vulkan: fix submission batching size, add debug tools for diagnosing causes of DeviceLost drivers errors - #26371
Merged
Merged
Conversation
github-actions
Bot
added
Vulkan
jeffbolznv
approved these changes
Jul 31, 2026
Contributor
Author
|
I fixed an issue with the queue owning a device shared pointer. Should be fine now. Can you reapprove if it's fine @jeffbolznv ? |
jeffbolznv
approved these changes
Aug 5, 2026
0cc4m
added
the
merge ready
Contributor
|
It seems this PR fixed the issue with my AMD cards. For quants above Q5, llama-server would silently crash without any errors, and sometimes the AMD driver would crash too. I just had a long agent session with a Q8 model without any errors. |
Maxritz
pushed a commit
to Maxritz/LLAMA-ALL-INCLUSIVE
that referenced
this pull request
Aug 8, 2026
…causes of DeviceLost drivers errors (ggml-org#26371) * vulkan: add debug tooling to get more information about a DeviceLost error * fix submission threshold applied too late * use logging macros, throw instead of aborting * clean up circular dependency
Maxritz
added a commit
to Maxritz/LLAMA-ALL-INCLUSIVE
that referenced
this pull request
Aug 8, 2026
No shared git history exists with upstream (our root commit is parentless, a fresh source import rather than a real fork/clone), so a normal merge/rebase isn't possible -- cherry-picked commits individually instead, each verified independently. Landed 7 of 8 identified commits; skipped upstream's own Q2_0 (788e07d) since it collides with our already-verified, benchmarked custom Q2_0 (type 42) implementation -- same feature name, two independently-developed and incompatible kernels. - e34e9c6 vulkan: Refactor vk_queue to use per-instance mutexes and unique handles (ggml-org#23570) -- required updating our custom paged-KV/radix Vulkan code (pkv_vulkan_init/pkv_dispatch_init call sites), which accessed the old plain vk_queue struct fields directly. - 27b7fcf vulkan: add iq4_nl support back to FA (ggml-org#24585) - 29f1212 vulkan: Support quantized concat (ggml-org#25684) - 3cb4d39 vulkan: add POOL_1D op (ggml-org#25431) - a329897 vulkan: extend topk_moe fusion to support sqrt(softplus) (ggml-org#26124) - a72c87f vulkan backend ops: implemented GATED_LINEAR_ATTN (ggml-org#25601) - 0756c85 vulkan: fix submission batching size, add debug tools for diagnosing causes of DeviceLost drivers errors (ggml-org#26371) -- AMD-specific driver-timeout workaround, directly relevant to this fork's RDNA4 focus. Verified: full Release build of build-vulkan-rdna4-fresh (0 errors), real model load + inference request through the Vulkan backend produced correct output. docs/ops.md left as the incoming (upstream) version for now since it's auto-generated (scripts/create_ops_docs.py) from real test-backend-ops CSV dumps per backend -- needs a proper regeneration pass, not hand-editing, to reflect our custom ops too. Backup of pre-sync main state: branch + tag pre-upstream-vulkan-sync (pushed to origin). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
HarshDaniel
pushed a commit
to HarshDaniel/koboldcpp-rocm
that referenced
this pull request
Aug 9, 2026
…causes of DeviceLost drivers errors (ggml-org#26371) * vulkan: add debug tooling to get more information about a DeviceLost error * fix submission threshold applied too late * use logging macros, throw instead of aborting * clean up circular dependency
miltos22
pushed a commit
to miltos22/llama.cpp-wackMall-merge-request
that referenced
this pull request
Aug 10, 2026
…causes of DeviceLost drivers errors (ggml-org#26371) * vulkan: add debug tooling to get more information about a DeviceLost error * fix submission threshold applied too late * use logging macros, throw instead of aborting * clean up circular dependency
satindergrewal
pushed a commit
to satindergrewal/llama.cpp
that referenced
this pull request
Aug 11, 2026
…causes of DeviceLost drivers errors (ggml-org#26371) * vulkan: add debug tooling to get more information about a DeviceLost error * fix submission threshold applied too late * use logging macros, throw instead of aborting * clean up circular dependency
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
4 participants
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.
Overview
We get a lot of DeviceLost error reports, especially from AMD users on Linux. Most of these are a driver submission timeout. It is currently not easy to figure out what tensor or tensors timed out, to improve that I added two things in this PR:
VK_EXT_device_faultextension to get more information about an error from the driver. Sadly this does not provide any information about the timeouts, but it may still be useful for other issues.GGML_VK_SERIALIZE_SUBMISSIONSenv var that submits all submission sequentially and stores what tensors were last submitted. If the driver then throws a DeviceLost error, it reports what was submitted. That way you can see if too much was submitted at once, or if a specific operator is running too slow.This already showed an issue with submission batching, currently it only stops a batch once it has exceeded the flops threshold, not before a tensor would push it over the limit. I patched this as well to make submissions smaller where necessary, as was intended. Possibly this already reduces the amount of driver timeouts.
Requirements