Skip to content

feat: add conv1d_fn op for Qwen3.5 linear attention on NPU.#1295

Merged
yingxudeng merged 1 commit intojd-opensource:mainfrom
maojunx99:conv1d_fn
Apr 19, 2026
Merged

feat: add conv1d_fn op for Qwen3.5 linear attention on NPU.#1295
yingxudeng merged 1 commit intojd-opensource:mainfrom
maojunx99:conv1d_fn

Conversation

@maojunx99
Copy link
Copy Markdown
Contributor

@maojunx99 maojunx99 commented Apr 16, 2026

add new conv1d_fn op interface

Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request implements the causal_conv1d kernel for NPU and integrates it into the Qwen3GatedDeltaNetBase layer, alongside adding a comprehensive NPU helper utility. The review feedback highlights several style guide violations regarding member variable naming, type aliases, casting, and assertion macros. Critical issues were identified concerning an uninitialized variable in the layer implementation and the presence of debugging logs in production code. Furthermore, performance improvements are suggested to address synchronous device-to-host transfers and inefficient data handling in the worker implementation.

Comment thread xllm/core/framework/model/model_input_params.h Outdated
Comment thread xllm/core/kernels/npu/pytorch_npu_helper.hpp Outdated
Comment thread xllm/core/kernels/npu/pytorch_npu_helper.hpp Outdated
Comment thread xllm/core/kernels/npu/pytorch_npu_helper.hpp Outdated
Comment thread xllm/core/kernels/npu/pytorch_npu_helper.hpp Outdated
Comment thread xllm/core/layers/npu_torch/qwen3_gated_delta_net_base.cpp Outdated
Comment thread xllm/core/runtime/worker_impl.cpp Outdated
Comment thread xllm/core/runtime/worker_impl.cpp Outdated
Copy link
Copy Markdown
Collaborator

@yingxudeng yingxudeng left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

为什么加这个文件

@yingxudeng
Copy link
Copy Markdown
Collaborator

image 格式检查麻烦改下

Comment thread xllm/core/layers/npu_torch/qwen3_gated_delta_net_base.cpp Outdated
Comment thread xllm/core/kernels/npu/causal_conv1d.cpp Outdated
@maojunx99 maojunx99 force-pushed the conv1d_fn branch 2 times, most recently from 56e4ad5 to 7ce6e47 Compare April 17, 2026 03:59
Comment thread xllm/core/layers/npu_torch/CMakeLists.txt Outdated
@yingxudeng
Copy link
Copy Markdown
Collaborator

xllm_ops 仓库的 commit 貌似需要更新下

@yingxudeng yingxudeng changed the title update conv1d_fn op feat: add conv1d_fn op. Apr 17, 2026
Comment thread xllm/core/kernels/npu/xllm_ops/causal_conv1d.cpp Outdated
@yingxudeng yingxudeng changed the title feat: add conv1d_fn op. feat: add conv1d_fn op for Qwen3.5 linear attention on NPU. Apr 17, 2026
@yingxudeng yingxudeng merged commit 183d0eb into jd-opensource:main Apr 19, 2026
24 of 55 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants