14

Closed

subodh-dubey-amd added 2 commits February 19, 2026 06:53

Enhance sccache setup in setup_sccache_rocm.py and `build_prod_whee…

e01ae7c

…ls.py` to include backup and restore functionality for ROCm compilers. Update usage instructions and add a flag to skip compiler wrapping.

Detailed usage instructions for --use-sccache and `--sccache-no-wra…

2143819

…p` flags, clarifying their functionalities and mutual exclusivity with `--use-ccache`.

subodh-dubey-amd requested a review from ScottTodd February 19, 2026 07:42

ScottTodd approved these changes Feb 19, 2026

View reviewed changes

subodh-dubey-amd merged commit 2a15973 into main Feb 20, 2026
103 of 108 checks passed

subodh-dubey-amd deleted the users/subodh-dubey-amd/pytorch-build-sccache-support branch February 20, 2026 08:39

github-project-automation Bot moved this from TODO to Done in TheRock Triage Feb 20, 2026

This was referenced Feb 20, 2026

Wire sccache into PyTorch wheel build workflows #3532

Merged

Add sccache support for PyTorch builds #3171

Closed

rahulc-gh mentioned this pull request Mar 4, 2026

Bump rocm-systems submodule from 93bc019 to 093b66caa3 #3754

Merged

ScottTodd mentioned this pull request Mar 4, 2026

Use/evaluate HIP_CLANG_LAUNCHER with ccache/sccache #3760

Open

Conversation

subodh-dubey-amd commented Feb 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Test Result

Submission Checklist

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

subodh-dubey-amd Feb 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

GZGavinZhao commented Feb 18, 2026

Uh oh!

subodh-dubey-amd commented Feb 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

GZGavinZhao commented Feb 19, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

subodh-dubey-amd commented Feb 18, 2026 •

edited

Loading

subodh-dubey-amd Feb 19, 2026 •

edited

Loading

subodh-dubey-amd commented Feb 19, 2026 •

edited

Loading