Upload scripts/training/train_flux_lora.py with huggingface_hub b476a30 verified memoryai commited on 4 days ago
Upload scripts/serving/inference_4k.py with huggingface_hub 94e279b verified memoryai commited on 4 days ago
Upload scripts/training/train_flux_lora.py with huggingface_hub faacfe9 verified memoryai commited on 4 days ago
Upload scripts/training/train_flux_lora.py with huggingface_hub 69bd807 verified memoryai commited on 4 days ago
Upload scripts/training/train_flux_lora.py with huggingface_hub 5a193ef verified memoryai commited on 4 days ago
v11: JOB_ID for multi-VPS parallel training, auto-detect NVMe 18970a4 verified memoryai commited on 9 days ago
v10: auto-detect all NVMe disks + GPU count, no hardcoded devices d3810d0 verified memoryai commited on 9 days ago
v9: download 100K COYO images (5 partitions), 4 processes, 256 threads 252cd07 verified memoryai commited on 9 days ago
Fix NaN: autocast forward, NaN skip, show cur_loss for debug 32ae43b verified memoryai commited on 10 days ago
Fix NaN: rewrite training loop - proper grad accum, zero_grad before loop, correct loss averaging 11b11fa verified memoryai commited on 10 days ago
Add NaN debug prints to identify source of NaN loss 22f612e verified memoryai commited on 10 days ago
Fix NaN loss: cast LoRA params to fp32 for stable bf16 backward dd759d4 verified memoryai commited on 10 days ago
Fix: num_workers=0 to avoid multiprocessing crash with few shards 6e8b5bd verified memoryai commited on 10 days ago
Fix: num_workers <= num_shards, add empty_check=False for webdataset bdb80f0 verified memoryai commited on 10 days ago
v8: mount NVMe properly (no sudo), use train_flux_lora.py (no precompute needed) e8a9daa verified memoryai commited on 10 days ago
Fix train_flux_lora.py: add cache-dir, resume, 2GPU split, infinite train 85eb3ff verified memoryai commited on 10 days ago
v7: explicit Flux model download with progress, no pipe to prevent timeout kill 88eee8e verified memoryai commited on 10 days ago
v6: fix shard threshold, separate Flux model download, better error logs 3f6f388 verified memoryai commited on 10 days ago
Job v5: train forever, auto-backup 30min, resume on resubmit 9bed334 verified memoryai commited on 10 days ago
Job v4: full A-Z with foreground data download + auto-backup every 30min e52a784 verified memoryai commited on 10 days ago
Job v3: run autosetup.sh first, then precompute + train a6bad17 verified memoryai commited on 10 days ago
Fix wc -l whitespace bug, add auto-download data fallback a1a1386 verified memoryai commited on 10 days ago
Add complete Azure ML job script (token removed from comments) cb21559 verified memoryai commited on 10 days ago
Fix train_flux_deepspeed.py: single/multi GPU compat, resume logic, stable training b1c12a9 verified memoryai commited on 10 days ago
memoryai commited on 