Skip to content

[Feature]: Support DeepSeek-V3 with PyTorch-based graph construction on Ascend NPU #2407

Open

Description

@XuZhang99

🚀 The motivation and feature

Implement DeepSeek-V3 using xLLM's PyTorch model execution path on Ascend NPU, reusing the existing runtime and graph infrastructure to simplify model development and optimization.

Expected support:

  • Register deepseek_v3 for --model_impl=python, adapting the existing DeepSeek-V3.2 configuration and W8A8 checkpoint loader.
  • Use dense MLA with index_topk=0, reusing YaRN RoPE and MLA preprocessing without requiring indexer weights.
  • Reuse dense FFN, routed/shared experts, and TP/DP/EP communication; validate supported parallel configurations.
  • Integrate paged KV cache, chunked prefill, and prefix-cache reuse with a compatible NPU cache layout.
  • Reuse eager/ACLGraph runners and extend prepared-metadata/task-pipeline support to DeepSeek-V3, including sequence-length updates during replay.

Additional context

Provide a minimal deployment example and validate accuracy and eager-versus-graph performance on real Ascend NPUs.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions