[GPU] Recognize parameters as valid inputs for compressed weights #32276
Conversation
github-actions
bot
added
category: GPU
mdvoretc-intel
force-pushed
the
param_quant_weight
branch
from
522f237 to
6a6a649
Compare
|
build_jenkins |
mdvoretc-intel
force-pushed
the
param_quant_weight
branch
from
6a6a649 to
16d03b4
Compare
mdvoretc-intel
requested review from
CuriousPanCake
and removed request for
a team
|
build_jenkins |
mdvoretc-intel
force-pushed
the
param_quant_weight
branch
2 times, most recently
from
b82b902 to
476f80f
Compare
mklimenk
left a comment
mklimenk
left a comment
There was a problem hiding this comment.
Two branches of the if (pattern_map.count(weights_const_m)) { condition share a lot of similarities, please consider refactoring it to avoid code duplication
src/plugins/intel_gpu/src/plugin/transformations/convert_fc_to_compressed.cpp
Outdated
Show resolved
Hide resolved
This change enables use of quantized LoRA weights, passed as parameters during execution, to be recognized by the transformaions that produce FullyConnectedCompressed nodes for QGEMM execution.
The test previously expected the transformation to fail due to the use of input2 as a weight. The new logic allows use of parameters as weights, so the test has been adjusted to expect a successful transformation.
mdvoretc-intel
force-pushed
the
param_quant_weight
branch
from
e9e889a to
1c45696
Compare
mklimenk
left a comment
mklimenk
left a comment
There was a problem hiding this comment.
Looks much cleaner now, thanks!
|
@CuriousPanCake please review. |
github-actions
bot
removed
the
category: transformations
|
@CuriousPanCake please review. |
|
@Lyamin-Roman please take a look |
This has required addressing a previously not covered use case where a reshape follows float conversion, as well as adding logic for reshaping the param weight if the reshape covers the entire decompression pattern.
Tests added. |
|
@Lyamin-Roman please review. |
src/plugins/intel_gpu/src/plugin/transformations/convert_fc_to_compressed.cpp
Outdated
Show resolved
Hide resolved
| } | ||
| } | ||
|
|
||
| TEST_F(TransformationTestsF, ConvertFCToCompressed11) { |
There was a problem hiding this comment.
[random spot] Please add functional accuracy tests
It looks like you can extend an existing test with additional parameters
src/plugins/intel_gpu/tests/functional/subgraph_tests/dynamic/matmul_weights_decompression.cpp
There was a problem hiding this comment.
Attempted this. The tests have a singular parameter for input precision which prevents proper creation of weight parameters in compressed types.
There was a problem hiding this comment.
Tests added with a configure_model() override to ensure weight parameters are provided in appropriate type. Tests currently fail due to the lack of u4 transposition support, WIP to fix.
src/plugins/intel_gpu/src/plugin/transformations/compressed_weights_pattern.hpp
Outdated
Show resolved
Hide resolved
mryzhov
left a comment
mryzhov
left a comment
There was a problem hiding this comment.
Please address the requested changes
|
build_jenkins |
|
build_jenkins |
1 similar comment
|
build_jenkins |
| void configure_model() override { | ||
| ov::preprocess::PrePostProcessor p(function); | ||
| { | ||
| auto& params = function->get_parameters(); |
There was a problem hiding this comment.
Please remove this unused variable, it causes build to fail.
@ZackyLake
|
This PR will be closed in a week because of 2 weeks of no activity. |
mdvoretc-intel
force-pushed
the
param_quant_weight
branch
from
e2a50ee to
b510b91
Compare
|
@mryzhov All requested changes have been addressed. |
9 participants
Details:
Description of the issue:
At present, the FC_COMPRESSED_WEIGHT_PATTERN macro contains a pattern for dequantization of a constant integer weight. This pattern is used to recognize and fold cases where fused weight dequantization can be used, replacing them with FullyConnectedCompressed nodes. Due to expecting a constant weight input, this pattern fails to recognize quantized LoRA weights, which are provided as parameters:
With the changes in this patch, these weights can be recognized, and the transformations can proceed and produce nodes that would then leverage oneDNN fused QGEMM for execution:
Tickets: