[V1] Defragmentation support

[madamczyk-intel]

extension PR: HabanaAI/vllm-hpu-extension#275

[madamczyk-intel]

/run-gaudi-tests

[madamczyk-intel]

/run-gaudi-tests

[madamczyk-intel]

/run-gaudi-tests

[Copilot]

Pull Request Overview

This PR adds support for memory defragmentation in the HPU model runner, introduces step-based profiling in the HPU worker, and ensures runtime configuration is finalized.

• Invoke and import finalize_config to apply updated runtime settings.
• Add setup_step_profiler and debug logging to track per-step profiling in HPUWorker.
• Integrate OnlineDefragmenter into HPUModelRunner to resolve, track, and defragment block IDs.
• Update HPU extension dependency to the defragmentation-enabled branch.

Reviewed Changes

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

File	Description
vllm/worker/hpu_model_runner.py	Import and call finalize_config after setting VLLM config.
vllm/v1/worker/hpu_worker.py	Add setup_step_profiler, per-step profiling, and debug logging.
vllm/v1/worker/hpu_model_runner.py	Initialize and use OnlineDefragmenter throughout model runner.
requirements/hpu.txt	Point vllm-hpu-extension to dev/madamczyk/v1_defrag branch.

Comments suppressed due to low confidence (5)

vllm/v1/worker/hpu_worker.py:36

• [nitpick] The helper setup_step_profiler lacks a docstring; add a brief description of its purpose and the meaning of its parameters.

def setup_step_profiler(steps):

vllm/v1/worker/hpu_worker.py:104

• [nitpick] The attribute name step_debug may not clearly convey its purpose; consider renaming it to step_logger or similar for clarity.

        self.step_debug = init_debug_logger('steps')

vllm/v1/worker/hpu_worker.py:101

• [nitpick] The variable step is quite generic; renaming to current_step or step_counter could improve readability.

        self.step = 0

vllm/v1/worker/hpu_model_runner.py:619

• The integration of OnlineDefragmenter is significant but currently untested; consider adding unit tests for its resolve, update_state, and defragment flows.

        self.defragmenter = OnlineDefragmenter()

vllm/worker/hpu_model_runner.py:990

• [nitpick] The indentation of finalize_config() is inconsistent with the surrounding block; align it with the environment.set_vllm_config call for clarity.

        finalize_config()

[madamczyk-intel]

/run-gaudi-tests

[madamczyk-intel]

/run-gaudi-tests

[madamczyk-intel]

/run-gaudi-tests

[madamczyk-intel]

/skip-gaudi-tests

[madamczyk-intel]

CI tests already passed before sha update: ef7cbbc