hongshaorou/dreamer4
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

pass the attend kwargs for the block causal masking in tokenizer

Browse Source
This commit is contained in:
lucidrains 2025-10-16 08:33:26 -07:00
parent 517ef6b94b
commit 2ccb290e26
2 changed files with 23 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
dreamer4/dreamer4.py
View File

@ -1278,6 +1278,8 @@ class VideoTokenizer(Module):
tokens, packed_latent_shape = pack((decoder_pos_emb, latent_tokens), 'b t * d')
space_seq_len = tokens.shape[-2]
# pack time
tokens, inverse_pack_time = pack_one(tokens, 'b * d')
@ -1286,7 +1288,16 @@ class VideoTokenizer(Module):
# decoder attend
decoder_attend_fn = get_attend_fn(use_flex, seq_len, seq_len, causal = True, num_special_tokens = self.num_latent_tokens, special_attend_only_itself = True)
decoder_attend_fn = get_attend_fn(
use_flex,
seq_len, seq_len,
causal = True,
causal_block_size = space_seq_len,
softclamp_value = self.attn_softclamp_value,
block_size_per_special = space_seq_len,
num_special_tokens = self.num_latent_tokens,
special_attend_only_itself = True # different than encoder
)
# decoder attention
@ -1373,14 +1384,6 @@ class VideoTokenizer(Module):
# attend hyper parameters
attend_kwargs = dict(
causal = True,
causal_block_size = space_seq_len,
softclamp_value = self.attn_softclamp_value,
block_size_per_special = space_seq_len,
num_special_tokens = 1
)
use_flex = tokens.is_cuda and exists(flex_attention)
# encoder attend
@ -1388,7 +1391,16 @@ class VideoTokenizer(Module):
# modality can only attend to itself while latents can attend to everything
# similar to agent token in dynamics model
encoder_attend_fn = get_attend_fn(use_flex, seq_len, seq_len, causal = True, num_special_tokens = self.num_latent_tokens, special_attend_only_itself = False)
encoder_attend_fn = get_attend_fn(
use_flex,
seq_len, seq_len,
causal = True,
causal_block_size = space_seq_len,
softclamp_value = self.attn_softclamp_value,
block_size_per_special = space_seq_len,
num_special_tokens = self.num_latent_tokens,
special_attend_only_itself = False # different than decoder
)
# encoder

2
pyproject.toml
View File

@ -1,6 +1,6 @@
[project]
name = "dreamer4"
version = "0.0.22"
version = "0.0.23"
description = "Dreamer 4"
authors = [
{ name = "Phil Wang", email = "lucidrains@gmail.com" }