basil-k-aji-dev 37ac206e9a [perf] skip the per-forward RoPE buffer broadcast in the RL trainers ... Follow-up to the review on #854. DDP re-broadcasts every registered buffer from rank 0 on each forward. The only buffers here are freqs_cos / freqs_sin, which each rank computes deterministically from config and never mutates from data, so the broadcast moves identical bytes every step for nothing. Verified with a two-rank CPU/Gloo run: with broadcast_buffers=False the gradients stay identical across ranks and so do the buffers. Also document the critic's tied-embedding assumption. CriticModel.forward never calls lm_head, which is only safe under DDP because tie_word_embeddings (default True) makes lm_head.weight the same Parameter as embed_tokens.weight. Untied, it becomes a genuinely unused parameter and DDP raises.		2026-09-14 22:22:00 +05:30
..
rollout_engine.py	[fix] keep agent rollout token-aligned #850	2026-09-10 11:44:05 +08:00
train_agent.py	[perf] skip the per-forward RoPE buffer broadcast in the RL trainers	2026-09-14 22:22:00 +05:30
train_distillation.py	[fix] ddp exit hang	2026-06-01 17:50:39 +08:00
train_dpo.py	[fix] ddp exit hang	2026-06-01 17:50:39 +08:00
train_full_sft.py	[update] remove expert args	2026-09-04 22:16:38 +08:00
train_grpo.py	[perf] skip the per-forward RoPE buffer broadcast in the RL trainers	2026-09-14 22:22:00 +05:30
train_lora.py	[fix] ddp exit hang	2026-06-01 17:50:39 +08:00
train_ppo.py	[perf] skip the per-forward RoPE buffer broadcast in the RL trainers	2026-09-14 22:22:00 +05:30
train_pretrain.py	[update] remove expert args	2026-09-04 22:16:38 +08:00
train_tokenizer.py	[update] minimind-3	2026-03-25 23:57:45 +08:00
trainer_utils.py	[update] simplify safe math eval	2026-09-14 23:55:32 +08:00