# Reproduce the DeepSeek evaluation

Build the runtime and scorer, then prepare a workspace for the tokenizer checks, MMLU-Pro evaluation, PPL and chat diagnostics.

## Python environment

Use Python 3.12 and the dependencies pinned in `requirements-minimal.txt`. The oracle uses CPU PyTorch 2.14.0+cpu. The C++ runtime requires a CUDA toolkit and a compatible NVIDIA driver. The `native` build commands below require CMake 3.24 or newer and CUDA 11.6 or newer.

From the model repository root:

```bash
cd runtime
python3.12 -m venv .venv
.venv/bin/python -m pip install --index-url https://download.pytorch.org/whl/cpu 'torch==2.14.0+cpu'
.venv/bin/python -m pip install -r requirements-minimal.txt
export NLTK_DATA="$PWD/nltk_data"
```

The included English Punkt/Punkt-tab resources are in `nltk_data`; their hashes and sources are in `nltk-resources.json`.

`official-source/manifest.json` pins the tokenizer, encoder, reference model/Engram equations and supporting files to `deepseek-ai/DeepSeek-V4.1-Flash` revision `dba1be0a40aa45a94ad051997016db3960a90277`, with sizes, SHA256 hashes and source URLs. Original model shards are not included.

## Build from the retained archive

From `runtime/`:

```bash
mkdir llama-deepseek41
tar -xzf deepseek-runtime-base-source.tar.gz -C llama-deepseek41
git -C llama-deepseek41 apply ../deepseek-runtime.patch
git -C llama-deepseek41 apply ../deepseek-perplexity-sidecar.patch
cmake -S llama-deepseek41 -B llama-deepseek41/build -G Ninja \
  -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native
cmake --build llama-deepseek41/build --target llama-server llama-perplexity -j 24
bash build_helpers.sh ./llama-deepseek41 ./bin
```

`native` targets the CUDA devices visible during the build. Replace it with the target CUDA architecture when building for another machine. `provenance.json` records the source archive and patch hashes; `helper-build-validation.json` records the helper source and binary hashes.

## Fresh diagnostic workspace

The helper creates a new directory, links the inputs and binaries, and copies the diagnostic driver with its `ROOT` path set to the workspace. `--check-request-plan` checks all 24 encoded requests against the supplied request plan without loading the model.

Still from `runtime/`:

```bash
.venv/bin/python prepare_replay_workspace.py \
  --workspace "$PWD/replay" \
  --runtime-checkout "$PWD/llama-deepseek41" \
  --helper-dir "$PWD/bin" \
  --model "$PWD/../DeepSeek-V4.1-Flash-GSQ-RCO-3.0bit.gguf" \
  --check-request-plan
```

Before inference, download the model at repository revision `100f4d51b43805557945f94b73a7f9b667948a7b` and verify SHA256 `11f46543370256ef616b6e458b6950e148625b5f8b7545173d124445d3393c53`. Allow CPU memory for the approximately 209 GB Engram mapping and sufficient GPU memory for the backbone.

Set `DEEPSEEK_CUDA_DEVICES` to the comma-separated device IDs to use for scoring and capture. In `replay/deepseek-run-diagnostics.relocated.py`, set both `CUDA_VISIBLE_DEVICES` in the `ENV` assignment and the device string in the first `replicas` entry to that same list. These assignments override the shell's device selection. Keep the default single replica and the F32 GEMM/cache settings with flash attention disabled. The driver uses loopback port 8093.

Run the prerequisite checks:

1. Run `bin/dump-deepseek-vocab MODEL NATIVE_DUMP` and `.venv/bin/python verify_vocab.py official-source/tokenizer.json NATIVE_DUMP replay/results/deepseek-vocab-parity.json`, replacing `MODEL` and `NATIVE_DUMP` with the model and output paths.
2. Run `.venv/bin/python test_reference_components.py replay` to check tokenizer mapping, Engram/history constants and candidate-mask behavior.
3. Run `bin/capture-deepseek-serial MODEL replay/results/deepseek-capture-few-ref-f32kv 'The capital of France is' 128 128` with `CUDA_VISIBLE_DEVICES="$DEEPSEEK_CUDA_DEVICES"`, `NVIDIA_TF32_OVERRIDE=0`, `GGML_CUDA_MMF_F32_DISABLE=1`, `GGML_CUDA_REFERENCE_F32=1`, `DEEPSEEK_CAPTURE_NO_FLASH=1`, and `DEEPSEEK_CAPTURE_F32_KV=1`. Unset `DEEPSEEK_CAPTURE_SERIAL` for full prefill. Run the [independent oracle comparison](oracle/README.md) on that capture.
4. Run the [raw MMLU command](README.md#evaluation-and-diagnostics) with `-o replay/results/deepseek3.0-f32.tsv`, first adding `--preflight` to check the tokenizer and context bounds. Then run `.venv/bin/python ../eval/2026-09-13/analyze_mmlu.py --config replay/mmlu-analysis-config.json --output-prefix replay/results/mmlu-analysis` to validate the 2,000 rows.

Then run `.venv/bin/python replay/deepseek-run-diagnostics.relocated.py` without additional flags. It checks serial decoding against the prefill, evaluates eight PPL contexts, saves the native cache and lossless F32 sidecar, and runs the 24 chat items. Results and raw responses are written to `replay/results/diagnostics/`; the metadata records the driver hash and numerical settings.