vulkan: fix missing view-alias dependencies in ggml_vk_graph_optimize - #27812
Merged
0cc4m merged 3 commits intoAug 28, 2026
Conversation
is_src_of doesn't treat two views of one tensor as dependent, so the optimizer reorders nodes across aliased reads and writes. Result: silently wrong tokens under greedy decoding, different output on every server start, and invalid speculative-decoding acceptance, with nothing logged. Hits Qwen3.8's recurrent state (and any model with view-aliased state) on AMD and NVIDIA Vulkan. CUDA is clean. Compare view_src bases on both sides. Fixes ggml-org#27805
github-actions
Bot
added
Vulkan
jeffbolznv
reviewed
Aug 27, 2026
Nodes whose op is NONE, RESHAPE, TRANSPOSE, VIEW or PERMUTE execute nothing, so aliasing through them is not a real dependency. The previous base comparison matched them anyway, which only costs the optimizer reordering freedom. Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
…c_of Code will not compile without these changes. is_src_of has an empty capture list, so is_empty was not visible inside it, and is_empty took a non-const pointer, while is_src_of receives const ones. Other call sites pass non-const pointers, which still convert as usual.
jeffbolznv
approved these changes
Aug 28, 2026
0cc4m
approved these changes
Aug 28, 2026
jbooth
pushed a commit
to jbooth/llama.cpp
that referenced
this pull request
Aug 30, 2026
…ggml-org#27812) * vulkan: fix missing view-alias dependencies in ggml_vk_graph_optimize is_src_of doesn't treat two views of one tensor as dependent, so the optimizer reorders nodes across aliased reads and writes. Result: silently wrong tokens under greedy decoding, different output on every server start, and invalid speculative-decoding acceptance, with nothing logged. Hits Qwen3.8's recurrent state (and any model with view-aliased state) on AMD and NVIDIA Vulkan. CUDA is clean. Compare view_src bases on both sides. Fixes ggml-org#27805 * vulkan: don't treat view/no-op nodes as aliasing dependencies Nodes whose op is NONE, RESHAPE, TRANSPOSE, VIEW or PERMUTE execute nothing, so aliasing through them is not a real dependency. The previous base comparison matched them anyway, which only costs the optimizer reordering freedom. Co-authored-by: Jeff Bolz <jbolz@nvidia.com> * vulkan: make the lambda parameter const and capture is_empty in is_src_of Code will not compile without these changes. is_src_of has an empty capture list, so is_empty was not visible inside it, and is_empty took a non-const pointer, while is_src_of receives const ones. Other call sites pass non-const pointers, which still convert as usual. --------- Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
This was referenced Aug 31, 2026
Nathanw1014
pushed a commit
to Nathanw1014/llama.cpp
that referenced
this pull request
Aug 31, 2026
…ggml-org#27812) * vulkan: fix missing view-alias dependencies in ggml_vk_graph_optimize is_src_of doesn't treat two views of one tensor as dependent, so the optimizer reorders nodes across aliased reads and writes. Result: silently wrong tokens under greedy decoding, different output on every server start, and invalid speculative-decoding acceptance, with nothing logged. Hits Qwen3.8's recurrent state (and any model with view-aliased state) on AMD and NVIDIA Vulkan. CUDA is clean. Compare view_src bases on both sides. Fixes ggml-org#27805 * vulkan: don't treat view/no-op nodes as aliasing dependencies Nodes whose op is NONE, RESHAPE, TRANSPOSE, VIEW or PERMUTE execute nothing, so aliasing through them is not a real dependency. The previous base comparison matched them anyway, which only costs the optimizer reordering freedom. Co-authored-by: Jeff Bolz <jbolz@nvidia.com> * vulkan: make the lambda parameter const and capture is_empty in is_src_of Code will not compile without these changes. is_src_of has an empty capture list, so is_empty was not visible inside it, and is_empty took a non-const pointer, while is_src_of receives const ones. Other call sites pass non-const pointers, which still convert as usual. --------- Co-authored-by: Jeff Bolz <jbolz@nvidia.com> (cherry picked from commit b387ddf)
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
3 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
Fixes #27805
is_src_ofinggml_vk_graph_optimizechecks whether one node depends on another before the optimizer reorders them. It knows two cases: a node depends on its direct inputs, and two views of the same tensor share memory. It misses a third case: a node that reads or writes memory through one view depends on a node that writes or reads the same memory through a different view. The optimizer can then run those two nodes in the wrong order.Result: wrong tokens at temperature 0, a different answer on every server start, and speculative-decoding acceptance numbers that mean nothing. Nothing is written to the log.
Found with the DFlash 2 verify graph from #27342, which does not touch ggml-vulkan.cpp. Any model that keeps state in views is exposed (Qwen3.8's recurrent state where I encountered it). AMD and NVIDIA Vulkan affected; CUDA on the same NVIDIA GPU is not.
Fix: compare the underlying tensor (
view_src) on both sides of the check, in addition to the existing tests.Additional information
Verified 12/12 fresh runs match the no-speculation output on RTX 5000 (Vulkan) and RX 7800 XT, where the unpatched build gave 4 distinct outputs and 0/12 matches. Plain and MTP output unchanged.
Requirements