card: correct the input_scale claim, add the DFlash2 tile/mem findings from #6, note the repo-id processor_config trap from #2, point to the compressed-tensors branch
vLLM's ModelOptNvFp4FusedMoE.create_weights() registers w13_input_scale / w2_input_scale unconditionally (torch.empty, uninitialised) and folds them into the dequant alphas. A weight-only NVFP4 checkpoint that ships no *.input_scale therefore gets garbage (observed 0.0) alphas and every expert is zeroed, producing degenerate output.
Fix: ship a float32 scalar 1.0 input_scale for each of the 37,152 quantized expert tensors, in one extra shard (model-input-scales.safetensors) referenced from model.safetensors.index.json. Weights are unchanged. This mirrors nvidia/GLM-5.2-NVFP4, whose input_scale tensors are all exactly 1.0.
See https://github.com/vllm-project/vllm/issues/54189
Scope reasoning-parser guidance: glm45 is SGLang-only; vLLM needs deepseek_r1
Links the LibertAI Labs recipe that makes vLLM serve this checkpoint on GB10, and documents a fault that affects vLLM users on ANY architecture: vLLM's ModelOptNvFp4FusedMoE reads an activation scale this weight-only checkpoint does not carry, zeroing every expert.
fix: tool-call parser is glm47, not glm (glm fails SILENTLY on GLM-5.x)
