| --- |
| license: mit |
| pretty_name: DeepSeek-V4-Flash-0731 quantization measurements |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - evaluation |
| - perplexity |
| - kl-divergence |
| - gguf |
| - quantization |
| - llama.cpp |
| - deepseek-v4 |
| - reproducibility |
| size_categories: |
| - n<1K |
| --- |
| |
| # DeepSeek-V4-Flash-0731 — quantization measurements |
|
|
| Everything needed to reproduce, audit or extend the numbers published in |
| [AtomicChat/DeepSeek-V4-Flash-0731-GGUF](https://huggingface.co/AtomicChat/DeepSeek-V4-Flash-0731-GGUF): |
| the reference logits, the evaluation corpus, the raw tool output for every quant we |
| measured, and the parsed results. |
|
|
| Every GGUF of this model that we could find on the Hub was measured here — ours, |
| unsloth's, bartowski's, ggml-org's, antirez's and others — on one machine, against one |
| reference, with one command. Publishers normally report numbers from their own harness, |
| which makes cross-vendor comparison meaningless. These files exist so that anyone can |
| check ours instead of trusting them. |
|
|
| > [!IMPORTANT] |
| > Measured using `8x5090` |
|
|
| ## Files |
|
|
| | File | Size | What it is | |
| |---|---:|---| |
| | `wiki-alt.txt` | 1.29 MB | Evaluation corpus: wikitext-2 test split from `Salesforce/wikitext`, parquet rows concatenated | |
| | `ref5632.kld` | 37.1 GB | Reference logits from the lossless `AD-BF16` quant over that corpus at ctx 5632 | |
| | `RESULTS-0731.jsonl` | small | Parsed results, AtomicChat and unsloth ladders | |
| | `RIVALS-B1.jsonl`, `RIVALS-B2.jsonl`, `RIVALS-B3.jsonl` | small | Parsed results, other publishers, split by the machine that produced them | |
| | `logs/*.log` | few MB | Full unedited `llama-perplexity` output for every quant, nothing filtered | |
|
|
| The `.kld` file stores the reference model's full probability distribution at every scored |
| token position — roughly 258 KB per token at this vocabulary size. It is what makes the |
| KL-divergence numbers comparable: every quant is compared against these exact logits. |
|
|
| ## Result schema |
|
|
| ```json |
| {"repo": "bartowski/DeepSeek-V4-Flash-0731-GGUF", |
| "name": "MXFP4", |
| "bytes": 145678901234, |
| "ppl": "4.5446", |
| "kld": "0.156403", |
| "rms": "12.686", |
| "top1": "87.369"} |
| ``` |
|
|
| `ppl` is `Mean PPL(Q)` from the KL-divergence block, `kld` is `Mean KLD`, `rms` is |
| `RMS Δp`, `top1` is `Same top p` — the share of positions where the quant picks the same |
| next token as the reference. Note that `Mean PPL(Q)` and the standalone `Final estimate: |
| PPL` printed by the same tool are different aggregations and do not match; the logs |
| contain both. |
|
|
| ## Measurement setup |
|
|
| | | | |
| |---|---| |
| | Reference | `AtomicChat/DeepSeek-V4-Flash-0731-GGUF` → `AD-BF16` (bit-exact with the official weights) | |
| | Corpus | `Salesforce/wikitext`, `wikitext-2-raw-v1`, test split, rows concatenated | |
| | Context | 5632, batch 5632, 51 chunks | |
| | llama.cpp | PR [#24162](https://github.com/ggml-org/llama.cpp/pull/24162), commit `f180ae8b2`, built with `-DCMAKE_CUDA_ARCHITECTURES=120` | |
| | GPU | 8× RTX 5090 | |
|
|
| ## Hardware matters here, and it is not optional |
|
|
| The routed experts of this model are 96% of its weights and they are stored in MXFP4. |
| llama.cpp has two paths for that format — unpack to BF16 and use a normal tensor-core |
| matmul, or feed the packed 4-bit data into block-scaled instructions. The second is gated |
| on compute capability ≥ 12.0, which covers consumer Blackwell only. H100 and H200 are 9.0, |
| B200 is 10.0, B300 is 10.3; all take the first path despite having FP4 hardware. |
|
|
| Same file, same corpus, same commit, reference model: |
|
|
| | GPU | ctx 512 | ctx 5632 | |
| |---|---:|---:| |
| | RTX 5090 | 5.4312 | 4.5381 | |
| | H100 | 5.1554 | 4.3406 | |
|
|
| A 4–5% difference from the GPU alone. Reproducing these numbers requires consumer |
| Blackwell **and** a build that targets it — compiling for `sm_90` on a 5090 gives the |
| H100 numbers, because the native kernel never lands in the binary. |
|
|
| ## Reproducing |
|
|
| ```bash |
| git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp |
| git fetch origin pull/24162/head:dsv4 && git checkout dsv4 |
| cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=120 |
| cmake --build build -j --target llama-perplexity |
| ``` |
|
|
| ```bash |
| hf download AtomicChat/dsv4-eval-artifacts --repo-type dataset --local-dir . |
| |
| ./build/bin/llama-perplexity \ |
| -m <any-quant>-00001-of-*.gguf \ |
| -f wiki-alt.txt --kl-divergence-base ref5632.kld --kl-divergence \ |
| -ngl 99 -c 5632 -b 5632 |
| ``` |
|
|
| To rebuild the reference from scratch instead of downloading it, run the same command |
| against `AD-BF16` with only `--kl-divergence-base` and no `--kl-divergence`. Takes about |
| ten minutes and should print `Final estimate: PPL = 4.5381`. |
|
|
| ## Caveats |
|
|
| - Absolute values are not comparable to numbers published elsewhere. Other publishers use |
| different corpora, context lengths and hardware. Compare within one table. |
| - 51 chunks at ctx 5632 gives roughly ±0.003 on mean KLD. Differences smaller than that |
| are noise. |
| - Quants of derived models — expert-pruned, abliterated, distilled — are deliberately |
| excluded. KL-divergence against this reference would measure the difference between |
| models, not the cost of quantization. |
|
|
| ## License |
|
|
| MIT. Derived from `deepseek-ai/DeepSeek-V4-Flash-0731`. Produced by |
| [Atomic Chat](https://huggingface.co/AtomicChat). |
