v7.1 training: update header, base config BACKPROP_K default

v7.1: chunked CE loss to fix OOM on base config, T4-friendly base defaults

v7 config: add backprop_depth, use_rope

v7: fix LR schedule, backprop depth k=4, RoPE, eps=1e-8

v7 training: fix LR schedule mismatch, k=4 backprop, eps=1e-8

v6: Replace CausalConv with standard causal attention (Huginn-style), sandwich norm, proper state injection"

v5 training: LR=5e-5, bf16 on T4, eps=1e-15, NaN detection+recovery, wd=0.1

v5: Fix all 5 training instabilities — inter-iteration norm, lower LR, z-loss, bf16 on T4, eps=1e-15"

Update ARCHITECTURE.md for v4.1 depth-recurrent design

Update config.json for v4.1 architecture

Complete README rewrite for v4.1 — accurate architecture, all env vars, training guide

v4.1 model: fix gated residual (was blocking gradient flow), add proper pre-norm residual pattern

v4.1 Colab notebook: fix total_memory, add SEQ_LEN/SAVE_EVERY/EVAL_EVERY, correct architecture description

v4.1: Add periodic sample generation, fix notebook (total_memory, SEQ_LEN, architecture desc), add EVAL/SAVE env vars

v4 training script: match new architecture API, proper depth recurrence training

v4: Complete redesign — proper depth recurrence with transposed MLP (TRM-style), not broken sequence recurrence"

Fix scheduler warning: initial step before loop, keep optimizer step order correct

Fix NaN: entire recurrence in float32 (disable autocast), numerically stable formulation, fix scheduler ordering

Major fix: replace sequential WKV loop with chunked parallel processing, fix T4 OOM and hang"

Fix train defaults for T4: BS=4, seq=256, N_sup=2, add PYTORCH_CUDA_ALLOC_CONF

Fix T4 OOM: gradient checkpointing, disable autocast in WKV loop, reduce seq_len default to 256, explicit cache clearing

Fix OOM on T4: reduce defaults (BS=4,N_sup=2), fix trackio API (dir→project), remove trust_remote_code"

Fix: proper AMP for T4(fp16)/A100(bf16)/CPU(fp32), GradScaler only with fp16

Fix: RMSNorm compat (PyTorch <2.4), dtype handling, GradScaler logic, WKV mixed precision

Fix: total_mem → total_memory (PyTorch API fix)

Add detailed architecture documentation

Add comprehensive README with architecture documentation

Add model config

Add Colab/Kaggle training notebook

Add training script (Colab/Kaggle compatible)

Add NEXUS architecture core module

initial commit