Module qwen3_attention 

Qwen3 full attention layer.

Q/K/V projection -> Q/K norms -> RoPE -> KV cache write -> paged decode attention -> O projection, then MoE FFN.

Split into submodules:

• types: MlaWeights + Qwen3AttentionLayer struct definitions
• init: new, new_ungated, new_with_gating (kernel loading)
• helpers: setters + apply_layer_scalar + effective_attn_scale
• prefill_weights: prefill weight setup + W4A16 M128 dispatcher
• decode: single-token attention forward + KV cache helpers
• prefill: batched prefill with paged attention
• trait_impl: TransformerLayer trait implementation

Re-exports

pub use innerq_driver::InnerQDriver;

Modules

innerq_driver

Host-side driver for TurboQuant+ InnerQ per-channel K equalization.

Structs

CompressorWeights

DeepSeek-V4 compressed-attention compressor weights (one per compressed layer). Produces n_win = usable/ratio compressed KV entries that are concatenated to the raw sliding-window KV before core attention. CSA (ratio 4) uses a 2×ratio overlap window (Ca/Cb); HCA (ratio 128) uses a single non-overlapping window.

HcHeadWeights

Both HC sites for a DeepSeek-V4 block: the attention site runs before/after attention, the FFN site before/after the MoE FFN. Model-level HC head parameters (final collapse before LM head). Loaded once, attached to every layer, but only used by the last layer.

HcLowRank

Qwen3.8-Flash-Next’s LOW-RANK hyper-connection parameters for one site.

HcSiteWeights

Per-block Manifold-Constrained Hyper-Connection (mHC) parameters for one site (attention or FFN). All buffers are float32 device pointers, matching the checkpoint dtype. See ops::hc_pre / ops::hc_post.

HcWeights

MlaWeights

MLA (Multi-head Latent Attention) weight components for 2-step decode.

Qwen3AttentionLayer

Qwen3-Next full attention layer (12 of 48 layers).

Statics

ATTN_PHASE_US

Per-phase host-time accumulators for the prefill ATTENTION path (ATLAS_PREFILL_HOST_TIMING=1). Index: 0=qkv projections, 1=everything between qkv and the attention call (deinterleave + per-head norms + RoPE + KV write), 2=the attention kernel call itself, 3=o_proj + head gate. Summed across layers; read and reset once per prefill.

FFN_HOST_US

Host-time accumulator for the FFN/MoE half of prefill layers (ATLAS_PREFILL_HOST_TIMING=1). Summed across layers and read+reset once per prefill by the layer loop, so the attention half can be derived as loop_wall - ffn.

Functions

add_attn_phase_us

add_ffn_host_us

take_attn_phase_us

take_ffn_host_us

validate_required_kv_kernels

Startup fail-fast for --kv-cache-dtype: resolve every kernel handle the dtype’s dispatch arms require (chunked-prefill kernel, WHT bookends) and error with the full missing list — BEFORE the multi-minute weight load, instead of at first dispatch. See kernel_requirements.rs.