CUDA: fix bad asserts for partial offload #13337
Merged
Conversation
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
github-actions
bot
added
Nvidia GPU
slaren
approved these changes
May 6, 2025
gabe-l-hart
added a commit
to gabe-l-hart/llama.cpp
that referenced
this pull request
May 6, 2025
* origin/master: (27 commits) llama : fix build_ffn without gate (ggml-org#13336) CUDA: fix bad asserts for partial offload (ggml-org#13337) convert : qwen2/3moe : set yarn metadata if present (ggml-org#13331) CUDA: fix --split-mode row for MMQ (ggml-org#13323) gguf-py : avoid requiring pyside6 for other scripts (ggml-org#13036) CUDA: fix logic for clearing padding with -ngl 0 (ggml-org#13320) sampling : Integrate Top-nσ into main sampling chain (and add it to the server) (ggml-org#13264) server : Webui - change setText command from parent window to also send the message. (ggml-org#13309) mtmd : rename llava directory to mtmd (ggml-org#13311) clip : fix confused naming ffn_up and ffn_down (ggml-org#13290) convert : bailingmoe : set yarn metadata if present (ggml-org#13312) SYCL: Disable mul_mat kernels for noncontiguous tensor b (ggml-org#13308) mtmd : add C public API (ggml-org#13184) rpc : use backend registry, support dl backends (ggml-org#13304) ggml : activate s390x simd for Q3_K (ggml-org#13301) llava/mtmd : fixes to fully support dl backends (ggml-org#13303) llama : build windows releases with dl backends (ggml-org#13220) CUDA: fix race condition in MMQ stream-k fixup (ggml-org#13299) CUDA: fix race condition in MMQ ids_dst (ggml-org#13294) vulkan: Additional type support for unary, binary, and copy (ggml-org#13266) ...
Nexesenex
added a commit
to Nexesenex/croco.cpp
that referenced
this pull request
May 9, 2025
This reverts commit 2356fb1.
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.
Fixup to #13320 .
The requirements for clearing the padding is that it should not be done for views since that would risk clearing valid tensor data and that the tensor memory should be in one contiguous block. The latter is to avoid having to handle the edge case of potentially having to clear memory inside the tensor.
The asserts in
mmq.cuandmmvq.cuwere too strict becauseggml_is_contiguousdisallows both views and permutations. I added a new functionggml_is_contiguously_allocatedthat only disallows views but allows permutations (to my knowledge we do not yet have a utility function for this). I changed the asserts to check that the memory is allocated contiguously and that the tensor in question is not a view of another tensor (to avoid potentially overwriting valid tensor data).The assert in
ggml_mul_mat_idwas wrong because I forgot that forMUL_MAT_IDthe number of tokens is stored in dimension 2 instead of dimension 1. But since the generic MoE code effectively creates views ofsrc0I think it's better to mark the slices as such and to remove that assert entirely.I remembered that in the FlashAttention code the conversion of quantized KV data to FP16 implicitly assumes a contiguous block of memory and added asserts with the new function.