DeepSeek V4 Conversion

The repository-local scripts/convert_deepseek_v4_to_w8a8.py utility converts the DeepSeek V4 Flash source checkpoint variant validated by this repository to the W8A8 layout expected by PyPTO Serving.

The conversion can run on CPU and does not require torch_npu. The source and output directories must be different, and the host must have enough free disk space for both copies.

Prepare Dependencies

python -m pip install --upgrade huggingface_hub safetensors
python -c "import torch, safetensors; print(torch.__version__)"

Use the PyTorch build already validated for the active Ascend environment.

Download or Locate the Source Checkpoint

hf download deepseek-ai/DeepSeek-V4-Flash \
  --local-dir /path/to/DeepSeek-V4-Flash

If an official mirror is already available locally, use that snapshot directory as --input-dir.

Arguments

Argument	Required	Description
--input-dir PATH	yes	DeepSeek V4 Flash source checkpoint directory.
--output-dir PATH	yes	Output directory for the converted PyPTO W8A8 checkpoint.
--resume	no	Validate and skip completed output shards after an interrupted conversion.
--dry-run	no	Validate the source and print the conversion plan without writing files.

Dry Run

python scripts/convert_deepseek_v4_to_w8a8.py \
  --input-dir /path/to/DeepSeek-V4-Flash \
  --output-dir /path/to/dsv4-flash-w8a8 \
  --dry-run

Convert

python scripts/convert_deepseek_v4_to_w8a8.py \
  --input-dir /path/to/DeepSeek-V4-Flash \
  --output-dir /path/to/dsv4-flash-w8a8

If the process is interrupted, rerun with --resume:

python scripts/convert_deepseek_v4_to_w8a8.py \
  --input-dir /path/to/DeepSeek-V4-Flash \
  --output-dir /path/to/dsv4-flash-w8a8 \
  --resume

A successful run prints Conversion complete and leaves a converted config.json, model.safetensors.index.json, safetensors shards, and a .pypto-w8a8-conversion.json marker in the output directory.