Metal: Fix flash_attn_ext assertion by applying consistent has_kvpad logic#17294
Closed
xi-guo-0 wants to merge 1 commit intoggml-org:masterfrom
Closed
Metal: Fix flash_attn_ext assertion by applying consistent has_kvpad logic#17294xi-guo-0 wants to merge 1 commit intoggml-org:masterfrom
xi-guo-0 wants to merge 1 commit intoggml-org:masterfrom
Conversation
Member
|
Fixed with #17295 |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Description:
This PR resolves an assertion failure in
ggml_metal_op_flash_attn_extthat occurs when thehas_kvpadvariable evaluates tofalse.Root Cause
I observed an inconsistency in how the
has_kvpadvariable is determined withinggml-metal-ops.cpp.In one section of the file (around line 1980), the logic was intentionally hard-coded to
true, likely as a temporary fix:However, later in the file (around line 2156), the original (and now inconsistent) logic was still being used:
Command:
./llama-simple -m ~/Library/Caches/llama.cpp/ggml-org_gemma-3-1b-it-GGUF_gemma-3-1b-it-Q4_K_M.ggufThis inconsistency was causing the following assertion to fail in my case, as the code was hitting the un-patched
elseblock: