Background

GLM-5.3-Flash introduces hybrid attention (KDA linear attention alternating with DSA sparse attention) on top of the existing GLM5/GLM5.2 DSA stack. The sparse-attention groundwork (KPool indexer, kvallgather CP path for DSA, index-share reuse across shared decoder layers) is shared infrastructure that should land and be reviewed independently of the GLM53Flash model itself.

Expected Changes

1. KPool DSA indexer (kpool.py, kpool_ascend.py) with fixed-width top-k layout and Ascend fused kernels.
2. dsa_kvallgather_context_parallel.py: kvallgather CP path for DSA, including contiguous/lightning-indexer localization branches.
3. dsa.py/dsa_utils/dsa_fused/dsa_absorption enhancements: index-share holder reuse across shared decoder layers, dense-TND fallback, metadata validation; glm_dsa_index_share.py holder module.
4. Feature modules reorganized under features/megatron/{glm,glm_vl,qwen_vl,common} and features/megatron_bridge/{glm_vl,common}; OptimizeSendRecvCommFeature (dual-group ordinary P2P for PP send/recv) and Qwen3ASRAutoConfigFeature added.
5. Reduced-layer packed-TND validation scripts under tests/poc/glm5 and tests/poc/glm52.

The GLM53Flash model, data pipeline and ST land in follow-up PRs that depend on this one.

Acceptance Criteria

• glm5_744b / glm52_744b CI ST cases keep passing (no regression on the existing DSA stack).
• python -m py_compile passes on all new/changed files; repository pre-commit hooks pass on the PR diff.
• POC scripts run against the reduced-layer checkpoints with the documented parallel configurations.