Key dimensions
hidden
H = 7168 · HC width
4×7168 = 28672 (n=4 streams)
layers
×61 · vocab
129280
attn:
128 heads × 512 (fc_qb 1536×65536) · MQA w/ sink
o-proj: reshape
[S,16,4096] → bmm_oa → fc_ob 16384×7168
KV: SWA window
128 + compressed pages (
128 tokens/page)
MoE:
384 experts, router [S,384],
top-6 (ArgTop6/tid2eid)
expert FFN
7168→3072→7168, MXFP4×MXFP8 · shared expert FP8
FC / GEMM (FP8 · BF16 · FP32 · MXFP4×MXFP8)
elementwise / norm / reshape
attention core (MQA)
KV cache / persistent state
hyper-connection (mHC) ops
embeddings / logits / block refs
─ solid: dataflow · ┄ dashed: control/decode-step · ┄ orange: cache update