MLA paged decode h8 ckv512 kpe64 ps1
0 eligible runs
mla-paged-attention
Batched Multi-head Latent Attention decode with a paged KV cache. Captured from Kimi K2 / Kimi K2.5 with tensor parallel size 8 (64/8=8 query heads). The Kimi K2.5 text backbone (text_config.model_type=kimi_k2, DeepseekV3ForCausalLM) shares the same MLA shape as Kimi K2: kv_lora_rank=512, qk_rope_head_dim=64, qk_nope_head_dim=128, v_head_dim=128, num_attention_heads=64 → h=8 at TP=8.
Current records
No published measurement for the selected workload.
Implementations
Implementation
Runtime
Best latency
Evidence
Availability
Semantics
Inputs and outputs
q_nopebf16 [batch_size, num_qo_heads, head_dim_ckv]
q_pebf16 [batch_size, num_qo_heads, head_dim_kpe]
ckv_cachebf16 [num_pages, page_size, head_dim_ckv]
kpe_cachebf16 [num_pages, page_size, head_dim_kpe]
kv_indptrint32 [len_indptr]
kv_indicesint32 [num_kv_indices]
sm_scalefp32 scalar
outputbf16 [batch_size, num_qo_heads, head_dim_ckv]
lsefp32 [batch_size, num_qo_heads]
Axes and behavior
num_pagesvariable
page_sizeconstant = 1
batch_sizevariable
len_indptrvariable
head_dim_ckvconstant = 512
head_dim_kpeconstant = 64
num_qo_headsconstant = 8
num_kv_indicesvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
No source imports for this operation yet.How records are decidedJSON