Best speedup
4.62x
17 accepted edits
Run detail
Live optimizer showcase - multi-op curve
modal-b200-live on NVIDIA B200 with Fireworks coding agent + 1M controller. Git c8e96db.
runningOpen HUD
Best latency
0.0310 ms
Lower is better
Trials
24
7 rejected
Artifacts
1
Blob-backed source and logs
Improvement curve
Preparing chart...
GPU samples
Preparing chart...
Reward
raw reward preservedPreparing chart...
Trial audit log
Best source| Trial | Op | Candidate | Verdict | Reward | Speedup | Latency | Cost |
|---|---|---|---|---|---|---|---|
| #24 | sum_reduction | two_stage_reduce | accepted | 4.620 | 4.62x | 0.0310 ms | $0.0043 |
| #23 | prefix_scan | warp_scan_v2 | accepted | 4.310 | 4.31x | 0.0360 ms | $0.0041 |
| #22 | prefix_scan | bad_scan_order incorrect prefix order | incorrect | 0.000 | n/a | n/a | $0.0040 |
| #21 | embedding_lookup | coalesced_cache_v1 | accepted | 4.080 | 4.08x | 0.0390 ms | $0.0038 |
| #20 | quantize_dequant | vectorized_pack_v1 | accepted | 3.760 | 3.76x | 0.0440 ms | $0.0036 |
| #19 | fused_mlp | persistent_blocks_v3 | accepted | 3.420 | 3.42x | 0.0520 ms | $0.0034 |
| #18 | fused_mlp | persistent_blocks_v2 correct but not best | rejected | 2.440 | 2.44x | 0.2130 ms | $0.0032 |
| #17 | fused_mlp | fuse_gelu_matmul | accepted | 2.910 | 2.91x | 0.1790 ms | $0.0031 |
| #16 | fused_mlp | default_seed | accepted | 1.000 | 1.00x | 0.5200 ms | $0.0029 |
| #15 | attention_softmax | overfit_train_shape continuous held-out shape failed | incorrect | 0.000 | n/a | n/a | $0.0027 |
| #14 | attention_softmax | online_max_sum | accepted | 2.630 | 2.63x | 0.0840 ms | $0.0025 |
| #13 | attention_softmax | default_seed | accepted | 1.000 | 1.00x | 0.2210 ms | $0.0023 |
| #12 | matmul_tile | tensor_core_64x64 | accepted | 2.480 | 2.48x | 0.1250 ms | $0.0022 |
| #11 | matmul_tile | tensor_core_32x64 | accepted | 2.120 | 2.12x | 0.1460 ms | $0.0020 |
| #10 | matmul_tile | bad_boundary_mask held-out correctness failed | incorrect | 0.000 | n/a | n/a | $0.0018 |
| #9 | matmul_tile | default_seed | accepted | 1.000 | 1.00x | 0.3100 ms | $0.0016 |
| #8 | softmax_rows | warp_reduce_v2 | accepted | 1.860 | 1.86x | 0.0880 ms | $0.0014 |
| #7 | softmax_rows | warp_reduce_v1 correct but slower than best | rejected | 0.880 | 0.88x | 0.1860 ms | $0.0013 |
| #6 | softmax_rows | torch_passthrough PyTorch passthrough rejected by anti-hack check | incorrect | 0.000 | n/a | n/a | $0.0011 |
| #5 | softmax_rows | default_seed | accepted | 1.000 | 1.00x | 0.1640 ms | $0.0009 |
| #4 | rmsnorm | block256_warp4 | accepted | 1.540 | 1.54x | 0.0760 ms | $0.0007 |
| #3 | rmsnorm | block128_vectorized | accepted | 1.280 | 1.28x | 0.0920 ms | $0.0005 |
| #2 | rmsnorm | shape_hack_wrong_stride shape/stride mismatch blocked | incorrect | 0.000 | n/a | n/a | $0.0004 |
| #1 | rmsnorm | default_seed | accepted | 1.000 | 1.00x | 0.1180 ms | $0.0002 |