llama.cpp src/llama-quant.cpp (llama_ftype to tensor-type mapping and per-tensor promotion rules)

Fields

id

sources/qzgw-llama-quant-cpp

URL

https://raw.githubusercontent.com/ggml-org/llama.cpp/master/src/llama-quant.cpp

Publisher

ggml-org

Kind

repo

Accessed

2026-10-04

notes

Authoritative source for WHICH llama_ftype name maps to WHICH GGML_TYPE, i.e. the mixture/promotion rules. Establishes that IQ2_M is not a block type (maps to GGML_TYPE_IQ2_S), IQ3_M and IQ3_XS both map to GGML_TYPE_IQ3_S, and that token-embedding / output tensors get promoted (Q5_K or Q6_K or Q2_K depending on ftype). Also confirms MXFP4_MOE is a real, current ftype not covered by any existing record. Accessed 2026-10-04.

Notes

Authoritative source for WHICH llama_ftype name maps to WHICH GGML_TYPE, i.e. the mixture/promotion rules. Establishes that IQ2_M is not a block type (maps to GGML_TYPE_IQ2_S), IQ3_M and IQ3_XS both map to GGML_TYPE_IQ3_S, and that token-embedding / output tensors get promoted (Q5_K or Q6_K or Q2_K depending on ftype). Also confirms MXFP4_MOE is a real, current ftype not covered by any existing record. Accessed 2026-10-04.

Cross-references

This record references no other records.