ggml-cuda: add mem check for fusion#19916
Merged
am17an merged 3 commits intoggml-org:masterfrom Mar 6, 2026
Merged
Conversation
github-actions
bot
added
Nvidia GPU
ORippler
reviewed
Mar 2, 2026
Comment on lines
+122
to
+132
| // Sanitize NaN to -FLT_MAX so the iterative argmax produces unique expert IDs. | ||
| // NaN comparisons always return false, which would cause the same expert to be | ||
| // selected repeatedly. -FLT_MAX compares normally and is still excluded by the | ||
| // -INFINITY sentinel used after each selection round. | ||
| // More relevant for the cuBLAS path. See https://github.com/ggml-org/llama.cpp/issues/19659 | ||
| #pragma unroll | ||
| for (int i = 0; i < experts_per_thread; i++) { | ||
| if (__isnanf(wt[i])) { | ||
| wt[i] = -FLT_MAX; | ||
| } | ||
| } |
Collaborator
There was a problem hiding this comment.
- If the issue is in llama.cpp and not cuBLAS, I feel we should use
fmaxfas a NaN-safe comparator: https://docs.nvidia.com/cuda/cuda-math-api/cuda_math_api/group__CUDA__MATH__SINGLE.html#_CPPv45fmaxfff (I presume we are talking aboutval_s > max_val_slater on in this kernel?) - If the issue is in cuBLAS, I'd love more details so I can ask the cuBLAS team/take a look myself
Contributor
Author
There was a problem hiding this comment.
- Yes but it's not just
val_s > max_val_sit'sval_s > max_val_s || (val_s == max_val_s && expert < max_expert) - The linked issue has a repro. It's cuBLAS + Nemotron, so think it would be fun for you guys to look at :)
Collaborator
There was a problem hiding this comment.
Yes but it's not just val_s > max_val_s it's val_s > max_val_s || (val_s == max_val_s && expert < max_expert)
Shouldn't we be fine with fmaxf, so long as max_val & max_val_s are initialized to -FLT_MAX instead of -INFINITY at the beginning of the selection-loop over n_expert_used? At least for the case where k non-NAN values exist inside the logits for a given row. But at this point we are just pulling your proposal into the loop itself 😄
Contributor
Author
|
@JohannesGaessler can you review this PR? Apart from the NaN check it also fixes a latent bug |
JohannesGaessler
approved these changes
Mar 6, 2026
Comment on lines
+3382
to
+3386
| if ((b_start <= a_start && a_start < b_end) || (a_start <= b_start && b_start < a_end)) { | ||
| return true; | ||
| } | ||
|
|
||
| return false; |
Contributor
There was a problem hiding this comment.
Suggested change
| if ((b_start <= a_start && a_start < b_end) || (a_start <= b_start && b_start < a_end)) { | |
| return true; | |
| } | |
| return false; | |
| return (b_start <= a_start && a_start < b_end) || (a_start <= b_start && b_start < a_end); |
This would maybe be slightly simpler but either way is fine I think.
bartowski1182
pushed a commit
to bartowski1182/llama.cpp
that referenced
this pull request
Mar 10, 2026
* ggml-cuda: add mem check for fusion * Replace NaNs with -FLT_MAX * fix typo Co-authored-by: Johannes Gäßler <johannesg@5d6.de> --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Ethan-a2
pushed a commit
to Ethan-a2/llama.cpp
that referenced
this pull request
Mar 20, 2026
* ggml-cuda: add mem check for fusion * Replace NaNs with -FLT_MAX * fix typo Co-authored-by: Johannes Gäßler <johannesg@5d6.de> --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
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
3 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.
Fixes #19659