[Transformation] move the place of ConvertWeightCompressedConv1x1ToMatmul pattern #33103
Conversation
github-actions
bot
added
category: GPU
|
This is just a code movement work for #32224 |
bopeng1234
force-pushed
the
conv1x1tomatmul_move
branch
from
0712b55 to
841f925
Compare
...ormations/src/transformations/op_conversions/convert_weight_compressed_conv1x1_to_matmul.cpp
Outdated
Show resolved
Hide resolved
mdvoretc-intel
left a comment
mdvoretc-intel
left a comment
There was a problem hiding this comment.
Confirmed that Phi-Silica use case is not affected, LGTM.
…d its test from intel_gpu plugin to common
bopeng1234
force-pushed
the
conv1x1tomatmul_move
branch
from
5c06c94 to
50c547c
Compare
| #include "openvino/opsets/opset7_decl.hpp" | ||
| #include "transformations/rt_info/decompression.hpp" | ||
|
|
||
| using namespace ov; |
There was a problem hiding this comment.
The test coverage seems limited. Currently there are only two test cases: one positive (1x1 conv transforms to matmul) and one negative (3x3 conv does not transform). Are there plans to add tests for:
asymmetric quantization path (with zero-point subtraction)
bias addition path
Parameter weights instead of Constant
Reshape on output instead of Transpose
Convert on output
There was a problem hiding this comment.
Added more test cases.
| public: | ||
| OPENVINO_GRAPH_REWRITE_RTTI("ConvertWeightCompressedConv1x1ToMatmul"); | ||
| ConvertWeightCompressedConv1x1ToMatmul() { | ||
| add_matcher<ov::pass::ConvertWeightCompressedConv1x1ToMatmulMatcher>(); |
There was a problem hiding this comment.
the transformation contains only one matcher transformation, do we really need to include in GraphRewrite manually?
There was a problem hiding this comment.
removed the graph rewriter, only keep one match pass ConvertWeightCompressedConv1x1ToMatmul
...tions/include/transformations/op_conversions/convert_weight_compressed_conv1x1_to_matmul.hpp
Outdated
Show resolved
Hide resolved
5 participants
Move the ConvertWeightCompressedConv1x1ToMatmul pattern and its test from the intel_gpu plugin to the common transformation folder. The purpose is to reuse it on both the CPU and GPU sides.
Tickets: