fix: restore grad_ckpt and safe micro 8 accum 8 for zero-OOM guarantee on T4

perf: disable grad_ckpt on T4 and enlarge loss chunk_size to 2048 for 10k+ tok/s

config: update Stage 1 batch_size to 16 for micro=16 accum=8 per GPU

refactor: consolidate pure 32L Dense architecture, purge MoE files, audit hardening

Fix model/scripts/train.py: batch math comments and find_unused_parameters=False

Audit fix: Update model/scripts/train.py for ViuMini-Dense-360M

Update model/scripts/train.py for ViuMini-Dense-360M

Upgrade architecture to ViuMini-Dense-350M (100% Dense, 360M active)

Restore find_unused_parameters=True: prevent MoE DDP reducer freeze

Speed optimization: set DDP find_unused_parameters=False

Add UTF-8 stdout support to train.py

P0 fix: ensure train.py targets viu1_moe 242M with updated defaults

Add live elapsed time counter (fmt_duration) and step pace (s/step, step/h) to console logs

Fix MoE Muon lag: route 4032 sparse expert weights to PagedAdamW8bit (eliminates 68,000 matmuls per step, 15x faster)

RTX 3090 Max Speed: Zero-sync GPU loss accum, pinned-memory batch slicing, Muon in VRAM, 262k tok/step curriculum

RTX 3090 Max Speed: Zero-sync GPU loss accum, pinned-memory batch slicing, Muon in VRAM, 262k tok/step curriculum

Fix CUDA OOM on T4: cast model to amp_dtype directly, batch_size=2, safe RMSNorm

Fix CUDA OOM on T4: cast model to amp_dtype directly, batch_size=2, safe RMSNorm

Fix CUDA OOM on T4: cast model to amp_dtype directly, batch_size=2, safe RMSNorm

Fix CUDA OOM on T4: cast model to amp_dtype directly, batch_size=2, safe RMSNorm

Fix CUDA OOM on T4: cast model to amp_dtype directly, batch_size=2, safe RMSNorm

Sync model/scripts/train.py with 3-stage sequence length curriculum (512->1024->2048)

Sync model/scripts/train.py with adult erotica stories priority integration

Sync model/scripts/train.py with adult erotica stories priority integration

feat: step-1 priority queue integration for real Indian history & cricket shards

feat: 100% zero-exclusion streaming architecture (2,081 shards) and updated docs

feat: add CLI overrides and manual batch lock support

fix: auto-download tokenizer.json if LFS pointer detected

fix: mute DDP barrier and find_unused_parameters stderr warnings

feat: add RTX 5090 & 4090 configs (model/scripts/train.py)

feat: smart auto-download Hugging Face checkpoint resume

fix: add --hf_split argument and remove invalid attribute access in streaming logger

fix: pass token to list_repo_files and HfFileSystem to eliminate unauthenticated warnings

feat: non-blocking asynchronous checkpoint upload and local disk pruning

Add --hf_safety argument and default full

Replace load_dataset with direct PyArrow streaming generator: Eliminate CastError permanently

Clean pretraining logging: Streamlined single-line logs (loss, lr, speed, ETA) with zero console clutter

Fix CastError permanently: Direct parquet loader via hf:// paths + clean step 1-20 logging with pbar.write

Fix CastError permanently: exclude code/ prefix + print every step for first 10 steps

Speed optimizations: no_sync DDP communication + cuDNN/TF32 + 10k ckpt sync

Enhance live training logs: print at step 1 and show all 8 MoE expert percentages

Update model/scripts/train.py: Exclude distilled/ from streaming + set log_every to 10

Fix train.py argparse: Support --max_steps alias and --resume flag

feat: wipe-proof HF persistence (full ckpt + live log + newest-2 cleanup)

feat: full-ckpt HF push every 10k steps (resume safety net) + docs

feat: hf_round corpus-slice rotation (Option A full-data: file rotate + row skip + per-round seed)

fix: exclude non-5col schema paths from HF stream (wikipedia/grammar/gsm8k CastError on shuffle)

fix: probe OOM reserve 4GB + expandable_segments + natural-mode stream shuffle (Hindi-only trap)

fix: DDP auto_tune probe-before-wrap + broadcast micro/accum, R0-only logs, empty-shard guard, safety import path

DDP multi-GPU support (torchrun, stride-shard, rank0 ckpt, ckpt without module prefix)