[core] Add natvie stream and stream buff#36511
Merged
Merged
Conversation
Add file utils functions
…across file handling
olpipi
reviewed
Jun 29, 2026
olpipi
reviewed
Jun 29, 2026
olpipi
approved these changes
Jul 1, 2026
…penvinotoolkit#36571) ### Details: - *Create compiler in plugin handle using device properties if the device exists* ### Tickets: - *N/A* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).* --------- Signed-off-by: Bogdan Pereanu <bogdan.pereanu@intel.com>
…cted (openvinotoolkit#36592) Fixes regression detected in openvinotoolkit#36584
…envinotoolkit#36602) ### Details: Reuse `compose_loop_args()` in the AArch64 loop end emitter so pointer increments and finalization offsets arescaled consistently with x64/RISC-V. This removes duplicate byte-offset math while keeping dynamic loop arguments loaded from runtime data. ### Tickets: - N/A ### AI Assistance: - *AI assistance used: yes* - Implementation is agentically assisted and then manually validated
### Details: - Isolate GQA subgraph properly with the new decomposition update ( openvinotoolkit#34980 ) ### Tickets: - EISW-220403 ### AI Assistance: - *AI assistance used: yes* --------- Co-authored-by: Ekaterina Shiryaeva <ekaterina.shiriaeva@intel.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…umentation (openvinotoolkit#36542) ### Details: - *Added documentation with best practices for transformation tests writing* - *Added a corresponding skill to ensure automatic LLM agent routing to the new documentation in transformation tests writing scenario* - *The new skill was applied to a few existing legacy tests files to check how it works. Prompt example: `/ov-transformation-tests refactor @src/common/transformations/tests/common_optimizations/convert_convertlike.cpp tests`. The skill has been confirmed to work on lightweight models as well (e.g. Haiku).* ### Tickets: - *N\A* ### AI Assistance: - *AI assistance used: yes* - *AI was used to test the new skill and for the new documentation adjustments*
### Details: PR openvinotoolkit#35720 fixed a Phi accuracy bug (garbage output) by disabling the onednn `sum` post-op for any fused eltwise add whose addend is an `input_layout` node: ``` && !(dep_node.is_type<input_layout>()) ``` This was over-broad. The onednn `sum` post-op aliases the addend buffer in-place as the node's output (see `get_reused_eltwmem_idx` → `primitive_inst.cpp`), so the primitive overwrites the addend. That is only unsafe when the `input_layout` addend has another live consumer - which is exactly the Phi `input_hidden_states` residual case: the network input feeds both the FullyConnected (reaching `sum` via `is_direct_ancestor`) and downstream blocks (`users > 1`), so the in-place write corrupts the value the residual still needs. When the `input_layout` addend has a single user, there is no surviving reader of the original buffer, so the in-place `sum` is safe. Blanket-disabling it forced these cases onto a slower `binary_add` post-op (extra addend read, lost buffer reuse, possible extra reorder), which regressed per-token decode latency. This caused a 2nd-token (decode) latency regression for distil-large-v2 INT8-CW on GPU (b_fs_yx_fsv16/compressed-FC path). Measured on a Battlemage dGPU (arls): INT8-CW 2nd-token went from ~2.54 ms to ~2.96 ms (+16.6%), with 1st-token unaffected. FP16/INT4 were unaffected because they dispatch through a different FC path. ### Fix: Narrow the guard so sum is only disabled for the unsafe multi-consumer case: ``` && !(dep_node.is_type<input_layout>() && dep_node.get_users().size() > 1) ``` - Multi-user input_layout addend (Phi residual): users > 1 → falls back to binary add → accuracy preserved. - Single-user input_layout addend (distil-large-v2 INT8-CW FC): keeps the in-place sum → regression fixed. The new clause is only reachable in the multi-user case through the existing is_direct_ancestor branch (a single-user dependency already satisfies get_users().size() == 1 on the preceding line), so single-user behavior is unchanged everywhere else. fuse_nodes() appends the fused addend edge to the dependency's user list without dedup, so a residual addend reliably has users > 1 after fusing. ### Results (arls, distil-large-v2, 2nd-token AVG latency) ``` ┌────────────┬─────────────────────────┬───────────────────────┬─────────────────────────────────┐ │ precision │ before openvinotoolkit#35720 (healthy) │ on openvinotoolkit#35720 (regressed) │ with this fix │ ├────────────┼─────────────────────────┼───────────────────────┼─────────────────────────────────┤ │ INT8-CW │ 2.537 ms │ 2.959 ms │ 2.695 ms (back in healthy band) │ ├────────────┼─────────────────────────┼───────────────────────┼─────────────────────────────────┤ │ FP16 │ 3.158 ms │ 3.042 ms │ 3.010 ms (unchanged) │ ├────────────┼─────────────────────────┼───────────────────────┼─────────────────────────────────┤ │ INT4-MIXED │ 2.715 ms │ 2.545 ms │ 2.525 ms (unchanged) │ └────────────┴─────────────────────────┴───────────────────────┴─────────────────────────────────┘ ``` ### Tickets: - CVS-170687
…SequenceArrayLowering (openvinotoolkit#36445) ### Details: - SequenceIfReplacer (common_translators): - Rewrites the "If-branch lazy sequence initialization" idiom (past-KV cache first produced on loop iteration 0 inside a Loop body) into a form the array lowering can process. - SequenceArrayLowering (common_translators): - Lowers loop-carried sequence-of-N patterns (SequenceConstruct / SequenceInsert / SequenceErase / SequenceAt / SequenceLength) into N parallel slot tensors, producing a fully numeric IR with no sequence residuals. Reader gate: true no-op when no SequenceAt/SequenceLength reader exists (prevents SequenceConcatReplacer corruption). Two-phase Loop slot resolver: breaks nested-Loop chicken-and-egg dependencies. - Per-body slot-count reconciliation: pads mismatched slot counts with zero-dummies. - SequenceLength static-N fallback: axis_ambiguous flag correctly defers only sequences with genuinely runtime-varying length; all others emit Constant(slots->size()). ### Tickets: - CVS-187784 ### AI Assistance: - AI assistance used: yes - GitHub Copilot was used to prototype --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
### Details: - Detect if current system supports interoperability. - If check is passed, then create OCL remote context over L0 GPU runtime. - If check fails, then create L0 remote context over L0 GPU runtime. - Remote context and remote tensors query internal GPU plugin objects fur runtime specific handles. - OCL remote context will query for OCL handles. - L0 remote context will query for L0 handles. - Enable L0 internal GPU plugin objects to return converted OCL handles. - Add `resource_owner` template for gpu resource management. - Add `ze_resource` template that will hold Level Zero objects and optionally corresponding OpenCL objects. - Add operators for conversion between Level Zero objects and OpenCL objects. ### Benchmarks: Peak memory usage measured on benchmark_app running resnet-50 on BMG Linux | | VmPeak (kB) | MaxRSS (kB) | | --- | --- | --- | | This PR (runtime=L0, interop=ON) | 3,262,536 | 190,924 | | This PR (runtime=L0, interop=OFF) | 3,263,688 | 193,200 | | Master (runtime=L0) | 3,293,048 | 192,188 | | [REF] Master (runtime=OCL) | 3,281,412 | 193,196 | ### Tickets: - [CVS-185378](https://jira.devtools.intel.com/browse/CVS-185378) ### AI Assistance: - AI assistance used: yes - AI help was used to create tests --------- Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
github-actions
Bot
added
category: C API
github-actions
Bot
removed
category: inference
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
19 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Details:
Tickets: