Fix seg fault when splitting non-fusible MLIR modules

[justinrosner]

Motivation

When compiling fused gpu::mlir_op modules (gemm/conv + pointwise), MIGraphX checks whether the module can remain fused for a given rocMLIR tuning config via is_module_fusible(). If the config is not fusible, most notably when splitKFactor > 1 + certain output fusions, the compiler falls back to splitting the module at a boundary computed by find_final_split(). When trying to do this split MIGraphX could dereference an invalid iterator that would lead to a seg fault.

Technical Details

find_final_split() could dereference an invalid iterator when:

• get_output_path() returned fewer than two instructions
• std::adjacent_find() found no valid split edge and returned end().

This is separate from #5144, which validates the pointwise submodule after a successful split. This PR fixes the crash that occurred during split-point selection.

Changelog Category

Add a CHANGELOG.md entry for any option other than Not Applicable

• • Added: New functionality.
• • Changed: Changes to existing functionality.
• • Removed: Functionality or support that has been removed. (Compared to a previous release)
• • Optimized: Component performance that has been optimized or improved.
• • Resolved Issues: Known issues from a previous version that have been resolved.
• • Not Applicable: This PR is not to be included in the changelog.

Follow the LLVM AI Tool Use Policy for contributions using AI.

[Copilot]

Pull request overview

Fixes a GPU compilation crash when rocMLIR reports a fused gpu::mlir_op module as non-fusible and MIGraphX falls back to splitting the fused module. The change hardens split-point selection (find_final_split) to avoid invalid iterator dereferences, and adds a GPU regression test that exercises the previously-crashing split paths.

Changes:

• Add guards in find_final_split() to handle short output paths and the “no split edge found” case without dereferencing end().
• Add a new GPU test that compiles fused MLIR modules under a non-fusible split-K tuning configuration to ensure no split-time crash.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/targets/gpu/jit/mlir.cpp	Adds defensive handling in split-point selection to prevent iterator invalidation/dereference crashes during MLIR module splitting.
test/gpu/mlir_split_fusion.cpp	Adds GPU regression coverage for split-K non-fusible MLIR module splitting paths that previously could segfault.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[gh-app-migraphx-bot-pr-write]

Test	Batch	New Rate (358629)	Old Rate (3a503c)*	Diff	Status
torchvision-resnet50	64	3,299.19	3,264.92	1.05%	✅
torchvision-resnet50_fp16	64	7,885.34	7,548.67	4.46%	✅
torchvision-densenet121	32	2,490.19	2,483.99	0.25%	✅
torchvision-densenet121_fp16	32	5,014.97	5,004.24	0.21%	✅
torchvision-inceptionv3	32	2,059.43	2,058.51	0.04%	✅
torchvision-inceptionv3_fp16	32	4,449.48	4,416.99	0.74%	✅
cadene-inceptionv4	16	818.04	820.61	-0.31%	✅
cadene-resnext64x4	16	784.85	782.78	0.26%	✅
slim-mobilenet	64	8,376.76	8,386.36	-0.11%	✅
slim-nasnetalarge	64	229.30	228.86	0.19%	✅
slim-resnet50v2	64	3,229.71	3,180.91	1.53%	✅
bert-mrpc-onnx	8	1,170.14	1,168.84	0.11%	✅
bert-mrpc-tf	1	496.30	498.63	-0.47%	✅
pytorch-examples-wlang-gru	1	386.53	473.35	-18.34%	🔴
pytorch-examples-wlang-lstm	1	547.82	384.83	42.35%	🔆
torchvision-resnet50_1	1	1,053.97	1,046.63	0.70%	✅
cadene-dpn92_1	1	448.70	437.32	2.60%	✅
cadene-resnext101_1	1	363.70	365.89	-0.60%	✅
onnx-taau-downsample	1	842.47	844.09	-0.19%	✅
dlrm-criteoterabyte	1	32.27	32.42	-0.48%	✅
dlrm-criteoterabyte_fp16	1	51.61	51.80	-0.37%	✅
agentmodel	1	14,863.35	9,209.12	61.40%	🔆
unet_fp16	2	58.36	58.80	-0.74%	✅
resnet50v1_fp16	1	1,432.62	1,366.11	4.87%	✅
resnet50v1_int8	1	1,773.26	1,883.96	-5.88%	🔴
bert_base_cased_fp16	64	1,099.18	1,098.16	0.09%	✅
bert_large_uncased_fp16	32	347.28	345.59	0.49%	✅
bert_large_fp16	1	207.04	206.59	0.22%	✅
distilgpt2_fp16	16	2,099.20	2,092.89	0.30%	✅
yolov5s	1	553.86	558.33	-0.80%	✅
tinyllama	1	45.83	45.83	-0.01%	✅
vicuna-fastchat	1	44.25	44.20	0.11%	✅
whisper-tiny-encoder	1	413.42	411.87	0.37%	✅
whisper-tiny-decoder	1	410.72	408.48	0.55%	✅
llama2_7b	1	20.86	20.84	0.13%	✅
qwen1.5-7b	1	23.64	23.58	0.26%	✅
phi3-3.8b	1	28.33	26.72	6.03%	🔆
llama3-8b	1	22.73	21.80	4.25%	✅
whisper-large-encoder	1	10.17	10.18	-0.08%	✅
whisper-large-decoder	1	106.96	105.30	1.57%	✅
mistral-7b	1	23.81	23.78	0.16%	✅
FLUX.1-schnell	1	792.31	755.22	4.91%	✅

Regressions detected 🔴

* No develop baseline was found for this PR's branch point; compared against the latest available develop run instead.

[gh-app-migraphx-bot-pr-write]

Test	Status	Result
bert-mrpc-onnx	✅	PASSED: MIGraphX meets tolerance
bert-mrpc-tf	❌	ERROR - check error output traceback Traceback (most recent call last): File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 377, in main() File "/src/AMDMIGraphX/tools/accuracy/accuracy_checker.py", line 313, in main import tensorflow as tf File "/usr/local/lib/python3.12/dist-packages/tensorflow/init.py", line 40, in from tensorflow.python import pywrap_tensorflow as _pywrap_tensorflow # pylint: disable=unused-import ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/lib/python3.12/dist-packages/tensorflow/python/pywrap_tensorflow.py", line 37, in self_check.preload_check() File "/usr/local/lib/python3.12/dist-packages/tensorflow/python/platform/self_check.py", line 63, in preload_check from tensorflow.python.platform import _pywrap_cpu_feature_guard ImportError: libnuma.so.1: cannot open shared object file: No such file or directory
pytorch-examples-wlang-gru	🔴	FAILED: MIGraphX is not within tolerance - check verbose output
pytorch-examples-wlang-lstm	🔴	FAILED: MIGraphX is not within tolerance - check verbose output
dlrm-criteoterabyte	✅	PASSED: MIGraphX meets tolerance
agentmodel	✅	PASSED: MIGraphX meets tolerance
unet	✅	PASSED: MIGraphX meets tolerance
resnet50v1	✅	PASSED: MIGraphX meets tolerance
bert_base_cased_fp16	✅	PASSED: MIGraphX meets tolerance
bert_large_uncased_fp16	🔴	FAILED: MIGraphX is not within tolerance - check verbose output
bert_large	✅	PASSED: MIGraphX meets tolerance
yolov5s	✅	PASSED: MIGraphX meets tolerance
tinyllama	✅	PASSED: MIGraphX meets tolerance
vicuna-fastchat	✅	PASSED: MIGraphX meets tolerance
whisper-tiny-encoder	✅	PASSED: MIGraphX meets tolerance
whisper-tiny-decoder	✅	PASSED: MIGraphX meets tolerance
distilgpt2_fp16	🔴	FAILED: MIGraphX is not within tolerance - check verbose output
llama2_7b	✅	PASSED: MIGraphX meets tolerance
qwen1.5-7b	✅	PASSED: MIGraphX meets tolerance
phi3-3.8b	✅	PASSED: MIGraphX meets tolerance
llama3-8b	✅	PASSED: MIGraphX meets tolerance
whisper-large-encoder	✅	PASSED: MIGraphX meets tolerance
whisper-large-decoder	✅	PASSED: MIGraphX meets tolerance
mistral-7b	✅	PASSED: MIGraphX meets tolerance
FLUX.1-schnell	✅	PASSED: MIGraphX meets tolerance

[pfultz2]

I dont know about the unit tests, but can you add the verify tests that you had in the other PR? That got us coverage as well for this.

[justinrosner]

I dont know about the unit tests, but can you add the verify tests that you had in the other PR? That got us coverage as well for this.

I assume you are talking about the tests from #5064? I'm not sure if they would actually give us coverage for the fixes in this PR. Those tests cover picking the right split point when one exists, whereas this PR covers the cases where there isn't one.

[pfultz2]

I dont know about the unit tests, but can you add the verify tests that you had in the other PR? That got us coverage as well for this.

I assume you are talking about the tests from #5064? I'm not sure if they would actually give us coverage for the fixes in this PR. Those tests cover picking the right split point when one exists, whereas this PR covers the cases where there isn't one.

OK but I dont see how this would happen as it should always have a gemm or convolution. And the unit tests show cases that are pointwise-only which would never happen.

[justinrosner]

I dont know about the unit tests, but can you add the verify tests that you had in the other PR? That got us coverage as well for this.

I assume you are talking about the tests from #5064? I'm not sure if they would actually give us coverage for the fixes in this PR. Those tests cover picking the right split point when one exists, whereas this PR covers the cases where there isn't one.

OK but I dont see how this would happen as it should always have a gemm or convolution. And the unit tests show cases that are pointwise-only which would never happen.

The unit tests all have a GEMM (they all call the make_dot_graph helper).