accelerate

ffc870d6 - Fix next-token labels and accumulation target weighting (#4303)

Fix next-token labels and accumulation target weighting (#4303) * Compare autoregressive example updates with a global-token reference Replace the exit-only example check with offline two-rank FP32 validation of its real training loop. Unequal shifted target counts expose microbatch-mean and rank-local normalization mistakes against an independent full-batch objective. Co-Authored-By: OpenAI GPT-6 Astra <noreply@openai.com> * Preserve the autoregressive example smoke alongside token weighting Retain the existing CLI, configured launcher, pretrained loading and mocked-data integration coverage unchanged. Keep the distinct numerical contract in one additional CPU DDP case. Co-Authored-By: OpenAI GPT-6 Astra <noreply@openai.com> * Fix next-token labels and normalization in accumulation example * Fix next-token labels and accumulation target weighting Keep labels aligned for the model to shift once, and count only predicted targets in the loss denominator. Check the real example against an independent full-batch update, alongside both collators and the existing CLI path. * Keep accumulation example changes focused on correctness * Limit autoregressive correction to the example and testing loader --------- Co-authored-by: OpenAI GPT-6 Astra <noreply@openai.com>