Asha AnooshehandClaude Opus 5.5 2ad0e70ae8 Address review: free teacher fp32 copy early, input-only skip_lm_loss, dtype-tied ghost floor ...

- Compute each full-vocab normalizer once and free the teacher's fp32 copy
  before materializing the student's; document the memory/communication cost.
- Stop writing back into the removed skip_lm_loss field so configs survive
  dataclasses.replace()/asdict() round-trips; read kd_loss_alpha == 1.0 instead.
- Tie the ghost-token clamp floor to the dtype's eps and document its effects.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com>

2026-10-01 19:12:44 +02:00

..

_test_utils

[2/2] Track every Megatron-Bridge script with MLflow (#2514)

2026-09-28 21:29:25 +00:00

examples

[4/5] Add the IQ2_S CUDA encoder and register the format (#2565)

2026-09-29 20:35:38 +00:00

gpu

[4/5] Add the IQ2_S CUDA encoder and register the format (#2565)

2026-09-29 20:35:38 +00:00

gpu_megatron

Address review: free teacher fp32 copy early, input-only skip_lm_loss, dtype-tied ghost floor

2026-10-01 19:12:44 +02:00

gpu_trtllm

Fix 2-GPU test_model_load_utils hang; test import and fixture cleanup (#2079)

2026-08-05 21:00:42 +05:30

gpu_vllm

Fix vLLM compilation guard for models without marker (#2518)

2026-09-29 14:28:32 -07:00

regression/torch/speculative

Speed up slow unit/gpu/example tests (#1616)

2026-06-04 17:13:10 +00:00

unit

[4/5] Add the IQ2_S CUDA encoder and register the format (#2565)

2026-09-29 20:35:38 +00:00

conftest.py

[1/2] One MLflow tracking core behind a Tool record (#2544)

2026-09-28 19:57:03 +00:00