Commit
·
3649bbb
1
Parent(s):
4f25dda
Got output to match Phi2 exactly
Browse files- attention.py +25 -5
attention.py
CHANGED
@@ -370,11 +370,29 @@ class MHA(nn.Module):
|
|
370 |
dtype=kv.dtype,
|
371 |
device=kv.device,
|
372 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
373 |
kv_cache.kv_block_map[block_n][
|
374 |
-
|
375 |
-
|
376 |
...
|
377 |
] = kv
|
|
|
|
|
|
|
|
|
|
|
|
|
378 |
|
379 |
def _forward_cross_attn(
|
380 |
self,
|
@@ -396,8 +414,9 @@ class MHA(nn.Module):
|
|
396 |
],
|
397 |
dim=2,
|
398 |
)
|
399 |
-
self._update_kv_cache(kv, kv_cache, self.block_n)
|
400 |
-
|
|
|
401 |
|
402 |
if self.using_flash_attn and unpad_input and pad_input: # not touching flash attention code
|
403 |
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
@@ -528,4 +547,5 @@ class ParallelAttentionBlock(nn.Module):
|
|
528 |
key_padding_mask=key_padding_mask,
|
529 |
)
|
530 |
mlp_outputs = self.mlp(x)
|
531 |
-
|
|
|
|
370 |
dtype=kv.dtype,
|
371 |
device=kv.device,
|
372 |
)
|
373 |
+
|
374 |
+
batch_start = kv_cache.batch_size_offset
|
375 |
+
batch_end = batch_start + kv.shape[0]
|
376 |
+
sequence_start = kv_cache.seqlen_offset
|
377 |
+
sequence_end = sequence_start + kv.shape[1]
|
378 |
+
|
379 |
+
# TODO: figure out why they're doing this
|
380 |
+
if sequence_end >= kv_cache.max_seqlen:
|
381 |
+
kv_cache.kv_block_map[block_n] = torch.concatenate(
|
382 |
+
(kv_cache.kv_block_map[block_n], kv),
|
383 |
+
dim=1,
|
384 |
+
)
|
385 |
kv_cache.kv_block_map[block_n][
|
386 |
+
batch_start:batch_end,
|
387 |
+
sequence_start:sequence_end,
|
388 |
...
|
389 |
] = kv
|
390 |
+
kv = kv_cache.kv_block_map[block_n][
|
391 |
+
batch_start:batch_end,
|
392 |
+
:sequence_end,
|
393 |
+
...
|
394 |
+
]
|
395 |
+
return kv
|
396 |
|
397 |
def _forward_cross_attn(
|
398 |
self,
|
|
|
414 |
],
|
415 |
dim=2,
|
416 |
)
|
417 |
+
kv = self._update_kv_cache(kv, kv_cache, self.block_n)
|
418 |
+
|
419 |
+
causal = (kv_cache.seqlen_offset == 0)
|
420 |
|
421 |
if self.using_flash_attn and unpad_input and pad_input: # not touching flash attention code
|
422 |
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
|
|
547 |
key_padding_mask=key_padding_mask,
|
548 |
)
|
549 |
mlp_outputs = self.mlp(x)
|
550 |
+
x = self.dropout(attn_outputs + mlp_outputs) + residual
|
551 |
+
return x
|