Match the mapping to the layer.
Weights stay. Inputs flow.
WO-S · 16 weights/PE · inputs ↓ columns · outputs accumulate
A · Token parallel
4 weight replicas · 32 channels
B · Channel parallel
2 weight replicas · 64 channels
First tile · setup and pipeline gaps omitted
Runtime PE reconfiguration
Full viewLayer details
Why A or B?
| Mapping | Tile M×K×N | 16 columns | Predicted cycles |
|---|
Buffer layouts
16 banks · 16 scalars/cell · selected PE vector outlined
Address details
bank = L mod 16; scalar row = 16⌊L/16⌋ + lane. W/I vectors span K; O vectors span N. StaB/StrB/OB depths: 128/64/64 scalar rows/bank.
Commands & sources
Recorded workload: logits_to_keep=1. Each operation loads new weights. Residuals, normalization, row selection, sampling, embedding and repacking are graph context, excluded from GEMM costs. No automatic buffer handoff is assumed.
A/B match the bounded mapper's winners. Search excludes graph-wide layout-conversion costs; global optimality is not claimed.
serialized_minisa_v2 predicts loads, preload, streaming, gaps/drain, dispatch and stores. These are predictions, not new RTL measurements. Numerical samples are synthetic; the graph is symbolic.
Mapping data · Generator · Qwen3 config · Qwen3 implementation