ggml-hip: enable -ffast-math for HIP builds - #23862
Merged
Merged
Conversation
Mirrors the -use_fast_math flag already applied to CUDA builds (ggml-cuda/CMakeLists.txt). Previously this was missing from HIP builds because fast-math had no effect in older HIP/ROCm toolchains, but this has since been resolved upstream in ROCm/LLVM. Benchmarked on gfx1151 (RDNA3.5, Strix Halo) against latest master: +5-10% prompt processing throughput on FA=1 across all context lengths up to 32K with no measurable quality impact.
github-actions
Bot
added
the
ggml
Contributor
|
Looks like you closed the wrong PR. |
This comment was marked as resolved.
This comment was marked as resolved.
Closed
Contributor
Author
Indeed, my bad, I was wondering where it went :/, thanks |
|
I think there needs to be an exception -fno-finite-math-only otherwise it can lead to undefined behavior: warning: use of infinity via a macro is undefined behavior due to the currently enabled floating-point options [-Wnan-infinity-disabled] |
IMbackK
approved these changes
Jul 6, 2026
am17an
approved these changes
Jul 6, 2026
iacopPBK
pushed a commit
to DENEB1312/mx-llama.cpp
that referenced
this pull request
Jul 7, 2026
1 task done
iacopPBK
pushed a commit
to DENEB1312/mx-llama.cpp
that referenced
this pull request
Jul 13, 2026
satindergrewal
pushed a commit
to satindergrewal/llama.cpp
that referenced
this pull request
Aug 12, 2026
zbrad
pushed a commit
to zbrad/llama.cpp
that referenced
this pull request
Sep 10, 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
This add
-ffast-mathto HIP builds, mirroring the-use_fast_mathflag already applied to CUDA builds inggml-cuda/CMakeLists.txt.The flag was previously missing from HIP because it had no effect in older toolchains. It was changed by ROCm/LLVM: https://reviews.llvm.org/D154790
I tried disuccing this in #23339, @IMbackK confirmed the historical reason and @am17an suggested adding the flag rather than using
__expfdirectly.To prove the changes I ran some benchmarks.
Benchmarks
Benchmarked on gfx1151 (RDNA3.5/Strix Halo, 40 CUs) against latest master (2f6c815). These two models cover both the FA paths.
Qwen3.5-27B Q4_K_M — head_dim=256 (TILE path), FA=1, prompt t/s:
Qwen3-0.6B BF16 — head_dim=64 (MMA path), FA=1, prompt t/s:
FA=0 results were essentially the same across both builds.
Requirements