llama.cpp b10067 fixes DeepSeek-V4 quantization crash in llama-quantize
llama.cpp build b10067 patches a bug where the DeepSeek-V4 routing table tensor was being incorrectly quantized, causing a type conversion failure that blocked anyone using llama-quantize on the model.
llama.cpp build b10067, released July 18, fixes one user-facing bug: the ffn_gate_tid2eid tensor in DeepSeek-V4 was being passed through quantization when it should not have been. That tensor is a routing table, an i32 index mapping token IDs to expert IDs, not a weight matrix. Because it was never added to the name-based exclusion list alongside ffn_gate_inp.weight, llama-quantize tried to quantize it, hit the i32-to-float type conversion wall, and failed.
If you tried to quantize a DeepSeek-V4 GGUF with llama-quantize before this build, you would have seen a type conversion error and gotten nowhere. The fix is simple in hindsight: exclude the routing table from quantization the same way the weight gate is excluded. The PR is #25787, authored by Yash Raj Pandey.
This is a small build, one targeted fix. But DeepSeek-V4 is one of the more-used models in the local-LLM crowd right now, and if you were hitting this error and assuming the model file was corrupt, it was not. Update to b10067 and try again.