Default relaunch to dblock MoE trial
Browse files
relaunch_agillm4_dblock_sg2.sh
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
#!/usr/bin/env bash
|
| 2 |
-
# Relaunch AGILLM-4 dblock with SG2's tuned config +
|
| 3 |
set -Eeuo pipefail
|
| 4 |
cd /workspace/agillm-4
|
| 5 |
export TOKENIZERS_PARALLELISM=false
|
|
@@ -9,17 +9,22 @@ export AGILLM_ATTN_BACKEND=sublinear
|
|
| 9 |
[ -f /root/.cache/huggingface/token ] && { export HF_TOKEN="$(tr -d '\r\n' </root/.cache/huggingface/token)"; export HUGGING_FACE_HUB_TOKEN="$HF_TOKEN"; }
|
| 10 |
SAVE_DIR=/workspace/agillm4_4090_ckpts
|
| 11 |
TOKEN_PARAM_RATIO="${TOKEN_PARAM_RATIO:-55}"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
CKPT="$(ls -1t "$SAVE_DIR"/pretrain_step*.pt 2>/dev/null | head -1)"
|
| 13 |
exec >> /workspace/agillm4_floor_train.log 2>&1
|
| 14 |
-
echo "
|
| 15 |
exec python -u nB300_agillm4.py train --preset agillm4_floor --resume "$CKPT" \
|
| 16 |
-
--dblock --dblock_blocks
|
| 17 |
--dblock_sigma_curriculum_steps 2000 --dblock_log_every 25 --dblock_objective_mode stochastic \
|
| 18 |
-
--dblock_ar_prob 0.
|
| 19 |
-
--dblock_ar_loss_tokens 512 --dblock_sat_loss_tokens 0 --dblock_nat_loss_tokens
|
|
|
|
| 20 |
--tie_weights --batch_size 6 --block 1024 --amp --attn_backend sublinear \
|
| 21 |
--sublinear_window 128 --sublinear_stride 128 --sublinear_max_anchors 128 --sublinear_chunk 128 \
|
| 22 |
--sublinear_sinks 4 --sublinear_recent_anchors 64 --no-sublinear_pooled_landmarks \
|
| 23 |
-
--grad_checkpoint --dblock_checkpoint_stride 1 --optimizer paged_adamw8bit --sat_every 4 --nat_every
|
| 24 |
--token_param_ratio "$TOKEN_PARAM_RATIO" --save_dir "$SAVE_DIR" --save_every_sec 86400 --heartbeat_every_sec 300 \
|
| 25 |
--empty_cache_every_steps 0 --delta_every_steps 25000 --delta_max_keep 1 --max_ckpts 1
|
|
|
|
| 1 |
#!/usr/bin/env bash
|
| 2 |
+
# Relaunch AGILLM-4 dblock with SG2's tuned config + optional MoE FFN upgrade.
|
| 3 |
set -Eeuo pipefail
|
| 4 |
cd /workspace/agillm-4
|
| 5 |
export TOKENIZERS_PARALLELISM=false
|
|
|
|
| 9 |
[ -f /root/.cache/huggingface/token ] && { export HF_TOKEN="$(tr -d '\r\n' </root/.cache/huggingface/token)"; export HUGGING_FACE_HUB_TOKEN="$HF_TOKEN"; }
|
| 10 |
SAVE_DIR=/workspace/agillm4_4090_ckpts
|
| 11 |
TOKEN_PARAM_RATIO="${TOKEN_PARAM_RATIO:-55}"
|
| 12 |
+
DBLOCK_BLOCKS="${DBLOCK_BLOCKS:-4}"
|
| 13 |
+
MOE_EXPERTS="${MOE_EXPERTS:-2}"
|
| 14 |
+
MOE_TOP_K="${MOE_TOP_K:-1}"
|
| 15 |
+
MOE_MLP_MULT="${MOE_MLP_MULT:-4}"
|
| 16 |
CKPT="$(ls -1t "$SAVE_DIR"/pretrain_step*.pt 2>/dev/null | head -1)"
|
| 17 |
exec >> /workspace/agillm4_floor_train.log 2>&1
|
| 18 |
+
echo "RELAUNCH_AGILLM4_DBLOCK_MOE_SG2 $(date -u +%Y-%m-%dT%H:%M:%SZ) resume=$CKPT (quality ratio55 + B6/L1024 + dblock_blocks=$DBLOCK_BLOCKS + moe_experts=$MOE_EXPERTS top_k=$MOE_TOP_K)"
|
| 19 |
exec python -u nB300_agillm4.py train --preset agillm4_floor --resume "$CKPT" \
|
| 20 |
+
--dblock --dblock_blocks "$DBLOCK_BLOCKS" --dblock_schedule loss_balanced --dblock_warmup_steps 16 \
|
| 21 |
--dblock_sigma_curriculum_steps 2000 --dblock_log_every 25 --dblock_objective_mode stochastic \
|
| 22 |
+
--dblock_ar_prob 0.70 --dblock_sat_prob 0.15 --dblock_nat_prob 0.15 \
|
| 23 |
+
--dblock_ar_loss_tokens 512 --dblock_sat_loss_tokens 0 --dblock_nat_loss_tokens 512 \
|
| 24 |
+
--moe_ffn --moe_experts "$MOE_EXPERTS" --moe_top_k "$MOE_TOP_K" --moe_mlp_mult "$MOE_MLP_MULT" \
|
| 25 |
--tie_weights --batch_size 6 --block 1024 --amp --attn_backend sublinear \
|
| 26 |
--sublinear_window 128 --sublinear_stride 128 --sublinear_max_anchors 128 --sublinear_chunk 128 \
|
| 27 |
--sublinear_sinks 4 --sublinear_recent_anchors 64 --no-sublinear_pooled_landmarks \
|
| 28 |
+
--grad_checkpoint --dblock_checkpoint_stride 1 --optimizer paged_adamw8bit --sat_every 4 --nat_every 4 --nat_max_tokens 768 --nat_mask_ratio 0.5 \
|
| 29 |
--token_param_ratio "$TOKEN_PARAM_RATIO" --save_dir "$SAVE_DIR" --save_every_sec 86400 --heartbeat_every_sec 300 \
|
| 30 |
--empty_cache_every_steps 0 --delta_every_steps 25000 --delta_max_keep 1 --max_ckpts 1
|