[AIRADSW-567] Fix int8 models qlinearconv

[urpetkov-amd]

Motivation

QLinearConv models that have a bias and use per-tensor (scalar) weight quantization fail to parse. The parser throws while building the bias dequantizelinear:

.../migraphx/check_shapes.hpp:220: same_dims: dequantizelinear: Dimensions do not match

This blocks loading common quantized models such as resnet50_int8. The goal of this PR is to make QLinearConv with bias parse correctly for both per-tensor and per-axis (per-channel) weight scales, and to add a regression test for the previously-uncovered per-tensor case.

This is a regression introduced by #4521. Before that change, the int32 bias was added to the convolution result without being dequantized (numerically incorrect, but it parsed). #4521 added the correct bias dequantization (bias_scale = x_scale * w_scale), but only handled the per-axis weight-scale shape, so per-tensor models began throwing at parse time.

Technical Details

When QLinearConv has a bias, the parser dequantizes the int32 bias using a scale computed as x_scale * w_scale:

auto bcast_scale_x = info.add_instruction(
    migraphx::make_op("multibroadcast", {{"out_lens", in_scale_w->get_shape().lens()}}),
    in_scale_x);

auto bias_scale =
    info.add_instruction(migraphx::make_op("mul"), bcast_scale_x, in_scale_w);

auto dquant_bias =
    info.add_instruction(migraphx::make_op("dequantizelinear"), in_b, bias_scale);

bias_scale inherits the shape of the weight scale (in_scale_w):

• Per-axis weight scale → in_scale_w is {out_channels}, so bias_scale is {out_channels} and matches the bias {out_channels}. Works.
• Per-tensor weight scale → in_scale_w is a scalar {1}, so bias_scale is {1} while the bias is {out_channels}.

Fix

Broadcast bias_scale to the bias shape before the bias dequantizelinear:

auto bias_scale =
    info.add_instruction(migraphx::make_op("mul"), bcast_scale_x, in_scale_w);

// dequantizelinear requires the scale to match the bias dims. Broadcast handles
// per-tensor weight scale (scalar); it is a no-op for per-axis weight scale.
auto bcast_bias_scale = info.add_instruction(
    migraphx::make_op("multibroadcast", {{"out_lens", in_b->get_shape().lens()}}),
    bias_scale);

auto dquant_bias = info.add_instruction(
    migraphx::make_op("dequantizelinear"), in_b, bcast_bias_scale);

This is a minimal, single-instruction change confined to the if(args.size() > 8) (bias-present) branch of parse_qlinearconv:

• Per-tensor: {1} → {out_channels} — fixes the crash.
• Per-axis: {out_channels} → {out_channels} — an identity broadcast, so numerics are unchanged and the existing qlinearconv_perchannel_weightbias_test still passes.

Related

• Regression introduced by Correct QLinearMatMul broadcasting and QLinearConv bias quantization #4521 (commit cf56c399b).

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.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files

@@           Coverage Diff            @@
##           develop    #4969   +/-   ##
========================================
  Coverage    92.73%   92.73%           
========================================
  Files          592      592           
  Lines        31289    31289           
========================================
  Hits         29015    29015           
  Misses        2274     2274           

Files with missing lines	Coverage Δ
src/onnx/parse_qlinearconv.cpp	74.39% <100.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[CharlieL7]

The fix looks correct. Please add an ONNX parse test for the per tensor parsing case in test/onnx/parse/qlinearconv_test.cpp

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

Test	Batch	New Rate (91abf8)	Old Rate (241f7a)	Diff	Status
torchvision-resnet50	64	3,165.98	3,158.05	0.25%	✅
torchvision-resnet50_fp16	64	5,991.67	6,678.10	-10.28%	🔴
torchvision-densenet121	32	2,670.25	2,705.33	-1.30%	✅
torchvision-densenet121_fp16	32	4,555.74	4,480.31	1.68%	✅
torchvision-inceptionv3	32	680.96	1,772.37	-61.58%	🔴
torchvision-inceptionv3_fp16	32	2,731.77	2,613.31	4.53%	✅
cadene-inceptionv4	16	823.83	444.89	85.17%	🔆
cadene-resnext64x4	16	785.04	412.50	90.32%	🔆
slim-mobilenet	64	6,555.02	8,419.17	-22.14%	🔴
slim-nasnetalarge	64	228.89	nan	nan	❌
slim-resnet50v2	64	3,330.18	1,641.82	102.83%	🔆
bert-mrpc-onnx	8	1,171.47	1,165.20	0.54%	✅
bert-mrpc-tf	1	496.54	492.10	0.90%	✅
pytorch-examples-wlang-gru	1	329.56	334.96	-1.61%	✅
pytorch-examples-wlang-lstm	1	451.34	469.54	-3.88%	✅
torchvision-resnet50_1	1	765.74	760.64	0.67%	✅
cadene-dpn92_1	1	452.94	457.65	-1.03%	✅
cadene-resnext101_1	1	363.74	363.68	0.01%	✅
onnx-taau-downsample	1	401.40	401.93	-0.13%	✅
dlrm-criteoterabyte	1	32.70	12.36	164.54%	🔆
dlrm-criteoterabyte_fp16	1	52.61	29.66	77.39%	🔆
agentmodel	1	9,751.34	10,523.13	-7.33%	🔴
unet_fp16	2	57.26	57.23	0.05%	✅
resnet50v1_fp16	1	975.37	959.87	1.62%	✅
resnet50v1_int8	1	935.31	942.92	-0.81%	✅
bert_base_cased_fp16	64	1,102.62	1,092.47	0.93%	✅
bert_large_uncased_fp16	32	347.60	345.19	0.70%	✅
bert_large_fp16	1	204.74	203.23	0.74%	✅
distilgpt2_fp16	16	2,096.30	2,096.77	-0.02%	✅
yolov5s	1	571.79	564.74	1.25%	✅
tinyllama	1	46.01	46.00	0.02%	✅
vicuna-fastchat	1	43.99	44.04	-0.11%	✅
whisper-tiny-encoder	1	419.66	419.55	0.03%	✅
whisper-tiny-decoder	1	414.20	414.85	-0.16%	✅
llama2_7b	1	20.47	20.42	0.23%	✅
qwen1.5-7b	1	23.67	21.75	8.82%	🔆
phi3-3.8b	1	26.84	26.81	0.12%	✅
llama3-8b	1	21.83	21.83	0.00%	✅
whisper-large-encoder	1	10.32	6.50	58.78%	🔆
whisper-large-decoder	1	106.71	106.80	-0.08%	✅
mistral-7b	1	23.86	23.85	0.05%	✅
FLUX.1-schnell	1	754.56	763.88	-1.22%	✅

Regressions detected 🔴

[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 2026-06-17 09:11:18.762790 [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

[urpetkov-amd]

@CharlieL7

I've added it. Thanks.