Increase performance for Gemma3n models on NVGPUs by enabling CUDA Graph execution #11525
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
Similar to ggml-org/llama.cpp#14741, though ollama has a slightly different model graph than llama.cpp which requires different workaround checks.
This should make the heuristics more robust
mxyng
approved these changes
Jul 29, 2025
Contributor
mxyng
left a comment
mxyng
left a comment
There was a problem hiding this comment.
this is awesome! thanks for sharing
gabe-l-hart
added a commit
to gabe-l-hart/ollama
that referenced
this pull request
Jul 30, 2025
* origin/main: Revert "CI: switch back to x86 macos builder" (ollama#11588) mac: disable bf16 on unsupported OS versions (ollama#11585) CI: switch back to x86 macos builder (ollama#11572) Increase performance for Gemma3n models on NVGPUs by enabling CUDA Graph execution (ollama#11525) kvcache: Don't shift empty batches docs: fix typos and remove trailing whitespaces (ollama#11554) Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
3 tasks
rick-github
pushed a commit
to rick-github/ollama
that referenced
this pull request
Aug 20, 2025
…aph execution (ollama#11525) * Enable CUDA Graphs for gemma3n. Similar to ggml-org/llama.cpp#14741, though ollama has a slightly different model graph than llama.cpp which requires different workaround checks. * Remove residual check by reshaping differently in gemma3n model This should make the heuristics more robust
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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Outdated suggestions cannot be applied.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.
This PR enables the execution of Gemma3n as CUDA Graphs on NVGPUs by porting ggml-org/llama.cpp#14741 to ollama. Since the model graph is defined differently in ollama compared to llama.cpp, the heuristic used to identify and exclude the
per_layer_projectionfrom batch-size determination needed to be modified a bit. As a consequence, the patch will need to be maintained even after llama.cpp is updated to a commit that contains ggml-org/llama.cpp#14741.On a RTX PRO 6000 Max-Q under Windows, this PR improves perf by ~2.5x, see
Thanks @mxyng for providing changes to gemma3n model graph definition in c4de3ea that make the checking more robust.