Batched k×k conv (any stride-1 extent) → inference BN → relu.

Batched stride-2 conv, SYMMETRIC padding (the stem and the fused stage) → inference BN → relu.

Batched depthwise → inference BN, no activation (timm's dw_start, the pre-DW).

Batched depthwise → inference BN → relu (the stride-1 post-DW).

Batched stride-2 depthwise, symmetric → inference BN → relu (the strided post-DW).

A depthwise slot at inference: nothing at k = 0, a Mnv4DWEval otherwise — DWSlot's eval twin.

The slot's parameters at any k: the stored ones at k > 0, a zero placeholder at k = 0 that the absent slot never reads.

The pre-DW slot at inference: identity at k = 0, depthwise → BN otherwise.

The stride-1 post-DW slot at inference: identity at k = 0, depthwise → BN → relu otherwise.

A stride-1 UIB body at inference, read off its row: pre-DW? → expand-BN-relu → post-DW? → project-BN, at side h. The skip is added by the caller.

A stride-2 UIB block at inference: the pre-DW slot and the expand at the input side 2h, the post-DW carrying the stride to h (timm's dw_mid), the project at h. No skip.

The fused stage at inference: stride-2 k×k conv-BN-relu, then the 1×1 project-BN.

The head at inference, timm's order: 1×1 conv-BN-relu at h, GAP, conv_head 1×1 conv-BN-relu on the pooled [N, mid, 1, 1], dense — with mnv4Head's two relabellings.

The inference MobileNetV4-Conv-M at final feature side f, N·(3·32f·32f) → N·nCls: stem, fused stage, the 21 table rows (the eighteen stride-1 ones under residual), head. Every BN reads frozen statistics, so the whole net is per-example: each stage is batchMap N of a per-example map or a pointwise map.

One inference BN node, named as mnv4Bn names it at .eval: γ, β by their parameter names, μ, σ² as %{site}mu / %{site}var.

The pre-DW slot's graph: no tokens at k = 0, depthwise → BN otherwise — uibFwdSkipB's and uibFwdStridedB's if preDWk > 0.

The stride-1 post-DW slot's graph: no tokens at k = 0, depthwise → BN → relu otherwise.

A stride-1 UIB body's inference graph at side h, every name read off the row.

A stride-2 UIB block's inference graph: pre-DW slot and expand at 2h, the strided post-DW (.depthwiseStrided, symmetric) to h, project at h.

Stem inference graph: 3×3/s2 symmetric conv → BN (stn) → relu. Generic in the widths, for the reason mnv4StemGraphB records.

Fused-stage inference graph: 3×3/s2 conv → BN (f0cn) → relu → 1×1 project → BN (f0pn).

Head inference graph, mnv4HeadGraphB's tokens with the two BNs at frozen statistics (h1n, hn).

The MobileNetV4-Conv-M inference forward graph at final feature side f, at mnv4FwdChainB's .eval tokens and names: mnv4_fwd_eval.mlir and mnv4in_fwd_eval.mlir at f = 7, mnv4in_fwd_eval_s256.mlir at f = 8.