[feature]: add spargeattn search in infer stage

[zhiwei-dong]

This commit add sparge attn search support.

[Copilot]

Pull Request Overview

This PR introduces support for sparse attention search in the infer stage via new configuration options and corresponding code updates.

• Added a new shell script to run the sparge tuning job.
• Updated transformer weight loading logic to support sparse tuning.
• Enhanced the infer stage to extract and save sparse attention parameters.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
scripts/run_wan_t2v_sparge_tune.sh	Added a new bash script to run inference with sparge tuning.
lightx2v/models/networks/wan/weights/transformer_weights.py	Introduced a new flag (sparge_tune) and modified weight-loading logic based on its value.
lightx2v/models/networks/wan/model.py	Implemented a final infer stage that extracts and saves sparse attention parameters when tuning is enabled.
configs/wan_t2v_sparge_tune.json	Provided configuration options for enabling sparge and sparse tuning.

Comments suppressed due to low confidence (1)

lightx2v/models/networks/wan/model.py:196

• Ensure that 'SparseAttentionMeansim' is properly imported in this module, as it is used for type checking but not visible in the diff.

if isinstance(v, SparseAttentionMeansim):

[Copilot]

Pull Request Overview

This PR adds support for sparge attention search during inference, including the ability to tune sparse attention weights. Key changes include:

• A new shell script (scripts/run_wan_t2v_sparge_tune.sh) for running inference with the new feature.
• Updates in transformer_weights.py and model.py to load and optionally tune sparge attention weights.
• A new configuration file (configs/wan_t2v_sparge_tune.json) to enable and control these features.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
scripts/run_wan_t2v_sparge_tune.sh	New script to set up the environment and run inference.
lightx2v/models/networks/wan/weights/transformer_weights.py	Added handling for sparse tuning mode and environment variable setup.
lightx2v/models/networks/wan/model.py	Integrated saving of tuned sparse attention parameters during inference.
configs/wan_t2v_sparge_tune.json	New configuration to enable sparge attention search with tuning.

Comments suppressed due to low confidence (2)

lightx2v/models/networks/wan/weights/transformer_weights.py:30

• [nitpick] Consider aligning the naming for the tuning flag (e.g., using 'sparge_tune') to more clearly relate to the 'sparge' flag, as the current mix of 'sparge' and 'sparse_tune' may be confusing.

self.sparge_tune = config.get("sparse_tune", False)

lightx2v/models/networks/wan/model.py:196

• The class 'SparseAttentionMeansim' is referenced but not imported; please ensure that an appropriate import statement is added if it is defined in another module.

if isinstance(v, SparseAttentionMeansim):

[Copilot]

Pull Request Overview

This PR introduces support for sparge attention search in the inference stage of the model.

• Added a new shell script for running the sparge attention tune.
• Updated transformer weights and model inference to support a tuning mode for sparsified attention.
• Added a new JSON configuration file with relevant parameters.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
scripts/run_wan_t2v_sparge_tune.sh	New shell script to run inference with sparge attention tuning, managing environment variables and input paths.
lightx2v/models/networks/wan/weights/transformer_weights.py	Extended to include a 'sparse_tune' flag to determine weight loading behavior for sparge attention tuning.
lightx2v/models/networks/wan/model.py	Enhanced the infer stage to save weights when tuning mode is enabled using the 'sparse_tune' flag from the configuration.
configs/wan_t2v_sparge_tune.json	Added configuration entries for enabling sparge attention support and tuning, including checkpoint specification.

Comments suppressed due to low confidence (2)

lightx2v/models/networks/wan/weights/transformer_weights.py:30

• There is slight inconsistency in naming between 'sparge' and 'sparse_tune'; consider aligning the naming convention across the codebase for clarity.

self.sparge_tune = config.get("sparse_tune", False)

lightx2v/models/networks/wan/model.py:193

• The use of 'sparse_tune' here together with 'sparge' in other parts may cause confusion; consider unifying the naming to improve code clarity.

self.sparge_tune = self.config.get("sparse_tune", False)