Repository navigation
Add hipSOLVER - #63
Merged
Merged
Conversation
stellaraccident
added a commit
that referenced
this pull request
Feb 11, 2025
* Prefetch from #63 to get docker container ready. * Temporary while working out deps.
Collaborator
|
I pushed f69fd0d individually to prep the docker container. |
stellaraccident
force-pushed
the
hipsolver
branch
from
February 11, 2025 02:29
29daf9c to
82b6c9d
Compare
stellaraccident
approved these changes
Feb 11, 2025
Collaborator
|
Still seems to need some work... |
stellaraccident
force-pushed
the
hipsolver
branch
2 times, most recently
from
February 12, 2025 01:57
6524eec to
c967ba5
Compare
stellaraccident
force-pushed
the
hipsolver
branch
from
February 12, 2025 07:00
7c63dcd to
5d5a3a5
Compare
1 task done
powderluv
added a commit
to powderluv/TheRock
that referenced
this pull request
Jul 9, 2026
…uards + diagnostics (WIP, not yet a fix) Investigation scaffolding + defensive guards for the torch process-teardown hang (matmul computes correctly, then the process hangs at interpreter shutdown -> the pytest subprocess times out, rc=-999). Mechanism (fully diagnosed): a torch finalizer launches a kernel during Python interpreter finalization; the first-use code-object load does a CPU write over the PCIe BAR that wedges un-killably in a non-alertable kernel-mode wait on the by-then-dying driver (survives os._exit / TerminateProcess / taskkill /F). These wddm-layer guards do NOT fix that -- the wedge is the CPU-BAR memory access, not an escape. The real fix is torch-side (don't launch a kernel at finalization) or a deeper loader change (SDMA code-object copy). Kept as defensive/diagnostic. gpu_init.cpp: - g_gpuBroughtUpOnce: skip a 2nd full PSP/GFX bring-up per process (recipeBootload + wddmGfxBringUp) so a teardown re-init can't re-run the recipe on a live GPU. - ROCR_LITE_SKIP_POST_BRINGUP_MEMSET (env, default off): skip the post-bring-up rcpAllocVram zeroing memset (itself a CPU-over-BAR write). - ROCR_LITE_WATCHDOG (env): native all-thread stack dumper (Toolhelp + StackWalk64 + dbghelp) that survives CPython finalization where faulthandler is blind. - ROCR_LITE_BACKTRACE (env): symbolized backtrace at rcpAllocVram-fail / wddmGfxBringUp-entry. wddm_lite.cpp: - escape-guard: fast-fail WddmLite::escape() when RtlDllShutdownInProgress() or Py_IsFinalizing() -- don't issue GPU escapes during process teardown. Defensive. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
powderluv
added a commit
to powderluv/TheRock
that referenced
this pull request
Jul 9, 2026
…utdown guards + diagnostics (WIP, not yet a fix)" This reverts commit 3b368e8.
powderluv
added a commit
to powderluv/TheRock
that referenced
this pull request
Jul 9, 2026
…nd driver guards Points the superproject at the rocm-systems commits that fix the Windows lite:: torch teardown hang (skip SyncAllStreams drain at process exit) and revert the disproven driver-object bring-up guards. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Includes workarounds for #82, #83