Downloads · 30 days
823
22% of all-time downloads
RemySkye/rwkv7-g1h-7.2b-i1-GGUF
rwkv7-g1h-7.2b-i1-GGUF is a text generation model from RemySkye. Use it when you need the model to write or continue text. It is set up for llama.cpp. The card lists the license as apache-2.0.
Model-specific imatrix quants generated from rwkv7-g1h-7.2b-20260710-ctx10240-BF16.gguf. The BF16 master remains in the linked static-GGUF repository and is not duplicated here.
Downloads · 30 days
823
22% of all-time downloads
All-time downloads
3.7K
Public
Repo size
120 GB
Likes
0
Public
Click a slice to open those files.
.gguf93 GB · 100%
From the Hugging Face model README
Model-specific imatrix quants generated from rwkv7-g1h-7.2b-20260710-ctx10240-BF16.gguf. The BF16 master remains in the linked static-GGUF repository and is not duplicated here.
lemon07r/pile-calibration-v5ea863bb930b9959dd78095c165aa1376d14e698bc92e806d1c81091c9035edce99c35374da1b465e1024 tokens51252428854d1f2bd6a80cc75a72e7f8a62e23438e23f06287aab3855c5a7826127fae7be8c7de8adad55f0be5e00f394b7c7efca7c6d13fdd4a2f476bc08c0491fe98027The source dataset is diverse and duplicate-free, but contains a few book-length outliers. Before calibration, corrupt/severely repetitive records are removed, long records are capped using four separated excerpts, records are deterministically shuffled, and rare scripts are lightly interleaved into the early calibration window.
rwkv7-g1h-7.2b-20260710-ctx10240.i1-imatrix.gguf — the model-specific GGUF importance matrix, uploaded firstrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ1_M.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ1_S.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ2_M.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ2_S.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ2_XS.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ2_XXS.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ3_S.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ3_XXS.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ4_NL.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-IQ4_XS.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q2_K.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q3_K_S.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q4_0.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q4_1.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q4_K_S.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q5_0.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q5_1.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q5_K_S.ggufrwkv7-g1h-7.2b-20260710-ctx10240.i1-Q6_K.gguf| Quant | PPL | BF16 retained |
|---|---|---|
i1-Q6_K | 6.0916 | 99.71% |
i1-Q5_1 | 6.1122 | 99.37% |
i1-Q5_K_M | 6.1081 | 99.44% |
i1-Q5_K_S | 6.1115 | 99.38% |
i1-Q5_0 | 6.1424 | 98.88% |
i1-Q4_1 | 6.2244 | 97.58% |
i1-Q4_K_M | 6.1812 | 98.26% |
i1-Q4_K_S | 6.2299 | 97.49% |
i1-Q4_0 | 7.5525 | 80.42% |
i1-IQ4_NL | 6.3564 | 95.55% |
i1-IQ4_XS | 6.3423 | 95.77% |
i1-Q3_K_L | 6.3309 | 95.94% |
i1-Q3_K_M | 6.3737 | 95.29% |
i1-IQ3_S | 6.7405 | 90.11% |
i1-Q3_K_S | 10.686 | 56.84% |
i1-Q2_K | 26.7059 | 22.74% |
i1-IQ3_XXS | 7.0303 | 86.39% |
i1-IQ2_M | 8.0809 | 75.16% |
i1-IQ2_S | 9.2989 | 65.32% |
i1-IQ2_XS | 78.4712 | 7.74% |
i1-IQ2_XXS | 77.9658 | 7.79% |
i1-IQ1_M | 240.6085 | 2.52% |
i1-IQ1_S | 485.3095 | 1.25% |
The Q3_K_M, Q3_K_L, Q4_K_M, and Q5_K_M files use custom RWKV-aware recipes with explicit tensor assignments. Higher precision is used for the token embeddings and selected value, time-mix output, and channel-mix tensors where it is expected to preserve the most quality.
Earlier automated files with these names were removed after verification showed that llama.cpp's generic mixed recipes did not recognize RWKV's time_mix_* and channel_mix_* tensor roles. Because of that, the automated M and L variants had collapsed to the same effective layouts as the retained S variants.
These replacement files have genuinely different tensor layouts, providing additional size and quality choices between the existing S variants and the larger quantizations.
The files in this i1 repository use the same RWKV-aware tensor layouts together with this model's existing importance matrix.