chore: import upstream snapshot with attribution

YOCO/scripts/eval_needle.sh

@@ -0,0 +1,11 @@
+cd yoco/
+torchrun --master-port=29504 --nproc_per_node=1 validate.py \
+    --task pseudo \
+    --criterion multi_needle --needle-num 4 \
+    --batch-size 1 \
+    --max-epoch 1 \
+    --no-save \
+    --tiktoken-model cl100k_base \
+    --bf16 \
+    --arch yoco_3b_new --tiktoken-model cl100k_base --load-ckpt /data/yutao/ckpt_opensource/YOCO-3B-1M/checkpoint.pth --yoco-model /data/yutao/ckpt_opensource/YOCO-3B-1M --tokens-per-sample 1048576 --interval 1048576
+

YOCO/scripts/eval_task.sh

@@ -0,0 +1,17 @@
+TASK='harness_boolq'
+# TASK='hendrycksTest-abstract_algebra'
+
+cd yoco/
+torchrun --master-port=29505 --nproc_per_node=1 validate.py \
+    --data-dir ../harness_data/ \
+    --criterion harness_eval \
+    --task harness_eval \
+    --batch-size 4 \
+    --eval-data ${TASK}  \
+    --log-format simple  --log-interval 10 \
+    --bf16 \
+    --tokenizer-pad-to-multiple 8 \
+    --arch yoco_3b_new --tiktoken-model cl100k_base --load-ckpt /data/yutao/ckpt_opensource/YOCO-3B-1M/checkpoint.pth --yoco-model /data/yutao/ckpt_opensource/YOCO-3B-1M  --tokens-per-sample 4096
+    # --arch llama_from_ckpt --llama-model /data/yutao/llama/llama-2-7b --load-ckpt /data/yutao/llama/llama-2-7b/consolidated.00.pth --tokens-per-sample 4096
+
+

YOCO/scripts/train.sh

@@ -0,0 +1,27 @@
+cd yoco/
+torchrun --master-port=29501 --nproc-per-node=1 train.py /mnt/nlcredstone/shaohanh/data/redstone_v4_21_config \
+    --save-interval-updates 5000 \
+    --no-epoch-checkpoints \
+    --arch yoco_base \
+    --criterion cross_entropy \
+    --task gpt \
+    --tokens-per-sample 2048 \
+    --tokenizer-pad-to-multiple 8 \
+    --pad-to-max-len \
+    --optimizer adam --adam-betas "(0.9, 0.95)" \
+    --adam-eps 1e-06 \
+    --clip-norm 2.0 \
+    --lr 0.00015 \
+    --lr-scheduler polynomial_decay \
+    --warmup-updates 50 \
+    --weight-decay 0.05 \
+    --batch-size 1  \
+    --model-parallel-size 1 \
+    --update-freq 1 \
+    --batch-read-ahead 1000 \
+    --total-num-update 300000 \
+    --log-format simple      --log-interval 10    --disable-validation \
+    --tiktoken-model cl100k_base \
+    --no-save \
+    --bf16 \
+