Fix final checkpoint RoPE buffers and lm-eval inference

Save 100% checkpoint at 30,000,807,936 tokens

Save 95% checkpoint at 28,501,344,256 tokens

Save 90% checkpoint at 27,000,832,000 tokens

Save 85% checkpoint at 25,501,368,320 tokens

Save 80% checkpoint at 24,000,856,064 tokens

Save 75% checkpoint at 22,501,392,384 tokens

Save 70% checkpoint at 21,000,880,128 tokens

Save 65% checkpoint at 19,501,416,448 tokens

Save 60% checkpoint at 18,000,904,192 tokens

Save 55% checkpoint at 16,501,440,512 tokens

Save 50% checkpoint at 15,000,928,256 tokens

Update modeling_tiger_alpha.py

Save 45% checkpoint at 13,500,416,000 tokens

Save 40% checkpoint at 12,000,952,320 tokens

Save 35% checkpoint at 10,500,440,064 tokens

Save 30% checkpoint at 9,000,976,384 tokens

Load persistent RoPE buffers in lm-eval

Fix persistent RoPE buffers for Hugging Face lm-eval

Reapply finite lm-eval inference fix to 25 percent checkpoint

Save 25% checkpoint at 7,500,464,128 tokens

Fix NaN logits in padded batched inference

Save 20% checkpoint at 6,001,000,448 tokens

Save 15% checkpoint at 4,500,488,192 tokens

Save 10% checkpoint at 3,001,024,512 tokens

Save 5% checkpoint at 1,500,512,256 tokens

Add tokenizer

Initialize Tiger Alpha

initial commit