[GPU] fix dGPU func testcases in smoke_ScaledAttnDynamic4D_GPU and smoke_MatMulCompressedWeights_extra_multiply

[yuanxion]

Details

fixes 2 Intel dGPU functional testcases:

1. smoke_ScaledAttnDynamic4D_GPU/ScaledAttnLayerGPUTest.CompareWithRefs
2. smoke_MatMulCompressedWeights_extra_multiply/MatmulWeightsDecompression.Inference

Description of the issue

Symptom

1. smoke_ScaledAttnDynamic4D_GPU/ScaledAttnLayerGPUTest.CompareWithRefs failed on Intel dGPU when the graph carried a scalar or rank-1 placeholder input in the attention-mask slot. The SDPA OCL path treated that placeholder as a real runtime attention mask, which changed kernel configuration and input binding unexpectedly.
2. smoke_MatMulCompressedWeights_extra_multiply/MatmulWeightsDecompression.Inference failed on dGPU for the group_size=2 + extra_multiply=1 + param_weights=1 configuration. The runtime failure surfaced during GPU program build / implementation selection.

Root cause

1. The SDPA GPU path only checked whether an attention-mask input slot existed, but did not distinguish a real runtime mask tensor from a scalar / rank-1 placeholder. As a result, placeholder inputs were propagated through the real attention-mask path and affected JIT constants, kernel arguments, and execution logic.
2. ConvertMatMulToFullyConnected converted a MatMul with parameter-based compressed weights into FullyConnected even when the decompressed weights still had non-trivial batch dimensions. That conversion is unsafe for this pattern: after extra multiply and transpose, the weights remained effectively per-batch / 3D, but were still fed into the FC path.

How to fix it

1. Add a shared helper to identify whether the attention-mask input is a real runtime mask. Scalar and rank-1 placeholders are now excluded from the runtime attention-mask path, and the same logic is reused by the SDPA OCL implementations.
2. Detect parameter-based compressed weights explicitly in the matcher result and block MatMul -> FullyConnected conversion when the aligned weight shape still contains non-1 batch dimensions. In that case, keep the original MatMul path instead of forcing FC.

The code and line that caused this issue

1. smoke_ScaledAttnDynamic4D_GPU
   

   openvino/src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa/sdpa_gen_opt.cpp

   Lines 126 to 127 in 6c7a684

   	if (i == attn_mask_idx && desc->attn_mask_val.has_value())
   	continue;

2. smoke_MatMulCompressedWeights_extra_multiply
   

   openvino/src/plugins/intel_gpu/src/plugin/transformations/convert_matmul_to_fc.cpp

   Lines 122 to 124 in 6c7a684

   	} else if (!is_compressed_weight || !supports_immad) {
   	return std::make_tuple(false, std::move(shape_a_aligned), std::move(shape_b_aligned));
   	}

Reproduction step and snapshot

1. smoke_ScaledAttnDynamic4D_GPU
   ./ov_gpu_func_tests --device_suffix=1 --gtest_filter='smoke_ScaledAttnDynamic4D_GPU/ScaledAttnLayerGPUTest.CompareWithRefs/netPRC=f16_IS=[?.5.?.128]_[?.5.?.128]_[?.5.?.32]_[?.1.?.?]_TS=(2.5.100.128)_(2.5.1.128)_(2.5.387.128)_(2.5.100.128)_(2.5.1.128)_(2.5.387.128)_(2.5.100.32)_(2.5.1.32)_(2.5.387.32)_(1.1.100.100)_(1.1.1.1)_(2.1.387.387)_is_causal=0_has_attn=0_is_attn_const=1_has_scale=1_is_scale_const=1_with_transpose0_has_sink=0_'
2. smoke_MatMulCompressedWeights_extra_multiply
   ./ov_gpu_func_tests --device_suffix=1 --gtest_filter='smoke_MatMulCompressedWeights_extra_multiply/MatmulWeightsDecompression.Inference/data_shape=[]_[1.4.16]__weights_shape=[16,32]_group_size=2_weights_precision=u8_activations_precision=f32_transpose_weights=0_decompression_subtract=0_reshape_on_decompression=0_extra_multiply=1_per_tensor_zp=0_param_weights=1_dyn_quan_group_size=0'

Problematic graph

N/A

Checklist

• Is it a proper fix? (not a workaround)
• Did you include test case for this fix, if necessary?
• Did you review existing test that can be extended to cover this scenario? Which test did you review?

Tickets:

• CVS-182520

AI Assistance:

• AI assistance used: yes
• Used it to reproduce the failing dGPU function tests, inspect the dumped GPU graphs, find the root cause, and also add testcases.

[Copilot]

Pull request overview

Fixes two Intel dGPU functional test failures by (1) treating scalar/rank-1 SDPA attention-mask inputs as placeholders (not real runtime masks) in OCL SDPA implementations, and (2) preventing unsafe MatMul→FullyConnected conversion for parameter-based compressed weights that still carry non-trivial batch dimensions.

Changes:

• Add shared SDPA helper to detect whether the attention-mask input is a real runtime mask, and reuse it across SDPA OCL implementations.
• Update MatMul→FC transformation to block conversion for parameter-based compressed weights with non-1 aligned batch dimensions.
• Add/extend unit tests for SDPA placeholder-mask behavior and MatMul→FC “no-convert” scenario.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/plugins/intel_gpu/tests/unit/transformations/convert_matmul_to_fc_test.cpp	Adds a regression unit test ensuring MatMul→FC does not occur for parameter-based compressed weights with per-batch dimensions.
src/plugins/intel_gpu/tests/unit/test_cases/sdpa_gpu_test.cpp	Adds a unit test asserting scalar placeholder mask behaves like “no runtime mask” in SDPA OCL path.
src/plugins/intel_gpu/src/plugin/transformations/convert_matmul_to_fc.cpp	Extends aligned-shape check to detect parameter-based compressed weights and block unsafe FC conversion.
src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa/sdpa_utils.hpp	Introduces has_runtime_attn_mask_input() helper to filter out scalar/1D placeholder masks.
src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa/sdpa_ref.cpp	Uses the helper to set JIT constants and to skip binding placeholder mask inputs.
src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa/sdpa_gen_opt.cpp	Uses the helper for JIT config and argument binding in the optimized SDPA generator.
src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa/sdpa_gen_micro.cpp	Uses the helper to control mask-related JIT constants/args for the micro-kernel generator.

[yuanxion]

This PR is split into 2 PRs:
#35437 for smoke_ScaledAttnDynamic4D_GPU
#35442 for smoke_MatMulCompressedWeights_extra_multiply

No needed anymore, close it.