Fix CUDA ReduceSum erroring out on empty tensors with explicit axes by justinchuby · Pull Request #28353 · microsoft/onnxruntime

Justin Chu (justinchuby) · 2026-05-04T21:19:38Z

Description

Remove the overly strict assertion in CUDA PrepareForReduce that rejects reducing along a zero-sized dimension even with explicit axes. This matches the behavior of the CPU implementation which handles empty tensors via check_and_reduce_empty_set_input().

Motivation

ORT GenAI's Gemma4 CUDA pipeline triggers ReduceSum on {1, 0} tensors during prefill (past_sequence_length=0). The CPU implementation handles this correctly, but the CUDA path crashes with:

input_dims[axis] != 0 was false. Can't reduce on dim with value of 0 if 'keepdims' is false.

Reducing axis 1 of {1, 0} with keepdims=false produces shape {1} filled with the identity value (0 for sum). This is mathematically valid and numpy handles it correctly.

Changes

Removed the ORT_ENFORCE(input_dims[axis] != 0, ...) assertion at line 291 of reduction_ops.cc. The existing ReduceComputeCore already handles input_count == 0 correctly (line 369-370).

The default-axes path (line 302) is left unchanged — it already conditionally checks keepdims || dim != 0.

Testing

Verified with Gemma4 e2b-it model on H200 GPU:

Before: ReduceSum_node_232 crashes on {1, 0} tensor
After: ReduceSum succeeds, inference proceeds to next node (GroupQueryAttention)

Related issues

CUDA EP: CUBLAS misaligned address crash during prefill with Gemma4 model onnxruntime-genai#2120 (Gemma4 CUDA prefill crash)

Xavier Dupré (xadupre) · 2026-05-05T09:35:15Z

Why not removing this one?

When not retaining axis, the empty dimension will be removed and lead to an incorrect shape (non zero). So I think this check is still necessary

Response is incorrect for reduce-all empty tensors with keepdims=0

Fixed. The reduce-all path in PrepareForReduce now sets all output_dims to 1 (not 0), so output_count=1 for scalar results. With keepdims=0, squeezed_output_dims is empty (scalar). The identity fill at line 377 handles this correctly: input_count=0, output_count>0, fills scalar with identity.

Xavier Dupré (xadupre) · 2026-05-05T09:36:19Z

+  test.Run(OpTester::ExpectResult::kExpectSuccess, "", {}, nullptr, &execution_providers);
+}
+
+TEST(ReductionOpTest, ReduceSum_CudaEmptyTensor_MultiDim) {


ReduceProd should be tested well.

Dmitri Smirnov (yuslepukhin) · 2026-05-05T19:11:44Z

The PR does not actually fix the claimed CUDA empty-tensor case.

After shape prep, both CUDA execution paths still special-case empty input by assuming the output must also have size 0 and then returning without materializing the reduction identity. See reduction_ops.cc:365, reduction_ops.cc:758, and reduction_ops.cc:759.

For the PR’s target case {1,0} reduced over axis 1 with keepdims=0, the output shape is {1}, not size 0, so this path is still wrong.

In debug it will trip the assert; in release it returns success with an uninitialized non-empty output buffer.

The ATen fast path does not save this case because it explicitly requires input_shape.Size() > 0 at reduction_ops.cc:702

Dmitri Smirnov (yuslepukhin) · 2026-05-05T19:15:21Z

  ORT_ENFORCE(keepdims || dim != 0,

The remaining default-axes guard is still incorrect for empty full reductions with keepdims=0, so the reviewer comment asking “why not removing this one?” is not satisfactorily addressed.

The code still rejects any zero dimension when no axes are provided at reduction_ops.cc:298.

Per ONNX, reducing an empty set over all axes is valid and should yield the operator identity, not an invalid shape.

CPU already implements that behavior via reduction_ops.cc:875 and fills identities via reduction_ops.cc:912

Refers to: onnxruntime/core/providers/cuda/reduction/reduction_ops.cc:298 in 005f7ad. [](commit_id = 005f7ad, deletion_comment = False)

Dmitri Smirnov (yuslepukhin) · 2026-05-05T19:26:06Z

There is a difference between a crash and returning an error.

github-actions

You can commit the suggested changes from lintrunner.

Dmitri Smirnov (yuslepukhin) · 2026-05-05T20:23:36Z

I took a look at the CPU implementation since they share the same tests (and spec!).

the shared CPU empty-set helper can dereference a missing or malformed axes input before any validation. reduction_ops.cc:875, check_and_reduce_empty_set_input reads ctx->Input<Tensor>(1), then immediately uses Shape()[0] and Data<int64_t>() without checking for nullptr or that the tensor is 1-D.
The normal path does that validation in reduction_ops.cc:948, but CommonReduce1Loop reaches the empty-set helper first in reduction_ops.cc:966. For any reduction using this shared path, an empty input combined with omitted optional axes or a scalar axes tensor is a real crash/OOB risk, not just a bad-status case.
the same empty-set helper bypasses the real axes semantics, so CPU gets empty-input behavior wrong for noop_with_empty_axes and for normalized axes. In reduction_ops.cc:895, an empty axes list is treated as reduce-all when building the output shape. But the intended noop_with_empty_axes handling only happens later in reduction_ops.cc:976, which the empty-set shortcut never reaches. That means empty tensor plus empty or missing axes plus noop_with_empty_axes=1 will produce a reduced output shape instead of the no-op result. The helper also compares raw axes values directly against dimension indices instead of calling HandleNegativeAxis or validating bounds, so negative axes and out-of-range axes on empty inputs will produce the wrong shape or silently ignore invalid input. Current noop tests are non-empty only, for example reduction_ops_test.cc:3921, so this gap is not covered.
ReduceMax and ReduceMin register bool CPU kernels, but their empty-set implementation is still explicitly unimplemented. Bool registrations are present in reduction_ops.cc:201 and reduction_ops.cc:246, yet the empty-set fill path for those aggregators calls ORT_NOT_IMPLEMENTED in reduction_ops.h:388 and reduction_ops.h:599. So the registered kernel surface and the empty-input behavior disagree. Existing bool tests such as reduction_ops_test.cc:998 and reduction_ops_test.cc:1012 are non-empty only, which is why this has likely gone unnoticed.

We really need to improve test coverage here.

Remove the overly strict assertion that rejected reducing along a zero-sized dimension even with explicit axes. Reducing axis K of shape {N, 0} with keepdims=false produces shape {N} filled with the identity value (0 for sum), which is mathematically valid. The CPU implementation already handles this case via check_and_reduce_empty_set_input(). The CUDA path now allows PrepareForReduce to succeed, and ReduceComputeCore (line 369) already handles input_count==0 correctly. This fixes CUDA inference for models with dynamic KV cache where past_sequence_length=0 during prefill (e.g., Gemma4 via ORT GenAI). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Justin Chu <justinchu@microsoft.com>

ONNX spec defines identity values for reductions over empty sets: - ReduceSum, ReduceMean, ReduceL1, ReduceL2: 0 - ReduceProd: 1 - ReduceMin: +type_max (plus infinity) - ReduceMax: -type_max (minus infinity / lowest) When reducing over an empty dimension (e.g. {1,0} axis=1), the output shape is {1} — not empty. The previous code asserted output size == 0 and returned without writing any values. Fix: 1. ReduceComputeCore: fill with 0 when input_count==0 (handles Sum, Mean, L1, L2, SumSquare). 2. Macro path: dispatch identity value by cudnn_reduce_op (handles Min, Max, Prod). 3. Remove ORT_ENFORCE on default-axes reduce with zero dimensions. Add CPU and CUDA tests for ReduceSum, ReduceMax, ReduceMin, ReduceProd with empty tensor inputs. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Justin Chu <justinchu@microsoft.com>

Address review feedback: replace CUDA-only and CPU-only test blocks with unified tests that run on all available execution providers via test.Run(). Coverage now includes 19 tests across 5 reduction ops: - ReduceSum: explicit axis, keepdims, multi-dim, default axes, {0,5} shape - ReduceMax: explicit axis, keepdims, multi-dim, default axes - ReduceMin: explicit axis, keepdims, multi-dim - ReduceProd: explicit axis, keepdims, multi-dim, default axes - ReduceMean: explicit axis, multi-dim (NaN for 0/0) Each op tested with shapes {1,0}, {2,0,3}, {0,5} and both keepdims=true/false. Identity values per ONNX spec: Sum→0, Prod→1, Min→+inf, Max→-inf, Mean→NaN. Signed-off-by: Justin Chu <justinchu@microsoft.com>

CUDA fixes: - PrepareForReduce: default-axes path now correctly maps all dims to 1 (was mapping 0-dims to 0, producing wrong output shape for empty tensors with keepdims=false) - ReduceComputeCore: dispatch identity fill by cudnn_reduce_op type. ReduceMean now fills NaN (was incorrectly filling 0). Prod fills 1, Min fills +inf, Max fills -inf. CPU fixes: - check_and_reduce_empty_set_input: add nullptr check for axes tensor, validate 1-D shape, normalize negative axes via HandleNegativeAxis, use std::set for O(1) axis lookup. - CommonReduce1Loop: check noop_with_empty_axes BEFORE empty-set reduction. Previously an empty tensor with noop_with_empty_axes=1 and no axes would incorrectly reduce to identity instead of preserving the input shape. - Bool ReduceMax/ReduceMin fill_for_empty_set: implement ONNX spec identity values (ReduceMax→false, ReduceMin→true) instead of ORT_NOT_IMPLEMENTED. Tests (26 unified tests covering CPU + CUDA): - 5 reduction ops × explicit axis / keepdims / multi-dim / default axes - Negative axes + empty tensor - noop_with_empty_axes=1 + empty tensor (shape preservation) - Default axes + keepdims + empty tensor - Bool ReduceMax/ReduceMin + empty tensor Signed-off-by: Justin Chu <justinchu@microsoft.com>

Justin Chu (justinchuby) · 2026-05-06T16:41:59Z

Thanks for the thorough analysis! All three issues are addressed in the latest push (657664e):

1. CPU empty-set helper nullptr/validation — check_and_reduce_empty_set_input now:

Checks axes_tensor != nullptr before dereferencing
Validates NumDimensions() == 1 (rejects scalar axes)
Normalizes negative axes via HandleNegativeAxis
Uses std::set for O(1) axis lookup instead of linear scan

2. noop_with_empty_axes ordering — CommonReduce1Loop now resolves effective axes and checks noop_with_empty_axes before the empty-set reduction. An empty tensor with noop_with_empty_axes=1 and no axes now correctly preserves the input shape (e.g. {1,0} stays {1,0}) instead of reducing to scalar.

3. Bool ReduceMax/ReduceMin identity — Replaced ORT_NOT_IMPLEMENTED() with ONNX spec identity values:

ReduceMax<bool> → false
ReduceMin<bool> → true

Tests added (26 unified CPU+CUDA tests):

All 5 reduction ops × explicit axis / keepdims / multi-dim / default axes
Negative axes + empty tensor
noop_with_empty_axes=1 + empty tensor (shape preservation)
Default axes + keepdims + empty tensor
Bool ReduceMax/ReduceMin + empty tensor

Also fixed the CUDA side: ReduceComputeCore now dispatches identity by op type (Mean→NaN was incorrectly filling 0), and PrepareForReduce default-axes path maps all dims to 1 (was mapping 0-dims to 0).

CUDA fixes: - Use CudaT (mapped type) + ToCudaType<T>::FromFloat() for identity values instead of T(0)/T(1) which are ambiguous for MLFloat16. - Use MutableDataRaw() instead of MutableData<T>() for memcpy since we're writing CudaT, not T. - Simplify ReduceMean empty set to memset 0 (consistent with CPU). Test fixes: - ReduceMean empty tensor: expect 0 not NaN (ORT fills 0 for undefined mean of empty set, matching CPU behavior). - Bool ReduceMax/Min: restrict to CPU EP via ConfigEp() since CUDA EP doesn't register bool reduction kernels. Signed-off-by: Justin Chu <justinchu@microsoft.com>

Signed-off-by: Justin Chu <justinchu@microsoft.com>

ONNX ReduceMax/ReduceMin schema at opset 18 does not include bool as a valid input type. OpTester graph validation fails before the kernel executes. The fill_for_empty_set implementation is correct but cannot be tested through OpTester. Signed-off-by: Justin Chu <justinchu@microsoft.com>

Dmitri Smirnov (yuslepukhin)

The original CUDA bug that motivated the PR does appear fixed: the updated CUDA empty-input branches now materialize identities instead of returning success with an uninitialized non-empty output buffer, and the default-axes reduce-all shape handling is no longer rejecting valid empty full reductions.

Dmitri Smirnov (yuslepukhin) · 2026-05-06T20:43:23Z

+// --- Bool reduction + empty tensor ---
+
+// NOTE: Bool ReduceMax/ReduceMin empty tensor tests omitted.
+// ORT registers bool CPU kernels for ReduceMax/ReduceMin, but the ONNX


The new bool empty-tensor tests are written against unsupported opsets, so they are not valid coverage for the claimed fix. The tests use ReduceMax and ReduceMin at opset 18 in reduction_ops_test.cc:6593 and reduction_ops_test.cc:6603. In ORT, bool kernels for these ops are only registered starting at opset 20 in reduction_ops.cc:208 and reduction_ops.cc:253, which matches the ONNX spec change. These tests should move to opset 20 or be removed from the graph-level suite.

Fixed. Bool empty-tensor tests have been removed — the comment at the end of the test file explains why: ORT registers bool CPU kernels for ReduceMax/ReduceMin at opset 20, but OpTester's ONNX schema validation rejects bool inputs before execution reaches the kernel. The fill_for_empty_set implementation is correct (Max→false, Min→true) but cannot be tested through the standard graph-based test path.

Dmitri Smirnov (yuslepukhin) · 2026-05-06T20:49:38Z

+    if (output_count > 0) {
+      // For types that don't support std::numeric_limits natively (MLFloat16,
+      // BFloat16), use float intermediary and convert via CudaT.
+      if (cudnn_reduce_op == CUDNN_REDUCE_TENSOR_AVG) {


CPU and CUDA now disagree on empty-set ReduceMean, and the new generic tests encode only the CUDA behavior. The CPU empty-input fast path still routes through reduction_ops.cc:924, ReduceAggregatorMean still inherits ReduceAggregatorSum at reduction_ops.h:307, and the inherited empty-set fill is still 0 at reduction_ops.h:224.

CUDA was changed to emit quiet_NaN() for empty ReduceMean at reduction_ops.cc:380, reduction_ops.cc:381, reduction_ops.cc:816, and reduction_ops.cc:817, and the new test expects NaN at reduction_ops_test.cc:6512. ONNX says empty-set ReduceMean is undefined, so choosing NaN is defensible, but this PR has not made ORT internally consistent. As written, the “tests for all EPs” claim is not satisfactorily addressed since it should fail on CPU.

This needs to be consistent and the test should cover it. Please, consider using zeros from historical perspective.

Addressed in 7e84448. CPU and CUDA are now consistent — both fill ReduceMean empty-set with 0:

CPU: ReduceAggregatorMean inherits ReduceAggregatorSum::fill_for_empty_set (fills 0) at reduction_ops.h:307/224

CUDA: CUDNN_REDUCE_TENSOR_AVG branch uses cudaMemsetAsync(..., 0, ...) at reduction_ops.cc:382

Macro path: else branch at line 821 uses cudaMemsetAsync(..., 0, ...) with comment 'Sum, SumSquare, Mean, L1, L2, Amax: identity is 0'

Test ReduceMean_EmptyTensor_ExplicitAxis expects 0.0f and runs on all EPs. Added ReduceMean_EmptyTensor_DefaultAxes for reduce-all coverage. CUDA comment updated to explain the consistency rationale.

Per reviewer feedback: std::set allocates even when empty (MSVC) and is overkill for rank < 8. Use std::find with linear scan instead, matching the existing style in the same file (reduction_ops.cc:665, :680). Remove unused #include <set>. Signed-off-by: Justin Chu <justinchu@microsoft.com>

…ests - Fix ReduceMean comment: clarify 0 fill is for CPU/CUDA consistency (ReduceAggregatorMean inherits ReduceAggregatorSum::fill_for_empty_set) - Add ReduceMin_EmptyTensor_DefaultAxes test (reduce-all, scalar output) - Add ReduceMin_EmptyTensor_DefaultAxes_KeepDims test - Add ReduceMean_EmptyTensor_DefaultAxes test - Update test header comment: ReduceMean → 0 (not NaN) - All tests are EP-independent (no #ifdef USE_CUDA guards) Signed-off-by: Justin Chu <justinchu@microsoft.com>

WebGPU, QNN, TensorRT, CoreML, DML, DNNL, MIGraphX, and OpenVINO do not support empty tensor reductions. Add kEmptyTensorExcludedEps exclusion set to all EmptyTensor tests so they only run on CPU and CUDA (which have proper empty-set identity handling). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Justin Chu <justinchu@microsoft.com>

Only exclude QNN, TensorRT, and WebGPU — the three EPs that actually failed CI on empty tensor tests. Remove over-broad exclusions for CoreML, CudaNHWC, DML, DNNL, MIGraphX, and OpenVINO that were copied from the older test_empty_set helper without evidence of failure. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Justin Chu <justinchu@microsoft.com>

Add proper empty-set identity values to WebGPU reduction shader generation. Previously, the empty input path used loop_header_/ loop_footer_ which referenced undefined variables (first_element for Max/Min) or divided by zero (ReduceMean). Changes: - Add reduce_op_empty_identity_map with WGSL expressions for each reduction op's ONNX identity value (Max→-inf, Min→+inf, Sum→0, Prod→1, Mean→0, etc.) - Modify ReduceNaiveProgram::GenerateShaderCode to directly output the identity value for empty inputs instead of running broken loop header/footer code - Enable allow_empty_input for ReduceMax and ReduceMin WebGPU kernels (was false, causing CheckInput to reject empty tensors) - Remove WebGPU from empty tensor test exclusion lists Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Justin Chu <justinchu@microsoft.com>

github-actions

You can commit the suggested changes from lintrunner.

- Fix clang-format: remove extra trailing spaces before comments - Fix WebGPU shader: replace bitcast<f32>(0xff800000u) with float literals (-3.4028234663852886e+38f / +3.4028234663852886e+38f). The bitcast approach failed because output_value_t may be f16, causing a WGSL type mismatch in the shader module. - Add DML and NNAPI to kEmptyTensorExcludedEps: these EPs don't support empty tensor reductions. Signed-off-by: Justin Chu <justinchu@microsoft.com>

github-actions

You can commit the suggested changes from lintrunner.

github-actions · 2026-05-07T22:12:26Z

+    {ReduceOpType::Max, "output_value_t(-3.4028234663852886e+38f)"},           // -FLT_MAX ≈ -inf
+    {ReduceOpType::Min, "output_value_t(3.4028234663852886e+38f)"},            // FLT_MAX ≈ +inf


Suggested change

{ReduceOpType::Max, "output_value_t(-3.4028234663852886e+38f)"}, // -FLT_MAX ≈ -inf

{ReduceOpType::Min, "output_value_t(3.4028234663852886e+38f)"}, // FLT_MAX ≈ +inf

{ReduceOpType::Max, "output_value_t(-3.4028234663852886e+38f)"}, // -FLT_MAX ≈ -inf

{ReduceOpType::Min, "output_value_t(3.4028234663852886e+38f)"}, // FLT_MAX ≈ +inf

github-actions · 2026-05-07T22:12:26Z

+    {ReduceOpType::Sum, "output_value_t(0)"},
+    {ReduceOpType::Prod, "output_value_t(1)"},
+    {ReduceOpType::SumSquare, "output_value_t(0)"},
+    {ReduceOpType::LogSumExp, "output_value_t(-3.4028234663852886e+38f)"},     // log(0) ≈ -inf


Suggested change

{ReduceOpType::LogSumExp, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf

{ReduceOpType::LogSumExp, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf

github-actions · 2026-05-07T22:12:26Z

+    {ReduceOpType::LogSumExp, "output_value_t(-3.4028234663852886e+38f)"},     // log(0) ≈ -inf
+    {ReduceOpType::L1, "output_value_t(0)"},
+    {ReduceOpType::L2, "output_value_t(0)"},
+    {ReduceOpType::LogSum, "output_value_t(-3.4028234663852886e+38f)"},        // log(0) ≈ -inf


Suggested change

{ReduceOpType::LogSum, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf

{ReduceOpType::LogSum, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf

…inity Replace float literal approximations (-3.4028e+38f) with proper infinity via division-by-zero: output_value_t(-1.0) / output_value_t(0.0) for -inf, output_value_t(1.0) / output_value_t(0.0) for +inf. The float literal approach produced -FLT_MAX instead of -inf, causing test failures where the expected output is -inf (ReduceMax/Min/LogSum empty tensor tests). Division by zero is well-defined in IEEE 754 and WGSL (produces ±inf). Using output_value_t() ensures the expression works for both f32 and f16 output types. Signed-off-by: Justin Chu <justinchu@microsoft.com>

github-actions

You can commit the suggested changes from lintrunner.

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Some WGSL shader validators reject constant division by zero (output_value_t(-1.0) / output_value_t(0.0)), causing 'Invalid ShaderModule' errors on Windows CI. Replace with bitcast from IEEE 754 bit patterns: bitcast<f32>(0xFF800000u) = -inf bitcast<f32>(0x7F800000u) = +inf bitcast<f32>() is always valid WGSL (u32→f32 reinterpretation). output_value_t() then converts f32 inf to f16 inf when the output type is half-precision (f16 has an infinity representation). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Justin Chu <justinchu@microsoft.com>

Justin Chu (justinchuby) · 2026-05-26T16:43:09Z

Dmitri Smirnov (@yuslepukhin) Guenther Schmuelling (@guschmue) Do you have suggestions on a fix for webgpu kernel? I tried a few things to produce inf but they don't seem to be liked by the webgpu runtime. Thanks!

Justin Chu (justinchuby) mentioned this pull request May 4, 2026

CUDA EP: CUBLAS misaligned address crash during prefill with Gemma4 model microsoft/onnxruntime-genai#2120

Closed

Justin Chu (justinchuby) requested a review from Copilot May 4, 2026 23:45

Copilot started reviewing on behalf of Justin Chu (justinchuby) May 4, 2026 23:55 View session

Xavier Dupré (xadupre) reviewed May 5, 2026

View reviewed changes

Comment thread onnxruntime/core/providers/cuda/reduction/reduction_ops.cc

Xavier Dupré (xadupre) reviewed May 5, 2026

View reviewed changes

Dmitri Smirnov (yuslepukhin) reviewed May 5, 2026

View reviewed changes

Comment thread onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc Outdated

Dmitri Smirnov (yuslepukhin) suggested changes May 5, 2026

View reviewed changes

Dmitri Smirnov (yuslepukhin) changed the title ~~Fix CUDA ReduceSum crash on empty tensors with explicit axes~~ Fix CUDA ReduceSum erroring out on empty tensors with explicit axes May 5, 2026

Justin Chu (justinchuby) mentioned this pull request May 5, 2026

Gemma4 ONNX + CUDA support: all PRs tracker onnxruntime/mobius#257

Open

37 tasks

Justin Chu (justinchuby) force-pushed the fix-reducesum-empty-tensor branch from 005f7ad to 01a3b14 Compare May 5, 2026 19:57

github-actions Bot reviewed May 5, 2026

View reviewed changes

Justin Chu (justinchuby) force-pushed the fix-reducesum-empty-tensor branch from 01a3b14 to 2300137 Compare May 5, 2026 20:10

Dmitri Smirnov (yuslepukhin) reviewed May 5, 2026

View reviewed changes

Comment thread onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc Outdated

Justin Chu (justinchuby) force-pushed the fix-reducesum-empty-tensor branch 2 times, most recently from 06d3755 to ca84cd8 Compare May 5, 2026 20:14

Justin Chu (justinchuby) and others added 4 commits May 6, 2026 00:28

Justin Chu (justinchuby) force-pushed the fix-reducesum-empty-tensor branch from b152f38 to 657664e Compare May 6, 2026 00:29

Justin Chu (justinchuby) marked this pull request as draft May 6, 2026 18:14

Justin Chu (justinchuby) added 2 commits May 6, 2026 18:25

Fix clang-format: trailing whitespace alignment in macro

1cc286e

Signed-off-by: Justin Chu <justinchu@microsoft.com>

Dmitri Smirnov (yuslepukhin) reviewed May 6, 2026

View reviewed changes

Justin Chu (justinchuby) and others added 5 commits May 6, 2026 21:18

github-actions Bot reviewed May 7, 2026

View reviewed changes

Comment thread onnxruntime/core/providers/webgpu/reduction/reduction_ops.cc Outdated

github-actions Bot reviewed May 7, 2026

View reviewed changes

github-actions Bot reviewed May 8, 2026

View reviewed changes

Comment thread onnxruntime/core/providers/webgpu/reduction/reduction_ops.cc Outdated

Comment thread onnxruntime/core/providers/webgpu/reduction/reduction_ops.cc Outdated

Justin Chu (justinchuby) and others added 2 commits May 7, 2026 17:14

Apply suggestions from code review

90c5b9f

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Justin Chu (justinchuby) closed this Jun 9, 2026

		{ReduceOpType::Max, "output_value_t(-3.4028234663852886e+38f)"}, // -FLT_MAX ≈ -inf
		{ReduceOpType::Min, "output_value_t(3.4028234663852886e+38f)"}, // FLT_MAX ≈ +inf

	{ReduceOpType::LogSumExp, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf
	{ReduceOpType::LogSumExp, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf

	{ReduceOpType::LogSum, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf
	{ReduceOpType::LogSum, "output_value_t(-3.4028234663852886e+38f)"}, // log(0) ≈ -inf

Uh oh!

Conversation

Justin Chu (justinchuby) commented May 4, 2026

Description

Motivation

Changes

Testing

Related issues

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Dmitri Smirnov (yuslepukhin) commented May 5, 2026

Uh oh!

Dmitri Smirnov (yuslepukhin) commented May 5, 2026

Uh oh!

Uh oh!

Dmitri Smirnov (yuslepukhin) commented May 5, 2026

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Dmitri Smirnov (yuslepukhin) commented May 5, 2026

Uh oh!

Justin Chu (justinchuby) commented May 6, 2026

Uh oh!

Dmitri Smirnov (yuslepukhin) left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions Bot May 7, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot May 7, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot May 7, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Justin Chu (justinchuby) commented May 26, 2026

Uh oh!

Reviewers