diff --git a/gdpa_megakernel/src/tlx_gdpa_megakernel.py b/gdpa_megakernel/src/tlx_gdpa_megakernel.py index a367ed2..208df70 100644 --- a/gdpa_megakernel/src/tlx_gdpa_megakernel.py +++ b/gdpa_megakernel/src/tlx_gdpa_megakernel.py @@ -7102,7 +7102,7 @@ def gdpa_backward_tlx( qlen, start_n, BLOCK_M1, BLOCK_N1, WINDOW_SIZE ) - for i in tl.range(0, num_steps, 1, num_stages=0): + for i in tl.range(0, num_steps, 1, num_stages=1): q_buf_id, q_phase = _get_bufidx_phase( accum_cnt_inner + i, NUM_BUFFERS_Q )