Repository navigation
Conversation
Hybrid graphs append bias and an NPU packed weight after the microsoft MatMulNBits inputs, and fuse attention and MLP as GQO and SSMLP. Apply zero-points whenever they are present so the scoped model compiles on GPU.
pfultz2
reviewed
Oct 19, 2026
| } | ||
|
|
||
| // Lower A @ dequant(B) [+ bias]. | ||
| inline instruction_ref add_matmul_nbits(onnx_parser::node_info& info, |
Collaborator
There was a problem hiding this comment.
This should really be an op builder.
Contributor
There was a problem hiding this comment.
🟡 Changes recommended
Input-layout errors, dropped attention inputs, FP16 normalization clipping, and missing fused-operator regression tests block approval.
6 open issues
What changed in this PR
Extends MIGraphX ONNX parsing to lower Ryzen AI hybrid-model operators for GPU compilation.
Changes:
- Shares MatMulNBits lowering and adds optional bias handling.
- Adds fused GQO and SSMLP parsers.
- Adds a Ryzen AI MatMulNBits fixture with parse and reference tests.
| File | Description |
|---|---|
| test/onnx/verify/matmulnbits_tests.cpp | Verifies biased MatMulNBits results. |
| test/onnx/parse/matmulnbits_tests.cpp | Checks the expected lowered graph. |
| test/onnx/matmulnbits_ryzenai_bias_test.onnx | Adds the regression fixture. |
| test/onnx/gen_onnx.py | Generates the Ryzen AI fixture. |
| src/onnx/parse_ssmlp.cpp | Lowers fused normalization and SwiGLU. |
| src/onnx/parse_matmulnbits.cpp | Delegates lowering and handles optional inputs. |
| src/onnx/parse_group_query_attention.cpp | Adds fused attention/output projection parsing. |
| src/onnx/include/migraphx/onnx/matmulnbits.hpp | Provides shared quantized matrix-multiplication lowering. |
Review effort: Balanced
Comment on lines
+348
to
+349
| // First 9 inputs match microsoft GroupQueryAttention layout. | ||
| std::vector<instruction_ref> gqa_args(args.begin(), args.begin() + 9); |
| const int block_size = get_i("o_proj_block_size"); | ||
|
|
||
| // o_proj weights begin at input 12 after three optional empty slots (9-11). | ||
| const std::size_t w = args.size() >= 17 ? 12 : args.size() - 5; |
Comment on lines
+58
to
+60
| std::optional<instruction_ref> bias; | ||
| if(args.size() > 4 and not args[4]->is_undefined()) | ||
| bias = args[4]; |
Comment on lines
+53
to
+58
| rms = info.add_instruction(make_op("convert", {{"target_type", x_dtype}}), rms); | ||
| epsilon = | ||
| (x_dtype == migraphx::shape::half_type and std::abs(epsilon) < 1e-7) ? 1e-7 : epsilon; | ||
| auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_dtype}, {epsilon}}); | ||
| rms = info.add_common_op("add", rms, eps); | ||
| auto rrms = info.add_instruction(make_op("rsqrt"), rms); |
| // First 9 inputs match microsoft GroupQueryAttention layout. | ||
| std::vector<instruction_ref> gqa_args(args.begin(), args.begin() + 9); | ||
| parse_group_query_attention gqa; | ||
| auto gqa_outs = gqa.parse(opd, parser, info, gqa_args); |
Comment on lines
+135
to
+136
| auto [out, skip_sum] = skip_simplified_ln(info, down, residual, args[12], epsilon); | ||
| return {out, skip_sum}; |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
com.ryzenai::MatMulNBitswhen the hybrid schema adds optional bias and an NPU-only packed weight after the microsoft inputs. Zero-points are applied whenever they are present, not only when the node has exactly four inputs.com.ryzenai::GQO(GroupQueryAttention + o_proj MatMulNBits) andcom.ryzenai::SSMLP(SkipSimplifiedLayerNorm + SwiGLU MatMulNBits) so hybridmodel_jitparses and GPU-compiles.matmulnbits_ryzenai_bias_testparse and reference verification.Test plan
migraphx-driver read --onnx model_jit.onnxmigraphx-driver compile --gpu --onnx model_jit.onnx(gfx1152, ROCm 10.1)test_onnx_parse/test_onnx_verifymatmulnbits_*includingmatmulnbits_ryzenai_bias_testSWDEV-604598