[ROCm] Remove stale SDPA and skinny GEMM workarounds (#50907)
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com> Signed-off-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
4f2f31b82e
commit
1ab2801dde
@@ -140,7 +140,7 @@ def test_rocm_wvsplitkrc_kernel(xnorm, n, k, m, dtype, seed, padded_a, bias_mode
|
||||
GrpsShrB = min(N_p2 // 16, 4)
|
||||
# Given the above, how many CUs would we need?
|
||||
CuNeeded = rndup_cus * GrpsShrB
|
||||
# candidate for atomic reduce count splitk?
|
||||
# Deterministic reduction stores one float workspace value per K shard.
|
||||
fits_wvsplitkrc = (N_p2 * m * ((k + 512 - 1) // 512)) <= 128 * 1024 * 12
|
||||
fits_wvsplitkrc &= CuNeeded <= cu_count
|
||||
|
||||
|
||||
Reference in New Issue
Block a user