[kahlun]

What does this PR do?

Extends PlatformBase (verl/plugin/platform/platform_base.py) with two optional hook groups so a hardware backend that can't reuse CUDA's exact code paths plugs in cleanly, instead of the rest of verl growing more if device == "..." branches. These were surfaced while bringing up an Intel XPU backend (companion PR: verl-project#7371) and all default to today's behavior, so CUDA/ROCm/NPU are unaffected. The XPU plugin (verl-hardware-plugin) depends on this PR for the new hooks below.

Companion PRs in verl-hardware-plugin

1. Implementation of these hooks for XPU (PlatformXPU.attention_utils_module() / profiler_markers() / dist_profiler_cls()), plus docs — verl-hardware-plugin#22
2. Docker image for Intel XPU + verl — verl-hardware-plugin#23 (after this PR merges, [fsdp, vllm] fix: drop redundant tied lm_head.weight from bucketed we… #23 will be updated to clone verl from the official repo instead of this fork)
3. AVG-reduce + attention support, without any core hook — verl-hardware-plugin#32 (draft). This PR originally also proposed an is_reduce_avg_supported() hook (xccl doesn't implement ReduceOp.AVG), but per the maintainer's ruling on verl-hardware-plugin#26, new hardware backends should default to plugin-side monkeypatching and only ask verl-core for a hook when something genuinely can't be done from the plugin.

---

1. Attention fallback — attention_utils_module(). Lets a platform point verl/utils/attention_utils.py at its own flash-attn-equivalent module (a dotted path exposing index_first_axis/pad_input/rearrange/unpad_input), checked after the NPU path and before falling back to the flash_attn pip package or the pure-PyTorch implementation. Also moved torch_functional.py's two remaining hardcoded flash_attn / get_device_name() == "cuda"/"npu" branches onto this same dispatcher.
2. Profiler — profiler_markers() / dist_profiler_cls(). profiler_markers() lets a platform supply its own (mark_start_range, mark_end_range, mark_annotate, marked_timer) tuple, picked up by profiler/__init__.py when neither the nvtx package nor NPU's built-in markers apply. dist_profiler_cls(tool) lets a platform register a DistProfiler subclass under its own profiler.tool name; wired into DistProfiler.__init__ (profile.py) and into the three rollout HTTP servers (vllm_async_server.py, sglang's async_sglang_server.py, trtllm_async_server.py) so a plugin-supplied tool name isn't rejected there either.

No public API changes — every new PlatformBase method is an optional override with a safe default (-> None).

Test

pre-commit run ruff ruff-format --files verl/plugin/platform/platform_base.py \
  verl/utils/attention_utils.py verl/utils/profiler/__init__.py \
  verl/utils/profiler/performance.py verl/utils/profiler/profile.py \
  verl/utils/torch_functional.py verl/workers/engine_workers.py \
  verl/workers/rollout/sglang_rollout/async_sglang_server.py \
  verl/workers/rollout/trtllm_rollout/trtllm_async_server.py \
  verl/workers/rollout/vllm_rollout/vllm_async_server.py

PYTHONPATH=. pytest tests/plugin/test_platform_abstraction.py \
  tests/utils/test_attention_utils.py tests/utils/test_profiler_plugin_hooks.py -q

New unit tests, which inject a fake platform so no accelerator is needed:

• tests/utils/test_attention_utils.py: checks that attention_utils_module() sends all four helpers (index_first_axis/pad_input/rearrange/unpad_input) through the plugin module, that NPU still takes precedence, and that verl falls back when the hook returns None.
• tests/utils/test_profiler_plugin_hooks.py: checks that profiler_markers() is resolved lazily on first use and the returned markers are forwarded, that dist_profiler_cls() selects the plugin tool and passes it rank/config/tool_config, and that an unrecognized tool falls back to the no-op profiler.

Updated existing tests:

• tests/models/test_transformer.py, tests/models/test_transformers_ulysses.py: import the padding helpers from verl.utils.attention_utils instead of branching on device to pick flash_attn.
• tests/plugin/test_platform_abstraction.py: accept intel (registered by the XPU plugin) as an auto-detected platform name.

The XPU attention/profiler paths were also validated end-to-end on real Intel Arc B-series hardware as part of verl-project#7371's GRPO/PPO/SFT runs.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved profiler lifecycle issues and missing focused tests for the new plugin dispatch paths.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 3

Open (3)

• Add regression test for platform plugin module dispatch · New
• Test lazy profiler marker resolution and forwarding · New
• Test custom dist profiler plugin selection and arguments · New

What changed in this PR

Adds optional platform hooks for hardware-specific attention utilities and profiling, enabling Intel XPU integration without device-specific branches.

Changes:

• Adds attention and profiler hooks to PlatformBase.
• Routes attention and profiler dispatch through platform plugins.
• Updates engine, rollout servers, and model tests for plugin support.

File	Review summary
verl/​workers/​rollout/​vllm_rollout/​vllm_async_server.py	Plugin profiler lifecycle is not invoked; custom tools may be inert.
verl/​workers/​rollout/​trtllm_rollout/​trtllm_async_server.py	Plugin profiler lifecycle is not invoked; custom tools may be inert.
verl/​workers/​rollout/​sglang_rollout/​async_sglang_server.py	Plugin profiler lifecycle is not invoked; custom tools may be inert.
verl/​workers/​engine_workers.py	Accepts plugin profiler tools.
verl/​utils/​torch_functional.py	Uses unified attention dispatch.
verl/​utils/​profiler/​profile.py	Supports plugin profiler classes; focused dispatch tests are needed.
verl/​utils/​profiler/​__init__.py	Adds lazy plugin marker resolution; focused tests are needed.
verl/​utils/​attention_utils.py	Dispatches attention helpers through plugins; regression tests are needed.
verl/​plugin/​platform/​platform_base.py	Defines new hooks; plugin-path coverage is needed.
tests/​plugin/​test_platform_abstraction.py	Updates platform abstraction expectations.
tests/​models/​test_transformers_ulysses.py	Updates tests to use unified attention utilities.
tests/​models/​test_transformer.py	Updates tests to use unified attention utilities; PR description should reflect modified tests.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.