DecoderGgufMemSegConverter: unsupported quant type Q4_1 (silent pass-through → forward pass fails at matmul) #654

Summary

When loading a GGUF model that contains Q4_1-quantized tensors, DecoderGgufMemSegConverter does not handle the Q4_1 quant type. It logs a warning and passes the tensor through unchanged instead of dequantizing/converting it, so the forward pass later fails at matmul (dtype/layout mismatch).

Warning (verbatim)

DecoderGgufMemSegConverter: unsupported quant type Q4_1 for 'blk.0.ffn_down.weight'; passing through unchanged. Forward pass may fail at matmul.

Version / revision

Observed on SKaiNET @ 80500204a1120c9d91f45535da51aa9eccf8d656.

How to reproduce

Load a GGUF model whose tensors use Q4_1 quantization (e.g. *.ffn_down.weight) and run a forward pass. Encountered via the SKaiNET-examples KllamaDemo (Qwen3 playground) loading a Q4_1 GGUF.

Expected

• DecoderGgufMemSegConverter should support Q4_1 dequantization (as it does for Q4_0 and other supported types), or
• fail fast with a clear error at load time instead of silently passing the tensor through and crashing later at matmul.

Impact

Q4_1-quantized models cannot run, and the failure surfaces deep in the forward pass at matmul, far from the root cause (the unsupported quant type at load/convert time).