Skip to content

Commit afd23fd

Browse files
authored
Full finetune FSDP2 recipe (#1287)
1 parent 3955b33 commit afd23fd

File tree

21 files changed

+196
-178
lines changed

21 files changed

+196
-178
lines changed

recipes/configs/gemma/2B_full.yaml

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -59,7 +59,6 @@ device: cuda
5959

6060
# Memory management
6161
enable_activation_checkpointing: True
62-
memory_efficient_fsdp_wrap: False
6362

6463
# Reduced precision
6564
dtype: bf16

recipes/configs/gemma/7B_full.yaml

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -61,7 +61,6 @@ device: cuda
6161

6262
# Memory management
6363
enable_activation_checkpointing: True
64-
memory_efficient_fsdp_wrap: False
6564

6665
# Reduced precision
6766
dtype: bf16

recipes/configs/llama2/7B_full.yaml

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -62,7 +62,6 @@ device: cuda
6262

6363
# Memory management
6464
enable_activation_checkpointing: True
65-
memory_efficient_fsdp_wrap: False
6665

6766
# Reduced precision
6867
dtype: bf16

recipes/configs/llama3/70B_full.yaml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -94,7 +94,7 @@ device: cuda
9494

9595
# Memory management
9696
enable_activation_checkpointing: True
97-
memory_efficient_fsdp_wrap: True
97+
custom_sharded_layers: ['tok_embeddings', 'output']
9898
fsdp_cpu_offload: True
9999

100100
# Reduced precision

recipes/configs/llama3/8B_full.yaml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -64,7 +64,7 @@ device: cuda
6464

6565
# Memory management
6666
enable_activation_checkpointing: True
67-
memory_efficient_fsdp_wrap: True
67+
custom_sharded_layers: ['tok_embeddings', 'output']
6868

6969
# Reduced precision
7070
dtype: bf16

recipes/configs/llama3_1/70B_full.yaml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -94,7 +94,7 @@ device: cuda
9494

9595
# Memory management
9696
enable_activation_checkpointing: True
97-
memory_efficient_fsdp_wrap: True
97+
custom_sharded_layers: ['tok_embeddings', 'output']
9898
fsdp_cpu_offload: True
9999

100100
# Reduced precision

recipes/configs/llama3_1/8B_full.yaml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -67,7 +67,7 @@ device: cuda
6767

6868
# Memory management
6969
enable_activation_checkpointing: True
70-
memory_efficient_fsdp_wrap: True
70+
custom_sharded_layers: ['tok_embeddings', 'output']
7171

7272
# Reduced precision
7373
dtype: bf16

recipes/configs/mistral/7B_full.yaml

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -64,7 +64,6 @@ device: cuda
6464

6565
# Memory management
6666
enable_activation_checkpointing: True
67-
memory_efficient_fsdp_wrap: False
6867

6968
# Reduced precision
7069
dtype: bf16

recipes/configs/phi3/mini_full.yaml

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -61,7 +61,6 @@ device: cuda
6161

6262
# Memory management
6363
enable_activation_checkpointing: True
64-
memory_efficient_fsdp_wrap: False
6564
dtype: bf16
6665

6766
# Logging

recipes/configs/qwen2/0.5B_full.yaml

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -61,7 +61,6 @@ device: cuda
6161

6262
# Memory management
6363
enable_activation_checkpointing: False
64-
memory_efficient_fsdp_wrap: False
6564

6665
# Reduced precision
6766
dtype: bf16

0 commit comments

Comments
 (0)