llama.cpp
CUDA: keep the MUL_MAT_ID precision in the per-expert fallback
#31345
Merged

CUDA: keep the MUL_MAT_ID precision in the per-expert fallback #31345

JonSnow1807
JonSnow1807 tests: add CUDA-only MUL_MAT/MUL_MAT_ID cases with results above the …
bfa229a7
JonSnow1807 CUDA: use an fp32 result for quantized src0 in the cuBLAS fallback
9f7c2188
JonSnow1807 JonSnow1807 requested a review from ggerganov ggerganov 23 days ago
JonSnow1807 JonSnow1807 requested a review 23 days ago
github-actions github-actions added testing
github-actions github-actions added ggml
github-actions github-actions added CUDA
ggml-gh-bot
JonSnow1807
JohannesGaessler
JohannesGaessler commented on 2026-11-11
JonSnow1807 Merge remote-tracking branch 'origin/master' into fix-cuda-cublas-qua…
76be1389
JonSnow1807 JonSnow1807 changed the title CUDA: use an fp32 result for quantized src0 in the cuBLAS fallback CUDA: keep the MUL_MAT_ID precision in the per-expert fallback 9 days ago
JonSnow1807 CUDA: keep the MUL_MAT_ID precision in the per-expert fallback
c61aba2b
JonSnow1807 JonSnow1807 force pushed from 523524eb to c61aba2b 9 days ago
JonSnow1807
JohannesGaessler
JohannesGaessler approved these changes on 2026-11-17
JohannesGaessler
JohannesGaessler JohannesGaessler added merge ready
ggerganov ggerganov merged 9fbe30c2 into master 1 day ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone