56 lines
1.4 KiB
YAML
56 lines
1.4 KiB
YAML
|
# FRANKENSTALLM-H 3B: Hybrid Mamba-2 + Transformer
|
||
|
|
#
|
||
|
|
# [설계 근거 — 2026-03-05]
|
||
|
|
# - 아키텍처: Nemotron-H 8B Dense 참고, 3B 스케일 적용
|
||
|
|
# - 40 layers: 37 Mamba-2 + 3 Attention (layer 13, 26, 39)
|
||
|
|
# - 파라미터: ~2.9B (embedding 포함)
|
||
|
|
# - 데이터: 3b_train.bin (기존 Pure Transformer 동일 데이터)
|
||
|
|
# - lr=2e-4: Mamba-2 논문 참고, Transformer보다 약간 높음
|
||
|
|
# - Attention 3개: 초반(13), 중반(26), 후반(39) 균등 배치
|
||
|
|
# - Mamba 장점: O(n) 시퀀스 처리, 추론 시 constant memory
|
||
|
|
#
|
||
|
|
# 실행: bash scripts/launch_hybrid_3b.sh
|
||
|
|
model:
|
||
|
|
vocab_size: 64000
|
||
|
|
d_model: 3072
|
||
|
|
n_layers: 40
|
||
|
|
n_heads: 24
|
||
|
|
n_kv_heads: 8
|
||
|
|
d_ffn: 9216
|
||
|
|
max_seq_len: 4096
|
||
|
|
rope_theta: 500000.0
|
||
|
|
dropout: 0.0
|
||
|
|
bias: false
|
||
|
|
use_flash_attn: true
|
||
|
|
use_fp8: true
|
||
|
|
# Hybrid settings
|
||
|
|
use_hybrid: true
|
||
|
|
hybrid_pattern: "M M M M M M M M M M M M M A M M M M M M M M M M M M A M M M M M M M M M M M M A"
|
||
|
|
mamba_d_state: 128
|
||
|
|
mamba_head_dim: 64
|
||
|
|
mamba_expand: 2
|
||
|
|
mamba_conv_kernel: 4
|
||
|
|
mamba_n_groups: 1
|
||
|
|
mamba_chunk_size: 256
|
||
|
|
|
||
|
|
train:
|
||
|
|
max_steps: 57000
|
||
|
|
batch_size: 4
|
||
|
|
grad_accum_steps: 8
|
||
|
|
lr: 2e-4
|
||
|
|
weight_decay: 0.1
|
||
|
|
warmup_steps: 2000
|
||
|
|
max_grad_norm: 1.0
|
||
|
|
log_interval: 10
|
||
|
|
save_interval: 2000
|
||
|
|
eval_interval: 500
|
||
|
|
use_amp: false
|
||
|
|
compile_model: false
|
||
|
|
fp8_amax_history_len: 16
|
||
|
|
fp8_amax_compute_algo: "max"
|
||
|
|
fp8_format: "MXFP8"
|
||
|
|
|
||
|
|
tokenizer:
|
||
|
|
vocab_size: 64000
|
||
|
|
type: sentencepiece_unigram
|