🔥 Jadwal & Hasil: microsoft/onnxruntime/pull/

adrastogi · 2026-10-21T23:43:14Z

Description

This cherry-picks the following commits for the release:

CI fixes:

[CUDA] Guard FP8 XQA test by compute capability #32559 (8ec7981a7439546a54196b9cbc777ae5de708b6b): guard FP8 XQA test cases by GPU compute capability. The Windows CUDA runner is an SM86 A10, while FP8 XQA requires SM89 or SM90+. INT4/INT8 XQA coverage remains enabled on supported SM80+ devices. This fixes the two FP8 dispatch assertions without changing product behavior.
Raise the Android minimal baseline binary-size threshold by 1 KiB, from 1,589,248 to 1,590,272 bytes (release branch only). CI reports a section total of 1,590,022 bytes, exceeding the previous threshold by 774 bytes. The adjusted threshold leaves 250 bytes of headroom.
Move back to main A10 image #32676, then Fix CUDA CICD. Remove unnecessary image override. #33489: rename the old A10-Test image and remove the unnecessary image override, in upstream order.
[CICD Stability] Start transition to scaleset based providers #33492: backport the scale-set runner migration across the release branch's 41 affected workflows. The Linux CUDA, CUDA Plugin, and TensorRT test jobs now use ort-github-linux-6-a10-scus-ss. Their previous retries were cancelled by 1es-resource-management[bot] before any steps executed.

The first ten cherry-picks, plus #32676 and #33489, applied cleanly from the original mainline commits, with source commit provenance preserved. #32607 includes the full mainline fix and regression tests rather than the reduced 1.28.3 backport.

#33492 required conflict resolution because the release branch has an older workflow layout:

Preserve the existing Linux CI and Linux WebGPU job sets; do not introduce main-only no-exceptions or static-plugin jobs.
Leave windows_arm64_python.yml absent. It was introduced after 1.30 branched by Add github action for arm64 windows build #32791; its runner migration does not require adding that workflow to this release.
Preserve Windows WebGPU's combined build/test jobs and their check names, keeping them on the new GPU scale set rather than importing main's split build/test topology. The build-only minimal Edge job uses the new CPU scale set.

Motivation and Context

Bring the product fixes from #32806 into ORT 1.30.1, address the Windows CUDA test and Android size-gate failures, and migrate CI runner selection to unblock the cancelled GPU jobs.

The merged EP API version correction from #33525 was brought in by merging the updated rel-1.30.1 base. The uuid-dev CI fix in #32491 was already present in the 1.30 base and is not duplicated.

The OpenVINO changes in #33112 and #33327 are intentionally excluded: the namespace regression fixed by #33327 was introduced by #33112, and neither change is present in the 1.30 base.

Validation:

All nine original product cherry-pick patch IDs were verified against their mainline commits; the additional [CUDA] Guard FP8 XQA test by compute capability #32559 patch ID also matches upstream.
The upstream test_fp8_xqa_cuda_capability regression passed in an isolated standard-library harness using mocked SM86, SM89, and SM90 capabilities. Additional no-CUDA and external-runner cases passed.
Python syntax and Android threshold consistency/arithmetic checks passed.
git diff --check 3b0f46f82c HEAD passed.
All 41 migrated workflow YAML files parse. Structural comparisons against the pre-migration release workflows confirm that only runner labels, caller pool names, and the reusable pool-input description changed; job sets, steps, triggers, permissions, matrices, and build flags are preserved.
All 13 callers of the reusable Linux build workflow use scale-set pool names; all three previously cancelled GPU test jobs use the new A10 scale set.
The Android release-specific size threshold remains 1,590,272 bytes.
The full helper pytest invocation was not runnable because pytest is not installed in the local virtual environment. No local GPU execution or Android rebuild was performed; CI will validate the integrated changes.

… scalar beta (#32610) ### Description `EmbedLayerNormalizationShapeInference` validated the `beta` input's rank using the wrong shape: `beta_dims` was aliased to `gamma_shape.dim()` instead of `beta_shape.dim()`. As a result, a rank-1 `gamma` would let the rank check pass even when `beta` is a 0-dim scalar, and execution reached `beta_shape.dim(0).dim_value()` — an unchecked protobuf repeated-field access on an empty `dim` list, causing an out-of-bounds read during session load, before provider/kernel assignment. ### Fix - Alias `beta_dims` to `beta_shape.dim()` so the rank check guards the correct input; a scalar `beta` is now rejected via `fail_shape_inference` instead of reaching the OOB access. This matches the adjacent `gamma` validation. - Correct the `gamma` error message ("2 dimension" → "1 dimension") to reflect the actual check (cherry picked from commit 7054657) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description  Add an explicit validation in `Node::UpdateInputArgCount()` to reject a node that supplies actual inputs while its bound operator/function schema declares no formal input parameters. ### Motivation and Context  When `op.inputs()` is empty but the node has ≥1 input, the arg-count adjustment loop was skipped yet the trailing `input_arg_count.push_back(arg_count_left)` still ran unconditionally, producing `InputArgCount().size() == 1` against `op.inputs().size() == 0`. This size-invariant violation later caused an out-of-bounds read at `op.inputs()[i]` in `InferAndVerifyTypeMatch` during `Graph::Resolve()`. This is reachable only via a model-local function in a custom domain: for registered ops, `OpSchema::Verify` in the ONNX checker rejects the extra inputs first, but for custom-domain function calls the checker performs no arity validation. (cherry picked from commit 613bc03) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

…te finalization (#32641) ### Description `SessionState::FinalizeSessionStateImpl` iterates over the subgraphs attached to each node and unconditionally downcasts the node's kernel to `controlflow::IControlFlowKernel`, then calls `SetupSubgraphExecutionInfo` on it: ```cpp // Downcast is safe, since only control flow nodes have subgraphs auto& control_flow_kernel = static_cast<controlflow::IControlFlowKernel&>(*p_op_kernel); ``` The "only control flow nodes have subgraphs" invariant is not enforced anywhere. Node::Init materializes a subgraph for any attribute of type GRAPH, with no schema gate, so an ordinary (non-control-flow) node can carry a subgraph. When it does, the downcast above is invalid: IControlFlowKernel adds a vtable slot that a plain OpKernel does not have, so the call reads an out-of-bounds vtable slot — an undefined-behavior type confusion that crashes (observed as an access violation inside FinalizeSessionStateImpl). This is reachable by loading a crafted/malformed model whose non-control-flow node has a GRAPH-typed attribute (for example, an op that permits unchecked attributes). ORT should reject such a model with a clear error instead of executing the bad cast. ### Fix Gate the downcast on a virtual predicate: - Add `OpKernel::IsControlFlowKernel()` returning `false` by default. - Override it to `true` on `IControlFlowKernel`, which covers If/Loop/Scan and their CUDA derivatives. - Check it in `FinalizeSessionStateImpl` before the cast and return an error otherwise. A virtual predicate is used rather than `dynamic_cast` because `onnxruntime_DISABLE_RTTI` is on by default. There is no behavior change for valid models: only the control-flow kernels inherit `IControlFlowKernel`, and they always return `true`. Adds `InferenceSessionTests.SubgraphAttributeOnNonControlFlowNodeIsRejected` test case, which builds a model whose non-control-flow node carries a `GRAPH` attribute and asserts that session initialization fails gracefully instead of triggering the downcast. Existing If/Loop/Scan subgraph tests cover the no-regression path. (cherry picked from commit b95cf78) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

…full-byte tensor (#32611) ### Description The allocation planner's `SameSize()` decided buffer reuse by comparing the C++ carrier size of the element type (`elt_type->Size()`) plus the logical shape. Packed sub-byte types (`int4`/`uint4`) have the same 1-byte carrier size as `int8`/`uint8`, but one carrier stores 2 logical elements. As a result a `uint4[N]` tensor (physical `ceil(N/2)` bytes) and a `uint8[N]` tensor (physical `N` bytes) were treated as the same size, and the smaller `uint4` buffer was reused for the `uint8` output. Writing the `uint8` tensor into that half-sized buffer overflows it. ### Fix - `SameSize()` now also requires the sub-element packing density (`GetNumSubElems()`) to match, so tensors are only considered the same size when their physical storage bytes are actually equal. - `ExecutionFrame::AllocateMLValueTensorPreAllocateBuffer()` adds a defense-in-depth capacity check that rejects a reuse when the requested tensor needs more storage bytes than the buffer being reused (the previous check only compared logical element counts). - Added `AllocationPlannerTest.AvoidReuseOfPackedSubByteBufferForFullByteTensor`, which builds `X(float) → uint4 → float → uint8 → float` so a dead `uint4[1024]` output would be reused by a `uint8[1024]` output. Without the fix the planner reuses the buffer (`alloc_kind == kReuse`) and the output is corrupted; with the fix the test passes and the model runs correctly (cherry picked from commit 061eecb) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

…izer (#32609) ### Problem `OuterScopeNodeArgLocationAccumulator` (session_state.cc) maps a Loop/Scan node's explicit inputs onto the subgraph's inputs using `GraphViewer::GetInputs()`, which *excludes* initializer-backed inputs. When a body declares an input that is also a body initializer, `GetInputs()` is shorter than the parent node's input list, so indexing `subgraph_inputs[arg_idx]` reads past the end of the vector and dereferences an invalid `NodeArg` pointer during session initialization. This causes a crash / heap out-of-bounds read (observable under ASan) when loading a malformed model containing such a Loop/Scan. ### Fix In the Loop/Scan≥9 branch of `OuterScopeNodeArgLocationAccumulator`: - If `subgraph.GetInputs().size()` does not equal the parent node's input count, return an `INVALID_GRAPH` status naming the parent node and both counts, instead of indexing. When the size difference is explained by initializer-backed inputs (`GetInputsIncludingInitializers()` is larger), the message adds a hint pointing at that as the cause. - Add `Loop.BodyInputAlsoInitializerIsRejected` test: builds a Loop whose body input `state_in` is also a body initializer, and asserts `InferenceSession::Initialize()` fails with `INVALID_GRAPH` and the expected message — i.e. rejected cleanly rather than crashing. The test asserts on the status code and message text on purpose, so that it keeps covering this path rather than silently falling through to the `ORT_ENFORCE` downstream. (cherry picked from commit 09dfa6a) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Summary Strengthens attribute/input validation in three contrib-op shape-inference functions in bert_defs.cc so malformed or malicious models are rejected at `Graph::Resolve` instead of triggering out-of-bounds reads or signed-integer overflow during shape inference. ### Changes `CausalConvWithState`: validate the ndim attribute is in [1, 3] and cross-check tensor ranks against it (weight == ndim+2, channels-first input == ndim+2, channels-last input >= 3) before the spatial-dim loop indexes input.dim(2+i). Previously only a rank >= 2 guard existed, so ndim=2/3 with a low-rank input read past the shape's dimensions. `GatedDeltaNet`: require the head counts/sizes to be positive and add step-by-step overflow guards before computing the state_update capsule width, preventing signed int64 overflow (UB) in state_update_capacity * (num_heads_v + num_heads_k*head_size_qk + num_heads_v*head_size_v). `GroupQueryAttention` / `SparseAttention`: check the parsed `total_sequence_length` initializer is non-empty before indexing data[0]. (cherry picked from commit bca9d15) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description Add an explicit rank-consistency check in `convTransposeWithDynamicPadsShapeInference()` so that a derived `kernel_shape` (built from the weight tensor's rank when the `kernel_shape` attribute is absent) must have the same length as `n_input_dims` (derived from the input tensor's rank) before it is used, exactly like the existing check on the explicit-attribute path. ### Motivation and Context `n_input_dims` is computed from input `X`'s rank. When `kernel_shape` is not given as an attribute, it is instead derived from weight `W`'s rank. If `W`'s rank disagrees with `X`'s rank, the resulting `kernel_shape` (and `effective_kernel_shape`, sized identically) is shorter or longer than `n_input_dims`. Both directions are unsafe: - **Shorter** (e.g. X rank 5, W rank 3): the output-shape loop iterates `n_input_dims` times over `effective_kernel_shape`/`pads`, reading past the end of the shorter vector. - **Longer** (e.g. X rank 3, W rank 5): the dilation loop iterates `kernel_shape.size()` times over `dilations`, which is always sized to `n_input_dims`, reading past the end of `dilations`. This runs during `Graph::Resolve()` (model load), before any kernel `Compute()` validation is ever reached. ### Testing Added `ConvTransposeWithDynamicPads_MismatchedInputWeightRank` (X rank 5, W rank 3, no `kernel_shape` attribute), asserting the model now fails cleanly at the existing kernel-level rank check (`"X num_dims does not match W num_dims."`) instead of hitting the shape-inference OOB read during load. (cherry picked from commit 2ecddca) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description `InferenceContextImpl::getInputData` (graph.cc) indexed `node_.InputDefs()` directly by the schema input index during shape inference, with no bounds check. Add the same bounds check already used by `DataPropagationContextImpl::getInputData` in the same file, so `getInputData` returns `nullptr` (treated as "input not available") instead of indexing out of bounds. ### Motivation and Context ONNX schemas can declare trailing inputs as optional (e.g. `ConvTransposeWithDynamicPads`'s `Pads`), so a node can validly omit them entirely, shrinking `Node::InputDefs()` below the schema's full input count. A `TypeAndShapeInferenceFunction` that queries such an omitted optional input's data (e.g. via `ctx.getInputData(index)`) on such a node reads past the end of the vector, causing an out-of-bounds read. (cherry picked from commit e5f272c) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description Forward the caller's stream through the Plugin EP single-tensor `CopyTensorImpl` path. This matches the existing batched-copy behavior while preserving a null stream for synchronous copies. Adds focused callback regression coverage for asynchronous, synchronous, and batched copies. ### Motivation and Context The single-tensor Plugin EP data-transfer path discarded the supplied stream, preventing asynchronous copies from receiving the execution stream. Fixes #32643 Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> (cherry picked from commit 5899c7c) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

Tianlei Wu (tianleiwu) · 2026-10-22T00:37:49Z

It seems many CI failures. Might need cherry-pick some build related changes.

adrastogi · 2026-10-22T01:21:20Z

It seems many CI failures. Might need cherry-pick some build related changes.

Tianlei Wu (@tianleiwu) yeah, there's a versioning fix that didn't make it through. 😢 #33525 should hopefully take care of most of these.

Tianlei Wu (tianleiwu) · 2026-10-22T03:07:58Z

adrastogi, you can rebase since that fix is merged. Let's see whether there are other issues.

Bring in the merged EP API version assertion correction from #33525. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description The Windows CUDA CI runner for PR #32515 uses an SM86 A10-4Q. `PagedAttention` intentionally enables FP8 XQA only on SM89 or SM90+, so the FP8 cases in `test_xqa_large_attention_scale_and_k_scale` fell back to portable kernels and failed the XQA dispatch assertion. Mirror the runtime capability predicate in the test and include the FP8 variant only on supported GPUs. UINT8 and INT8 remain covered on SM80 and newer devices. This is a capability guard, not a memory guard: the failed job completed pytest normally and reported two assertion failures rather than a CUDA allocation failure or process crash. Failed job: https://github.com/microsoft/onnxruntime/actions/runs/34435175610/job/102757939916?pr=32515 ### Motivation and Context Fixes the Windows GPU CUDA CI regression introduced by #32515 without skipping valid XQA coverage on SM86. ### Testing - `TestPagedAttentionInt4Helpers`: 4 tests passed, including simulated SM86, SM89, and SM90 capability cases. - `lintrunner onnxruntime/test/python/transformers/test_paged_attention_int4.py` - `python -m py_compile onnxruntime/test/python/transformers/test_paged_attention_int4.py` - `git diff --check` (cherry picked from commit 8ec7981) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

Raise the release baseline by 1 KiB, from 1,589,248 to 1,590,272 bytes. CI reports a deterministic section total of 1,590,022 bytes, exceeding the old threshold by 774 bytes. The adjusted limit leaves 250 bytes of headroom. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

This pull request updates the GitHub Actions workflow configurations for several Linux CUDA-related CI jobs. The main change is to update the image used for the CI runners from `onnxruntime-ubuntu2204-CUDA-A10-Test` to `onnxruntime-ubuntu2204-CUDA-A10`, likely to standardize or align with the correct runner image. **CI runner image updates:** * Updated the `1ES.ImageOverride` setting from `onnxruntime-ubuntu2204-CUDA-A10-Test` to `onnxruntime-ubuntu2204-CUDA-A10` in the following workflow files: - `.github/workflows/linux_cuda_ci.yml` - `.github/workflows/linux_cuda_no_cudnn.yml` - `.github/workflows/linux_cuda_plugin_ci.yml` - `.github/workflows/linux_tensorrt_ci.yml` (cherry picked from commit a97735f) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

This pull request updates several GitHub Actions workflow files related to Linux CUDA and TensorRT CI jobs. The main change is the removal of the `1ES.ImageOverride=onnxruntime-ubuntu2204-CUDA-A10` setting from the `runs-on` configuration in multiple workflows. This may allow the jobs to use the default image for the specified pool, potentially improving maintainability or compatibility. Workflow configuration updates: * Removed the `1ES.ImageOverride=onnxruntime-ubuntu2204-CUDA-A10` line from the `runs-on` section in the following workflow files: * `.github/workflows/linux_cuda_ci.yml` * `.github/workflows/linux_cuda_no_cudnn.yml` * `.github/workflows/linux_cuda_plugin_ci.yml` * `.github/workflows/linux_tensorrt_ci.yml` (cherry picked from commit ef1d811) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description  ### Motivation and Context  Backport notes for rel-1.30.1: - Preserve existing Linux and WebGPU job sets; do not introduce main-only no-exceptions or static-plugin jobs. - Keep the main-only Windows ARM64 Python workflow absent. - Preserve combined Windows WebGPU build/test jobs on the GPU scale set; use the CPU scale set for the build-only minimal Edge job. (cherry picked from commit 973f680) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description  This cherry-picks the following commit for the release: * #33561 (`7ec270791d92a789db39c868e565d05ce7d1654b`): set up Python 3.12 before running package validation in the combined packaging stage. The cherry-pick applied cleanly, with source commit provenance preserved. This branch is based on the current `rel-1.30.1`, including the servicing payload merged in #33524. ### Motivation and Context  Bring the packaging-stage fix into ORT 1.30.1 so package validation has an explicitly selected Python interpreter before the next internal release build. Validation: * The cherry-pick patch ID matches the original mainline commit. * YAML parsing and assertions passed: Python 3.12 setup precedes the existing package-validation task. * `git diff --cached --check` passed before committing. * The internal release pipelines have not been rerun with this change. Co-authored-by: eserscor <erscor@microsoft.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

### Description  This cherry-picks the following commit for the release: * #33142 (`38dc99fba3f5bd72e201d1d942235bf4f600a1b3`): work around the Windows certificate regression in build-tools setup by installing and injecting `truststore`, inspecting proxy settings, and sanitizing `NO_PROXY` before invoking the GitHub Action wrapper. The full two-file cherry-pick applied cleanly, with source commit provenance preserved. This branch is based on current `rel-1.30.1`, including #33524 and #33641. ### Motivation and Context  The internal release build failed in six Windows packaging jobs while downloading the build-tools action through `run_gh_action.py`: `CERTIFICATE_VERIFY_FAILED: unable to get local issuer certificate` This upstream fix enables Python to use system-trusted certificates without disabling certificate verification. The previous Python-version fix (#33561, backported through #33641) is present and package validation passed in both release runs. This PR does not address the separate macOS checkout timeout or the Maven feed connection timeout in the CUDA13 build. Those failures need targeted infrastructure retries; no additional cherry-picks are currently indicated. Validation: * The cherry-pick patch ID matches the original mainline commit. * Python AST parsing and YAML parsing/assertions passed: `truststore` installation/injection and the PowerShell wrapper invocation are present. * `git diff --cached --check` passed before committing. * No end-to-end execution on the affected release agents was performed. No release pipelines were restarted. Co-authored-by: eserscor <erscor@microsoft.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

Wei Wang (wangw-1991) and others added 9 commits October 21, 2026 16:18

Tianlei Wu (tianleiwu) previously approved these changes Oct 22, 2026

View reviewed changes

Aditya Rastogi and others added 3 commits October 21, 2026 20:43

Merge rel-1.30.1 version fix into servicing cherry-picks

3b0f46f

Bring in the merged EP API version assertion correction from #33525. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 3eb3fc9b-6147-4706-b7ad-47bb3476377b

adrastogi dismissed Tianlei Wu (tianleiwu)’s stale review via ef2fc04 October 23, 2026 18:04

eserscor and others added 3 commits October 26, 2026 13:32

Tianlei Wu (tianleiwu) approved these changes Oct 26, 2026

View reviewed changes

Akshay Sonawane (apsonawane) approved these changes Oct 27, 2026

View reviewed changes

adrastogi merged commit 77e55fc into rel-1.30.1 Oct 27, 2026
95 of 99 checks passed

adrastogi deleted the adrastogi/1.30.1-cherry-picks branch October 27, 2026 06:16

adrastogi mentioned this pull request Oct 27, 2026

ORT 1.30.1 Cherry Picks - 2 #33641

Merged

adrastogi mentioned this pull request Oct 27, 2026

ORT 1.30.1 Cherry Picks - 3 #33651

Merged

Navigation Menu

Repository navigation

ORT 1.30.1 Cherry Picks - #33524

ORT 1.30.1 Cherry Picks#33524
adrastogi merged 15 commits into
rel-1.30.1from
adrastogi/1.30.1-cherry-picks

adrastogi commented Oct 21, 2026 •

edited

Loading

Uh oh!

Tianlei Wu (tianleiwu) commented Oct 22, 2026

Uh oh!

adrastogi commented Oct 22, 2026

Uh oh!

Tianlei Wu (tianleiwu) commented Oct 22, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

Conversation

adrastogi commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Motivation and Context

Uh oh!

Tianlei Wu (tianleiwu) commented Oct 22, 2026

Uh oh!

adrastogi commented Oct 22, 2026

Uh oh!

Tianlei Wu (tianleiwu) commented Oct 22, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

adrastogi commented Oct 21, 2026 •

edited

Loading