88 kB

Ctrl+K

2 contributors

out-of-core rework: embed_cluster_acts streams resids to disk memmap + streaming multi-GPU k-means (--kmeans-sample); build_data reads unified emb.f32 + GPU-batched Newton/IRLS probes

320b589 2 days ago

build_data.py

12.8 kB
out-of-core rework: embed_cluster_acts streams resids to disk memmap + streaming multi-GPU k-means (--kmeans-sample); build_data reads unified emb.f32 + GPU-batched Newton/IRLS probes 2 days ago
build_sae_data.py

4.39 kB
sae: loader + build_sae_data + eval_sae (cross-uplift metric); weights_only=False for torch 2.13 3 days ago
cache_resids.py

3.48 kB
scripts: attn sdpa (no flash-attn sm_103 build on Blackwell) 3 days ago
embed_cluster.py

4.42 kB
cluster: faiss-gpu has no Blackwell kernels — CPU faiss + torch-GPU recluster_torch.py (reuses cached emb) 3 days ago
embed_cluster_acts.py

9.57 kB
out-of-core rework: embed_cluster_acts streams resids to disk memmap + streaming multi-GPU k-means (--kmeans-sample); build_data reads unified emb.f32 + GPU-batched Newton/IRLS probes 2 days ago
eval_sae.py

8.61 kB
sae: loader + build_sae_data + eval_sae (cross-uplift metric); weights_only=False for torch 2.13 3 days ago
pretrain.py

12 kB
pretrain: default pack-len 0 (per-example+compile = validated 57% MFU; packing needs Blackwell flash-attn) 3 days ago
recluster_torch.py

2.55 kB
cluster: faiss-gpu has no Blackwell kernels — CPU faiss + torch-GPU recluster_torch.py (reuses cached emb) 3 days ago
rl.py

20.6 kB
rl: pin TRITON_ATTN backend (FLASHINFER silently breaks vllm-lens injection on Blackwell) 3 days ago
sbatch_rl.sh

2.01 kB
maxact-fast: hardening fixes (6 bugs) + Dr.GRPO vllm-lens rl.py + sbatch; Probe = universal direction 3 days ago
test_packing.py

7.49 kB
pretrain: sequence packing (block-diag attn, per-marker injection) — no-leak + inject proofs pass; MFU bench pending 3 days ago