Issues

• audit_filtered_corpora: TokenCorpus reads sequences as documents, so an empty text shifts every later document

  GeodesicResearch/geodesic-megatron#61 · Filter by author Kyle1668 opened on Oct 10, 2026

• Make the fp32 SSM-state setting a training-config field instead of an environment variable

  enhancement
  GeodesicResearch/geodesic-megatron#55 · Filter by author Kyle1668 opened on Oct 8, 2026

• Extend native chunked linear cross-entropy to vocabulary tensor parallelism

  GeodesicResearch/geodesic-megatron#54 · Filter by author puria-radmard opened on Oct 1, 2026

• Investigate HybridEP lifetime and variable-capacity regression coverage for RL consumers

  GeodesicResearch/geodesic-megatron#53 · Filter by author puria-radmard opened on Oct 1, 2026

• GRAM: dynamic per-request module gating in vLLM — serve all postures from one checkpoint

  enhancement
  GeodesicResearch/geodesic-megatron#37 · Filter by author Kyle1668 opened on Aug 12, 2026

• Megatron→HF export fails on torch_grouped checkpoints: function-local _target_ in saved run_config

  bug
  GeodesicResearch/geodesic-megatron#36 · Filter by author Kyle1668 opened on Aug 12, 2026

• GR checkpoint files can exceed their own torch_dist index ~4x on disk (237 GB for a 59 GiB plan)

  bug
  GeodesicResearch/geodesic-megatron#35 · Filter by author Kyle1668 opened on Aug 12, 2026

• 30B torch_dist checkpoint load transient peaks ~91 GiB for eval/probe (DP-heavy PP=1) loads — OOM cliff

  bug
  GeodesicResearch/geodesic-megatron#34 · Filter by author Kyle1668 opened on Aug 12, 2026

• GRAM: in-process restart unsupported (optimizer rebuild invalidates gater discovery)

  enhancement
  GeodesicResearch/geodesic-megatron#33 · Filter by author Kyle1668 opened on Aug 12, 2026

• GRAM: optimizer CPU offload unsupported (HybridDeviceOptimizer defeats param-group gating)

  enhancement
  GeodesicResearch/geodesic-megatron#32 · Filter by author Kyle1668 opened on Aug 12, 2026

• GRAM: recovery-strength sweep (module width / aux LR / p_cr) before forget_on-dependent big runs

  enhancement
  GeodesicResearch/geodesic-megatron#31 · Filter by author Kyle1668 opened on Aug 12, 2026

• GRAM: PP-aware gating — extend gradient routing beyond PP=1 to unblock Super/Ultra (GEOD-171d)

  enhancement
  GeodesicResearch/geodesic-megatron#30 · Filter by author Kyle1668 opened on Aug 12, 2026