初始化项目，由ModelHub XC社区提供模型

Model: pathcosmos/frankenstallm Source: Original Platform
2026-07-14 04:21:16 +08:00
commit d4abdb70fa
516 changed files with 5584477 additions and 0 deletions
--- a/source/configs/3b_pretrain.yaml
+++ b/source/configs/3b_pretrain.yaml
@@ -0,0 +1,60 @@
+# Korean LLM 3B parameters — FP8 (B200 TransformerEngine MXFP8)
+#
+# [아키텍처 근거 — 2026-02-27]
+#   - 저스티스리그 제안 기반: d_model=2560, 32L, 32H, 8KV
+#   - 파라미터: ~2.39B ("3B급" — Llama-3.2-3B 대비 경량, 한국어 64K vocab 효율)
+#   - d_ffn=6912: 2.7×d_model, 16배수 FP8 정렬
+#   - GQA 4:1 (32H:8KV) — 추론 효율 + KV cache 절약
+#   - head_dim=80 (2560/32) — Flash Attention 효율적
+#
+# [데이터/학습 설계]
+#   - 데이터: korean_train.bin 8.91B tokens
+#   - Chinchilla 최적: 2.4B × 20 = 48B tokens
+#   - 실제 목표: 60B tokens (6.7 에포크) — 한국어 단일 언어 특성상 추가 학습 유리
+#   - max_steps 57000 = 60B tokens / 1,048,576 tok/step
+#
+# [GPU 메모리 예측 — 8× B200 183GB]
+#   - 모델 FP8: 2.4 GB
+#   - Optimizer (bf16 master + fp32 mom/var): 23.9 GB
+#   - Gradient (bf16): 4.8 GB
+#   - Activation (per GPU, bs=8): ~27 GB
+#   - 합계: ~58 GB/GPU (31.7% 활용) → 여유 충분
+#
+# 실행: bash scripts/launch_korean_3b.sh
+# 테스트: RUN_NAME=korean_3b_test bash scripts/launch_korean_3b.sh --max_steps 50
+
+model:
+  vocab_size: 64000
+  d_model: 2560
+  n_layers: 32
+  n_heads: 32
+  n_kv_heads: 8           # GQA 4:1 (K/V 파라미터 75% 절감)
+  d_ffn: 6912             # 2.7×d_model, 16배수 (FP8 alignment)
+  max_seq_len: 4096
+  rope_theta: 500000.0
+  dropout: 0.0
+  bias: false
+  use_flash_attn: true
+  use_fp8: true            # TransformerEngine MXFP8BlockScaling (B200 네이티브)
+
+train:
+  # 57k steps × 1,048,576 tok/step = 59.8B tokens ≈ 6.7 에포크
+  max_steps: 57000
+  batch_size: 4            # per GPU: 4 × 4096 = 16,384 토큰 | VRAM ~130 GB (183GB의 71%)
+  grad_accum_steps: 8      # eff_batch: 4 × 8GPU × 8 × 4096 = 1,048,576 tok/step
+  lr: 1.5e-4               # 3B 규모: GPT-3 scaling 기준 1B(2e-4) → 3B(1.5e-4)
+  weight_decay: 0.1
+  warmup_steps: 2000       # 57k steps의 3.5% — 안정적 warmup
+  max_grad_norm: 1.0
+  log_interval: 10
+  save_interval: 1000      # 57k steps 기준 ~57 체크포인트
+  eval_interval: 500       # val loss 모니터링
+  use_amp: false           # fp8_autocast가 대체
+  compile_model: false     # TE 2.10 + DDP graph break 위험
+  fp8_amax_history_len: 16
+  fp8_amax_compute_algo: "max"
+  fp8_format: "MXFP8"     # B200 Blackwell 네이티브 블록 스케일링
+
+tokenizer:
+  vocab_size: 64000
+  type: sentencepiece_unigram