Fix DSpark warmup without sparse index buffer (#50693)
Signed-off-by: xijiade.aihemaiti <3146335281@qq.com> Signed-off-by: xijiaat <52128022+xijiaat@users.noreply.github.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
co-authored by
Wentao Ye
OpenAI Codex
parent
1482d2e015
commit
d40c3e3c00
@@ -96,8 +96,11 @@ class DeepseekV4FlashMLAAttention(DeepseekV4Attention):
|
||||
// self.compress_ratio
|
||||
)
|
||||
M = N + self.window_size + self.max_num_batched_tokens
|
||||
assert self.topk_indices_buffer is not None
|
||||
top_k = 0 if swa_only else self.topk_indices_buffer.shape[-1]
|
||||
if swa_only:
|
||||
top_k = 0
|
||||
else:
|
||||
assert self.topk_indices_buffer is not None
|
||||
top_k = self.topk_indices_buffer.shape[-1]
|
||||
combined_topk = round_up(top_k + self.window_size, 128)
|
||||
current_workspace_manager().get_simultaneous(
|
||||
((self.PREFILL_CHUNK_SIZE, M, q.shape[-1]), torch.bfloat16),
|
||||
|
||||
Reference in New Issue
Block a user