Support for FLUX.2-klein-4B#36483
Closed
mpaulitsch wants to merge 33 commits into
Closed
Conversation
…inotoolkit#35844) ### Details: - Pick openvinotoolkit#35817 to the `releases/2026/2` ### Tickets: - EISW-215960 ### AI Assistance: - *AI assistance used: no*
Nightly documentation currently has built errors and need to resolve in master and 2026.2 branches. Updating formatting in document that is causing the error.
…nvinotoolkit#35925) ### Details: Duplicate openvinotoolkit#35875 ### Tickets: - *[EISW-216237](https://jira.devtools.intel.com/browse/EISW-216237)* ### 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: intelgaoxiong <xiong.gao@intel.com>
…oolkit#35950) ### Details: - Pick openvinotoolkit#35856 ### Tickets: - EISW-176128 - EISW-214507 ### AI Assistance: - AI assistance used: no, picked it myself
openvinotoolkit#35951) ### Details: - Picks openvinotoolkit#35910 ### Tickets: - EISW-216649 ### AI Assistance: - *AI assistance used: no* - Picked it myself
…n (GatedDeltaNet) models (openvinotoolkit#35965) Backport of (openvinotoolkit#35961) to releases/2026/2 --------- Signed-off-by: Andrew Park <andrew.park@intel.com>
…of warning for older drivers (openvinotoolkit#35977) ### Details: - *Forbids compilation for older drivers if user sets encryption callback e.g. `ov::cache_encryption_callbacks(ov::EncryptionCallbacks{ov::codec_xor, nullptr})`* - Applies to all ze graph extensions lower than version `1.17` ### Tickets: - *C148679* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
### Details
- Fixes FP16 overflow in `mvn_gpu_b_fs_yx_fsv16` by using F32 for
`MEAN_TYPE` and variance-related intermediate buffers.
- The previous implementation used the activation type for mean/variance
accumulation. For FP16 inputs, large spatial reductions can overflow
during `sum((x - mean)^2)` accumulation before variance normalization.
- The issue was observed with Torchvision RAFT Large FP16 on GPU, where
blocked `b_fs_yx_fsv16` MVN was selected for large-spatial instance
normalization nodes.
- The regression test uses shape `{1, 16, 257, 256}`. Its spatial
reduction size is `257 * 256 = 65792`, which exceeds the FP16 finite max
value `65504` when squared deviations are accumulated with values of
order 1 or larger.
- Keeps fused-op behavior consistent by converting the final normalized
value back to `ACTIVATION_TYPE` before passing it to fused ops.
- Adds a GPU unit test that forces the `mvn_gpu_b_fs_yx_fsv16`
implementation.
```mermaid
%%{init: {"flowchart": {"htmlLabels": true}}}%%
flowchart TB
classDef conv fill:#e8f1ff,stroke:#2b6cb0,stroke-width:1px,color:openvinotoolkit#111
classDef reorder fill:#fff4df,stroke:#b7791f,stroke-width:1px,color:openvinotoolkit#111
classDef mvn fill:#e8f7ee,stroke:#2f855a,stroke-width:1px,color:openvinotoolkit#111
subgraph GOOD["Good dump: 2026.2.0 / ffa272d"]
direction TB
GConv1["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: Convolution<br/>layout: b_fs_yx_fsv16<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: undef</div>"]
GReorder1["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: Reorder<br/>layout: bfyx<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: reorder_data__f16</div>"]
GMVN["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: MVN<br/>layout: bfyx<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: mvn_gpu_bfyx_opt__f16</div>"]
GReorder2["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: Reorder<br/>layout: b_fs_yx_fsv16<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: reorder_data__f16</div>"]
GConv2["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: Convolution<br/>layout: b_fs_yx_fsv16<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: undef</div>"]
GConv1 --> GReorder1 --> GMVN --> GReorder2 --> GConv2
end
subgraph BAD["Bad dump: 2026.3.0 / 48b684f"]
direction TB
BConv1["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: Convolution<br/>layout: b_fs_yx_fsv16<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: undef</div>"]
BMVN["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: MVN<br/>layout: b_fs_yx_fsv16<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: mvn_gpu_b_fs_yx_fsv16__f16</div>"]
BConv2["<div style='min-width: 420px; text-align: left; white-space: nowrap;'>type: Convolution<br/>layout: b_fs_yx_fsv16<br/>shape: dynamic [2, 64, -1, -1]<br/>runtime shape: [2, 64, 260, 480]<br/>primitive: undef</div>"]
BConv1 --> BMVN --> BConv2
end
class GConv1,GConv2,BConv1,BConv2 conv
class GReorder1,GReorder2 reorder
class GMVN,BMVN mvn
```
### Validation:
- Passed: (new test-case for ticket issue)
`bin\intel64\Release\ov_gpu_unit_tests.exe
--gtest_filter="*mvn_fsv16_f16_large_spatial*" --device_suffix=1`
### Tests:
- `mvn_fsv16_f16_large_spatial/mvn_random_test_bsv32.random/0`
- `mvn_fsv16_f16_large_spatial/mvn_random_test_bsv32.random_cached/0`
### Tickets:
- 186526
### AI Assistance:
- AI assistance used: yes
- AI: find a root-cause, fix issue, add tests
- User: validation, code review
---------
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
…olkit#36040) ### Details: Duplicate of [PR#36009](openvinotoolkit#36009) ### Tickets: - *C-186974* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
### Details: - *item1* - *...* ### Tickets: - *ticket-id* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
Initial formatting and install docs updates
Last min updates to models ### Details: - *item1* - *...* ### Tickets: - *ticket-id* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
## Summary This PR adds the Physical AI documentation to OpenVINO. ## Changes - Added new documentation section: Physical AI Framework - Integrated full documentation structure: - getting-started - how-to - explanation - reference - Added entry to documentation navigation (new top-level section) - Added homepage tile linking to Physical AI section ## Motivation This documentation introduces workflows for deploying VLA models on robots, expanding OpenVINO usage in robotics and physical AI scenarios. ## Notes - Documentation is integrated directly (no external dependencies) - Structure follows OpenVINO documentation guidelines - Links updated to comply with Sphinx linking rules ## Testing - Verified documentation locally (Sphinx build) - Checked navigation and links ## Checklist - [x] Changes are focused on documentation only - [x] PR follows naming conventions - [x] Documentation structure is consistent
### Details: add release notes for physical ai ### Tickets: - *ticket-id* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
### Details: - *new hardware added to system descriptions, WCL-7-350* - *new models: yolo26n, ltx-video, qwen3.6-27b and gpt-oss-120b* - *updated accuracy table* ### Tickets: - *na* ### AI Assistance: - *AI assistance used: no - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
…vinotoolkit#36136) (openvinotoolkit#36204) ### Details: After `kernels` got an update on PyPI, we started to have failures in PyTorch Models Tests job at PyTorch LLM Model Tests. This PR should fix it ### Tickets: - *CVS-187808* ### AI Assistance: - *AI assistance used: no* Co-authored-by: Andrey Babushkin <andrey.babushkin@intel.com>
…notoolkit#35917) (openvinotoolkit#36251) Cherry-pick of openvinotoolkit#35917 to releases/2026/2
…tachment on cache import. (openvinotoolkit#36196) ### Details: Duplicated openvinotoolkit#36195 and openvinotoolkit#36252 ### Tickets: - *[EISW-219826](https://jira.devtools.intel.com/browse/EISW-219826)* - *[EISW-220018](https://jira.devtools.intel.com/browse/EISW-220018)* ### 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: intelgaoxiong <xiong.gao@intel.com>
…36003) (openvinotoolkit#36276) ### Description - **Symptom:** yolo26n model fails to compile on GPU with `[GPU] ProgramBuilder build failed! — Requested activation is not supported for integer type` at the `Sign` node (`node_Sign_1348`, i64 input) - **Root cause:** `activations_int8` allowlist in `activation_inst::calc_output_layout()` did not include `activation_func::sign`. The list guards `i8/u8/i32` types, and the GPU plugin internally converts the model's `i64` Sign input to `i32`. Since `sign` was absent from the allowlist, it was unconditionally rejected. The Sign OCL kernel itself correctly handles integer types via a ternary/select expression (`input > 0 ? 1 : (input == 0 ? 0 : -1)`) without any type mismatch — the omission was an oversight. - **Resolution:** Added `activation_func::sign` to the `activations_int8` allowlist. Unlike the similar PR openvinotoolkit#29099 (`abs`), no change to activation_kernel_opt.cpp is needed because the Sign macro returns the same signed integer type as the input (no `int4 → uint4` mismatch). #### The code and line that caused this issue - `intel_gpu/src/graph/activation.cpp` — `activation_inst::calc_output_layout()`, `activations_int8` vector (line ~23) #### Reproduction step and snapshot - Model: `yolo26n-pytorch` FP16 (OMZ public) ```bash python -c " import openvino as ov core = ov.Core() model = core.read_model('yolo26n/openvino_model.xml') core.compile_model(model, 'GPU') # RuntimeError: sign:node_Sign_1348 — Requested activation is not supported for integer type " ``` #### Problematic graph - Single `Sign` node with `int64` input/output shape `[1, 300]`, fed from a `TopK` index output - GPU plugin converts `i64 → i32` internally, then the `i32` Sign hits the allowlist check and fails #### Checklist - [x] Is it a proper fix? (not a workaround) - [x] Did you include test case for this fix, if necessary? - [x] Did you review existing test that can be extended to cover this scenario? Which test did you review? - Extended `activation_i32_fw_gpu.basic_yxfb_i32_funcs` in activation_simple_gpu_test.cpp — added `activation_func::sign` to the tested functions list and its corresponding assertion (`val > 0 ? 1 : val == 0 ? 0 : -1`) ### Tickets - *187077* ### master PR - openvinotoolkit#36003
### Details: - *Disable shared command queue feature as default* - *Update workload type for current command queue immediately if shared command queue is disabled* ### Tickets: - *CVS-188132* ### AI Assistance: - *AI assistance used: yes* - *tests* --------- Signed-off-by: Bogdan Pereanu <bogdan.pereanu@intel.com>
…les (openvinotoolkit#36306) Summary Updates the main documentation index page to reflect the addition of the Physical AI Framework. Changes Updated text to reflect the increase from three to four models on the homepage Reordered existing tiles/boxes based on updated layout guidance Adjusted links to match the new ordering Motivation This change ensures the main documentation page accurately represents the current number of available models and aligns the layout with the updated structure and guidance. Notes Changes are limited to index.rst (homepage content) No new tiles were added as part of this change Update focuses on text accuracy and layout ordering Testing Verified changes in local Sphinx build Confirmed correct rendering and link navigation --------- Co-authored-by: whitneyfoster <whitney.foster@intel.com>
…ce doesn't support it (openvinotoolkit#36309) ### Details: - *Remove npu_enable_strides_for from supported properties if device doesn't support it* ### Tickets: - *ticket-id* ### 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>
… (openvinotoolkit#36311) ### Details: - *Fix dynamic quantization large group case. Add reduction between Sub_Groups, which handle the same quantization group.* - *cherry pick openvinotoolkit#36147 to release branch* ### Tickets: - *CVS-187724* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
…penvinotoolkit#36029) (openvinotoolkit#36274) This PR enhances the UnsqueezeBroadcastReshapeSDPAFusion pass to capture more flexible attention topologies, specifically targeting MQA and GQA patterns. Key Enhancements: 1) 3D Input Support: The pass now successfully intercepts and reshapes 3D Key/Value tensors feeding into Unsqueeze or Reshape nodes. 2) GQA Implicit Broadcasting: Previously, the fusion aborted if the model required expanding KV heads to match a larger number of Query heads. This PR introduces dynamic shape extraction to bypass explicit Broadcast nodes. The SDPA kernel will now perform an implicit broadcast in the registers—for example, natively mapping 2 KV heads across 32 Query heads (a 1:16 ratio)—saving massive memory bandwidth. ### Tickets: - CVS-186566, CVS-187072 ### AI Assistance: - *AI assistance used: yes - Debug and test generation
### Details: Minor product version bump ### AI Assistance: - *AI assistance used: yes, for creation of connected PRs to different repositories*
…oolkit#36128) ### Details: - With latest transformers (v5) LongRope pattern failed to match for Phi-3.5/Phi-4 models, updated pass according to new pattern - Port of openvinotoolkit#35981 from the master branch ### Tickets: - EISW-214108 ### AI Assistance: - *AI assistance used: yes* - Used for adding tests, after that checked them manually --------- Co-authored-by: Dmitry Matveev <dmitry.matveev@intel.com>
### Details: - *Fixed MoE sym-quant issue* - *Cherry pick the fix commit from openvinotoolkit#35901 to 26.2 release branch* ### Tickets: - *CVS-187724* ### AI Assistance: - *AI assistance used: no / yes* - *If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).*
…penvinotoolkit#36377) ## Summary Adds a reusable "Explore on GitHub" button to documentation pages and enables it for the Physical AI article. ## Changes * Added a new **"Explore on GitHub"** button to `docs/articles_en/physical-ai.md` * Displays in the right-hand sidebar * Links to the Physical AI GitHub repository * Extended Sphinx theme template: `docs/openvino_sphinx_theme/openvino_sphinx_theme/templates/edit-this-page.html` * Introduced a reusable HTML component that conditionally renders the "Explore on GitHub" button when `explore_github_url` is defined in page metadata * Preserved existing functionality by keeping the **"Edit on GitHub"** button rendered after the new component ## Implementation Notes * The new component checks: ```jinja2 meta.get("explore_github_url") ``` and renders the button only when the metadata field is present * The component is designed to be **reusable across documentation pages**, enabling easy linkage to corresponding GitHub repositories * Existing behavior remains unchanged when the metadata field is not provided ## Motivation This change improves discoverability of related GitHub repositories directly from documentation pages and introduces a reusable pattern that can be applied consistently across OpenVINO documentation. ## Testing * Verified rendering locally in Sphinx build * Confirmed: * button appears only when metadata is defined * correct placement in sidebar * existing "Edit on GitHub" button remains functional ## Checklist * [x] Changes are focused and scoped to documentation/theme * [x] Existing functionality preserved * [x] Reusable solution implemented * [x] Verified locally
openvinotoolkit#36429) ### Details: The PR https://github.com/openvinotoolkit/openvino/pull/36324/changes, made to the master branch, was reviewed and is awaiting merging. However, we also need those changes on the 2026/2 version of the page, and this PR is bringing them. Note: one change to docs/articles_en/get-started/install-openvino/install-openvino-windows.rst will be added in the next commit since it's missing from this PR.
### Details: - *Added WCL-5-330 AI-PC results and removed MTL-7-165H results* - *Updated results for PTL-7-368H and LNL-7-258V* - *Updated system descriptions* - *Updated date and release version on landing page for AI-PC results* ### Tickets: - ** ### AI Assistance: - *AI assistance used: no* - *-* --------- Co-authored-by: whitneyfoster <whitney.foster@intel.com>
Align MarkRopeInputsToKeepInMixedPrecision with disable_conversion(f16), add DisablePrecisionConversion rt_info (backed by legacy disable_fp16_compression wrappers), and run the pass on the GPU plugin after RoPEFusion so cos/sin subgraphs stay in FP32 under f16 compression. Co-authored-by: Cursor <cursoragent@cursor.com>
mpaulitsch
requested review from
cavusmustafa,
mfhansen and
mmikolajcz
and removed request for
a team
github-actions
Bot
added
category: inference
Contributor
|
this PR contains changes that are not a result of model enablement. |
Author
|
addressed comment. created new pull request (flux.2-klein-4b enablement #36544) with clean baseline and essential files. close this PR as it is not needed anymore. |
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
20 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:
AI Assistance:
yes, created and checked output aligning with expectations, created extensive performance tests, code review.