Skip to content

Remove no exp usage from logical rule Part III #3606

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 2 commits into
base: main
Choose a base branch
from
Open

Remove no exp usage from logical rule Part III #3606

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2a179f5
remoe no_exp from attentions
NuojCheng Apr 8, 2026
a0eaf98
remove no exp in moe
NuojCheng Apr 8, 2026
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
3 changes: 0 additions & 3 deletions src/maxtext/common/common_types.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,12 +34,10 @@
BATCH = "activation_batch"

ATTN_LENGTH = "activation_attn_length"
ATTN_LENGTH_NO_EXP = "activation_attn_length_no_exp"

LENGTH = "activation_length"
PREFILL_LENGTH = "prefill_activation_length"
Q_LENGTH = "activation_q_length"
Q_LENGTH_NO_EXP = "activation_q_length_no_exp"
Q_LORA_UP_PROJ = "q_lora_up_proj"
KV_LENGTH = "activation_kv_length"
KV_LORA_UP_PROJ = "kv_lora_up_proj"
Expand All @@ -48,7 +46,6 @@
HEAD = "activation_heads"
PREFILL_KV_BATCH = "activation_prefill_kv_batch"
KV_BATCH = "activation_kv_batch"
KV_BATCH_NO_EXP = "activation_kv_batch_no_exp"
KV_HEAD = "activation_kv_heads"
KV_HEAD_DIM = "activation_kv_head_dim"
D_KV = "activation_kv"
Expand Down
27 changes: 9 additions & 18 deletions src/maxtext/configs/base.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -454,8 +454,7 @@ 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_moe', ['data', 'fsdp', 'fsdp_transpose']],
['activation_batch_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 @@ -468,19 +467,18 @@ logical_axis_rules: [
['activation_length_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']],
['activation_q_length', ['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_mlp_moe', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_kv', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_prefill_kv_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
['activation_kv_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
['activation_kv_batch_no_exp', ['data', 'fsdp', 'fsdp_transpose']],
['activation_kv_batch', ['data', 'fsdp', 'fsdp_transpose']],
['activation_kv_head_dim', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_vocab', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_vocab', ['tensor', 'tensor_transpose']],
Expand All @@ -491,6 +489,7 @@ logical_axis_rules: [
['decode_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
['decode_length', ['sequence']],
['mlp', ['fsdp_transpose', 'tensor', 'tensor_sequence', 'autoregressive']],
['mlp_moe', ['fsdp_transpose', 'tensor', 'tensor_sequence', 'autoregressive']],
['mlp_no_fsdp', ['tensor', 'tensor_sequence', 'autoregressive']],
['vocab', ['tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive']],
['heads', ['tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive']],
Expand All @@ -500,18 +499,10 @@ logical_axis_rules: [
['embed', ['fsdp', 'sequence', 'tensor_transpose', 'context' , 'expert']],
['embed', ['fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert']],
['embed', ['fsdp', 'sequence', 'context', 'expert']],
['embed_no_exp', ['fsdp', 'fsdp_transpose', 'sequence', 'tensor_transpose', 'context']],
['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_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'tensor_transpose', 'context']],
['embed_moe', ['fsdp', 'sequence', 'tensor_transpose', 'context']],
['embed_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'context']],
['embed_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
18 changes: 8 additions & 10 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 @@ -30,38 +30,36 @@ mesh_axes: ['data', 'stage', 'fsdp', 'context', 'tensor', 'expert']
data_sharding: [['data', 'stage', 'fsdp', 'context', 'tensor', 'expert']]
logical_axis_rules: [
['activation_batch', ['data', 'fsdp', 'expert']],
['activation_batch_moe', ['data', 'fsdp', 'expert']],
['activation_batch_no_exp_moe', ['data', 'fsdp']],
['activation_batch_moe', ['data', 'fsdp']],
['activation_embed_and_logits_batch', ['data', 'stage', 'fsdp', 'expert']],
['activation_embed_and_logits_batch_sequence', ['data', 'stage', 'fsdp', 'context', 'expert']],
['activation_heads', ['tensor']],
['activation_kv_heads', ['tensor']],
['activation_length', ['context', 'expert']],
['activation_attn_length', ['context', 'expert']],
['activation_q_length', ['context', 'expert']],
['activation_length', ['context']],
['activation_attn_length', ['context']],
['activation_q_length', ['context']],
['activation_attn_embed', ['tensor']],
['activation_norm_length', ['context']],
['activation_norm_length_moe', ['context']],
['activation_embed', ['tensor']],
['activation_embed_moe', ['tensor']],
['activation_mlp', ['tensor']],
['activation_mlp_moe', ['tensor']],
['activation_kv', ['tensor']],
['activation_kv_batch', ['data', 'fsdp', 'expert']],
['activation_kv_batch_no_exp', ['data', 'fsdp']],
['activation_kv_batch', ['data', 'fsdp']],
['activation_kv_head_dim', ['tensor']],
['activation_vocab', ['tensor']],
['activation_stage', 'stage'],
['activation_exp', ['expert']],
['mlp', ['tensor']],
['mlp_moe', ['tensor']],
['mlp_no_fsdp', ['tensor']],
['vocab', ['tensor']],
['heads', ['tensor']],
['q_heads', ['tensor']],
['kv_heads', ['tensor']],
['embed', ['fsdp', 'expert']], # remove context from embed sharding
['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: 0 additions & 4 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 @@ -19,17 +19,13 @@ data_sharding: [['fsdp']]
logical_axis_rules: [
['activation_batch', ['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']],
['activation_kv_batch', ['fsdp']],
['activation_kv_batch_no_exp', ['fsdp']],
['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
10 changes: 6 additions & 4 deletions src/maxtext/configs/inference/inference.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,7 @@ logical_axis_rules: [
['activation_norm_length', ['tensor_sequence', 'sequence']],
['activation_embed', ['tensor_transpose']],
['activation_mlp', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_mlp_moe', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_kv', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['activation_prefill_kv_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert']],
['activation_kv_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert', 'context_autoregressive']],
Expand All @@ -25,6 +26,7 @@ logical_axis_rules: [
['decode_batch', ['data', 'fsdp', 'fsdp_transpose', 'expert', 'context_autoregressive']],
['decode_length', []],
['mlp', ['fsdp_transpose', 'tensor', 'tensor_sequence', 'autoregressive']],
['mlp_moe', ['fsdp_transpose', 'tensor', 'tensor_sequence', 'autoregressive']],
['vocab', ['tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive','context_autoregressive']],
['heads', ['tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive']],
['q_heads', ['tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive']],
Expand All @@ -33,10 +35,10 @@ logical_axis_rules: [
['embed', ['fsdp', 'sequence', 'tensor_transpose', 'expert']],
['embed', ['fsdp', 'fsdp_transpose', 'sequence', 'expert']],
['embed', ['fsdp', 'sequence', 'expert']],
['embed_no_exp', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_no_exp', ['fsdp', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_no_exp', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive']],
['embed_no_exp', ['fsdp', 'sequence', 'context_autoregressive']],
['embed_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_moe', ['fsdp', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive']],
['embed_moe', ['fsdp', 'sequence', 'context_autoregressive']],
['norm', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['layers', 'stage'],
['kv', []],
Expand Down
16 changes: 6 additions & 10 deletions src/maxtext/configs/inference/vllm.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,13 +31,11 @@ mesh_axes: ['data', 'attn_dp', 'model', 'expert', 'attn_dp_expert']
logical_axis_rules: [
['activation_batch', ['data']],
['activation_batch_moe', []],
['activation_batch_no_exp_moe', []],
['activation_embed_and_logits_batch', ['data', 'expert']],
['activation_embed_and_logits_batch_sequence', ['data', 'expert']],
['activation_heads', ['model', 'expert']],
['activation_kv_heads', ['model', 'expert']],
['activation_attn_length', ['expert']],
['activation_attn_length_no_exp', []],
['activation_attn_length', []],
['activation_length', ['data']],
['activation_length_moe', ['data', 'expert']],
['activation_length_moe', 'data'],
Expand All @@ -46,10 +44,10 @@ logical_axis_rules: [
['activation_embed', ['model', 'attn_dp']],
['activation_embed_moe', ['model', 'attn_dp']],
['activation_mlp', ['model', 'attn_dp']],
['activation_mlp_moe', ['model', 'attn_dp']],
['activation_kv', ['model']],
['activation_prefill_kv_batch', ['expert', 'attn_dp_expert']],
['activation_kv_batch', ['data', 'expert', 'attn_dp_expert']],
['activation_kv_batch_no_exp', ['data']],
['activation_kv_batch', ['data']],
['activation_kv_head_dim', ['model']],
['activation_vocab', ['model', 'attn_dp']],
['activation_norm_length', []],
Expand All @@ -58,8 +56,8 @@ logical_axis_rules: [
['decode_batch', ['expert', 'attn_dp_expert']],
['decode_length', []],
['mlp', ['model', 'attn_dp']],
['mlp_moe', ['model', 'attn_dp']],
['mlp_no_fsdp', ['model', 'attn_dp']],
['moe_mlp', ['model', 'attn_dp']],
['vocab', ['model', 'attn_dp']],
['heads', ['model']],
['q_heads', ['model', 'expert']],
Expand All @@ -68,11 +66,9 @@ logical_axis_rules: [
['kv', []],
['embed', ['expert', 'attn_dp_expert']],
['embed', ['attn_dp_expert']],
['embed_moe', ['expert', 'attn_dp_expert']],
['embed_moe', ['attn_dp_expert']],
['embed_moe', []],
['embed_moe', []],
['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
2 changes: 0 additions & 2 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 @@ -70,8 +70,6 @@ logical_axis_rules: [
['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
2 changes: 0 additions & 2 deletions src/maxtext/configs/models/deepseek3-671b-batchsplit.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,8 +70,6 @@ logical_axis_rules: [
['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
8 changes: 4 additions & 4 deletions src/maxtext/configs/post_train/rl_mt_jt.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,10 +49,10 @@ logical_axis_rules: [
['embed', ['fsdp', 'sequence', 'tensor_transpose', 'expert']],
['embed', ['fsdp', 'fsdp_transpose', 'sequence', 'expert']],
['embed', ['fsdp', 'sequence', 'expert']],
['embed_no_exp', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_no_exp', ['fsdp', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_no_exp', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive']],
['embed_no_exp', ['fsdp', 'sequence', 'context_autoregressive']],
['embed_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_moe', ['fsdp', 'sequence', 'context_autoregressive', 'tensor_transpose']],
['embed_moe', ['fsdp', 'fsdp_transpose', 'sequence', 'context_autoregressive']],
['embed_moe', ['fsdp', 'sequence', 'context_autoregressive']],
['norm', ['tensor', 'tensor_transpose', 'tensor_sequence']],
['layers', 'stage'],
['kv', []],
Expand Down
10 changes: 5 additions & 5 deletions src/maxtext/layers/attention_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,7 +63,7 @@
MODEL_MODE_PREFILL,
MODEL_MODE_TRAIN,
PREFILL_LENGTH,
Q_LENGTH_NO_EXP,
Q_LENGTH,
)
from maxtext.inference import page_manager
from maxtext.inference.kvcache import KVQuant, KVTensor
Expand Down Expand Up @@ -1134,13 +1134,13 @@ def tpu_flash_attention(
segment_axis_names_kv = None
sink_axis_names = self._logical_to_mesh_axes((HEAD,))
if decoder_segment_ids is not None:
segment_axis_names_q = self._logical_to_mesh_axes((BATCH, Q_LENGTH_NO_EXP))
segment_axis_names_q = self._logical_to_mesh_axes((BATCH, Q_LENGTH))
segment_axis_names_kv = self._logical_to_mesh_axes((BATCH, KV_LENGTH))

axis_names_splash_kernel = self._logical_to_mesh_axes(self.flash_axis_names_splash_kernel)
axis_names_q = self._logical_to_mesh_axes(self.flash_axis_names_q)
axis_names_kv = self._logical_to_mesh_axes(self.flash_axis_names_kv)
indexer_mask_axis_names = self._logical_to_mesh_axes((BATCH, Q_LENGTH_NO_EXP, KV_LENGTH))
indexer_mask_axis_names = self._logical_to_mesh_axes((BATCH, Q_LENGTH, KV_LENGTH))

global global_block_q, global_block_kv, global_block_kv_compute, global_block_q_dkv, global_block_kv_dkv
global global_block_kv_dkv_compute, global_block_q_dq, global_block_kv_dq, global_use_fused_bwd_kernel
Expand Down Expand Up @@ -1269,11 +1269,11 @@ def wrap_splash_kernel(single_head_mask):
return splash_kernel

splash_kernel = wrap_splash_kernel(single_head_mask)
segment_axis_names_splash_kernel = self._logical_to_mesh_axes((Q_LENGTH_NO_EXP,))
segment_axis_names_splash_kernel = self._logical_to_mesh_axes((Q_LENGTH,))
elif self.config.use_jax_splash and self.config.expert_shard_attention_option == EP_AS_FSDP:
if self.config.use_max_logit_estimate > 0:
sa_config = dataclasses.replace(sa_config, max_logit_const=self.config.use_max_logit_estimate)
segment_axis_names_splash_kernel = nn.logical_to_mesh_axes((Q_LENGTH_NO_EXP,))
segment_axis_names_splash_kernel = nn.logical_to_mesh_axes((Q_LENGTH,))
else:
# Create multi-head mask
multi_head_mask = splash_attention_mask.MultiHeadMask(masks=(mask,) * query.shape[1])
Expand Down
Loading
Loading