[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:
Andreas Karatzas
2026-08-11 14:19:16 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 4f2f31b82e
commit 1ab2801dde
56 changed files with 500 additions and 558 deletions
@@ -532,7 +532,9 @@ __global__ void moe_lora_align_block_size_kernel(
// Populate the token_mask based on the token-LoRA mapping
int num_tokens = numel / topk_num;
if (threadIdx.x == 0) {
// Only the even counting block owns per-LoRA metadata. The odd block only
// initializes sorted_token_ids and must not race the final count write.
if (blockIdx.x % 2 == 0 && threadIdx.x == 0) {
total_tokens_post_pad[lora_id] = 0;
for (int i = 0; i < num_tokens; i++) {