Enum FfnComponent 

pub enum FfnComponent {
    Moe(MoeLayer),
    Dense(DenseFfnLayer),
    None,
}

2026-09-25: A layer’s FFN: MoE, dense, or none.

Variants

Moe(MoeLayer)

Dense(DenseFfnLayer)

None

2026-09-25: No FFN: forward returns its input and the other passes launch nothing.

Implementations

impl FfnComponent

pub fn nvfp4_grouped_ok(&self, m: usize, ctx: &ForwardContext<'_>) -> bool

2026-09-27: Whether this is a MoE whose grouped NVFP4 decode serves m rows (MoeLayer::nvfp4_grouped_decode_ok).

pub fn forward_nvfp4_grouped( &self, input: DevicePtr, m: usize, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

2026-09-27: The grouped NVFP4 MoE decode over [m, H] rows into moe_output(). Errors for dense and none; callers gate on Self::nvfp4_grouped_ok first.

impl FfnComponent

pub fn is_none(&self) -> bool

pub fn is_dense(&self) -> bool

pub fn moe_grouped_decode_ok(&self) -> bool

2026-09-25: Whether this is a MoE whose MoeLayer::grouped_decode_ok holds, so multi-sequence decode may route it through forward_prefill (qwen3_attention/trait_impl/multi_seq/ffn.rs). False for dense and none.

pub fn fp32_routing_active(&self, levers: &ModelLevers) -> bool

2026-09-25: MoeLayer::fp32_routing_active for a MoE; false otherwise.

pub fn forward( &self, input: DevicePtr, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<DevicePtr>

pub fn forward_k2( &self, input: DevicePtr, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

pub fn forward_k3( &self, input: DevicePtr, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

pub fn dense_fixed_ok(&self, m: usize, ctx: &ForwardContext<'_>) -> bool

2026-09-30: Whether this is a dense FFN that runs m rows under a fixed --activation-quantization (DenseFfnLayer::fixed_ok). False for MoE and none.

pub fn forward_dense_fixed( &self, input: DevicePtr, m: usize, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

2026-09-30: DenseFfnLayer::forward_fixed over m rows into moe_output; call only when dense_fixed_ok(m) holds.

pub fn can_forward_km(&self, m: u32) -> bool

2026-09-25: Whether this is a dense FFN whose DenseFfnLayer::can_forward_km(m) holds: a batched-GEMV tier serves m rows and NVFP4 or FP8 weights are loaded. False for MoE and none. Multi-sequence attention decode (qwen3_attention/trait_impl/multi_seq/ffn.rs) asks before computing the pre-FFN norm.

pub fn narrow_rows(&self) -> u32

2026-09-28: Row edge of a dense FFN’s narrow decode arms (DenseFfnLayer::narrow_rows: its W4A4 edge under the weight-quantization tier, capped at 32, else the W4A16 edge). MoE and none answer the W4A16 edge; can_forward_km is false for them anyway.

pub fn try_forward_km( &self, input: DevicePtr, m: u32, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<bool>

2026-09-25: DenseFfnLayer::forward_km over m rows. Returns Ok(false) without launching when can_forward_km(m) is false.

pub fn forward_prefill( &self, input: DevicePtr, num_tokens: usize, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

pub fn forward_batched( &self, input: DevicePtr, num_tokens: usize, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

pub fn forward_token_major_decode( &self, input: DevicePtr, num_tokens: usize, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

pub fn forward_atomic_c4_decode( &self, input: DevicePtr, num_tokens: usize, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

pub fn fp8_grouped_decode_ok(&self, m: usize, ctx: &ForwardContext<'_>) -> bool

2026-09-25: Whether the target model’s cross-row grouped FP8 MoE decode serves m rows: ModelLevers::moe_fp8_grouped_decode_target (METRALE_FP8_MOE_GROUPED_DECODE) is on and this is a MoE whose MoeLayer::fp8_grouped_decode_ok(m) holds. The MTP drafter calls MoeLayer::fp8_grouped_decode_ok directly, without the lever.

pub fn fp8_grouped_routing_ok( &self, m: usize, routing: GroupedRouting, ctx: &ForwardContext<'_>, ) -> bool

2026-09-26: Whether this is a MoE whose grouped FP8 decode serves m rows with routing (MoeLayer::fp8_grouped_routing_ok). Unlike Self::fp8_grouped_decode_ok it does not read the moe_fp8_grouped_decode_target lever: the exact routings reproduce the path they replace byte for byte, so they need no opt-in.

pub fn forward_fp8_grouped_decode_routed( &self, input: DevicePtr, m: usize, routing: GroupedRouting, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

2026-09-26: Self::forward_fp8_grouped_decode with routing. Errors for dense and none; callers gate on fp8_grouped_routing_ok first.

pub fn forward_fp8_grouped_decode( &self, input: DevicePtr, m: usize, ctx: &ForwardContext<'_>, stream: u64, ) -> Result<()>

2026-09-25: Cross-row grouped FP8 MoE decode over [m, H] rows into moe_output(). Errors for dense and none; callers gate on fp8_grouped_decode_ok first.