GLM-5.2-NVFP4 — network + SGLang serving dataflow
Built from
nvidia/GLM-5.2-NVFP4 config +
checkpoint weight map, transformers
glm_moe_dsa, sglang
glm4_moe/
dsa_indexer/
nextn,
arXiv:2602.15763 (Gemini extraction, verified), and the serve command:
TP4 · modelopt_fp4 · EAGLE 5-1-6 · chunked prefill 8192.
Blue = GEMM, teal = norm/elementwise, red = attention, orange = caches, purple = routing/spec ops. Press ? for keys.
GLM-5.2 + serve command
hidden
6144 ·
78 layers (3 dense) +
MTP L78 · vocab
154880 · ctx
1M
MLA:
64 heads · q_lora 2048 · kv_lora 512 · nope 192 + rope 64 · v 256
indexer:
32×128 top-2048 ·
full @ 0,1,2,6,10,…,74 — shared layers reuse last full
MoE: 256 experts top-8 (sigmoid + bias) · shared ×1 · inter 2048 (dense 12288)
NVFP4:
routed experts only, W4A4 g16 · attn/shared/embed/head BF16 · KV
FP8
serve:
TP4 · EAGLE steps
5 topk
1 draft
6 · chunked prefill
8192 · mem 0.85 · :30000
FC / GEMM (precision as labeled)
elementwise / norm / reshape
attention / score core
KV cache / persistent state
routing / gating / speculative ops
embeddings / logits / block refs
─ solid: dataflow · ┄ dashed: control/index · ┄ orange: cache update