kingjones777/Xing4.0-29B-A4B-BF16-GGUF

Xing4.0-29B-A4B — BF16 GGUF

Full-precision BF16 GGUF conversion of `XingChen-AGI/Xing4.0-29B-A4B`, published as the quantization source for the ROCmFP4 tiers. This is the file every quantized tier in `kingjones777/Xing4.0-29B-A4B-ROCmFP4-GGUF` was produced from, and the file to re-quantize from if you want a different tier.

⛔ You need a fork that knows the xing4 architecture

Xing4.0-29B-A4B is not supported by upstream llama.cpp. The architecture is DeepSeek2-style MLA plus hyper-connections (hc), with a DeepSeek-style MTP layer 40 that carries no hc tensors. Loading this file with a stock build fails — it is not a normal GGUF that any llama.cpp can read.

Two correctness details that were wrong in the first working port and are worth knowing if you port it yourself, because both produced plausible but wrong output rather than an error:

• —Sinkhorn epsilon belongs in the denominator, not added to the value. An eps of 1e-6 added to entries that are legitimately around 1e-7 acts as a 5–17× floor on them.
• —`hc_expand` contracts `comb^T`, not comb.

With both fixed, perplexity went 80.05 → 8.365 against an HF reference of 8.425 (32 chunks × 512, -ub 64 -ngl 99 -fit off). Norm-matching was useless as a validator here — the transpose fix looked like a 6× regression while the epsilon bug masked it. Gate an hc port on perplexity, not on tensor norms.

Files

Reproduction

# convert with a fork that registers arch xing4 (conversion/xing4.py)
python convert_hf_to_gguf.py XingChen-AGI/Xing4.0-29B-A4B --outtype bf16 \
  --outfile Xing4.0-29B-A4B-BF16.gguf

# then quantize, e.g. the flagship tier
llama-quantize --output-tensor-type q6_K Xing4.0-29B-A4B-BF16.gguf OUT Q4_0_ROCMFP4_STRIX_LEAN 8

License and attribution

Apache-2.0, inherited from `XingChen-AGI/Xing4.0-29B-A4B`. Converted and measured on AMD Strix Halo (gfx1151, ROCm 7.2.4).