mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-09-05 19:13:01 +00:00
[Shardformer] Merge flash attention branch to pipeline branch (#4362)
* [shardformer] supported flash attention test dependency (#4158) * [shardformer] fix flash attention utils test (#4180) * [shardformer] opt support flash attention (#4163) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] add performance benchmark of shardformer (#4175) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] benchmark fix * [shardformer] benchmark fix * [shardformer] llama support flash attention (#4185) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] llama support flash attention * [shardformer] llama support flash attention * [shardformer] Move the import statement for xformer outside the forward function. * [shardformer] gpt2 support flash attention. (#4191) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] gpt2 support flash attention * [shardformer] gpt2 support flash attention * [shardformer] gpt2 support flash attention * [shardformer] bloom support flash attention (#4188) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] bloom suport flash attention * [shardformer] add assert to sequence length * [shardformer] fix * [shardformer] fix * [shardformer] fix * [shardformer] bert support flash attention. (#4206) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] bert support flash attention * [shardformer] t5 support flash attention. (#4216) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] t5 support flash attention * [shardformer] t5 support flash attention * fix typo * fix typo * fix typo * fix typo * fix typo * fix typo * [shardformer] support 'paddedcausal' type of attention mask in Coloattention. (#4215) * added padded causal attn mask type for ColoAttention * [shardformer]t5 flash attention fix (#4239) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] t5 flash attention fix * [shardformer] update gpt2 to use coloattention. (#4234) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] update gpt2 to use coloattention * [shardformer] update gpt2 to use coloattention * [shardformer] update gpt2 to use coloattention * [shardformer] update gpt2 to use coloattention * [shardformer] update gpt2 * [shardformer] update opt and llama to use coloattention. (#4226) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * update opt to use coloattention * [shardformer]update opt to use coloattention * [shardformer]update opt to use coloattention * [shardformer]update opt to use coloattention * [shardformer]update opt to use coloattention * [shardformer]update opt to use coloattention * [shardformer]update opt to use coloattention * [shardformer]update opt * [shardformer] shardformer support jit fused operator. (#4236) * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] opt support flash attention * [shardformer] move to modeling * [shardformer] move to modeling * [shardformer] bloom support jit fused operator * [shardformer] bloom support jit fused operator * [shardformer] bloom support jit fused operator * [shardformer] t5 support jit fused operator * [shardformer] t5 support jit fused operator * [shardformer] t5 support jit fused operator * [shardformer] add roadmap of flash attention * [shardformer] add roadmap of flash attention * [shardformer] add roadmap of flash attention * [shardformer] add type hint to 'self' param of forward * [shardformer] merge feature/shardformer-models branch to feature/flash-attention-shardformer branch. (#4290) * Feature/vit support (#4182) * [shardformer] added tests * [shardformer] vit test finish and support * fix attention dropout * [shardformer] support SAM (#4231) * 1.support sam 2.add fused qkv for nn.Linear * update utils support set element in list * overtwrite SamVisionAttention foward to use DropoutForParallelInput * remove unused code * [shardformer] support whisper (#4212) * support whisper * fix bug in vocabembedding * support downstream model of whisper * update readme * Feature/chatglm (#4240) * [shardformer] added tests * [shardformer] vit test finish and support * [shardformer] chatglm ready * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] chatglm shard without mlp sharding * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] fix chatglm configuration with pre-commit --------- Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> * [shardformer] whisper support flash attention (#4301) * Feature/vit support (#4182) * [shardformer] added tests * [shardformer] vit test finish and support * fix attention dropout * [shardformer] support SAM (#4231) * 1.support sam 2.add fused qkv for nn.Linear * update utils support set element in list * overtwrite SamVisionAttention foward to use DropoutForParallelInput * remove unused code * [shardformer] support whisper (#4212) * support whisper * fix bug in vocabembedding * support downstream model of whisper * update readme * Feature/chatglm (#4240) * [shardformer] added tests * [shardformer] vit test finish and support * [shardformer] chatglm ready * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] chatglm shard without mlp sharding * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] fix chatglm configuration with pre-commit * [shardformer] whisper support flash attention * [shardformer] whisper support flash attention * [shardformer]whisper support jit operator --------- Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> * [shardformer] sam support flash attention (#4316) * Feature/vit support (#4182) * [shardformer] added tests * [shardformer] vit test finish and support * fix attention dropout * [shardformer] support SAM (#4231) * 1.support sam 2.add fused qkv for nn.Linear * update utils support set element in list * overtwrite SamVisionAttention foward to use DropoutForParallelInput * remove unused code * [shardformer] support whisper (#4212) * support whisper * fix bug in vocabembedding * support downstream model of whisper * update readme * Feature/chatglm (#4240) * [shardformer] added tests * [shardformer] vit test finish and support * [shardformer] chatglm ready * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] chatglm shard without mlp sharding * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] fix chatglm configuration with pre-commit * [shardformer] sam support flash attention --------- Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> * [shardformer] merge blip2/chatglm (#4321) * Feature/vit support (#4182) * [shardformer] added tests * [shardformer] vit test finish and support * fix attention dropout * [shardformer] support SAM (#4231) * 1.support sam 2.add fused qkv for nn.Linear * update utils support set element in list * overtwrite SamVisionAttention foward to use DropoutForParallelInput * remove unused code * [shardformer] support whisper (#4212) * support whisper * fix bug in vocabembedding * support downstream model of whisper * update readme * Feature/chatglm (#4240) * [shardformer] added tests * [shardformer] vit test finish and support * [shardformer] chatglm ready * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] chatglm shard without mlp sharding * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] fix chatglm configuration with pre-commit * [shardformer] added tests * [shardformer] vit test finish and support * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] support ChatGLMForConditionalGeneration & add fusedlayernorm for vit * [shardformer] support Blip2 (#4243) * support base blip2 * add support for downstream blip2 model * update readme * add forward injection * skip not compatible models test * fix test for gemini and low_level_zero_pugin --------- Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> Co-authored-by: klhhhhh <1412841649@qq.com> * [shardformer] blip2 support flash attention and jit operator (#4325) * Feature/vit support (#4182) * [shardformer] added tests * [shardformer] vit test finish and support * fix attention dropout * [shardformer] support SAM (#4231) * 1.support sam 2.add fused qkv for nn.Linear * update utils support set element in list * overtwrite SamVisionAttention foward to use DropoutForParallelInput * remove unused code * [shardformer] support whisper (#4212) * support whisper * fix bug in vocabembedding * support downstream model of whisper * update readme * Feature/chatglm (#4240) * [shardformer] added tests * [shardformer] vit test finish and support * [shardformer] chatglm ready * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] chatglm shard without mlp sharding * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] fix chatglm configuration with pre-commit * [shardformer] added tests * [shardformer] vit test finish and support * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] support ChatGLMForConditionalGeneration & add fusedlayernorm for vit * [shardformer] support Blip2 (#4243) * support base blip2 * add support for downstream blip2 model * update readme * add forward injection * skip not compatible models test * fix test for gemini and low_level_zero_pugin * [shardformer] blip2 support flash attention and jit operator * [shardformer] blip2 support flash attention and jit operator * [shardformer] blip2 support flash attention and jit operator --------- Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> Co-authored-by: klhhhhh <1412841649@qq.com> * [shardformer] chatglm support flash attention and jit operator (#4330) * Feature/vit support (#4182) * [shardformer] added tests * [shardformer] vit test finish and support * fix attention dropout * [shardformer] support SAM (#4231) * 1.support sam 2.add fused qkv for nn.Linear * update utils support set element in list * overtwrite SamVisionAttention foward to use DropoutForParallelInput * remove unused code * [shardformer] support whisper (#4212) * support whisper * fix bug in vocabembedding * support downstream model of whisper * update readme * Feature/chatglm (#4240) * [shardformer] added tests * [shardformer] vit test finish and support * [shardformer] chatglm ready * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] chatglm shard without mlp sharding * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] fix chatglm configuration with pre-commit * [shardformer] added tests * [shardformer] vit test finish and support * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] support ChatGLMForConditionalGeneration & add fusedlayernorm for vit * [shardformer] support Blip2 (#4243) * support base blip2 * add support for downstream blip2 model * update readme * add forward injection * skip not compatible models test * fix test for gemini and low_level_zero_pugin * [shardformer] chatglm support flash attention and jit operator * [shardformer] chatglm support flash attention and jit operator * [shardformer] chatglm support flash attention and jit operator * [shardformer] chatglm support flash attention and jit operator --------- Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> Co-authored-by: klhhhhh <1412841649@qq.com> * [shardformer] vit support flash attention and jit operator (#4334) * Feature/vit support (#4182) * [shardformer] added tests * [shardformer] vit test finish and support * fix attention dropout * [shardformer] support SAM (#4231) * 1.support sam 2.add fused qkv for nn.Linear * update utils support set element in list * overtwrite SamVisionAttention foward to use DropoutForParallelInput * remove unused code * [shardformer] support whisper (#4212) * support whisper * fix bug in vocabembedding * support downstream model of whisper * update readme * Feature/chatglm (#4240) * [shardformer] added tests * [shardformer] vit test finish and support * [shardformer] chatglm ready * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] chatglm shard without mlp sharding * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] fix chatglm configuration with pre-commit * [shardformer] added tests * [shardformer] vit test finish and support * import chatglm * [shardformer] add test kit in model zoo for chatglm * [sharformer] add first version of policy of chatglm * [shardformer] polish chatglm code * [shardformer] polish code * [shardformer] support chatglm without layernorm * [shardformer] delete some file * [shardformer] ChatGLM support layernorm sharding * [shardformer] register without auto policy * [shardformer] pre-commit check files * [shardformer] support ChatGLMForConditionalGeneration & add fusedlayernorm for vit * [shardformer] support Blip2 (#4243) * support base blip2 * add support for downstream blip2 model * update readme * add forward injection * skip not compatible models test * fix test for gemini and low_level_zero_pugin * [shardformer] vit support flash attention and jit operator * [shardformer] vit support flash attention and jit operator --------- Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> Co-authored-by: klhhhhh <1412841649@qq.com> * [pipeline] merge flash attention branch * [pipeline] merge flash attention branch * [pipeline] merge flash attention branch * [pipeline] fix conflict * [pipeline] fix conflict * Merge branch 'feature/pipeline' into feature/pipeline * Merge branch 'feature/pipeline' into feature/pipeline * Merge branch 'feature/pipeline' into feature/pipeline * activate checks * activate checks * activate checks * activate checks * activate checks * activate checks * activate checks * activate checks * fix flash attention tests * gemini ignore whisper * fix vit * fix xformers import handle --------- Co-authored-by: Frank Lee <somerlee.9@gmail.com> Co-authored-by: Kun Lin <81014421+klhhhhh@users.noreply.github.com> Co-authored-by: FoolPlayer <45593998+FoolPlayer@users.noreply.github.com> Co-authored-by: klhhhhh <1412841649@qq.com>
This commit is contained in:
flybird1111
Hongxin Liu
@@ -1,4 +1,9 @@
|
||||
import math
|
||||
from typing import Tuple
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import Tensor
|
||||
|
||||
|
||||
def forward_fn():
|
||||
@@ -37,3 +42,162 @@ def forward_fn():
|
||||
return outputs
|
||||
|
||||
return forward
|
||||
|
||||
|
||||
def get_sam_flash_attention_forward():
|
||||
|
||||
from transformers.models.sam.modeling_sam import SamAttention
|
||||
try:
|
||||
from xformers.ops import memory_efficient_attention as me_attention
|
||||
except:
|
||||
raise ImportError("Error: xformers module is not installed. Please install it to use flash attention.")
|
||||
|
||||
def _separate_heads(hidden_states: Tensor, num_attention_heads: int) -> Tensor:
|
||||
batch, point_batch_size, n_tokens, channel = hidden_states.shape
|
||||
c_per_head = channel // num_attention_heads
|
||||
hidden_states = hidden_states.reshape(batch * point_batch_size, n_tokens, num_attention_heads, c_per_head)
|
||||
return hidden_states
|
||||
|
||||
def _recombine_heads(hidden_states: Tensor, point_batch_size: int) -> Tensor:
|
||||
batch, n_tokens, n_heads, c_per_head = hidden_states.shape
|
||||
return hidden_states.reshape(batch // point_batch_size, point_batch_size, n_tokens, n_heads * c_per_head)
|
||||
|
||||
def forward(self: SamAttention,
|
||||
query: Tensor,
|
||||
key: Tensor,
|
||||
value: Tensor,
|
||||
attention_similarity: Tensor = None) -> Tensor:
|
||||
# Input projections
|
||||
query = self.q_proj(query)
|
||||
key = self.k_proj(key)
|
||||
value = self.v_proj(value)
|
||||
|
||||
point_batch_size = query.shape[1]
|
||||
# Separate into heads
|
||||
query = _separate_heads(query, self.num_attention_heads)
|
||||
key = _separate_heads(key, self.num_attention_heads)
|
||||
value = _separate_heads(value, self.num_attention_heads)
|
||||
|
||||
# SamAttention
|
||||
_, _, _, c_per_head = query.shape
|
||||
bias = None
|
||||
if attention_similarity is not None:
|
||||
bias = attention_similarity
|
||||
|
||||
scale = 1.0 / math.sqrt(c_per_head)
|
||||
out = me_attention(query, key, value, attn_bias=bias, scale=scale)
|
||||
|
||||
out = _recombine_heads(out, point_batch_size)
|
||||
out = self.out_proj(out)
|
||||
|
||||
return out
|
||||
|
||||
return forward
|
||||
|
||||
|
||||
def get_sam_vision_flash_attention_forward():
|
||||
|
||||
from transformers.models.sam.modeling_sam import SamVisionAttention
|
||||
try:
|
||||
from xformers.ops import memory_efficient_attention as me_attention
|
||||
except:
|
||||
raise ImportError("Error: xformers module is not installed. Please install it to use flash attention.")
|
||||
|
||||
def add_decomposed_rel_pos(
|
||||
query: torch.Tensor,
|
||||
rel_pos_h: torch.Tensor,
|
||||
rel_pos_w: torch.Tensor,
|
||||
q_size: Tuple[int, int],
|
||||
k_size: Tuple[int, int],
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
|
||||
https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py
|
||||
|
||||
Args:
|
||||
attn (`torch.Tensor`):
|
||||
attention map.
|
||||
query (`torch.Tensor`):
|
||||
query q in the attention layer with shape (batch_size, query_height * query_width, channel).
|
||||
rel_pos_h (`torch.Tensor`):
|
||||
relative position embeddings (Lh, channel) for height axis.
|
||||
rel_pos_w (`torch.Tensor`):
|
||||
relative position embeddings (Lw, channel) for width axis.
|
||||
q_size (tuple):
|
||||
spatial sequence size of query q with (query_height, query_width).
|
||||
k_size (tuple):
|
||||
spatial sequence size of key k with (key_height, key_width).
|
||||
|
||||
Returns:
|
||||
attn (`torch.Tensor`):
|
||||
attention map with added relative positional embeddings.
|
||||
"""
|
||||
|
||||
query_height, query_width = q_size
|
||||
key_height, key_width = k_size
|
||||
relative_position_height = get_rel_pos(query_height, key_height, rel_pos_h)
|
||||
relative_position_width = get_rel_pos(query_width, key_width, rel_pos_w)
|
||||
|
||||
batch_size, _, nHead, dim = query.shape
|
||||
reshaped_query = query.transpose(1, 2).reshape(batch_size * nHead, query_height, query_width, dim)
|
||||
rel_h = torch.einsum("bhwc,hkc->bhwk", reshaped_query, relative_position_height)
|
||||
rel_w = torch.einsum("bhwc,wkc->bhwk", reshaped_query, relative_position_width)
|
||||
rel_pos = rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
|
||||
rel_pos = rel_pos.reshape(batch_size, nHead, query_height * query_width, key_height * key_width)
|
||||
return rel_pos
|
||||
|
||||
def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
Get relative positional embeddings according to the relative positions of
|
||||
query and key sizes.
|
||||
|
||||
Args:
|
||||
q_size (int):
|
||||
size of the query.
|
||||
k_size (int):
|
||||
size of key k.
|
||||
rel_pos (`torch.Tensor`):
|
||||
relative position embeddings (L, channel).
|
||||
|
||||
Returns:
|
||||
Extracted positional embeddings according to relative positions.
|
||||
"""
|
||||
max_rel_dist = int(2 * max(q_size, k_size) - 1)
|
||||
# Interpolate rel pos.
|
||||
rel_pos_resized = F.interpolate(
|
||||
rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
|
||||
size=max_rel_dist,
|
||||
mode="linear",
|
||||
)
|
||||
rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
|
||||
|
||||
# Scale the coords with short length if shapes for q and k are different.
|
||||
q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
|
||||
k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
|
||||
relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
|
||||
|
||||
return rel_pos_resized[relative_coords.long()]
|
||||
|
||||
def forward(self: SamVisionAttention, hidden_states: torch.Tensor, output_attentions=False) -> torch.Tensor:
|
||||
batch_size, height, width, _ = hidden_states.shape
|
||||
# qkv with shape (3, batch_size, nHead, height * width, channel)
|
||||
qkv = (self.qkv(hidden_states).reshape(batch_size, height * width, 3, self.num_attention_heads,
|
||||
-1).permute(2, 0, 1, 3, 4))
|
||||
|
||||
query, key, value = qkv.reshape(3, batch_size, height * width, self.num_attention_heads, -1).unbind(0)
|
||||
|
||||
rel_pos = None
|
||||
if self.use_rel_pos:
|
||||
rel_pos = add_decomposed_rel_pos(query, self.rel_pos_h, self.rel_pos_w, (height, width), (height, width))
|
||||
|
||||
attn_output = me_attention(query, key, value, attn_bias=rel_pos, p=self.dropout, scale=self.scale)
|
||||
|
||||
attn_output = attn_output.reshape(batch_size, height, width, -1)
|
||||
|
||||
attn_output = self.proj(attn_output)
|
||||
|
||||
outputs = (attn_output, None)
|
||||
|
||||
return outputs
|
||||
|
||||
return forward
|
||||
|
||||
Reference in New Issue
Block a user