[Kernel] Skip fully masked key blocks in windowed Triton prefill (#50776)
Signed-off-by: almogtavor <almogtavor@gmail.com>
This commit is contained in:
@@ -113,8 +113,17 @@ def _fwd_kernel(
|
||||
# Apply causal attention pruning and sliding window attention pruning
|
||||
end_n = tl.minimum(end_n, (start_m + 1) * BLOCK_M) if IS_CAUSAL else end_n
|
||||
|
||||
# Calculate the start position for backward sliding window
|
||||
# Calculate the start position for backward sliding window.
|
||||
# Keys outside the window are fully masked below, so skip those blocks
|
||||
# rather than loading them and discarding the result. Without this a
|
||||
# windowed pass still costs O(seq_len^2).
|
||||
start_n_limit = 0
|
||||
if SLIDING_WINDOW_Q > 0:
|
||||
first_needed = start_m * BLOCK_M - SLIDING_WINDOW_Q
|
||||
start_n_limit = tl.maximum(0, (first_needed // BLOCK_N) * BLOCK_N)
|
||||
if SLIDING_WINDOW_K > 0:
|
||||
last_needed = (start_m + 1) * BLOCK_M - 1 + SLIDING_WINDOW_K
|
||||
end_n = tl.minimum(end_n, last_needed + 1)
|
||||
end_n_limit = block_mask * end_n
|
||||
|
||||
for start_n in range(start_n_limit, end_n_limit, BLOCK_N):
|
||||
|
||||
Reference in New Issue
Block a user