Spaces:
Runtime error
Runtime error
feat: scan layers + gradient checkpointing (#161)
Browse files* scan layers for faster compilation
* support gradient checkpointing
src/dalle_mini/model/configuration.py
CHANGED
@@ -51,7 +51,8 @@ class DalleBartConfig(PretrainedFromWandbMixin, PretrainedConfig):
|
|
51 |
activation_dropout=0.0,
|
52 |
init_std=0.02,
|
53 |
scale_embedding=False,
|
54 |
-
gradient_checkpointing=
|
|
|
55 |
use_cache=True,
|
56 |
is_encoder_decoder=True,
|
57 |
forced_eos_token_id=None,
|
@@ -59,7 +60,7 @@ class DalleBartConfig(PretrainedFromWandbMixin, PretrainedConfig):
|
|
59 |
do_sample=True,
|
60 |
# transformer variants
|
61 |
use_bias=False, # use bias in attention and dense layers (except for lm_head)
|
62 |
-
ln_type="
|
63 |
ln_positions="normformer", # layer normalization positions, "normformer", "swinv2", "cogview", "postln", "preln", "deepnet" (same as postln)
|
64 |
use_head_scale=False, # used in NormFormer
|
65 |
use_cosine_attention=False, # used in Swin v2
|
@@ -67,7 +68,7 @@ class DalleBartConfig(PretrainedFromWandbMixin, PretrainedConfig):
|
|
67 |
use_absolute_position_embeddings=True, # default
|
68 |
use_swin_position_embeddings=False, # used in Swin v1/v2
|
69 |
use_deepnet_scaling=False, # used in Deepnet
|
70 |
-
use_glu=
|
71 |
use_alibi=False, # Not implemented yet - from "Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation"
|
72 |
sinkhorn_iters=1, # used in SinkFormers
|
73 |
use_final_ln_encoder=True, # final layer normalization in encoder
|
@@ -136,6 +137,11 @@ class DalleBartConfig(PretrainedFromWandbMixin, PretrainedConfig):
|
|
136 |
self.init_std = init_std
|
137 |
self.use_cache = use_cache
|
138 |
self.gradient_checkpointing = gradient_checkpointing
|
|
|
|
|
|
|
|
|
|
|
139 |
self.scale_embedding = (
|
140 |
scale_embedding # scale factor will be sqrt(d_model) if True
|
141 |
)
|
|
|
51 |
activation_dropout=0.0,
|
52 |
init_std=0.02,
|
53 |
scale_embedding=False,
|
54 |
+
gradient_checkpointing=True,
|
55 |
+
use_scan=None,
|
56 |
use_cache=True,
|
57 |
is_encoder_decoder=True,
|
58 |
forced_eos_token_id=None,
|
|
|
60 |
do_sample=True,
|
61 |
# transformer variants
|
62 |
use_bias=False, # use bias in attention and dense layers (except for lm_head)
|
63 |
+
ln_type="rmsnorm", # layer normalization type, "rmsnorm", "layernorm"
|
64 |
ln_positions="normformer", # layer normalization positions, "normformer", "swinv2", "cogview", "postln", "preln", "deepnet" (same as postln)
|
65 |
use_head_scale=False, # used in NormFormer
|
66 |
use_cosine_attention=False, # used in Swin v2
|
|
|
68 |
use_absolute_position_embeddings=True, # default
|
69 |
use_swin_position_embeddings=False, # used in Swin v1/v2
|
70 |
use_deepnet_scaling=False, # used in Deepnet
|
71 |
+
use_glu=True, # "GLU Variants Improve Transformer"
|
72 |
use_alibi=False, # Not implemented yet - from "Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation"
|
73 |
sinkhorn_iters=1, # used in SinkFormers
|
74 |
use_final_ln_encoder=True, # final layer normalization in encoder
|
|
|
137 |
self.init_std = init_std
|
138 |
self.use_cache = use_cache
|
139 |
self.gradient_checkpointing = gradient_checkpointing
|
140 |
+
# all layers are the same in most configurations
|
141 |
+
self.use_scan = use_scan if use_scan is not None else ln_positions != "swinv2"
|
142 |
+
assert not (
|
143 |
+
self.use_scan and ln_positions == "swinv2"
|
144 |
+
), "scan cannot be used with 'swinv2'"
|
145 |
self.scale_embedding = (
|
146 |
scale_embedding # scale factor will be sqrt(d_model) if True
|
147 |
)
|
src/dalle_mini/model/modeling.py
CHANGED
@@ -619,6 +619,9 @@ class FlaxBartEncoderLayer(nn.Module):
|
|
619 |
deterministic: bool = True,
|
620 |
) -> Tuple[jnp.ndarray]:
|
621 |
|
|
|
|
|
|
|
622 |
res_gain = (
|
623 |
deepnet_gain["encoder"]["alpha"](self.config)
|
624 |
if self.config.use_deepnet_scaling
|
@@ -679,12 +682,8 @@ class FlaxBartEncoderLayer(nn.Module):
|
|
679 |
)
|
680 |
hidden_states = ff_block(hidden_states, deterministic=deterministic)
|
681 |
hidden_states = residual * res_gain + hidden_states
|
682 |
-
if self.add_norm
|
683 |
-
use_scale =
|
684 |
-
self.use_scale
|
685 |
-
or self.config.ln_positions == "postln"
|
686 |
-
or self.config.force_ln_scale
|
687 |
-
)
|
688 |
hidden_states = norm(
|
689 |
self.config.ln_type,
|
690 |
dtype=self.dtype,
|
@@ -697,6 +696,9 @@ class FlaxBartEncoderLayer(nn.Module):
|
|
697 |
if output_attentions:
|
698 |
outputs += (attn_weights,)
|
699 |
|
|
|
|
|
|
|
700 |
return outputs
|
701 |
|
702 |
|
@@ -710,7 +712,7 @@ class FlaxBartDecoderLayer(nn.Module):
|
|
710 |
config: DalleBartConfig
|
711 |
dtype: jnp.dtype = jnp.float32
|
712 |
add_norm: bool = False
|
713 |
-
use_scale: bool =
|
714 |
|
715 |
@nn.compact
|
716 |
def __call__(
|
@@ -724,6 +726,9 @@ class FlaxBartDecoderLayer(nn.Module):
|
|
724 |
deterministic: bool = True,
|
725 |
) -> Tuple[jnp.ndarray]:
|
726 |
|
|
|
|
|
|
|
727 |
res_gain = (
|
728 |
deepnet_gain["decoder"]["alpha"](self.config)
|
729 |
if self.config.use_deepnet_scaling
|
@@ -831,12 +836,8 @@ class FlaxBartDecoderLayer(nn.Module):
|
|
831 |
)
|
832 |
hidden_states = ff_block(hidden_states, deterministic=deterministic)
|
833 |
hidden_states = residual * res_gain + hidden_states
|
834 |
-
if self.add_norm
|
835 |
-
use_scale =
|
836 |
-
self.use_scale
|
837 |
-
or self.config.ln_positions == "postln"
|
838 |
-
or self.config.force_ln_scale
|
839 |
-
)
|
840 |
hidden_states = norm(
|
841 |
self.config.ln_type,
|
842 |
dtype=self.dtype,
|
@@ -849,6 +850,9 @@ class FlaxBartDecoderLayer(nn.Module):
|
|
849 |
if output_attentions:
|
850 |
outputs += (attn_weights, cross_attn_weights)
|
851 |
|
|
|
|
|
|
|
852 |
return outputs
|
853 |
|
854 |
|
@@ -876,35 +880,80 @@ class FlaxBartEncoderLayerCollection(nn.Module):
|
|
876 |
|
877 |
n_layers = self.config.encoder_layers
|
878 |
layer = (
|
879 |
-
remat(
|
|
|
|
|
|
|
|
|
880 |
if self.config.gradient_checkpointing
|
881 |
else FlaxBartEncoderLayer
|
882 |
)
|
883 |
-
|
884 |
-
|
885 |
-
|
886 |
-
|
887 |
-
|
888 |
-
|
889 |
-
|
890 |
-
|
891 |
-
|
892 |
-
|
893 |
-
|
894 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
895 |
)(
|
896 |
hidden_states,
|
897 |
attention_mask,
|
898 |
output_attentions,
|
899 |
deterministic,
|
900 |
)
|
901 |
-
hidden_states =
|
902 |
-
|
903 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
904 |
|
905 |
-
|
906 |
-
|
907 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
908 |
|
909 |
outputs = [
|
910 |
hidden_states,
|
@@ -953,22 +1002,39 @@ class FlaxBartDecoderLayerCollection(nn.Module):
|
|
953 |
|
954 |
n_layers = self.config.decoder_layers
|
955 |
layer = (
|
956 |
-
remat(
|
|
|
|
|
|
|
|
|
957 |
if self.config.gradient_checkpointing
|
958 |
else FlaxBartDecoderLayer
|
959 |
)
|
960 |
-
|
961 |
-
|
962 |
-
|
963 |
-
|
964 |
-
|
965 |
-
|
966 |
-
|
967 |
-
|
968 |
-
|
969 |
-
|
970 |
-
|
971 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
972 |
)(
|
973 |
hidden_states,
|
974 |
attention_mask,
|
@@ -978,17 +1044,56 @@ class FlaxBartDecoderLayerCollection(nn.Module):
|
|
978 |
output_attentions,
|
979 |
deterministic,
|
980 |
)
|
|
|
981 |
|
982 |
-
|
983 |
-
|
984 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
985 |
|
986 |
-
|
987 |
-
|
|
|
988 |
|
989 |
-
#
|
990 |
-
if
|
991 |
-
|
|
|
|
|
|
|
|
|
|
|
992 |
|
993 |
outputs = [
|
994 |
hidden_states,
|
|
|
619 |
deterministic: bool = True,
|
620 |
) -> Tuple[jnp.ndarray]:
|
621 |
|
622 |
+
if self.config.use_scan:
|
623 |
+
hidden_states = hidden_states[0]
|
624 |
+
|
625 |
res_gain = (
|
626 |
deepnet_gain["encoder"]["alpha"](self.config)
|
627 |
if self.config.use_deepnet_scaling
|
|
|
682 |
)
|
683 |
hidden_states = ff_block(hidden_states, deterministic=deterministic)
|
684 |
hidden_states = residual * res_gain + hidden_states
|
685 |
+
if self.add_norm:
|
686 |
+
use_scale = self.use_scale or self.config.force_ln_scale
|
|
|
|
|
|
|
|
|
687 |
hidden_states = norm(
|
688 |
self.config.ln_type,
|
689 |
dtype=self.dtype,
|
|
|
696 |
if output_attentions:
|
697 |
outputs += (attn_weights,)
|
698 |
|
699 |
+
if self.config.use_scan:
|
700 |
+
outputs = (outputs, None)
|
701 |
+
|
702 |
return outputs
|
703 |
|
704 |
|
|
|
712 |
config: DalleBartConfig
|
713 |
dtype: jnp.dtype = jnp.float32
|
714 |
add_norm: bool = False
|
715 |
+
use_scale: bool = True
|
716 |
|
717 |
@nn.compact
|
718 |
def __call__(
|
|
|
726 |
deterministic: bool = True,
|
727 |
) -> Tuple[jnp.ndarray]:
|
728 |
|
729 |
+
if self.config.use_scan:
|
730 |
+
hidden_states = hidden_states[0]
|
731 |
+
|
732 |
res_gain = (
|
733 |
deepnet_gain["decoder"]["alpha"](self.config)
|
734 |
if self.config.use_deepnet_scaling
|
|
|
836 |
)
|
837 |
hidden_states = ff_block(hidden_states, deterministic=deterministic)
|
838 |
hidden_states = residual * res_gain + hidden_states
|
839 |
+
if self.add_norm:
|
840 |
+
use_scale = self.use_scale or self.config.force_ln_scale
|
|
|
|
|
|
|
|
|
841 |
hidden_states = norm(
|
842 |
self.config.ln_type,
|
843 |
dtype=self.dtype,
|
|
|
850 |
if output_attentions:
|
851 |
outputs += (attn_weights, cross_attn_weights)
|
852 |
|
853 |
+
if self.config.use_scan:
|
854 |
+
outputs = (outputs, None)
|
855 |
+
|
856 |
return outputs
|
857 |
|
858 |
|
|
|
880 |
|
881 |
n_layers = self.config.encoder_layers
|
882 |
layer = (
|
883 |
+
remat(
|
884 |
+
FlaxBartEncoderLayer,
|
885 |
+
static_argnums=(2, 3),
|
886 |
+
prevent_cse=not self.config.use_scan,
|
887 |
+
)
|
888 |
if self.config.gradient_checkpointing
|
889 |
else FlaxBartEncoderLayer
|
890 |
)
|
891 |
+
|
892 |
+
if self.config.use_scan:
|
893 |
+
# all blocks are the same so we use nn.scan
|
894 |
+
assert not output_attentions, "cannot scan with output_attentions"
|
895 |
+
assert not output_hidden_states, "cannot scan with output_hidden_states"
|
896 |
+
hidden_states = (hidden_states,)
|
897 |
+
# we use a scale on all norms (even last layer) to allow scanning
|
898 |
+
hidden_states, _ = nn.scan(
|
899 |
+
layer,
|
900 |
+
variable_axes={"params": 0},
|
901 |
+
split_rngs={"params": True, "dropout": True},
|
902 |
+
in_axes=(nn.broadcast, nn.broadcast, nn.broadcast),
|
903 |
+
length=n_layers,
|
904 |
+
)(
|
905 |
+
self.config,
|
906 |
+
dtype=self.dtype,
|
907 |
+
add_norm=self.config.ln_positions == "postln",
|
908 |
+
name="FlaxBartEncoderLayers",
|
909 |
)(
|
910 |
hidden_states,
|
911 |
attention_mask,
|
912 |
output_attentions,
|
913 |
deterministic,
|
914 |
)
|
915 |
+
hidden_states = hidden_states[0]
|
916 |
+
else:
|
917 |
+
for i in range(n_layers):
|
918 |
+
if output_hidden_states:
|
919 |
+
all_hidden_states += (hidden_states,)
|
920 |
+
# final layernorm on the output of the last layer
|
921 |
+
# or every 6 layers for Swin v2
|
922 |
+
add_norm = self.config.ln_positions == "postln" or (
|
923 |
+
self.config.ln_positions == "swinv2"
|
924 |
+
and ((i + 1) % 6 == 0)
|
925 |
+
and (i != n_layers - 1)
|
926 |
+
)
|
927 |
+
# we don't need to scale the norm for the last layer
|
928 |
+
use_scale = i != n_layers - 1
|
929 |
+
layer_outputs = layer(
|
930 |
+
self.config,
|
931 |
+
dtype=self.dtype,
|
932 |
+
add_norm=add_norm,
|
933 |
+
use_scale=use_scale,
|
934 |
+
name=f"FlaxBartEncoderLayer_{i}",
|
935 |
+
)(
|
936 |
+
hidden_states,
|
937 |
+
attention_mask,
|
938 |
+
output_attentions,
|
939 |
+
deterministic,
|
940 |
+
)
|
941 |
+
hidden_states = layer_outputs[0]
|
942 |
+
if output_attentions:
|
943 |
+
all_self_attns += (layer_outputs[1],)
|
944 |
|
945 |
+
# add hidden states from the last layer
|
946 |
+
if output_hidden_states:
|
947 |
+
all_hidden_states += (hidden_states,)
|
948 |
+
|
949 |
+
# postln is already applied in every layer
|
950 |
+
if self.config.use_final_ln_encoder and self.config.ln_positions != "postln":
|
951 |
+
hidden_states = norm(
|
952 |
+
self.config.ln_type,
|
953 |
+
dtype=self.dtype,
|
954 |
+
epsilon=1e-05,
|
955 |
+
use_scale=self.config.force_ln_scale,
|
956 |
+
)(hidden_states)
|
957 |
|
958 |
outputs = [
|
959 |
hidden_states,
|
|
|
1002 |
|
1003 |
n_layers = self.config.decoder_layers
|
1004 |
layer = (
|
1005 |
+
remat(
|
1006 |
+
FlaxBartDecoderLayer,
|
1007 |
+
static_argnums=(4, 5, 6),
|
1008 |
+
prevent_cse=not self.config.use_scan,
|
1009 |
+
)
|
1010 |
if self.config.gradient_checkpointing
|
1011 |
else FlaxBartDecoderLayer
|
1012 |
)
|
1013 |
+
|
1014 |
+
if self.config.use_scan:
|
1015 |
+
# all blocks are the same so we use nn.scan
|
1016 |
+
assert not output_attentions, "cannot scan with output_attentions"
|
1017 |
+
assert not output_hidden_states, "cannot scan with output_hidden_states"
|
1018 |
+
hidden_states = (hidden_states,)
|
1019 |
+
# we use a scale on all norms (even last layer) to allow scanning
|
1020 |
+
hidden_states, _ = nn.scan(
|
1021 |
+
layer,
|
1022 |
+
variable_axes={"params": 0},
|
1023 |
+
split_rngs={"params": True, "dropout": True},
|
1024 |
+
in_axes=(
|
1025 |
+
nn.broadcast,
|
1026 |
+
nn.broadcast,
|
1027 |
+
nn.broadcast,
|
1028 |
+
nn.broadcast,
|
1029 |
+
nn.broadcast,
|
1030 |
+
nn.broadcast,
|
1031 |
+
),
|
1032 |
+
length=n_layers,
|
1033 |
+
)(
|
1034 |
+
self.config,
|
1035 |
+
dtype=self.dtype,
|
1036 |
+
add_norm=self.config.ln_positions == "postln",
|
1037 |
+
name="FlaxBartEncoderLayers",
|
1038 |
)(
|
1039 |
hidden_states,
|
1040 |
attention_mask,
|
|
|
1044 |
output_attentions,
|
1045 |
deterministic,
|
1046 |
)
|
1047 |
+
hidden_states = hidden_states[0]
|
1048 |
|
1049 |
+
else:
|
1050 |
+
for i in range(n_layers):
|
1051 |
+
if output_hidden_states:
|
1052 |
+
all_hidden_states += (hidden_states,)
|
1053 |
+
# final layernorm on the output of the last layer
|
1054 |
+
# or every 6 layers for Swin v2
|
1055 |
+
add_norm = self.config.ln_positions == "postln" or (
|
1056 |
+
self.config.ln_positions == "swinv2"
|
1057 |
+
and ((i + 1) % 6 == 0)
|
1058 |
+
and (i != n_layers - 1)
|
1059 |
+
)
|
1060 |
+
# we don't need to scale the norm for the last layer
|
1061 |
+
use_scale = i != n_layers - 1
|
1062 |
+
layer_outputs = layer(
|
1063 |
+
self.config,
|
1064 |
+
dtype=self.dtype,
|
1065 |
+
add_norm=add_norm,
|
1066 |
+
use_scale=use_scale,
|
1067 |
+
name=f"FlaxBartDecoderLayer_{i}",
|
1068 |
+
)(
|
1069 |
+
hidden_states,
|
1070 |
+
attention_mask,
|
1071 |
+
encoder_hidden_states,
|
1072 |
+
encoder_attention_mask,
|
1073 |
+
init_cache,
|
1074 |
+
output_attentions,
|
1075 |
+
deterministic,
|
1076 |
+
)
|
1077 |
+
|
1078 |
+
hidden_states = layer_outputs[0]
|
1079 |
+
if output_attentions:
|
1080 |
+
all_self_attns += (layer_outputs[1],)
|
1081 |
+
|
1082 |
+
if encoder_hidden_states is not None:
|
1083 |
+
all_cross_attentions += (layer_outputs[2],)
|
1084 |
|
1085 |
+
# add hidden states from the last decoder layer
|
1086 |
+
if output_hidden_states:
|
1087 |
+
all_hidden_states += (hidden_states,)
|
1088 |
|
1089 |
+
# postln is already applied in every layer
|
1090 |
+
if self.config.use_final_ln_decoder and self.config.ln_positions != "postln":
|
1091 |
+
hidden_states = norm(
|
1092 |
+
self.config.ln_type,
|
1093 |
+
dtype=self.dtype,
|
1094 |
+
epsilon=1e-05,
|
1095 |
+
use_scale=self.config.force_ln_scale,
|
1096 |
+
)(hidden_states)
|
1097 |
|
1098 |
outputs = [
|
1099 |
hidden_states,
|
src/dalle_mini/model/partitions.py
CHANGED
@@ -55,7 +55,7 @@ def _get_partition_rules():
|
|
55 |
]
|
56 |
|
57 |
|
58 |
-
def set_partitions(in_dict):
|
59 |
rules = _get_partition_rules()
|
60 |
replace = _replacement_rules(rules)
|
61 |
initd = {k: _unmatched for k in flatten_dict(in_dict)}
|
@@ -63,5 +63,14 @@ def set_partitions(in_dict):
|
|
63 |
for k, v in result.items():
|
64 |
if v == _unmatched:
|
65 |
print(f"Unmatched -> {k}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
assert _unmatched not in result.values(), "Incomplete partition spec."
|
67 |
return freeze(unflatten_dict(result))
|
|
|
55 |
]
|
56 |
|
57 |
|
58 |
+
def set_partitions(in_dict, use_scan):
|
59 |
rules = _get_partition_rules()
|
60 |
replace = _replacement_rules(rules)
|
61 |
initd = {k: _unmatched for k in flatten_dict(in_dict)}
|
|
|
63 |
for k, v in result.items():
|
64 |
if v == _unmatched:
|
65 |
print(f"Unmatched -> {k}")
|
66 |
+
l = list(result.keys())
|
67 |
+
if use_scan:
|
68 |
+
# add None dimension to scanned layers
|
69 |
+
result = {
|
70 |
+
k: (P(*(None,) + v) if v is not None else None)
|
71 |
+
if any(x in k for x in ["FlaxBartEncoderLayers", "FlaxBartDecoderLayers"])
|
72 |
+
else v
|
73 |
+
for k, v in result.items()
|
74 |
+
}
|
75 |
assert _unmatched not in result.values(), "Incomplete partition spec."
|
76 |
return freeze(unflatten_dict(result))
|
tools/train/config/mega/config.json
CHANGED
@@ -7,14 +7,14 @@
|
|
7 |
"decoder_attention_heads": 32,
|
8 |
"decoder_ffn_dim": 4096,
|
9 |
"decoder_layerdrop": 0.0,
|
10 |
-
"decoder_layers":
|
11 |
"decoder_start_token_id": 16384,
|
12 |
"do_sample": true,
|
13 |
"dropout": 0.0,
|
14 |
"encoder_attention_heads": 32,
|
15 |
"encoder_ffn_dim": 4096,
|
16 |
"encoder_layerdrop": 0.0,
|
17 |
-
"encoder_layers":
|
18 |
"encoder_vocab_size": 50272,
|
19 |
"eos_token_id": 16385,
|
20 |
"force_ln_scale": false,
|
|
|
7 |
"decoder_attention_heads": 32,
|
8 |
"decoder_ffn_dim": 4096,
|
9 |
"decoder_layerdrop": 0.0,
|
10 |
+
"decoder_layers": 26,
|
11 |
"decoder_start_token_id": 16384,
|
12 |
"do_sample": true,
|
13 |
"dropout": 0.0,
|
14 |
"encoder_attention_heads": 32,
|
15 |
"encoder_ffn_dim": 4096,
|
16 |
"encoder_layerdrop": 0.0,
|
17 |
+
"encoder_layers": 26,
|
18 |
"encoder_vocab_size": 50272,
|
19 |
"eos_token_id": 16385,
|
20 |
"force_ln_scale": false,
|
tools/train/train.py
CHANGED
@@ -42,6 +42,7 @@ from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
|
|
42 |
from flax.serialization import from_bytes, to_bytes
|
43 |
from flax.training import train_state
|
44 |
from flax.training.common_utils import onehot
|
|
|
45 |
from jax.experimental import PartitionSpec, maps
|
46 |
from jax.experimental.compilation_cache import compilation_cache as cc
|
47 |
from jax.experimental.pjit import pjit, with_sharding_constraint
|
@@ -531,6 +532,54 @@ class TrainState(train_state.TrainState):
|
|
531 |
train_time: float = 0.0 # total time the model trained
|
532 |
train_samples: int = 0 # number of samples seen
|
533 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
def main():
|
536 |
# See all possible arguments by passing the --help flag to this script.
|
@@ -618,7 +667,7 @@ def main():
|
|
618 |
model_metadata = model_args.get_metadata()
|
619 |
|
620 |
# get PartitionSpec for model params (required to be a dict)
|
621 |
-
param_spec = set_partitions(model.params)
|
622 |
|
623 |
# convert params to frozen dict
|
624 |
model._params = freeze(model.params)
|
@@ -743,6 +792,23 @@ def main():
|
|
743 |
|
744 |
learning_rate_fn = create_learning_rate_fn()
|
745 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
746 |
# create adam optimizer
|
747 |
if training_args.optim == "distributed_shampoo":
|
748 |
# parameters from https://github.com/tensorflow/lingvo/blob/03ee9d7cd50764b0424c7c863733c91fc0b053ec/lingvo/jax/optimizers.py#L729
|
@@ -795,10 +861,12 @@ def main():
|
|
795 |
)
|
796 |
# get the real optimizer and helper functions
|
797 |
update_fn = optimizer.update
|
798 |
-
|
|
|
799 |
opt_fn = NamedTuple("opt_fn", pspec_fn=Any, shape_and_dtype_fn=Any)(
|
800 |
optimizer.pspec_fn, optimizer.shape_and_dtype_fn
|
801 |
)
|
|
|
802 |
optimizer = optax.GradientTransformation(optimizer.init_fn, update_fn)
|
803 |
|
804 |
elif training_args.optim == "adam":
|
@@ -819,7 +887,7 @@ def main():
|
|
819 |
# get PartitionSpec for optimizer state
|
820 |
def get_opt_state_spec_and_shape(param_spec):
|
821 |
# get opt_state shape without actual init
|
822 |
-
opt_state_shape = jax.eval_shape(optimizer.init,
|
823 |
|
824 |
if training_args.optim == "adam":
|
825 |
|
@@ -844,7 +912,7 @@ def main():
|
|
844 |
|
845 |
elif training_args.optim == "distributed_shampoo":
|
846 |
opt_state_spec = opt_fn.pspec_fn(
|
847 |
-
params=
|
848 |
params_partition_spec=param_spec,
|
849 |
partition_spec_for_statistics=PartitionSpec(None, "dp", None),
|
850 |
)
|
@@ -852,7 +920,7 @@ def main():
|
|
852 |
raise NotImplementedError
|
853 |
return opt_state_spec, opt_state_shape
|
854 |
|
855 |
-
opt_state_spec, opt_state_shape = get_opt_state_spec_and_shape(
|
856 |
|
857 |
# create a mesh
|
858 |
mesh_shape = (training_args.dp_devices, training_args.mp_devices)
|
|
|
42 |
from flax.serialization import from_bytes, to_bytes
|
43 |
from flax.training import train_state
|
44 |
from flax.training.common_utils import onehot
|
45 |
+
from jax import ShapeDtypeStruct
|
46 |
from jax.experimental import PartitionSpec, maps
|
47 |
from jax.experimental.compilation_cache import compilation_cache as cc
|
48 |
from jax.experimental.pjit import pjit, with_sharding_constraint
|
|
|
532 |
train_time: float = 0.0 # total time the model trained
|
533 |
train_samples: int = 0 # number of samples seen
|
534 |
|
535 |
+
def apply_gradients(self, *, grads, **kwargs):
|
536 |
+
params = self.unscan(self.params)
|
537 |
+
updates, new_opt_state = self.tx.update(
|
538 |
+
self.unscan(grads), self.opt_state, params
|
539 |
+
)
|
540 |
+
params = optax.apply_updates(params, updates)
|
541 |
+
return self.replace(
|
542 |
+
step=self.step + 1,
|
543 |
+
params=self.rescan(params),
|
544 |
+
opt_state=new_opt_state,
|
545 |
+
**kwargs,
|
546 |
+
)
|
547 |
+
|
548 |
+
@classmethod
|
549 |
+
def create(cls, *, apply_fn, params, tx, **kwargs):
|
550 |
+
opt_state = tx.init(cls.unscan(params))
|
551 |
+
return cls(
|
552 |
+
step=0,
|
553 |
+
apply_fn=apply_fn,
|
554 |
+
params=params,
|
555 |
+
tx=tx,
|
556 |
+
opt_state=opt_state,
|
557 |
+
**kwargs,
|
558 |
+
)
|
559 |
+
|
560 |
+
@staticmethod
|
561 |
+
def unscan(params):
|
562 |
+
params = unfreeze(params)
|
563 |
+
for l in ["encoder", "decoder"]:
|
564 |
+
params["model"][l]["layers"] = jax.tree_map(
|
565 |
+
lambda x: {f"{i}": x[i] for i in range(len(x))},
|
566 |
+
params["model"][l]["layers"],
|
567 |
+
)
|
568 |
+
params = freeze(params)
|
569 |
+
return params
|
570 |
+
|
571 |
+
@staticmethod
|
572 |
+
def rescan(params):
|
573 |
+
params = unfreeze(params)
|
574 |
+
for l in ["encoder", "decoder"]:
|
575 |
+
params["model"][l]["layers"] = jax.tree_map(
|
576 |
+
lambda x: jnp.stack([x[f"{i}"] for i in range(len(x))]),
|
577 |
+
params["model"][l]["layers"],
|
578 |
+
is_leaf=lambda x: "0" in x,
|
579 |
+
)
|
580 |
+
params = freeze(params)
|
581 |
+
return params
|
582 |
+
|
583 |
|
584 |
def main():
|
585 |
# See all possible arguments by passing the --help flag to this script.
|
|
|
667 |
model_metadata = model_args.get_metadata()
|
668 |
|
669 |
# get PartitionSpec for model params (required to be a dict)
|
670 |
+
param_spec = set_partitions(model.params, model.config.use_scan)
|
671 |
|
672 |
# convert params to frozen dict
|
673 |
model._params = freeze(model.params)
|
|
|
792 |
|
793 |
learning_rate_fn = create_learning_rate_fn()
|
794 |
|
795 |
+
# reshape params to split scanned layers for optimizers
|
796 |
+
if model.config.use_scan:
|
797 |
+
params_struct = unfreeze(model.params)
|
798 |
+
for l in ["encoder", "decoder"]:
|
799 |
+
params_struct["model"][l]["layers"] = jax.tree_map(
|
800 |
+
lambda x: {
|
801 |
+
f"{i}": ShapeDtypeStruct(shape=x.shape[1:], dtype=x.dtype)
|
802 |
+
for i in range(len(x))
|
803 |
+
},
|
804 |
+
params_struct["model"][l]["layers"],
|
805 |
+
)
|
806 |
+
params_struct = freeze(params_struct)
|
807 |
+
|
808 |
+
else:
|
809 |
+
params_struct = model.params
|
810 |
+
opt_param_spec = set_partitions(params_struct, False)
|
811 |
+
|
812 |
# create adam optimizer
|
813 |
if training_args.optim == "distributed_shampoo":
|
814 |
# parameters from https://github.com/tensorflow/lingvo/blob/03ee9d7cd50764b0424c7c863733c91fc0b053ec/lingvo/jax/optimizers.py#L729
|
|
|
861 |
)
|
862 |
# get the real optimizer and helper functions
|
863 |
update_fn = optimizer.update
|
864 |
+
|
865 |
+
optimizer = optimizer.init(params_struct)
|
866 |
opt_fn = NamedTuple("opt_fn", pspec_fn=Any, shape_and_dtype_fn=Any)(
|
867 |
optimizer.pspec_fn, optimizer.shape_and_dtype_fn
|
868 |
)
|
869 |
+
|
870 |
optimizer = optax.GradientTransformation(optimizer.init_fn, update_fn)
|
871 |
|
872 |
elif training_args.optim == "adam":
|
|
|
887 |
# get PartitionSpec for optimizer state
|
888 |
def get_opt_state_spec_and_shape(param_spec):
|
889 |
# get opt_state shape without actual init
|
890 |
+
opt_state_shape = jax.eval_shape(optimizer.init, params_struct)
|
891 |
|
892 |
if training_args.optim == "adam":
|
893 |
|
|
|
912 |
|
913 |
elif training_args.optim == "distributed_shampoo":
|
914 |
opt_state_spec = opt_fn.pspec_fn(
|
915 |
+
params=params_struct,
|
916 |
params_partition_spec=param_spec,
|
917 |
partition_spec_for_statistics=PartitionSpec(None, "dp", None),
|
918 |
)
|
|
|
920 |
raise NotImplementedError
|
921 |
return opt_state_spec, opt_state_shape
|
922 |
|
923 |
+
opt_state_spec, opt_state_shape = get_opt_state_spec_and_shape(opt_param_spec)
|
924 |
|
925 |
# create a mesh
|
926 |
mesh_shape = (training_args.dp_devices, training_args.mp_devices)
|