text stringlengths 1 90 |
|---|
# RL rollout pool, level4_huge members (staged into level_sparse/level4_huge/ by |
# stage_rl_pool.py). Only matrices whose SpGEMM baseline the harness can run |
# live on one GPU (cuSPARSE, 143 GB H200, C = A*A / A*A^T) -- see |
# spgemm_excluded.txt for the ten level4 matrices that were dropped and why. |
# Sources: test_set/level4_huge (4) and rl_pool/level4_huge regenerated by |
# build_rl_pool.py from level4_regen_list.txt (8). |
Bourchtein_atmosmodd |
DIMACS10_delaunay_n24 |
Freescale_circuit5M_dc |
Freescale_Freescale1 |
Freescale_memchip |
Hamrle_Hamrle3 |
Janna_CoupCons3D |
Schenk_AFE_af_shell10 |
Schmid_thermal2 |
SNAP_roadNet-CA |
SNAP_roadNet-TX |
Zaoui_kkt_power |
# Matrices dropped from the SpGEMM RL rollout pool (build_dataset.py --exclude-from). |
# Criterion: the harness must produce a cuSPARSE baseline live on one GPU |
# (benchmark_spgemm, 143 GB H200), and a cuSPARSE-backed reference kernel that |
# follows the kernel contract must run. Where that fails, every rollout scores |
# -1/-0.5 regardless of the policy, i.e. zero learning signal per GRPO group. |
# Checked 2026-09-12 with a cuSPARSE reference kernel (C = A*A, or A*A^T for |
# rectangular A via Csr2cscEx2). Error 11 = CUSPARSE_STATUS_INSUFFICIENT_RESOURCES. |
# |
# --- level4_huge: harness baseline itself fails (no external HSMU dump) --- |
DGL_reddit # cusparse_spgemm.cu:91 err 11 (A*A does not fit) |
OGB_ogbn-proteins # cusparse_spgemm.cu:91 err 11 |
SNAP_wiki-Talk # cusparse_spgemm.cu:111 err 11 (power-law, A*A explodes) |
SNAP_wiki-talk-temporal # cusparse_spgemm.cu:111 err 11 |
# --- level4_huge: baseline only via a precomputed dump in /mnt/data2/Dr.Sparse/spgemm_ref |
# (SPGEMM_EXT_REF_DIR, not on the remote); live cuSPARSE runs out of memory, |
# and the GPT-5.6 teacher never produced a >=1.05x kernel on any of them --- |
Belcastro_human_gene1 # C nnz 224M; ref kernel err 11 |
Belcastro_mouse_gene # C nnz 483M; ref kernel err 11 |
DIMACS10_asia_osm # 12M rows; ref kernel err 11 |
Gupta_gupta3 # ref kernel err 11 |
LAW_in-2004 # ref kernel err 11 |
QY_case39 # ref kernel err 11 |
# --- level3_large: harness baseline itself fails --- |
LAW_cnr-2000 # cusparse_spgemm.cu:111 err 11 (web graph, A*A explodes) |
Dr.Sparse RL rollout pool, level4_huge members
The SpGEMM online-RL rollout pool is 130 SuiteSparse matrices: the 118 level1-3
matrices in the Dr.Sparse git repo (level_sparse/level{1_small,2_medium,3_large})
plus the 12 level4 matrices in this dataset. It is disjoint from the 100-matrix
OTF test set.
The 12 level4 (>= 1.1M rows) matrices, in
Dr.Sparse, in the repo's raw .bin
layout ([int32 rows, cols, nnz][int32 row_ptr][int32 col_ind][float32 values][float32 x],
read by level_sparse/data_loader.h). The level1-3 members (118 matrices) ship
in the git repo; the 100-matrix OTF test set is KinGeorge/Dr.Sparse-OTF-test-set.
Stage into a checkout with:
python level_sparse/stage_rl_pool.py --download # -> level_sparse/level4_huge/
level4_list.txt is the member list; spgemm_excluded.txt records the ten
level4 matrices that were dropped because cuSPARSE cannot produce an SpGEMM
baseline for them on one GPU. Values are the real matrix values (float32);
x is a fixed standard-normal vector baked in at build time.
- Downloads last month
- 46