github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #15277] HIP build missing -ffast-math (CUDA has -use_fast_math) #56285

New Issue
Open
opened 2026-04-29 10:34:33 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-29 10:34:33 -05:00
Owner
Copy Link

Originally created by @jodendaal on GitHub (Apr 3, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15277

Description

The CUDA backend build uses -use_fast_math (ggml-cuda/CMakeLists.txt line 123) but the HIP/ROCm backend does not have the equivalent -ffast-math flag. This is a parity gap that leaves performance on the table for all AMD GPU users.

Impact

Benchmarked on an RX 7900 XTX (gfx1100, RDNA3) with ROCm 6.4, using qwen2.5-coder:14b (Q4_K_M):

Benchmark Without -ffast-math With -ffast-math Improvement
code-gen 60.81 t/s 64.20 t/s +5.6%
short-explanation 61.01 t/s 63.44 t/s +4.0%
list-generation 61.59 t/s 63.68 t/s +3.4%

This is a free ~4% generation speed improvement for all ROCm users with a one-line change.

Fix

Add to ml/backend/ggml/ggml/src/ggml-hip/CMakeLists.txt:

set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -ffast-math")

PR: #15276

Originally created by @jodendaal on GitHub (Apr 3, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15277 ## Description The CUDA backend build uses `-use_fast_math` ([ggml-cuda/CMakeLists.txt line 123](https://github.com/ollama/ollama/blob/main/ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt#L123)) but the HIP/ROCm backend does not have the equivalent `-ffast-math` flag. This is a parity gap that leaves performance on the table for all AMD GPU users. ## Impact Benchmarked on an RX 7900 XTX (gfx1100, RDNA3) with ROCm 6.4, using `qwen2.5-coder:14b` (Q4_K_M): | Benchmark | Without -ffast-math | With -ffast-math | Improvement | |---|---|---|---| | code-gen | 60.81 t/s | 64.20 t/s | +5.6% | | short-explanation | 61.01 t/s | 63.44 t/s | +4.0% | | list-generation | 61.59 t/s | 63.68 t/s | +3.4% | This is a free ~4% generation speed improvement for all ROCm users with a one-line change. ## Fix Add to `ml/backend/ggml/ggml/src/ggml-hip/CMakeLists.txt`: ```cmake set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -ffast-math") ``` PR: #15276
GiteaMirror commented 2026-04-29 10:34:35 -05:00
Author
Owner
Copy Link

@Jasdfgh commented on GitHub (Apr 9, 2026):

good catch. the CUDA/HIP parity gap on fast math is a clean example of free performance left on the table. ~4% generation speed for a one-liner is hard to argue against.

-ffast-math on HIP enables a similar class of optimizations to what -use_fast_math does on CUDA — FMA contraction, relaxed floating-point associativity, approximate reciprocals. the accuracy impact is usually small compared to quantization noise in typical inference workloads, though it can affect bitwise reproducibility.

the PR #15276 linked — has it gotten maintainer attention yet?

<!-- gh-comment-id:4214589983 --> @Jasdfgh commented on GitHub (Apr 9, 2026): good catch. the CUDA/HIP parity gap on fast math is a clean example of free performance left on the table. ~4% generation speed for a one-liner is hard to argue against. -ffast-math on HIP enables a similar class of optimizations to what -use_fast_math does on CUDA — FMA contraction, relaxed floating-point associativity, approximate reciprocals. the accuracy impact is usually small compared to quantization noise in typical inference workloads, though it can affect bitwise reproducibility. the PR #15276 linked — has it gotten maintainer attention yet?
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#56285
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?