Add support for TransformerEngine flash attention in WAN (#299)

* add flash attn te support for wan

* add gpu optimized sharding parallelism

* sharding bugfixes

* generalize across sharding parallelisms

* fix issue with inference using fsdp + te flash attention

* revert fsdp_tpu name change

* update readme with wan2.1 gpu notes

* re-order parallelism axes and revert dynamic context parallel axes selection

* remove unused max_utils imports

* change mesh names to more accurately reflect sharding

* cleanup

* fix lint errors

* update configs for unit tests.

---------

Co-authored-by: Juan Acevedo <juancevedo@gmail.com>

Author: 2 people

README.md

@@ -255,6 +255,9 @@ After installation completes, run the training script.
   - In Wan2.1, the ici_fsdp_parallelism axis is used for sequence parallelism, the ici_tensor_parallelism axis is used for head parallelism. 
     - You can enable both, keeping in mind that Wan2.1 has 40 heads and 40 must be evenly divisible by ici_tensor_parallelism.
     - For Sequence parallelism, the code pads the sequence length to evenly divide the sequence. Try out different ici_fsdp_parallelism numbers, but we find 2 and 4 to be the best right now.
+  - For use on GPU it is recommended to enable the cudnn_te_flash attention kernel for optimal performance.
+    - Best performance is achieved with the use of batch parallelism, which can be enabled by using the ici_fsdp_batch_parallelism axis. Note that this parallelism strategy does not support fractional batch sizes.
+    - ici_fsdp_batch_parallelism and ici_fsdp_parallelism can be combined to allow for fractional batch sizes. However, padding is not currently supported for the cudnn_te_flash attention kernel and it is therefore required that the sequence length is divisible by the number of devices in the ici_fsdp_parallelism axis.
 
   You should eventually see a training run as:
 

src/maxdiffusion/common_types.py

@@ -36,6 +36,7 @@
 # Physical axis names for device meshes.
 DATA = "data"
 FSDP = "fsdp"
+CONTEXT = "context"
 TENSOR = "tensor"
 # Logical axis names for model parameters and activations.
 BATCH = "activation_batch"
@@ -67,18 +68,18 @@
 ### Common axis rules for ring attention ###
 RING_ATTENTION_AXIS_RULES = [
     [SELF_ATTN_HEAD, None],
-    [SELF_ATTN_Q_LENGTH, FSDP],
-    [SELF_ATTN_KV_LENGTH, FSDP],
+    [SELF_ATTN_Q_LENGTH, CONTEXT],
+    [SELF_ATTN_KV_LENGTH, CONTEXT],
     [CROSS_ATTN_HEAD, None],
-    [CROSS_ATTN_Q_LENGTH, FSDP],
-    [CROSS_ATTN_KV_LENGTH, FSDP],
+    [CROSS_ATTN_Q_LENGTH, CONTEXT],
+    [CROSS_ATTN_KV_LENGTH, CONTEXT],
 ]
 
 SEQUENCE_PARALLEL_AXIS_RULES = [
     [SELF_ATTN_HEAD, None],
-    [SELF_ATTN_Q_LENGTH, FSDP],
+    [SELF_ATTN_Q_LENGTH, CONTEXT],
     [SELF_ATTN_KV_LENGTH, None],
     [CROSS_ATTN_HEAD, None],
-    [CROSS_ATTN_Q_LENGTH, FSDP],
+    [CROSS_ATTN_Q_LENGTH, CONTEXT],
     [CROSS_ATTN_KV_LENGTH, None],
 ]

src/maxdiffusion/configs/base14.yml

@@ -106,7 +106,7 @@ skip_jax_distributed_system: False
 base_output_directory: ""
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -131,17 +131,19 @@ logical_axis_rules: [
                       ['out_channels', 'tensor'],
                       ['conv_out', 'fsdp'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: -1  # recommended DCN axis to be auto-sharded
 dcn_fsdp_parallelism: 1
+dcn_context_parallelism: 1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: -1 # recommended ICI axis to be auto-sharded for TPUv5e
 ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
+ici_context_parallelism: 1
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False

src/maxdiffusion/configs/base21.yml

@@ -108,7 +108,7 @@ skip_jax_distributed_system: False
 base_output_directory: ""
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -133,17 +133,19 @@ logical_axis_rules: [
                       ['out_channels', 'tensor'],
                       ['conv_out', 'fsdp'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: -1  # recommended DCN axis to be auto-sharded
 dcn_fsdp_parallelism: 1
+dcn_context_parallelism: 1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: -1 # recommended ICI axis to be auto-sharded for TPUv5e
 ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
+ici_context_parallelism: 1
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False

src/maxdiffusion/configs/base_2_base.yml

@@ -121,7 +121,7 @@ skip_jax_distributed_system: False
 base_output_directory: ""
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -146,17 +146,19 @@ logical_axis_rules: [
                       ['out_channels', 'tensor'],
                       ['conv_out', 'fsdp'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: -1  # recommended DCN axis to be auto-sharded
 dcn_fsdp_parallelism: 1
+dcn_context_parallelism: 1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: -1 # recommended ICI axis to be auto-sharded for TPUv5e
 ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
+ici_context_parallelism: 1
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False

src/maxdiffusion/configs/base_flux_dev.yml

@@ -132,7 +132,7 @@ hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
 skip_jax_distributed_system: False
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -158,17 +158,19 @@ logical_axis_rules: [
                       ['out_channels', 'tensor'],
                       ['conv_out', 'fsdp'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: 1  # recommended DCN axis to be auto-sharded
 dcn_fsdp_parallelism: -1
+dcn_context_parallelism: 1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: -1
 ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
+ici_context_parallelism: 1
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False

src/maxdiffusion/configs/base_flux_dev_multi_res.yml

@@ -132,7 +132,7 @@ hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
 skip_jax_distributed_system: False
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -158,17 +158,19 @@ logical_axis_rules: [
                       ['out_channels', 'tensor'],
                       ['conv_out', 'fsdp'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: 1  # recommended DCN axis to be auto-sharded
 dcn_fsdp_parallelism: -1
+dcn_context_parallelism: 1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: -1
 ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
+ici_context_parallelism: 1
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False

src/maxdiffusion/configs/base_flux_schnell.yml

@@ -140,7 +140,7 @@ hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
 skip_jax_distributed_system: False
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -166,17 +166,19 @@ logical_axis_rules: [
                       ['out_channels', 'tensor'],
                       ['conv_out', 'fsdp'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: -1  # recommended DCN axis to be auto-sharded
 dcn_fsdp_parallelism: 1
+dcn_context_parallelism: 1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: -1
 ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
+ici_context_parallelism: 1
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False

src/maxdiffusion/configs/base_wan_14b.yml

@@ -151,7 +151,7 @@ hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
 skip_jax_distributed_system: False
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -166,31 +166,33 @@ mesh_axes: ['data', 'fsdp', 'tensor']
 # conv_in : conv.shape[2] weight
 # conv_out : conv.shape[-1] weight
 logical_axis_rules: [
-                      ['batch', 'data'],
-                      ['activation_batch', 'data'],
-                      ['activation_self_attn_heads', ['fsdp', 'tensor']], 
-                      ['activation_cross_attn_q_length', ['fsdp', 'tensor']],
-                      ['activation_length', 'fsdp'],
+                      ['batch', ['data', 'fsdp']],
+                      ['activation_batch', ['data', 'fsdp']],
+                      ['activation_self_attn_heads', ['context', 'tensor']],
+                      ['activation_cross_attn_q_length', ['context', 'tensor']],
+                      ['activation_length', 'context'],
                       ['activation_heads', 'tensor'],
                       ['mlp','tensor'],
-                      ['embed','fsdp'],
+                      ['embed', ['context', 'fsdp']],
                       ['heads', 'tensor'],
                       ['norm', 'tensor'],
-                      ['conv_batch', ['data','fsdp']],
+                      ['conv_batch', ['data', 'context', 'fsdp']],
                       ['out_channels', 'tensor'],
-                      ['conv_out', 'fsdp'],
+                      ['conv_out', 'context'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: 1  # recommended DCN axis to be auto-sharded
-dcn_fsdp_parallelism: -1
+dcn_fsdp_parallelism: 1
+dcn_context_parallelism: -1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: 1
-ici_fsdp_parallelism: -1  # recommended ICI axis to be auto-sharded
+ici_fsdp_parallelism: 1
+ici_context_parallelism: -1  # recommended ICI axis to be auto-sharded
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False

src/maxdiffusion/configs/base_wan_27b.yml

@@ -139,7 +139,7 @@ hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
 skip_jax_distributed_system: False
 
 # Parallelism
-mesh_axes: ['data', 'fsdp', 'tensor']
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
 
 # batch : batch dimension of data and activations
 # hidden :
@@ -154,30 +154,33 @@ mesh_axes: ['data', 'fsdp', 'tensor']
 # conv_in : conv.shape[2] weight
 # conv_out : conv.shape[-1] weight
 logical_axis_rules: [
-                      ['batch', 'data'],
-                      ['activation_batch', 'data'],
-                      ['activation_length', 'fsdp'],
-                      
+                      ['batch', ['data', 'fsdp']],
+                      ['activation_batch', ['data', 'fsdp']],
+                      ['activation_self_attn_heads', ['context', 'tensor']], 
+                      ['activation_cross_attn_q_length', ['context', 'tensor']],
+                      ['activation_length', 'context'],
                       ['activation_heads', 'tensor'],
                       ['mlp','tensor'],
-                      ['embed','fsdp'],
+                      ['embed', ['context', 'fsdp']],
                       ['heads', 'tensor'],
                       ['norm', 'tensor'],
-                      ['conv_batch', ['data','fsdp']],
+                      ['conv_batch', ['data', 'context', 'fsdp']],
                       ['out_channels', 'tensor'],
-                      ['conv_out', 'fsdp'],
+                      ['conv_out', 'context'],
                     ]
-data_sharding: [['data', 'fsdp', 'tensor']]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
 
 # One axis for each parallelism type may hold a placeholder (-1)
 # value to auto-shard based on available slices and devices.
 # By default, product of the DCN axes should equal number of slices
 # and product of the ICI axes should equal number of devices per slice.
 dcn_data_parallelism: 1  # recommended DCN axis to be auto-sharded
-dcn_fsdp_parallelism: -1
+dcn_fsdp_parallelism: 1
+dcn_context_parallelism: -1
 dcn_tensor_parallelism: 1
 ici_data_parallelism: 1
-ici_fsdp_parallelism: -1  # recommended ICI axis to be auto-sharded
+ici_fsdp_parallelism: 1
+ici_context_parallelism: -1  # recommended ICI axis to be auto-sharded
 ici_tensor_parallelism: 1
 
 allow_split_physical_axes: False