Kernel Judge

Challenges

Each score is the geometric mean of a submission's speedup over the baseline on every shape. The score ceiling is the score a submission would reach by hitting the floor on every shape. How to submit explains the rest.

ChallengeBest scoreScore ceilingDistance to ceiling
gdn_moe_layer
One Gated DeltaNet layer with its MoE block
no score yet 4.87x not available
qsa_moe_layer
One QSA sparse attention layer with its MoE block
no score yet 6.62x not available
embed_ple
Token embedding, all-reduce, n-gram PLE lookup and injection
no score yet 9.12x not available
lm_head
Final hyper-connection, LM head and logits all-gather
0.72x 4.60x 6.41x to go
verify_sampling
Top-k and top-p filtering with rejection sampling for MTP
no score yet 125.6x not available
step_metadata
Post-acceptance state commit and next-step metadata
no score yet not available not available
mtp_draft
MTP draft: one extend pass and two decode passes
no score yet 4.38x not available
dflash2_draft
Draft KV write, block preparation, five-layer draft model and candidate selector
no score yet 7.43x not available
dflash2_verify
Verify metadata, logits, acceptance and state commit
no score yet 9.66x not available