Repository navigation
ggml : fix ggml_backend_buft_get_alloc_size() guard - #28038
Merged
Merged
Conversation
github-actions
Bot
added
ggml
thecodacus
pushed a commit
to thecodacus/llama.cpp
that referenced
this pull request
Sep 7, 2026
SteelPh0enix
pushed a commit
to SteelPh0enix/llama.cpp-qwen4exp
that referenced
this pull request
Sep 8, 2026
(cherry picked from commit 6d1479c)
SteelPh0enix
pushed a commit
to SteelPh0enix/llama.cpp-qwen4exp
that referenced
this pull request
Sep 8, 2026
(cherry picked from commit 6d1479c)
zbrad
pushed a commit
to zbrad/llama.cpp
that referenced
this pull request
Sep 10, 2026
pl752
pushed a commit
to pl752/llama.cpp
that referenced
this pull request
Sep 15, 2026
zsogitbe
pushed a commit
to zsogitbe/llama.cpp
that referenced
this pull request
Sep 17, 2026
Jesssullivan
added a commit
to Jesssullivan/llama.cpp
that referenced
this pull request
Sep 20, 2026
PLACEHOLDER MESSAGE — rewrite before opening a pull request. llama.cpp CONTRIBUTING prohibits AI-written commit messages and PR descriptions; this text was prepared by tooling so the branch could be pushed, and the author is expected to replace it with their own words (git commit --amend). ggml_backend_buft_get_alloc_size guards the documented lower bound of the buffer-type contract with a plain assert(), which is compiled out under NDEBUG, while the upper bound two lines below is a GGML_ASSERT. A buffer type that under-reports therefore passes silently in release builds and only fails later, in the in-place-reuse path, at ggml-alloc.c GGML_ASSERT(parent_size >= node_size). Promote the lower bound to GGML_ASSERT so the failure lands on the buffer type that produced the wrong size. The neighbouring assert already evaluates ggml_nbytes in release builds, so there is no added cost. ref: ggml-org#27960 ref: ggml-org#28038 ref: ggml-org#17884 Assisted-by: Claude Opus 5
Jesssullivan
added a commit
to Jesssullivan/llama.cpp
that referenced
this pull request
Sep 20, 2026
… at ggml_nbytes PLACEHOLDER MESSAGE — rewrite before opening a pull request. llama.cpp CONTRIBUTING prohibits AI-written commit messages and PR descriptions; this text was prepared by tooling so the branch could be pushed, and the author is expected to replace it with their own words (git commit --amend). The process-global memo added in ggml-org#18626 keys RPC_CMD_GET_ALLOC_SIZE results on {device, type, op, op_params, ne} and omits nb. Two tensors with identical ne but different strides therefore share one entry, and the second caller receives the first tensor's size. When that size is smaller than ggml_nbytes, the buffer-type contract documented in ggml-org#27960 and ggml-org#28038 is violated and allocation fails downstream. Add nb to the cache key and never return less than ggml_nbytes. The memo is preserved: contiguous tensors, which are the common case, keep the same hit rate. rpc_tensor already carries nb over the wire, so there is no protocol or API change and RPC_PROTO_* is untouched. Reproduced against a local ggml-rpc-server -d CPU: a packed tensor reports 16384/16384 and a strided view with the same ne reports nbytes 32512, alloc_size 16384, then aborts; with this change the second call reports 32512 and the run completes. The regression is folded into tests/test-rpc-multi-server. ref: ggml-org#28360 ref: ggml-org#18626 ref: ggml-org#27960 Assisted-by: Claude Opus 5
Jesssullivan
added a commit
to Jesssullivan/llama.cpp
that referenced
this pull request
Sep 22, 2026
PLACEHOLDER MESSAGE — rewrite before opening a pull request. llama.cpp CONTRIBUTING prohibits AI-written commit messages and PR descriptions; this text was prepared by tooling so the branch could be pushed, and the author is expected to replace it with their own words (git commit --amend). ggml_backend_buft_get_alloc_size guards the documented lower bound of the buffer-type contract with a plain assert(), which is compiled out under NDEBUG, while the upper bound two lines below is a GGML_ASSERT. A buffer type that under-reports therefore passes silently in release builds and only fails later, in the in-place-reuse path, at ggml-alloc.c GGML_ASSERT(parent_size >= node_size). Promote the lower bound to GGML_ASSERT so the failure lands on the buffer type that produced the wrong size. The neighbouring assert already evaluates ggml_nbytes in release builds, so there is no added cost. ref: ggml-org#27960 ref: ggml-org#28038 ref: ggml-org#17884 Assisted-by: Claude Opus 5
Jesssullivan
added a commit
to Jesssullivan/llama.cpp
that referenced
this pull request
Sep 22, 2026
… at ggml_nbytes PLACEHOLDER MESSAGE — rewrite before opening a pull request. llama.cpp CONTRIBUTING prohibits AI-written commit messages and PR descriptions; this text was prepared by tooling so the branch could be pushed, and the author is expected to replace it with their own words (git commit --amend). The process-global memo added in ggml-org#18626 keys RPC_CMD_GET_ALLOC_SIZE results on {device, type, op, op_params, ne} and omits nb. Two tensors with identical ne but different strides therefore share one entry, and the second caller receives the first tensor's size. When that size is smaller than ggml_nbytes, the buffer-type contract documented in ggml-org#27960 and ggml-org#28038 is violated and allocation fails downstream. Add nb to the cache key and never return less than ggml_nbytes. The memo is preserved: contiguous tensors, which are the common case, keep the same hit rate. rpc_tensor already carries nb over the wire, so there is no protocol or API change and RPC_PROTO_* is untouched. Reproduced against a local ggml-rpc-server -d CPU: a packed tensor reports 16384/16384 and a strided view with the same ne reports nbytes 32512, alloc_size 16384, then aborts; with this change the second call reports 32512 and the run completes. The regression is folded into tests/test-rpc-multi-server. ref: ggml-org#28360 ref: ggml-org#18626 ref: ggml-org#27960 Assisted-by: Claude Opus 5
frostyautumnleaf
pushed a commit
to frostyautumnleaf/llama.cpp
that referenced
this pull request
Oct 5, 2026
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
1 participant
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
cont #27960
Didn't take into account that CUDA pads quantized tensors.
Fix: https://github.com/ggml-org/llama.cpp/actions/runs/33296587275/job/99217196889#step:3:31473
Requirements