Downloads · 30 days
222
14% of all-time downloads
robbintt/qwen3.8-27b-ud-iq2xxs-q8head
qwen3.8-27b-ud-iq2xxs-q8head is a machine learning model from robbintt. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Custom GGUF for Qwen3.8-27B built by grafting the Q80 output/lmhead from the Q80-tier Unsloth Dynamic quant onto the IQ2XXS (Unsloth Dynamic) body.
Downloads · 30 days
222
14% of all-time downloads
All-time downloads
1.6K
Public
Repo size
16.1 GB
Likes
2
Public
Click a slice to open those files.
.gguf8.1 GB · 100%
From the Hugging Face model README
Custom GGUF for Qwen3.8-27B built by grafting the Q8_0 output/lm_head from the Q8_0-tier Unsloth Dynamic quant onto the IQ2_XXS (Unsloth Dynamic) body.
unsloth/Qwen3.8-27B-GGUF -> Qwen3.8-27B-UD-IQ2_XXS.ggufoutput.weight): Q8_0 (1,288 MiB), up from the base Q3_K (521 MiB)output_norm: F32 (unchanged)Why: the lm_head is only ~5-9% of the file but is the single highest-leverage tensor for output quality. Holding it at Q8_0 on a 2-bit body yields better next-token logits for a small size cost.
How: header-offset GGUF surgery — output.weight (tensor #1, offset 0) replaced with a Q8_0 blob; all 850 following tensor offsets shifted +767 MiB. Body bytes are byte-identical to the base.
GGUF alignment note (the bug that mattered): GGUF tensor data does not begin at header_end — it begins at the 32-byte-aligned boundary align_up(header_end, 32) (llama.cpp reads it at GGUF_PAD(header_end, 32)). The Q8_0 head was sourced from Q6_K[align_up(Q6_header_end,32)] and the body from base[align_up(base_header_end,32) + old_head], with 3 bytes of alignment padding written after the header in the output.
Validated: loads clean in llama.cpp and produces coherent output:
output.weight row-0 matches the official BF16 lm_head (corr ~1.0, max abs err ~3.6e-4).Earlier revisions of this repo (before the alignment fix) produced garbage ("/////////") because the head was copied from the un-aligned
header_end. The current file is the corrected build.