[pytorchbot]

Fixes found while exporting models to ExecuTorch backends.

Fixes

• exir: constant_prop_pass keeps tensors that in-place ops write, also through views or in no_grad/autocast blocks. to_backend keeps the output node last. to_edge lowers torch._check_tensor_all and torch.sym_ite. GroupBasedPartitioner no longer makes partitions that depend on each other, unless caller node groups already do. Tensor constants get static shapes; deduplicated constants keep their data.
• AOTI (AOTInductor) and CUDA: the missing-kernel check catches scatter and index_put fallbacks. detach lowers, also inside torch.cond. Strided outputs are copied. Transposed convolutions (1-D, 2-D, no dilation) and torch._check_tensor_all lower.
• XNNPACK: the quantizer skips bfloat16 tensors, and cat respects scoped configs. ELU is delegated when XNNPACK can run it.
• Qualcomm: autocast regions are inlined correctly. Unquantized bfloat16 nodes, data-dependent shapes and HTP partitions without runtime inputs stay on the CPU. Fixes for stacks, 0-d selects, RMSNorm and int64 constants.
• OpenVINO: model data stays alive while OpenVINO 2026.4 reads it. Partitions are converted with their input types.
• Core ML: bfloat16 ops, and ops above rank 5 outside static layout chains, stay on the CPU.
• MLX, Arm, Cortex-M, Vulkan, NXP, portable: dtype checks for MLX tensor lists, one constant slot per weight, slice updates on a constant base; TOSA keeps unused delegate inputs for VGF; MatmulToBmmPass takes a constant operand; Vulkan constants above 4 dims; non-strict programs in the NXP quantizer; cummin.out.

This pick also carries the changes of #24316 (node names stay unique across the QNN and Samsung passes) and #24429 (on HTP, an op QNN keeps whole, such as linear, is decomposed when its constant-only partition would stay on the CPU; without it, to_edge_transform_and_lower fails with was not decomposed or delegated). It must merge after both of them land on main.

Behavior changes

• A CUDA .pte loads only the weights it was exported with. A .ptd with other weights, for example after fine-tuning, needs a newly exported .pte.
• CUDA graphs turn off when outputs do not match the ETensor layout. If only the capture call sees such a layout, that call returns an error.
• GroupBasedPartitioner is about 2x slower on large graphs, and some graphs get a different number of partitions. to_backend keeps the original order of top-level nodes.
• XNNPACK: weight-only quantized linears, and 8da4w or 8da8w linears (8-bit dynamic activations, 4-bit or 8-bit weights) lowered with to_backend, fail at export instead of giving wrong results. Delegated ELU flushes subnormal values to zero. Cats outside a scoped quantizer config are no longer quantized.
• Ethos-U raises an error for a delegate with unused inputs instead of returning wrong values.
• On HTP, an op QNN keeps whole is decomposed when its constant-only partition would stay on the CPU. Ops in delegated partitions stay whole.

We would like to cherry-pick this PR to release/1.6 after it merges. It needs #23915 there first.

Testing

Each fix has a test that fails without it. Run on an A100 (CUDA, Vulkan), an x86 CPU (OpenVINO, QNN SDK), macOS (Core ML, MLX), the TOSA reference model and VGF emulation (Arm), and the eIQ Neutron SDK (NXP). Not run: Metal, ROCm, Windows CUDA.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24298

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 42 Pending, 1 Unrelated Failure, 1 Unclassified Failure

As of commit 679565d with merge base 4db778e ():

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

• MLX / test-mlx-llm-offgraph (google/gemma-4-E2B-it, gemma4-e2b, gemma4, macos-15-xlarge) / test-mlx-llm-offgraph-gemma4-e2b (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  RuntimeError: Command bash /Users/runner/work/_temp/exec_script failed with exit code 1

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / nxp-mcuxpresso-test / linux-job (gh) (trunk failure)
  RuntimeError: Command bash /__w/_temp/exec_script failed with exit code 1

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[shoumikhin]

@claude /pr-review

[claude]

Claude finished @shoumikhin's task in 53m 0s —— View job

---

Approve. Nothing blocks this cherry-pick, but CUDA weight keys now include a SHA-256 of the weights, which stops weights-only .ptd updates, so 1.6 needs a release note and a clearer load error. The other findings can be follow-ups on main.

PR Review: #24298

Summary

This cherry-pick brings #24222, with #24316 and #24429, to release/1.6. It fixes crashes, wrong outputs and lowering failures in exir, the portable kernels (a new cummin.out), and the AOTI/CUDA, Qualcomm, Samsung, XNNPACK, Core ML, OpenVINO, Arm, MLX, Cortex-M, Vulkan and NXP backends.

• The diff was taken against ba053ba, the release/1.6 commit merged into HEAD, because the shallow checkout shares no merge base with origin/release/1.6. It matches the PR's 89 files and +3173/-216.
• The Arm, Cortex-M, MLX, NXP and Vulkan changes and exir/backend/test/test_group_partitioner.py were only read at the diff level.
• Problems found:
  • one compatibility break in CUDA weight loading;
  • a CUDA graph capture step that fails a valid call;
  • three possible lowering regressions (low confidence, not run);
  • a quantizer that skips bf16 without a message;
  • four performance follow-ups.

Backward Compatibility

• backends/cuda/cuda_weight_collector.py:604-616: the key of every CUDA FQN weight now ends in the SHA-256 of its bytes, and the .pte looks the weight up by that key.
  • A .ptd with other weights for the same .pte (fine-tuned, re-quantized, LoRA-merged) therefore no longer loads.
  • Yet docs/source/ptd-file-format.md:144 lists "async updates between program and data" as a use of .ptd files.
  • Loading fails with CUDA FQN storage '<key>:<sha256>' is missing from named data (NotFound, backends/cuda/runtime/cuda_weight_cache.cpp:457-460). The message doesn't say that the .ptd came from a different export.
  • On a release branch this needs a release note, and the error should name the likely cause.
  • If weights-only updates should keep working for CUDA, they need an explicit opt-in, for example a compile spec that keeps plain FQN keys.

Code Quality

• backends/cuda/runtime/cuda_backend.cpp:1080-1092: suppose an output is not laid out like its ETensor during the capture step but was during warmup. Then execute returns Error::NotSupported for a valid call.
  • Warmup handles the same case by switching to Disabled and copying the output (lines 1184-1192).
  • The capture step can't copy, because the captured work hasn't run yet.
  • Re-running the call eagerly with CUDA graphs disabled would avoid the failure. run() has already taken slim_inputs, so the retry has to rebuild them.
• backends/qualcomm/_passes/decompose_wrap_with_autocast.py:60-67 (low confidence; by reasoning, not run): make_fx(submod, tracing_mode="fake") traces after dispatch.
  • Inside an enabled region, CompositeImplicitAutograd ops such as aten.linear therefore decompose into t + addmm/mm, bypassing QNN's skip-decomp table.
  • The retraced nodes also lose nn_module_stack and source_fn_stack.
  • test_decompose_wrap_with_autocast_nested_regions (backends/qualcomm/tests/test_passes.py:154) checks only numerics.
  • Try pre_dispatch=True, or assert that the region's linear stays whole.
• backends/qualcomm/_passes/remove_0d_tensor.py:28-54 (low confidence; not run): a 0-D select_copy of a one-element tensor is now removed only when every user is add, sub, mul or div.
  • Before this PR, every 0-D select_copy was removed.
  • One that feeds where, maximum, minimum, pow or a comparison is just as safe to remove, but now stays 0-D. The pass docstring says QNN can't take 0-D tensors, so a model that lowered before can now fail.
  • Widen the allowlist to the other broadcasting elementwise ops, or add a test that lowers such a model end to end.
• backends/aoti/passes/replace_view_copy_with_view.py:49-76 (low confidence; not run): slice_copy, select_copy and split_copy inside torch.cond branches now become views. _has_blocking_user doesn't treat the branch output as blocking.
  • A branch can therefore return a view of its own input. torch.cond rejects that aliasing when it traces, and Inductor's Conditional may not handle it.
  • DetachInCondModule (backends/cuda/tests/test_cuda_export.py:345) is the only branch test. Its detach_copy becomes alias_copy, which is a copy, so no test has a branch that returns a view.
  • Treat submodule outputs as blocking, or limit the rewrite in submodules to detach_copy. Either way, add a test with a slice in a branch.
• backends/xnnpack/quantizer/xnnpack_quantizer.py:404: touches_bfloat16 leaves bf16 nodes unquantized without telling the user. A bf16 model therefore comes out with no quantized ops and no message. Log a warning once.

Performance

• backends/apple/coreml/partition/coreml_partitioner.py:455-477: one rejected node drops its op from the preserved list everywhere. For example, one unsupported SDPA call decomposes every SDPA call.
  • A per-node filter would keep the supported calls whole. It could re-run op_support.is_node_supported(None, node) (line 461), like QNN's filter_fn.
• backends/cuda/cuda_backend.py:485-548 (_conv_transpose_as_conv): inserting zeros costs about stride^n times the MACs and input memory of a direct transposed conv.
  • That is 4x for a stride-2 2-D upsampler and 8x for a stride-8 1-D vocoder stage.
  • A phase (sub-pixel) decomposition avoids the zeros.
• backends/cuda/runtime/utils.h:220-296: on every call with a strided output, the span is copied to the host, rearranged on the CPU and copied back, with a stream sync. This happens even when the destination is on the GPU. A strided copy on the device would avoid the round trip.
• backends/cuda/runtime/cuda_backend.cpp:95-98: output_is_raw_copyable requires a row-major contiguous output.
  • So a channels_last-dense output whose ETensor is also channels_last turns CUDA graphs off, although its bytes copy as-is.
  • The check would accept it if it tested whether the output is dense in the ETensor's dim order, with strides computed from the output's sizes.

Recommendation

Approve: none of the findings blocks the cherry-pick.

• Before 1.6 ships, the .ptd key change needs a release note and a clearer load error.
• The capture-step failure, the low-confidence regressions and the performance items can be follow-ups on main.

Progress

• Gather context. The shallow checkout has no merge base with origin/release/1.6, so the diff is against ba053ba (89 files, +3173/-216).
• Run three independent reviews (exir/XNNPACK, QNN/Samsung/Core ML/OpenVINO, CUDA/AOTI/Arm/others). Together they reported 33 candidate findings.
• Merge duplicate findings. That left about 25 distinct ones.
• Fact-check findings. 10 remain.
• Post review.