zR
commited on
Commit
•
ade85af
1
Parent(s):
1127073
fix device problem
Browse files- modeling_chatglm.py +126 -1
modeling_chatglm.py
CHANGED
@@ -332,6 +332,128 @@ class CoreAttention(torch.nn.Module):
|
|
332 |
|
333 |
return context_layer
|
334 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
335 |
|
336 |
class SelfAttention(torch.nn.Module):
|
337 |
"""Parallel self-attention layer abstract class.
|
@@ -697,6 +819,8 @@ class ChatGLMPreTrainedModel(PreTrainedModel):
|
|
697 |
config_class = ChatGLMConfig
|
698 |
base_model_prefix = "transformer"
|
699 |
_no_split_modules = ["GLMBlock"]
|
|
|
|
|
700 |
|
701 |
def _init_weights(self, module: nn.Module):
|
702 |
"""Initialize the weights."""
|
@@ -868,7 +992,8 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
868 |
self.config.eoi_token_id)
|
869 |
assert eoi_token_pos - boi_token_pos == 2
|
870 |
new_input_embeds.append(torch.cat(
|
871 |
-
(inputs_embeds[i, :boi_token_pos], images_features[i]
|
|
|
872 |
new_position_ids.append(torch.cat(
|
873 |
(position_ids[i, :boi_token_pos + 1], position_ids[i, boi_token_pos + 1].repeat(num_patches),
|
874 |
position_ids[i, eoi_token_pos:])
|
|
|
332 |
|
333 |
return context_layer
|
334 |
|
335 |
+
class SdpaAttention(CoreAttention):
|
336 |
+
def forward(self, query_layer, key_layer, value_layer, attention_mask):
|
337 |
+
if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
|
338 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
339 |
+
is_causal=True,
|
340 |
+
dropout_p=self.config.attention_dropout if self.training else 0.0)
|
341 |
+
else:
|
342 |
+
if attention_mask is not None:
|
343 |
+
attention_mask = ~attention_mask
|
344 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
345 |
+
attention_mask,
|
346 |
+
dropout_p=self.config.attention_dropout if self.training else 0.0)
|
347 |
+
context_layer = context_layer.transpose(1, 2).contiguous()
|
348 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
349 |
+
context_layer = context_layer.reshape(*new_context_layer_shape)
|
350 |
+
return context_layer
|
351 |
+
|
352 |
+
|
353 |
+
def _get_unpad_data(attention_mask):
|
354 |
+
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
355 |
+
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
356 |
+
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
357 |
+
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
|
358 |
+
return (
|
359 |
+
indices,
|
360 |
+
cu_seqlens,
|
361 |
+
max_seqlen_in_batch,
|
362 |
+
)
|
363 |
+
|
364 |
+
|
365 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2
|
366 |
+
class FlashAttention2(CoreAttention):
|
367 |
+
def __init__(self, *args, **kwargs):
|
368 |
+
super().__init__(*args, **kwargs)
|
369 |
+
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
370 |
+
|
371 |
+
def forward(self, query_states, key_states, value_states, attention_mask):
|
372 |
+
query_states = query_states.transpose(1, 2)
|
373 |
+
key_states = key_states.transpose(1, 2)
|
374 |
+
value_states = value_states.transpose(1, 2)
|
375 |
+
batch_size, query_length = query_states.shape[:2]
|
376 |
+
if not self._flash_attn_uses_top_left_mask:
|
377 |
+
causal = self.is_causal
|
378 |
+
else:
|
379 |
+
# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
|
380 |
+
causal = self.is_causal and query_length != 1
|
381 |
+
dropout = self.config.attention_dropout if self.training else 0.0
|
382 |
+
# Contains at least one padding token in the sequence
|
383 |
+
if attention_mask is not None:
|
384 |
+
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
|
385 |
+
query_states, key_states, value_states, attention_mask, query_length
|
386 |
+
)
|
387 |
+
|
388 |
+
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
389 |
+
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
390 |
+
|
391 |
+
attn_output_unpad = flash_attn_varlen_func(
|
392 |
+
query_states,
|
393 |
+
key_states,
|
394 |
+
value_states,
|
395 |
+
cu_seqlens_q=cu_seqlens_q,
|
396 |
+
cu_seqlens_k=cu_seqlens_k,
|
397 |
+
max_seqlen_q=max_seqlen_in_batch_q,
|
398 |
+
max_seqlen_k=max_seqlen_in_batch_k,
|
399 |
+
dropout_p=dropout,
|
400 |
+
softmax_scale=None,
|
401 |
+
causal=causal,
|
402 |
+
)
|
403 |
+
|
404 |
+
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
|
405 |
+
else:
|
406 |
+
attn_output = flash_attn_func(
|
407 |
+
query_states, key_states, value_states, dropout, softmax_scale=None, causal=causal
|
408 |
+
)
|
409 |
+
attn_output = attn_output.reshape(batch_size, query_length, self.hidden_size_per_partition).contiguous()
|
410 |
+
return attn_output
|
411 |
+
|
412 |
+
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
413 |
+
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
|
414 |
+
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
|
415 |
+
|
416 |
+
key_layer = index_first_axis(
|
417 |
+
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
418 |
+
)
|
419 |
+
value_layer = index_first_axis(
|
420 |
+
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
421 |
+
)
|
422 |
+
if query_length == kv_seq_len:
|
423 |
+
query_layer = index_first_axis(
|
424 |
+
query_layer.reshape(batch_size * kv_seq_len, self.num_attention_heads_per_partition, head_dim),
|
425 |
+
indices_k
|
426 |
+
)
|
427 |
+
cu_seqlens_q = cu_seqlens_k
|
428 |
+
max_seqlen_in_batch_q = max_seqlen_in_batch_k
|
429 |
+
indices_q = indices_k
|
430 |
+
elif query_length == 1:
|
431 |
+
max_seqlen_in_batch_q = 1
|
432 |
+
cu_seqlens_q = torch.arange(
|
433 |
+
batch_size + 1, dtype=torch.int32, device=query_layer.device
|
434 |
+
) # There is a memcpy here, that is very bad.
|
435 |
+
indices_q = cu_seqlens_q[:-1]
|
436 |
+
query_layer = query_layer.squeeze(1)
|
437 |
+
else:
|
438 |
+
# The -q_len: slice assumes left padding.
|
439 |
+
attention_mask = attention_mask[:, -query_length:]
|
440 |
+
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
|
441 |
+
|
442 |
+
return (
|
443 |
+
query_layer,
|
444 |
+
key_layer,
|
445 |
+
value_layer,
|
446 |
+
indices_q,
|
447 |
+
(cu_seqlens_q, cu_seqlens_k),
|
448 |
+
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
|
449 |
+
)
|
450 |
+
|
451 |
+
|
452 |
+
CORE_ATTENTION_CLASSES = {
|
453 |
+
"eager": CoreAttention,
|
454 |
+
"sdpa": SdpaAttention,
|
455 |
+
"flash_attention_2": FlashAttention2
|
456 |
+
}
|
457 |
|
458 |
class SelfAttention(torch.nn.Module):
|
459 |
"""Parallel self-attention layer abstract class.
|
|
|
819 |
config_class = ChatGLMConfig
|
820 |
base_model_prefix = "transformer"
|
821 |
_no_split_modules = ["GLMBlock"]
|
822 |
+
_supports_flash_attn_2 = True
|
823 |
+
_supports_sdpa = True
|
824 |
|
825 |
def _init_weights(self, module: nn.Module):
|
826 |
"""Initialize the weights."""
|
|
|
992 |
self.config.eoi_token_id)
|
993 |
assert eoi_token_pos - boi_token_pos == 2
|
994 |
new_input_embeds.append(torch.cat(
|
995 |
+
(inputs_embeds[i, :boi_token_pos], images_features[i].to(inputs_embeds.device),
|
996 |
+
inputs_embeds[i, eoi_token_pos + 1:])))
|
997 |
new_position_ids.append(torch.cat(
|
998 |
(position_ids[i, :boi_token_pos + 1], position_ids[i, boi_token_pos + 1].repeat(num_patches),
|
999 |
position_ids[i, eoi_token_pos:])
|