[FIX_FOR_VLLM_CUSTOM=f976e3b98ba45677a2213673a442c6cbff141e8e] Fix upstream regressions in attention, FP8, offloading and platform

[pawel-olejniczak]

Summary

Fixes five regressions introduced by recent upstream vLLM changes that break HPU unit tests and model execution.

Changes

1. Remove use_output guard from HPU attention patch — attribute removed upstream
2. Remove accept_output_buffer branching from HPU MLA attention — attribute removed upstream; unconditionally use output buffer in opaque path, direct call path manages output internally
3. Update KV offloading connector tests — field renames: block_hashes → keys, block_hashes_to_store → keys_to_store, config access via kv_group_configs[0]
4. Register HPU FP8 block-scaled kernel + add ops test conftest — new _POSSIBLE_FP8_BLOCK_KERNELS dict needs OOT entry; provide VllmConfig stub for ops unit tests
5. Add manual_seed_all to HpuPlatform — new required platform method for RNG seeding

Upstream PRs that introduced these regressions

• [Attention][V0 Deprecation] Deprecate accept output buffer vllm#39125 — removed accept_output_buffer and use_output from attention layer (fixes 1, 2)
• [kv_offload+HMA][5/N]: Track group block hashes and block IDs vllm#37109 — restructured OffloadingConnectorScheduler API (fix 3)
• [W8A8 Block Linear Refactor][2/N] Remove W8A8Fp8BlockLinearOp and adopt Fp8 block linear kernel selections. vllm#33892 — added model_config.dtype access in Fp8LinearMethod.__init__ and _POSSIBLE_FP8_BLOCK_KERNELS (fix 4)
• Add platform manual_seed_all API vllm#38468 — added manual_seed_all as required abstract method on Platform (fix 5)

[Copilot]

Pull request overview

Fixes multiple breakages caused by upstream vLLM API changes, keeping Gaudi/HPU attention, FP8, KV offloading tests, and platform integration compatible with the new interfaces.

Changes:

• Update HPU attention (regular + MLA) to align with upstream removal of use_output / accept_output_buffer.
• Register an HPU FP8 block-scaled kernel stub and add an ops test conftest providing a minimal VllmConfig context.
• Update KV offloading connector unit tests for upstream event/output field renames and config layout changes.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
vllm_gaudi/platform.py	Adds manual_seed_all required by upstream Platform API.
vllm_gaudi/ops/hpu_fp8.py	Registers an OOT entry for _POSSIBLE_FP8_BLOCK_KERNELS via an HPU stub kernel class.
vllm_gaudi/ops/hpu_attention.py	Removes dependency on upstream-removed use_output attribute in attention patching logic.
vllm_gaudi/attention/oot_mla.py	Removes accept_output_buffer branching and standardizes output-buffer usage in the opaque path.
tests/unit_tests/ops/conftest.py	Introduces a fixture that sets a minimal current VllmConfig for ops unit tests.
tests/unit_tests/kv_offload/offloading_connector/utils.py	Updates scheduler config assertions and adapts mocked PrepareStoreOutput to renamed fields.
tests/unit_tests/kv_offload/offloading_connector/test_scheduler.py	Updates offloading event tests to use OffloadKey/make_offload_key and new event fields.

[Copilot]

generate_store_output and MockLoadStoreSpec still use block_hashes naming, but the returned PrepareStoreOutput now uses the renamed keys_to_store / evicted_keys fields. Consider renaming the function parameter/local variables (and related mock spec fields, if appropriate) to keys to match the updated API and avoid confusion when reading or extending these tests.

[Copilot]

forward_impl still accepts output / output_scale / output_block_scale, but the implementation ignores these parameters and will overwrite output locally. Since the earlier explicit NotImplementedError guard was removed, callers that pass an output buffer could now get silently incorrect behavior. Consider either (a) restoring an explicit error when output is provided, or (b) implementing true output-buffer support (writing into the provided tensor) and documenting the contract.

Suggested change

	) -> torch.Tensor:
	) -> torch.Tensor:
	if (output is not None or output_scale is not None
	or output_block_scale is not None):
	raise NotImplementedError(
	"HPUMLAAttention.forward_impl does not support caller-"
	"provided output, output_scale, or output_block_scale.")

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
f976e3b98ba45677a2213673a442c6cbff141e8e