Skip to content

Split logical names in moe module #3473

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
NuojCheng wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Split logical names in moe module #3473

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 14 additions & 0 deletions src/maxtext/configs/base.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -435,7 +435,9 @@ custom_mesh_and_rule: "" # replace default mesh and logical rule by specifying y
mesh_axes: ['diloco', 'data', 'stage', 'fsdp', 'fsdp_transpose', 'sequence', 'context', 'context_autoregressive', 'tensor', 'tensor_transpose', 'tensor_sequence', 'expert', 'autoregressive']
logical_axis_rules: [
['activation_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
['activation_batch_moe', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
['activation_batch_no_exp', ['data', 'fsdp', 'fsdp_transpose']],
['activation_batch_no_exp_moe', ['data', 'fsdp', 'fsdp_transpose']],
['activation_embed_and_logits_batch', ['data', 'stage', 'fsdp', 'fsdp_transpose', 'expert']],
['activation_embed_and_logits_batch_sequence', ['data', 'stage', 'fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert']],
['activation_heads', ['tensor', 'tensor_transpose', 'sequence','tensor_sequence','autoregressive']],
Expand All @@ -448,14 +450,18 @@ logical_axis_rules: [
['activation_attn_length_no_exp', ['context']],
['activation_length_no_exp', ['sequence', 'context']],
['activation_length_no_exp', ['context']],
['activation_length_no_exp_moe', ['sequence', 'context']],
['activation_length_no_exp_moe', ['context']],
['activation_norm_length', ['tensor_sequence', 'context', 'sequence']],
['activation_norm_length_moe', ['tensor_sequence', 'context', 'sequence']],
['activation_q_length', ['context', 'expert']],
['activation_q_length_no_exp', ['context']],
['prefill_activation_length', ['sequence', 'context']],
['prefill_activation_norm_length', ['tensor_sequence', 'context', 'sequence']],
['activation_kv_length', []],
['activation_attn_embed', ['tensor', 'tensor_transpose']],
['activation_embed', ['tensor', 'tensor_transpose']],
['activation_embed_moe', ['tensor', 'tensor_transpose']],
['activation_mlp', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_kv', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_prefill_kv_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
Expand Down Expand Up @@ -484,6 +490,14 @@ logical_axis_rules: [
['embed_no_exp', ['fsdp', 'sequence', 'tensor_transpose', 'context']],
['embed_no_exp', ['fsdp', 'fsdp_transpose', 'sequence', 'context']],
['embed_no_exp', ['fsdp', 'sequence', 'context']],
['embed_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'tensor_transpose', 'context', 'expert']],
['embed_moe', ['fsdp', 'sequence', 'tensor_transpose', 'context' , 'expert']],
['embed_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert']],
['embed_moe', ['fsdp', 'sequence', 'context', 'expert']],
['embed_no_exp_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'tensor_transpose', 'context']],
['embed_no_exp_moe', ['fsdp', 'sequence', 'tensor_transpose', 'context']],
['embed_no_exp_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'context']],
['embed_no_exp_moe', ['fsdp', 'sequence', 'context']],
['embed_tensor_transpose', ['tensor_transpose']],
['q_lora', ['fsdp', 'fsdp_transpose', 'sequence', 'context', 'tensor_transpose', 'expert']],
['q_lora', ['fsdp', 'sequence', 'context', 'tensor_transpose', 'expert']],
Expand Down
6 changes: 6 additions & 0 deletions src/maxtext/configs/custom_mesh_and_rule/pipeline-large-moe.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,9 @@ mesh_axes: ['data', 'stage', 'fsdp', 'tensor', 'expert']
data_sharding: [['data', 'stage', 'fsdp', 'tensor', 'expert']]
logical_axis_rules: [
['activation_batch', ['data', 'fsdp', 'expert']],
['activation_batch_moe', ['data', 'fsdp', 'expert']],
['activation_batch_no_exp', ['data', 'fsdp']],
['activation_batch_no_exp_moe', ['data', 'fsdp']],
['activation_embed_and_logits_batch', ['data', 'stage', 'fsdp', 'expert']],
['activation_embed_and_logits_batch_sequence', ['data', 'stage', 'fsdp', 'expert']],
['activation_heads', ['tensor']],
Expand All @@ -38,6 +40,7 @@ logical_axis_rules: [
['activation_q_length', ['expert']],
['activation_attn_embed', ['tensor']],
['activation_embed', ['tensor']],
['activation_embed_moe', ['tensor']],
['activation_mlp', ['tensor']],
['activation_kv', ['tensor']],
['activation_prefill_kv_batch', ['data', 'fsdp', 'expert']],
Expand All @@ -55,7 +58,10 @@ logical_axis_rules: [
['q_heads', ['tensor']],
['kv_heads', ['tensor']],
['embed', ['fsdp', 'expert']],
['embed_moe', ['fsdp', 'expert']],
['embed_no_exp', ['fsdp']],
['embed_no_exp_moe', ['fsdp']],
['embed_moe', ['fsdp']],
['q_lora', ['fsdp']],
['kv_lora', ['fsdp']],
['norm', ['tensor']],
Expand Down
4 changes: 4 additions & 0 deletions src/maxtext/configs/custom_mesh_and_rule/pure-fsdp.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,8 @@ data_sharding: [['fsdp']]
logical_axis_rules: [
['activation_batch', ['fsdp']],
['activation_batch_no_exp', ['fsdp']],
['activation_batch_moe', ['fsdp']],
['activation_batch_no_exp_moe', ['fsdp']],
['activation_embed_and_logits_batch', ['fsdp']],
['activation_embed_and_logits_batch_sequence', ['fsdp']],
['activation_prefill_kv_batch', ['fsdp']],
Expand All @@ -27,6 +29,8 @@ logical_axis_rules: [
['decode_batch', ['fsdp']],
['embed', ['fsdp']],
['embed_no_exp', ['fsdp']],
['embed_moe', ['fsdp']],
['embed_no_exp_moe', ['fsdp']],
['q_lora', ['fsdp']],
['kv_lora', ['fsdp']],
['exp_with_fsdp', 'fsdp'],
Expand Down
8 changes: 8 additions & 0 deletions src/maxtext/configs/inference/vllm.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,18 +30,23 @@ weight_dtype: bfloat16
mesh_axes: ['data', 'attn_dp', 'model', 'expert', 'attn_dp_expert']
logical_axis_rules: [
['activation_batch', ['expert']],
['activation_batch_moe', ['expert']],
['activation_batch_no_exp', []],
['activation_batch_no_exp_moe', []],
['activation_embed_and_logits_batch', ['expert']],
['activation_embed_and_logits_batch_sequence', ['expert']],
['activation_heads', ['model']],
['activation_kv_heads', ['model']],
['activation_attn_length', ['expert']],
['activation_attn_length_no_exp', []],
['activation_length', ['data', 'expert']],
['activation_length_moe', ['data', 'expert']],
['activation_length_no_exp', 'data'],
['activation_length_no_exp_moe', 'data'],
['activation_q_length', ['expert', 'attn_dp_expert']],
['activation_attn_embed', 'model'],
['activation_embed', ['model', 'attn_dp']],
['activation_embed_moe', ['model', 'attn_dp']],
['activation_mlp', ['model', 'attn_dp']],
['activation_kv', ['model']],
['activation_prefill_kv_batch', ['expert', 'attn_dp_expert']],
Expand All @@ -50,6 +55,7 @@ logical_axis_rules: [
['activation_kv_head_dim', ['model']],
['activation_vocab', ['model', 'attn_dp']],
['activation_norm_length', []],
['activation_norm_length_moe', []],
['activation_exp', ['expert', 'attn_dp_expert']],
['decode_batch', ['expert', 'attn_dp_expert']],
['decode_length', []],
Expand All @@ -63,8 +69,10 @@ logical_axis_rules: [
['kv_head_dim', []],
['kv', []],
['embed', ['expert', 'attn_dp_expert']],
['embed_moe', ['expert', 'attn_dp_expert']],
['embed_tensor_transpose', ['attn_dp', 'model']],
['embed_no_exp', []],
['embed_no_exp_moe', []],
['q_lora', ['expert', 'attn_dp_expert']],
['kv_lora', ['expert', 'attn_dp_expert']],
['norm', []],
Expand Down
4 changes: 4 additions & 0 deletions src/maxtext/configs/models/deepseek3-671b-2dfsdp.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,14 +60,18 @@ mesh_axes: ['data', 'stage', 'fsdp', 'fsdp_transpose', 'expert', 'context']
data_sharding: [['data', 'stage', 'fsdp', 'fsdp_transpose', 'expert', 'context']]
logical_axis_rules: [
['activation_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert', 'context']],
['activation_batch_moe', ['data', 'fsdp', 'fsdp_transpose', 'expert', 'context']],
['activation_embed_and_logits_batch', ['data', 'stage', 'fsdp', 'fsdp_transpose', 'expert', 'context']],
['activation_kv_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert', 'context']],
['activation_embed_and_logits_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
['activation_norm_length', ['context']],
['activation_norm_length_moe', ['context']],
['activation_heads', []],
['activation_stage', 'stage'],
['embed', ['fsdp']],
['embed_moe', ['fsdp']],
['embed_no_exp', ['fsdp']],
['embed_no_exp_moe', ['fsdp']],
['q_lora', ['fsdp']],
['kv_lora', ['fsdp']],
['layers', 'stage'],
Expand Down
Loading
Loading