rtpurbo-stage2-wip

RTPurbo Stage-1 indexer — d_idx=32, Qwen3.5-9B, full 32-131K context

Trained by distilling the full self-attention's top-p=0.9 token set into a 32-dim retrieval index via forward KL (paper Eq.6), using a hybrid SVD initialisation that seeds rows 0-15 from a prior d_idx=16 model and rows 16-31 from SVD principal components of Vq[16:32].

Configuration

• base model: Qwen/Qwen3.5-9B (32 layers; 24 GDN + 8 GA at L3/7/11/15/19/23/27/31)
• GQA 16:4, head_dim=256, max_pos=262144
• seq_len=131072, chunked_teacher (cq=4096), teacher_query_sample=32768 (25%)
• teacher_query_min_pos=8192, reuse_k=5
• lr=1e-3 cosine, warmup=100, wd=0.01, batch=1, bf16
• init_from=hybrid_svd_dim32 (rows 0-15 from d_idx=16 ckpt, rows 16-31 SVD of Vq)
• steps=600, elapsed=341.4 min
• data: emozilla/pg19 train split, min_chars=120000 (docs roughly 30K-131K tokens)

Final probe metrics (real source-code text @ seq=131072, block=64)

• final_mean_kl = 1.7587
• final_max_kl = 4.4431

Compute reasoning (why d_idx=32, not 16)

Paper RTPurbo used head_dim=128 → d_idx=16 (1/8 ratio). Qwen3.5-9B has head_dim=256, so the paper-equivalent compression is 256/8 = d_idx=32. The earlier d_idx=16 run on this model was under-provisioned (equivalent to paper's d_idx=8). Probe confirms recall is sufficient (87% at top_p=0.9, 92% at top_p=0.95), so the bottleneck for tightening the sparsity gap is the training recipe (teacher-query-sample 25% vs paper's full-query), not d_idx.

Files

• indexer_final.pt — clean substate, 256 tensors of shape [32, 256] bf16
• state_dict.pt — same content, raw state_dict format
• stage1_summary.json— full per-head KL breakdown for all 8 GA layers
• loss_curve.csv — step,loss,lr,distill_kl,agree_kl,entropy,tok/s

Loading

import torch
from train.surgeries._rtpurbo_indexer import RetrievalIndexer

state = torch.load("indexer_final.pt", map_location="cpu")
# state is a flat dict of 256 tensors: q_heads.<layer_idx>.weight
# pass to RetrievalIndexer(d_idx=32, head_dim=256).load_state_dict(...)