festr2/GLM-5.2-BF16-AMDMXFP4experts

GLM-5.2 BF16 dense + AMD MXFP4 experts

Hybrid GLM-5.2 checkpoint spliced from two sources (pure byte copies, no requantization):

• Dense (attention, shared experts, dense MLPs, embeddings, MTP dense): BF16, bit-identical to lukealonso/GLM-5.2-NVFP4 (= original zai-org/GLM-5.2 BF16 passthrough)
• Routed experts (layers 3–77): MXFP4 (e2m1 + ue8m0 per-32 scales), bit-identical to amd/GLM-5.2-MXFP4 (AMD Quark calibrated)
• MTP layer 78 experts: MXFP4, RTN-requantized from the BF16 MTP experts (AMD checkpoint does not ship layer 78); model-mtp.safetensors

quantization_config = {"quant_method": "mxfp4"} — loads in vLLM without patches: dense goes through the unquantized BF16 path, routed experts through the MXFP4 MoE path (B12X backend on SM120).

Serving (vLLM)

vllm serve festr2/GLM-5.2-BF16-AMDMXFP4experts \
  --tensor-parallel-size 8 \
  --kv-cache-dtype fp8 \
  --trust-remote-code

MoE activation modes: default w4a16; on the B12X stack B12X_MOE_FORCE_A8=1 selects w4a8_mx (valid for MXFP4 experts; do not use A8 force on NVFP4-expert checkpoints).