Module layers 

Re-exports

pub use deepseek_v4_mtp::DeepseekV4MtpHead;

pub use deepseek_v4_mtp::DeepseekV4MtpProposerState;

pub use dense_ffn::DenseFfnLayer;

pub use dense_ffn::DenseFfnWeights;

pub use dense_ffn::FfnActivation;

pub use dflash_head::BlockDiffusionDraftHead;

pub use dflash_head::DflashLayer;

pub use dflash_head::DflashProposerState;

pub use dflash_head::DflashQuantization;

pub use dflash_head::dflash_ctx_cap;

pub use glm5next_mtp_head::Glm5NextMtpHead;

pub use moe::MoeLayer;

pub use mtp_head::MtpHead;

pub use mtp_head::MtpQuantization;

pub use mtp_head::mtp_drafter_prefill_enabled;

pub use nemotron_mamba2::NemotronMamba2Layer;

pub use nemotron_moe::NemotronMoeLayer;

pub use qwen3_attention::Qwen3AttentionLayer;

pub use qwen3_ssm::Qwen3SsmLayer;

pub use vision_encoder::MergerLayer;

pub use vision_encoder::ViTBlock;

pub use vision_encoder::VisionEncoder;

Modules

deepseek_v4_mtp

DeepSeek-V4-Flash Multi-Token-Prediction (MTP) draft proposer.

dense_ffn

Dense SwiGLU FFN component for non-MoE models.

dflash_head

DFlash block-diffusion draft head implementing DraftProposer.

ep_dispatch

EP=2 token dispatch/combine routing for MoE expert parallelism.

fp8_calibration

Online FP8 KV cache scale calibration.

glm5next_dsa

GLM-5.3-Flash KDA integrated layer (Slice 6 – one layer, no scheduler/cache wiring). GLM-5.3-Flash DSA (DeepSeek Sparse Attention) production surface.

glm5next_dsa_ref

GLM-5.3-Flash DSA + kpool indexer CPU reference (Slice 8 design artifact). GLM-5.3-Flash DSA (DeepSeek Sparse Attention) + kpool indexer CPU reference — Slice 8.

glm5next_kda

GLM-5.3-Flash KDA attention block — the reusable production component.

glm5next_kda_ref

GLM-5.3-Flash KDA CPU reference (Slice 2 design artifact – not a production forward path). GLM-5.3-Flash KDA (Kimi Delta Attention) CPU reference — Slice 2B.

glm5next_layer

GLM-5.3-Flash composite decoder layer – mixer (KDA|DSA) + MLP (dense|MoE) + mHC. Glm5NextLayer — the composite GLM-5.3 decoder layer that implements TransformerLayer.

glm5next_mlp

GLM-5.3-Flash MLP production surface – dense FFN + routed NVFP4 MoE (TP + EP sharded). GLM-5.3-Flash MLP production surface — dense FFN and routed MoE.

glm5next_mtp_head

Glm5NextMtpHead — GLM-5.3’s MTP block as a DraftProposer.

glm5next_skeleton

GLM-5.3-Flash 45-layer text-model skeleton (Slice 9 – topology, wiring, structural binding). GLM-5.3-Flash 45-layer text-model skeleton — Slice 9.

moe

MoE (Mixture of Experts) FFN component.

mtp_head

MTP (Multi-Token Prediction) head implementing DraftProposer.

mtp_multi

Multi-module MTP proposer (MiniMax M2, DeepSeek-V3 style).

nemotron_mamba2

Nemotron-H Mamba-2 SSM layer implementing TransformerLayer.

nemotron_moe

Nemotron-H standalone MoE FFN layer.

ngram_embed

N-gram embedding id math — LongCat-Flash-Lite / Qwen3.8-Flash-Next family.

ops

Shared kernel dispatch operations.

ple

PLE — hashed n-gram injection into the hyper-connection highway.

qsa

The Qwen3.8-Flash-Next QSA indexer — decode-side sparse-attention selection (#753 phase G).

qwen3_attention

Qwen3 full attention layer.

qwen3_ssm

Qwen3-Next SSM (Gated Delta Net) layer implementing TransformerLayer.

vision_encoder

Qwen3-VL vision encoder: 27-block ViT + DeepStack mergers.

w4a16_gemv_tiers

SSOT for the narrow w4a16_gemv_batch{M} tier family (M = 4..8).

Enums

FfnComponent

FFN component: MoE (expert routing), dense SwiGLU, or None (standalone attention).

Functions

k64_kernel

Resolve the k64 deep-K tile GEMM, preferring the 3-deep weight-pipeline variant. ON by default; ATLAS_NO_K64_PIPELINE3 (presence — =0 is NOT “off”) falls back to the 2-stage parent.

k64_n64_kernel

Resolve the NARROW-N (N_TILE=64) deep-K twin. KernelHandle(0) when the kernel is absent or the presence kill switch ATLAS_NO_K64_N64 is set (=0 is NOT “off”). Callers must store the handle — kernel() is an init-time lookup, not a per-launch one.

k64_n64_wins

Should the narrow-N deep-K twin serve this shape? See K64_N64_MAX_WIDE_CTAS for the derivation and the measured curve.

moe_grouped_decode_decide

Whether the grouped-GEMM MoE decode arm should run for n rows — PURE, so both polarities are testable without touching process env or the OnceLocks below (which latch, and would make the tests order-dependent).

moe_grouped_decode_enabled

Kill switch for the grouped-GEMM MoE decode arm. PRESENCE check per the house convention (ATLAS_NO_MOE_GROUPED_DECODE=0 is NOT off), read once per process — this predicate sits in the decode path, and the env::var it replaces ran on every dispatch for MoE models.

moe_grouped_decode_for

Whether the grouped-GEMM MoE decode arm should run for n rows.

moe_grouped_decode_forced

Force the grouped arm BELOW moe_grouped_decode_min_rows(). Diagnostic only — it exists so the n=5..15 gap can be measured without a rebuild, and it is the same var #415’s measurements used, kept working on purpose. Never a production setting: if forcing wins at a width, move the THRESHOLD.

moe_grouped_decode_min_rows

Optional kernel lookup: KernelHandle(0) instead of an error.

tgemm_kernel

Resolve the N128/M64 tile GEMM, preferring the 3-deep weight-pipeline variant. ON by default; ATLAS_NO_TGEMM_PIPELINE3 (presence — =0 is NOT “off”) falls back to the 2-stage parent. Falls back automatically on any target that does not ship _p3.

try_kernel

w4a16_v2_kernel

Try to load an optional kernel, logging at debug level if it’s not found. Returns KernelHandle(0) (null) on failure — callers must check before use.

w4a16_v3_kernel

Resolve the W4A16 m128 v3 GEMM. Opt-in ONLY, same contract as w4a16_v2_kernel: ATLAS_W4A16_VARIANT=v3 selects it, anything else resolves to a ZERO handle WITHOUT issuing a lookup.