Skip to content

force select self node #10331

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

Closed
phlrain wants to merge 1 commit into from
Closed

force select self node #10331

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
32 changes: 32 additions & 0 deletions paddlenlp/transformers/moe_gate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,10 @@
import paddle.nn.functional as F

from ..utils.log import logger
import math
import os

DSV3_FORCE_SELECT_SELF_NODE = os.getenv("DSV3_FORCE_SELECT_SELF_NODE", "False").lower() == "true"

class MoEGateMixin:
def gate_score_func(self, logits: paddle.Tensor) -> paddle.Tensor:
Expand Down Expand Up @@ -193,6 +196,9 @@ def __init__(self, config, num_experts, expert_hidden_size, **kwargs):
if self.global_aux_loss:
assert self.group is not None, "group is required when global_aux_loss is True"
self.rank = dist.get_rank(self.group)
else:
self.rank = dist.get_rank()


self.expert_drop = kwargs.pop("expert_drop", False)
self.noisy_gate_policy = kwargs.pop("noisy_gate_policy", None)
Expand All @@ -210,6 +216,29 @@ def __init__(self, config, num_experts, expert_hidden_size, **kwargs):
self.norm_topk_prob = kwargs.pop("norm_topk_prob", False)
self.routed_scaling_factor = kwargs.pop("routed_scaling_factor", 1.0)

if DSV3_FORCE_SELECT_SELF_NODE:
print( "self rank", self.rank )
ep_size = self.config.ep_size
self.ep_node_num = math.ceil(ep_size / 8 )
current_ep_node_id = (self.rank % self.ep_node_num)
# ep rang
groups_per_node = self.n_group // self.ep_node_num

start = groups_per_node * current_ep_node_id
end = groups_per_node * (current_ep_node_id+1)

print( "star end", start, end)

current_groupd_bias = paddle.zeros( [self.n_group], dtype="float32")
current_groupd_bias[start:end] = paddle.ones([1], dtype="float32") * 1000000


self.current_groupd_bias = current_groupd_bias
self.current_groupd_bias.stop_gradient = True
print( "group id")

print( current_groupd_bias )

def _priority(self, topk_idx: paddle.Tensor, capacity: int) -> paddle.Tensor:
"""_summary_
The priority is the cumulative sum of the expert indices.
Expand Down Expand Up @@ -305,6 +334,9 @@ def _topk_noaux_tc(
group_scores = (
scores_for_choice.reshape([bsz_seq_len, self.n_group, -1]).topk(2, axis=-1)[0].sum(axis=-1)
) # fmt:skip [n, n_group]
if DSV3_FORCE_SELECT_SELF_NODE:
group_scores = group_scores + self.current_groupd_bias

group_idx = paddle.topk(group_scores, k=topk_group, axis=-1, sorted=True)[1] # [n, top_k_group]
group_mask = paddle.zeros_like(group_scores).put_along_axis(group_idx, paddle.ones([], dtype="float32"), axis=-1) # fmt:skip
score_mask = (
Expand Down