Module mla 

Gated NoPE MLA CPU reference.

Do not reuse qwen3_attention blindly: K3 sets mla_use_nope=true while still allocating qk_rope_head_dim slots, and mla_use_output_gate=true applies a full-rank g_proj sigmoid gate on the attention output.

RoPE dims stay in the head (prod 128 nope + 64 rope = 192). NoPE means those slots are not rotated, not that they are dropped.

CUDA decode is kernels/gb10/kimi-k3/bf16/mla_decode.cu (k3_mla_*). BoundLayer FullAttention uses CUDA mla_decode unless K3_CUDA_MLA=0.

Structs

MlaConfig

Functions

apply_output_gate

Apply sigmoid(g) ⊙ attn when the output gate is on; otherwise identity.

cuda_mla_enabled

FullAttention BoundLayer uses CUDA mla_decode unless K3_CUDA_MLA=0. Same polarity as K3_CUDA_KDA. Projections stay on the host either way.

gated_mla_attend

Gated NoPE attend: optionally skip RoPE, SDPA, optional output gate.

maybe_rope

Optional RoPE on the rope slice of a packed [nope | rope] head. NoPE leaves q/k unchanged.

mla_decode_token

One decode token: optional RoPE, append K/V, SDPA, optional output gate. Projections stay in mla_mixer. CUDA k3_mla_* matches this order.

mla_from

sdpa

Causal scaled-dot-product attention. q/k: [T, H, dq], v: [T, H, dv].

sdpa_one

Decode SDPA: one query [H, dq] against cached K/V of length t.