Skip to content

[compat] fix FusedDSAIndexerLoss compat mcore<0.19 #142

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

Merged
Jintao-Huang merged 1 commit into from
Jul 9, 2026
Merged

[compat] fix FusedDSAIndexerLoss compat mcore<0.19 #142

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
11 changes: 8 additions & 3 deletions src/mcore_bridge/model/gpts/glm_moe_dsa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,8 @@
# Copyright (c) ModelScope Contributors. All rights reserved.
import megatron.core
import torch
from contextlib import contextmanager
from packaging import version
from typing import Optional

from ..constant import ModelType
Expand All @@ -15,6 +17,8 @@
except ImportError:
DSAttention = object

mcore_019 = version.parse(megatron.core.__version__) >= version.parse('0.19.0rc0')


class GlmMoeDsaDSAttention(DSAttention):
"""DSAttention with shared indexer support for GLM 5.2.
Expand Down Expand Up @@ -86,12 +90,13 @@ def forward(
if self.training and torch.is_grad_enabled():
q, k, weights = self.indexer.forward_before_topk(x, qr, packed_seq_params)
indexer_loss_coeff = getattr(self.config, 'dsa_indexer_loss_coeff', 0.0)

kwargs = {}
if mcore_019:
kwargs['calculate_per_token_loss'] = self.config.calculate_per_token_loss

@gemini-code-assist gemini-code-assist Bot Jul 9, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Using getattr(self.config, 'calculate_per_token_loss', False) is safer and more robust than direct attribute access. This prevents potential AttributeError if self.config is a mock object in unit tests or a custom configuration class that does not define calculate_per_token_loss.

Suggested change
kwargs['calculate_per_token_loss'] = self.config.calculate_per_token_loss
kwargs['calculate_per_token_loss'] = getattr(self.config, 'calculate_per_token_loss', False)

topk_indices, indexer_loss = FusedDSAIndexerLoss.apply(
q, weights, k, query.detach(), key.detach(), self.softmax_scale,
self.indexer.index_topk, indexer_loss_coeff, float_mask,
getattr(self.config, 'dsa_indexer_use_sparse_loss',
False), self.indexer.pg_collection, self.config.calculate_per_token_loss)
getattr(self.config, 'dsa_indexer_use_sparse_loss', False), self.indexer.pg_collection, **kwargs)
if indexer_loss_coeff > 0:
DSAIndexerLossLoggingHelper.save_loss_to_tracker(
loss=indexer_loss,
Expand Down
Loading