Copy AMDMLSS PDB file and install AMDMLSS with MIGraphX

[apwojcik]

Motivation

The PR applies only to Windows.

The AMDMLSS DLL is mandatory if MIGRAPHX_USE_AMDMLSS is ON. Because MIGraphX requires the library, the MIGraphX installation script must copy the required files to ensure the application's experience is out of the box.

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.

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

Test	Batch	New Rate (c2267a)	Old Rate (f54ca3)*	Diff	Status
torchvision-resnet50	64	2,376.99	3,151.54	-24.58%	🔴
torchvision-resnet50_fp16	64	3,478.90	6,672.57	-47.86%	🔴
torchvision-densenet121	32	1,042.74	2,669.98	-60.95%	🔴
torchvision-densenet121_fp16	32	4,423.95	4,553.52	-2.85%	✅
torchvision-inceptionv3	32	1,800.11	1,794.69	0.30%	✅
torchvision-inceptionv3_fp16	32	2,822.04	2,710.69	4.11%	✅
cadene-inceptionv4	16	811.62	224.10	262.18%	🔆
cadene-resnext64x4	16	784.23	409.63	91.45%	🔆
slim-mobilenet	64	8,211.14	8,288.90	-0.94%	✅
slim-nasnetalarge	64	229.36	229.24	0.05%	✅
slim-resnet50v2	64	3,321.36	3,329.55	-0.25%	✅
bert-mrpc-onnx	8	1,171.30	1,170.13	0.10%	✅
bert-mrpc-tf	1	480.82	482.99	-0.45%	✅
pytorch-examples-wlang-gru	1	327.75	335.86	-2.41%	✅
pytorch-examples-wlang-lstm	1	440.52	464.03	-5.07%	🔴
torchvision-resnet50_1	1	750.60	771.72	-2.74%	✅
cadene-dpn92_1	1	439.16	451.76	-2.79%	✅
cadene-resnext101_1	1	362.84	363.79	-0.26%	✅
onnx-taau-downsample	1	402.24	400.76	0.37%	✅
dlrm-criteoterabyte	1	31.32	32.67	-4.14%	✅
dlrm-criteoterabyte_fp16	1	43.83	52.58	-16.65%	🔴
agentmodel	1	9,754.48	9,679.70	0.77%	✅
unet_fp16	2	57.35	57.23	0.20%	✅
resnet50v1_fp16	1	585.93	972.36	-39.74%	🔴
resnet50v1_int8	1	919.08	960.27	-4.29%	✅
bert_base_cased_fp16	64	1,103.74	1,102.92	0.07%	✅
bert_large_uncased_fp16	32	347.50	347.52	-0.01%	✅
bert_large_fp16	1	208.12	205.28	1.38%	✅
distilgpt2_fp16	16	1,830.63	2,094.45	-12.60%	🔴
yolov5s	1	563.54	568.20	-0.82%	✅
tinyllama	1	37.25	46.01	-19.04%	🔴
vicuna-fastchat	1	43.75	44.04	-0.65%	✅
whisper-tiny-encoder	1	420.10	419.49	0.15%	✅
whisper-tiny-decoder	1	417.16	420.06	-0.69%	✅
llama2_7b	1	11.63	20.46	-43.16%	🔴
qwen1.5-7b	1	10.52	23.67	-55.54%	🔴
phi3-3.8b	1	5.30	26.98	-80.35%	🔴
llama3-8b	1	11.53	21.83	-47.18%	🔴
whisper-large-encoder	1	10.31	10.32	-0.01%	✅
whisper-large-decoder	1	104.03	107.05	-2.82%	✅
mistral-7b	1	23.58	23.86	-1.16%	✅
FLUX.1-schnell	1	38.56	761.61	-94.94%	🔴

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/python3.10/dist-packages/tensorflow/init.py", line 38, in from tensorflow.python.tools import module_util as _module_util File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/init.py", line 36, in from tensorflow.python import pywrap_tensorflow as _pywrap_tensorflow File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/pywrap_tensorflow.py", line 26, in self_check.preload_check() File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/platform/self_check.py", line 63, in preload_check from tensorflow.python.platform import _pywrap_cpu_feature_guard ImportError: libamdhip64.so.6: cannot open shared object file: No such file or directory
pytorch-examples-wlang-gru	✅	PASSED: MIGraphX meets tolerance
pytorch-examples-wlang-lstm	✅	PASSED: MIGraphX meets tolerance
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	✅	PASSED: MIGraphX meets tolerance
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	❌	ERROR - check error output traceback 2026-06-19 06:45:34.909264 [WARN] [/data/src/onnx/onnx_parser.cpp:282] Model has unbound symbolic dimension(s): batch_size, encoder_sequence_length, feature_size. These default to 1 and may cause unexpected behavior. Try setting --dim-param @<name> <value> or --input-dim @<input> <dims> if program compilation fails. 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 224, in main model = migraphx.parse_onnx(model_name, default_dim_value=batch) RuntimeError: /data/src/include/migraphx/op/convolution.hpp:113: normalize_compute_shape: CONVOLUTION: mismatched channel numbers: input channels (1) != weights channels (80) * group (1)
whisper-large-decoder	✅	PASSED: MIGraphX meets tolerance
mistral-7b	✅	PASSED: MIGraphX meets tolerance
FLUX.1-schnell	✅	PASSED: MIGraphX meets tolerance

[pfultz2]

The installation of AMDMLSS should be handle by the AMDMLSS component not by migraphx.

[apwojcik]

The installation of AMDMLSS should be handle by the AMDMLSS component not by migraphx.

The AMDMLSS is installing its components and binary files (as you can see by the imported targets). However, since AMDMLSS is now a hard dependency for MIGraphX and only MIGraphX knows the proper location of the installed AMDMLSS, we want MIGraphX to install it as well in the MIGraphX installation location chosen. Only for Windows, Linux is intact. Otherwise, for components that depend on MIGraphX, their build configurations are becoming increasingly difficult to parameterize, requiring developers to understand which component was used to build MIGraphX and which is the correct one to copy. The GPU EP is a complex stack of ML components that work together, and MIGraphX is one of them. It is very easy, even for the experienced engineer, to confuse things. We want to mitigate potential errors caused by using the wrong binaries and improve the development by avoiding such issues altogether.

[pfultz2]

The AMDMLSS is installing its components and binary files (as you can see by the imported targets). However, since AMDMLSS is now a hard dependency for MIGraphX and only MIGraphX knows the proper location of the installed AMDMLSS, we want MIGraphX to install it as well in the MIGraphX installation location chosen.

AMDMLSS should be installed into the same location as migraphx. It seems like there is a problem with your packaging/release scripts.

Otherwise, for components that depend on MIGraphX, their build configurations are becoming increasingly difficult to parameterize, requiring developers to understand which component was used to build MIGraphX and which is the correct one to copy.

This should all be managed by the packaging/release scripts. As I understand it, users arent building migraphx they are just installing the binaries, so the packaging script should be updated to install and include the binaries from AMDMLSS.

[apwojcik]

The AMDMLSS is installing its components and binary files (as you can see by the imported targets). However, since AMDMLSS is now a hard dependency for MIGraphX and only MIGraphX knows the proper location of the installed AMDMLSS, we want MIGraphX to install it as well in the MIGraphX installation location chosen.

AMDMLSS should be installed into the same location as migraphx. It seems like there is a problem with your packaging/release scripts.

Otherwise, for components that depend on MIGraphX, their build configurations are becoming increasingly difficult to parameterize, requiring developers to understand which component was used to build MIGraphX and which is the correct one to copy.

This should all be managed by the packaging/release scripts. As I understand it, users arent building migraphx they are just installing the binaries, so the packaging script should be updated to install and include the binaries from AMDMLSS.

It is not a concern of the packaging/release pipeline. The concern is the development of GPU EP. As I mentioned, it's a stack of components, and MIGraphX is one of them. On Windows, we chose to install each component separately rather than all components in the same directory. Such an approach is helping us debug individual components more easily. That means each component in the stack installs its dependencies so it can run out of the box from the installation directory. We are building a few variants of MIGraphX (in particular with and without AMDMLSS). For some components, we are even changing the default names of the resulting binaries depending on the provided build configuration. For consistency with the rest of the components and to ease debugging and configuration of components that depend on MIGraphX, we want MIGraphX to install some of its dependencies rather than only copy them into a build directory.

[pfultz2]

On Windows, we chose to install each component separately rather than all components in the same directory. Such an approach is helping us debug individual components more easily.

Sure, you just need to set the path to each component you want to use. Cget does this as well but it copies each file to a shared prefix and updates the path for the shared prefix only.

Of course having migraphx install another component breaks this scenario and you cannot easily swap in and out a different version of components for debugging purposes.

That means each component in the stack installs its dependencies so it can run out of the box from the installation directory.

This doesnt make sense with what you just said.

we want MIGraphX to install some of its dependencies rather than only copy them into a build directory.

I didnt realize we were copying it to the build directory. That should be removed. The dev env should be setup to find its dependencies correctly.