Add the option to use Boolean attention mask

[yangulei]

No description provided.

[Copilot]

Pull request overview

Adds a new runtime feature flag to optionally represent attention masks as boolean tensors (instead of additive 0/-inf bias tensors), and updates the HPU attention code paths to correctly apply boolean masks during attention score computation.

Changes:

• Add use_boolean_mask feature flag (env var VLLM_USE_BOOLEAN_MASK) to runtime config.
• Update attention-bias/mask construction in the HPU model runner to emit either boolean masks or additive -inf biases depending on the flag.
• Update HPU attention ops to apply boolean masks via masked_fill instead of adding an additive bias tensor.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
vllm_gaudi/v1/worker/hpu_model_runner.py	Adds use_boolean_mask wiring and conditionally builds boolean vs additive attention masks/biases across multiple mask construction helpers.
vllm_gaudi/extension/ops.py	Applies boolean attention masks via masked_fill in prompt attention and paged-attention softmax paths.
vllm_gaudi/extension/features.py	Introduces use_boolean_mask config value driven by VLLM_USE_BOOLEAN_MASK.

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
f296a1966dca96cd69e5c1fa1264edbf611a1bd6

[adobrzyn]

@yangulei please include this new flag in documentation

[afierka-intel]

@copilot open a new pull request to apply changes based on the comments in this thread

[yangulei]

A short description for the VLLM_USE_BOOLEAN_MASK was pushed.
More detailed documentation should be pushed in the PR for slicing of FusedSDPA (#1149 or #1155).

[yangulei]

Close as no obvious perf gain and have accuracy issue when using with INC fp8.