加载中…
加载中…
Clover quantization by IsValorum
[!IMPORTANT] Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.
If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.
Importance matrix for Diffuspace/Clover-35B-A3B-Preview at revision 4b0402b65e523d0183a14c431955438714cb16ff.
[!NOTE] What the prebuilt runtimes cover. Linux x86_64: any CPU, and NVIDIA CUDA 12 for sm_80/86/89/90. Windows x86_64, in a single universal package: NVIDIA CUDA 12.8 from sm_50 through sm_120, AMD and Intel GPUs through Vulkan 1.4.363, and any CPU with per-CPU backends. Not covered: macOS, ARM64, ROCm, and Blackwell on Linux; those need a build from the same branch and commit.
| Parameter | Value |
|---|---|
| Target | 1,250,000 tokens |
| Context windows | 611 x 2,048 = 1,251,328 positions |
| Mix | 35% code / 25% reasoning / 25% text / 15% tools |
| Languages | 11 natural languages, 10 programming languages |
| Source | Q8_0 single-file GGUF, SHA256 3d2ded3e5604c0e05c2956f95bd864a3bb6635c1442f23318484877e22cd52ed (declared deviation from the canonical BF16 source) |
| Backend | llama.cpp PR #31306, public CUDA build llama-clover-4838ffb-linux-x86_64-nvidia-cuda12.tar.gz |
| Acceleration | Flash Attention on (-fa on), -ngl 99, -t 12, --device CUDA0, batch/ubatch 2048, mmap on |
| Flags | --output-format gguf --process-output --no-ppl --no-escape -c 2048 -b 2048 -ub 2048 --chunks 611 -fa on --parse-special |
| Parameter | Value |
|---|---|
| Target | 1,250,000 tokens |
| Context | 2,048 tokens |
| Chunks | 611 |
| Planned positions | 1,251,328 (611 x 2,048) |
| Available pool | 1,261,939 |
| Processed positions | 1,251,328 (611 x 2,048) |
| Trailing unprocessed pool | 10,611 |
| BOS processed | 0 |
| Special tokens | parsed |
| Selector version | balanced-1.25M-35-25-25-15-no-hi-v4-unique |
| Source pool version | balanced-4M-35-25-25-15-v2 |
| Component | Target | Measured | Tokens | 2,048-token chunks |
|---|---|---|---|---|
| Code | 35.00% | 35.1614% | 439,985 | 214.84 |
| Reasoning | 25.00% | 24.8265% | 310,661 | 151.69 |
| Text | 25.00% | 24.9505% | 312,213 | 152.45 |
| Tools | 15.00% | 15.0615% | 188,469 | 92.03 |
| Total | 100.00% | 100.0000% | 1,251,328 | 611.00 |
Target and measured are percentages of the 1,251,328 processed positions. Weight is the selector weight; the code target is 35% split by that weight over 30.
| Language | Weight | Target | Measured | Tokens | Documents |
|---|---|---|---|---|---|
| Python | 8 | 9.3333% | 9.3310% | 116,762 | 45 |
| JavaScript | 5 | 5.8333% | 5.8449% | 73,139 | 28 |
| TypeScript | 3 | 3.5000% | 3.5108% | 43,931 | 26 |
| Java | 3 | 3.5000% | 3.5231% | 44,086 | 33 |
| C++ | 3 | 3.5000% | 3.5273% | 44,138 | 20 |
| C | 2 | 2.3333% | 2.3413% | 29,297 | 9 |
| Go | 2 | 2.3333% | 2.3485% | 29,387 | 17 |
| Rust | 2 | 2.3333% | 2.3525% | 29,437 | 13 |
| SQL | 1 | 1.1667% | 1.1963% | 14,970 | 11 |
| Shell | 1 | 1.1667% | 1.1858% | 14,838 | 9 |
| Code total | 30 | 35.0000% | 35.1614% | 439,985 | 211 |
Target and measured are percentages of the 1,251,328 processed positions. Weight is the selector weight; the 65% natural target (25 reasoning + 25 text + 15 tools) is split by that weight over 67.
| Language | Weight | Reasoning | Text | Tools | Total | Target | Measured |
|---|---|---|---|---|---|---|---|
| English | 14 | 64,842 | 64,641 | 39,328 | 168,811 | 13.5821% | 13.4905% |
| Spanish | 10 | 46,401 | 46,893 | 27,895 | 121,189 | 9.7015% | 9.6848% |
| French | 8 | 37,090 | 37,176 | 22,695 | 96,961 | 7.7612% | 7.7486% |
| Chinese | 10 | 46,310 | 46,244 | 27,830 | 120,384 | 9.7015% | 9.6205% |
| German | 5 | 23,195 | 23,426 | 14,131 | 60,752 | 4.8507% | 4.8550% |
| Portuguese | 5 | 23,207 | 23,397 | 13,904 | 60,508 | 4.8507% | 4.8355% |
| Japanese | 4 | 18,514 | 18,941 | 11,517 | 48,972 | 3.8806% | 3.9136% |
| Russian | 4 | 18,537 | 18,772 | 11,258 | 48,567 | 3.8806% | 3.8812% |
| Arabic | 3 | 13,902 | 13,838 | 8,332 | 36,072 | 2.9104% | 2.8827% |
| Korean | 2 | 9,303 | 9,614 | 5,730 | 24,647 | 1.9403% | 1.9697% |
| Italian | 2 | 9,360 | 9,271 | 5,849 | 24,480 | 1.9403% | 1.9563% |
| Natural total | 67 | 310,661 | 312,213 | 188,469 | 811,343 | 65.0000% | 64.8386% |
| Check | Value |
|---|---|
| Repeated corpus | True (max 8 uses per content; exact repeats in calibration-report.json) |
| Native quota adaptations | [] |
| HF/GGUF tokenizer full-id parity | False (Clover remaps token 151669 to <|mask|> and removes merges) |
| Native text parity | True |
| Special tokens parsed | True |
| Bundle SHA256 | 159f63e5d069a4541abd9c8a7b3d48ff13ecb171084f044313a608af2a557876 |
| Selected documents SHA256 | 8043b62d78e0dfc4151550e1000c36ac32bce003f35667cd62ba18a6c80895da |
| Calibration report SHA256 | 7b4ec74a2e58037c5c20941f2daf2d7a92f88e5dcb78fd1a4a8e3ae4fd3f0240 |
The exact per-group counts, documents and pool sizes are in calibration-report.json. The 1,251,328 processed positions are the full 611 x 2,048 windows; the 10,611-token tail is selected but outside the processed prefix.
clover_moe, Qwen3.5-MoE 35B-A3B backbone, 40 layers, 256 routed experts, top-8.ssm_bd_logit_scale, ssm_bd_history.weight, ssm_bd_history.bias and ssm_conv1d.weight_right are not exercised; they are excluded and must remain F32.VALIDATED_AGAINST_AGENT_CONTRACT; all required entries present: True. No PPL or benchmark claims.| File | Description |
|---|---|
Clover-VAL-1.25M-I-Matrix.gguf |
Matrix, 365.91 MiB, 383689280 bytes, SHA256 05d9669137d8ca843f0a7d3a8f00975e84d48644832d2734c564a1778b2f2d5d |
expert-coverage.tsv |
Observed rows per routed-expert tensor |
generation-command.json |
Exact command and plan |
generation.log |
llama-imatrix log |
calibration-report.json |
Corpus selection and quotas |
contract.json / inventory.json |
Agent-reviewed tensor contract and inventory |