GLM-4.5-Air-REAP-82B-A12B-NVFP4

GLM-4.5-Air-REAP-82B-A12B — NVFP4 (W4A16)

NVFP4 weight-only (W4A16) quantization of Cerebras' GLM-4.5-Air-REAP-82B-A12B — itself a REAP (router-weighted expert-activation pruning) prune of GLM-4.5-Air from 106B→82B total (12B active, 96 experts). This checkpoint quantizes the weights to NVFP4 (4-bit float E2M1, block-16 FP8-E4M3 micro-scale + FP32 per-tensor global scale, ≈4.5 bits/value) while keeping activations in BF16, so it loads and serves on a single 128 GB DGX Spark / GB10. It was produced from the BF16 source (Cerebras recommends BF16 for low-bit quant, avoiding stacked FP8→FP4 rounding) with a custom shard-by-shard streaming quantizer that reuses compressed-tensors' own pack/scale primitives, emitting the standard nvfp4-pack-quantized format — so it's a drop-in for vLLM with no custom code. On-disk size is 51 GB (down from 159 GB BF16); served runtime footprint ≈89 GB. Benchmarked on vLLM at prefill 163 / decode 158 tok/s (ShareGPT, concurrency 32, 64K context). The fused parallel projections (q/k/v→qkv, gate/up→gate_up, expert w1/w3) share one weight_global_scale per group — required for correct output under vLLM's fused MoE/attention layers.

Serving (vLLM)

vllm serve gauravmm/GLM-4.5-Air-REAP-82B-A12B-NVFP4 \
  --quantization compressed-tensors \
  --moe-backend marlin \
  --trust-remote-code \
  --max-model-len 65536 \
  --kv-cache-dtype fp8

--moe-backend marlin is mandatory on this NVFP4 W4A16 checkpoint — the flashinfer_cutlass MoE backend rejects W4A16 NVFP4 and crashes engine-core init.

Provenance

GLM-4.5-Air (zai-org, MIT) → REAP prune (Cerebras) → NVFP4 W4A16 (this repo). License inherited from upstream: MIT.