File size: 2,029 Bytes
517ef62 cae409b 517ef62 cae409b 517ef62 2bc89ac 517ef62 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | # Minimal inference
A readable reference implementation rather than a production serving engine. The
model code covers the vision encoder and aligner, sliding-window plus compressed
sparse attention with its two-level indexer, engram n-gram lookups, MoE,
Hyper-Connections, and the DSpark forward path. Generation itself is plain
autoregressive sampling.
## Install
```bash
python -m pip install -r requirements.txt
```
## Convert Hugging Face weights
The runtime uses one converted checkpoint file per tensor-parallel rank. From
this directory:
```bash
export HF_CKPT_PATH=/path/to/DeepSeek-V4.1-Flash-HF
export SAVE_PATH=/path/to/DeepSeek-V4.1-Flash-TP8
export MP=8
python convert.py \
--hf-ckpt-path "${HF_CKPT_PATH}" \
--save-path "${SAVE_PATH}" \
--model-parallel "${MP}" \
--expert-dtype fp4 \
--tokenizer-path "${HF_CKPT_PATH}"
```
Expert counts are inferred from the weight names, so they do not need to be
passed. `--tokenizer-path` points at whichever directory holds `tokenizer.json`
and `tokenizer_config.json`; they are copied into the converted checkpoint.
## Run the equivalent TXT and JSON examples
```bash
export CKPT_PATH=/path/to/DeepSeek-V4.1-Flash-TP8
export MP=8
INPUT_FILE=examples/example.txt ./run.sh
INPUT_FILE=examples/example_harmony.json ./run.sh
```
The two files express the same interleaved two-image prompt, so they produce
identical encoded prompts and input token IDs.
For interactive chat:
```bash
torchrun --nproc-per-node "${MP}" generate.py \
--ckpt-path "${CKPT_PATH}" \
--config config.json \
--interactive \
--temperature 0.6
```
For multi-node execution, pass the usual `torchrun --nnodes`, `--node-rank`,
`--master-addr`, and `--master-port` arguments before `generate.py`.
## Self-test
`model.py` builds a small model from the `ModelArgs` defaults and runs a prefill
plus 22 decode steps, exercising the real dense-fp8 / MoE-fp4 kernels. Weights
are uninitialized, so it checks shapes and kernel plumbing, not numerics:
```bash
python model.py
```
|