diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-21 16:38:45 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-21 16:38:45 +0200 |
| commit | a2907badacf6bb86e90d82f054419955be572d50 (patch) | |
| tree | 4bea85140cf0908f62cc18d87ad07dc3d11c0877 /tests | |
| parent | a09b41dc4824eb33e11b1e71d95ec94d2411ead8 (diff) | |
| download | ffwd-a2907badacf6bb86e90d82f054419955be572d50.tar.gz ffwd-a2907badacf6bb86e90d82f054419955be572d50.zip | |
cuda: fix bf16 BERT batched-attention buffer-layout mismatch
Diffstat (limited to 'tests')
| -rwxr-xr-x | tests/check_cuda_fast_gemm_parity.py | 6 |
1 files changed, 6 insertions, 0 deletions
diff --git a/tests/check_cuda_fast_gemm_parity.py b/tests/check_cuda_fast_gemm_parity.py index bc201f7..57adf72 100755 --- a/tests/check_cuda_fast_gemm_parity.py +++ b/tests/check_cuda_fast_gemm_parity.py @@ -42,6 +42,12 @@ DEFAULT_TEXTS = [ # (BERT/MiniLM/BGE max_position 512, XLM-R 514) accept it; larger 8x/14x inputs # overflowed their position tables and were (correctly) rejected, which looked # like a CUDA failure. Still long enough to stress packed-sequence attention. +# +# NOTE on coverage: these defaults are all DIFFERENT lengths, so they exercise +# only the per-sequence (ragged) attention path. The CUDA batched-attention path +# (cublasGemmBatchedEx) engages only when every sequence in the batch is the SAME +# length (see attn_batched_setup). To exercise it, pass >=2 identical texts as +# positional args, e.g. the same sentence four times with --batch-size 4. def run_cuda(binary: str, model_dir: str, mode: str, weights: str, |
