Bundles complete relocatable Python3.10 headers, checks Python C-extension and fresh Triton driver compilation, defines rollout is_lora, preserves symmetric training initialization for the evaluation-only driver and includes child traceback tails. Re-raises original collective errors before unwritten metadata is decoded; 58 related regressions and 32 tests under the restored runtime pass. Documents step60 77.2% as in-training eight-GPU validation and explicitly limits standalone/H200/EFA verification claims. The underlying H200 Gloo transport failure remains unconfirmed.
19 days ago
Publish validated reproduction guide, launch tools and frozen release manifest
