[ascend]Optimize moe

[yao-fengchen]

No description provided.

[Copilot]

Pull Request Overview

This PR updates the CI configuration to use a new branch for LMDEPLOY, aligning with the "[ascend]Optimize moe" effort.

• Updated LMDEPLOY_COMMIT_OR_BRANCH from 'main' to 'optimize_moe' in the GitHub workflow.

[Copilot]

Pull Request Overview

This PR introduces optimizations for the moe module by updating function signatures and configurations. Key changes include:

• Adding new parameters (head_size and head_size_v) to multiple attention-related functions.
• Removing the slicing of past key values in the DeepseekV2Attention_forward function.
• Updating the CI workflow to reference the "optimize_moe" branch and a new Git repository URL.

Reviewed Changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated no comments.

File	Description
dlinfer/ops/llm.py	Updated function signatures and documentation for attention operations.
dlinfer/framework/lmdeploy_ext/dynamo/graph_mode_patch.py	Modified parameter usage in the DeepseekV2Attention_forward function.
.github/workflows/main.yml	Adjusted CI environment variables and updated the Git repository URL.

Comments suppressed due to low confidence (3)

dlinfer/ops/llm.py:132

• [nitpick] Consider rephrasing the docstring to 'head_size (int): The size of each query head' for clarity.

head_size (int): The number of query head size.

dlinfer/framework/lmdeploy_ext/dynamo/graph_mode_patch.py:89

• Removing the slice '[:nope_size]' might change the intended behavior. Please verify that passing the full tensor is correct.

past_key_value[0][..., :nope_size]

.github/workflows/main.yml:77

• Ensure that the repository URL update from 'InternLM' to 'DeepLink-org' aligns with all relevant CI/CD configurations.

git clone https://github.com/DeepLink-org/lmdeploy.git ${{ env.LMDEPLOY_PATH }}