Lucebox/DeepSeek-V4.1-Flash-ROCMFP23-GGUF

DeepSeek V4.1 Flash for the Lucebox engine (ROCMFP block formats)

Lucebox builds of DeepSeek V4.1 Flash for the Lucebox DS4.1 engine (AMD R9700 + Strix Halo). Routed experts in the engine's ROCMFP block formats; attention, shared experts and head in Q8_0; both Engram tables embedded. Calibrated with the published 262,144-token imatrix (smalinin/DeepSeek-V4.1-Flash-GGUF, converted in imatrix/).

file	routed experts	expert bytes	KL top-256 vs teacher	top-1	ppl
teacher (fp4 checkpoint)		269 GiB	0	100%	8.45
antirez Q2 (IQ2_XXS / Q2_K)		142 GiB	0.260	78.3%	9.71
DeepSeek-V4.1-Flash-ROCMFP23.gguf	fp2 gate/up, fp3 down	179 GiB	0.272	77.9%	9.80
DeepSeek-V4.1-Flash-ROCMFP2S.gguf	fp2s everywhere	158 GiB	0.266	77.7%	9.59
DeepSeek-V4.1-Flash-ROCMFP2S-MIX10.gguf	fp2s, fp3 down on layers 0-9	164 GiB	0.246	78.3%	9.45

Formats: Q2_0_ROCMFP2 (ggml type 107, 10 bytes per 32 weights, 2.5 bits, codebook {-1, 0, 1, 2} x a ue4m3 scale per 16 weights), Q3_0_ROCMFPX (type 104, 14 bytes per 32, 3.5 bits). fp2s files use type 107 with bit 7 of a half-block's scale byte meaning "mirror the codebook"; a reader that ignores that bit produces garbage, so use an engine build that honors it. KL is measured on 8,184 held-out fineweb-edu tokens against the unquantized model through the same reference forward pass (eval/ holds the teacher log-probs and the tokens). Recipe, converter patches and harness: Lucebox repo, lucebox_training/quantization.