[Fix] dsv4: Skip DP-attention pad rows in the engram hasher and low-ratio sources (#40424)
Co-authored-by: jiashuai <jiashuai@baidu.com> Co-authored-by: hnyls2002 <lsyincs@gmail.com>
This commit is contained in:
co-authored by
jiashuai
hnyls2002
parent
68623cae92
commit
29e78045fb
@@ -3273,6 +3273,12 @@ class DeepseekV4AttnBackend(
|
||||
return
|
||||
hoisted_req = getattr(meta, "low_ratio_req_indices", None)
|
||||
hoisted_pos = getattr(meta, "low_ratio_pos_i64", None)
|
||||
if hoisted_pos is not None and hoisted_pos.shape[0] < positions.shape[0]:
|
||||
# DP attention pads an eager extend past the rows extend_seq_lens cover.
|
||||
num_real = hoisted_pos.shape[0]
|
||||
x, positions = x[:num_real], positions[:num_real]
|
||||
if q_lora is not None:
|
||||
q_lora = q_lora[:num_real]
|
||||
if (
|
||||
hoisted_req is not None
|
||||
and hoisted_pos is not None
|
||||
|
||||
@@ -308,13 +308,14 @@ class EngramHasher(nn.Module):
|
||||
assert self.history is not None, "EngramHasher.init_history was not called"
|
||||
n = self.max_ngram_size
|
||||
num_tokens = input_ids.shape[0]
|
||||
if num_tokens == 0:
|
||||
return torch.empty(
|
||||
(0, self.primes.shape[0], self.offsets.shape[1]),
|
||||
mode = forward_batch.forward_mode
|
||||
if num_tokens == 0 or mode.is_idle():
|
||||
# A DP-attention idle batch carries pad rows; nothing reads its hashes.
|
||||
return torch.zeros(
|
||||
(num_tokens, self.primes.shape[0], self.offsets.shape[1]),
|
||||
dtype=torch.int64,
|
||||
device=input_ids.device,
|
||||
)
|
||||
mode = forward_batch.forward_mode
|
||||
if (
|
||||
mode.is_extend()
|
||||
and not mode.is_target_verify()
|
||||
|
||||
Reference in New Issue
Block a user