[TF Quant Models] Prepares GroupConvolution weights #25641
Conversation
github-actions
bot
added
the
category: transformations
| auto reshaped_input = reshape_node->clone_with_new_inputs( | ||
| {limit_input, | ||
| ov::op::v0::Constant::create(element::i64, {new_limit_shape.size()}, new_limit_shape)})); | ||
| ov::op::v0::Constant::create(element::i64, {new_limit_shape.size()}, new_limit_shape)}); | ||
| if (auto constant = ov::util::get_constant_from_source(reshaped_input)) { | ||
| reshaped_input = constant; | ||
| } | ||
| renewed_inputs.push_back(reshaped_input); |
There was a problem hiding this comment.
Minor: can we use ov::op::util::clone_try_fold instead to make the code a bit shorter? The same at L87 and at L60 in pull_transpose_through_fq.cpp
There was a problem hiding this comment.
We can do that, sure. The only difference is that clone_try_fold only works for single node folding and get_constant_from_source checks all the necessary no_folding rt info and could fold a chain of nodes. For now, I won't change the code in the PR, however, in case more feedback arises -- I'll clean it up.
| if (pattern_map.count(convert_p)) { | ||
| const auto& convert_node = pattern_map.at(convert_p).get_node_shared_ptr(); | ||
| convert_node->input(0).replace_source_output(reshaped_input); | ||
| convert_node->validate_and_infer_types(); |
There was a problem hiding this comment.
I am just wondering: why should we manually call node validation here? It is usually automatically done by Validate pass, isn't it?
Upd. It seems like it is needed to keep a valid graph state during one GraphRewrite pass. Am I right?
There was a problem hiding this comment.
Yes, you ARE right. Since we are in a single Graph Rewrite -- it is safer to keep the graph in validated state, so that all the matchers could assume rank and shape is accurate
|
Performance validation is triggered for this particular PR with IR cache generation and calibration. No code is expected to be pushed in the branch. Merging is postponed until PR is proven to be non-regressive. |
|
This PR will be closed in a week because of 2 weeks of no activity. |
|
This PR was closed because it has been stalled for 2 week with no activity. |
src/common/transformations/tests/common_optimizations/fq_reshape_fusion.cpp
Show resolved
Hide resolved
|
This PR will be closed in a week because of 2 weeks of no activity. |
|
This PR was closed because it has been stalled for 2 week with no activity. |
|
I could not get the results for this PR. I propose merging the change to master and seeing if regressions arise in the regular validation. |
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
…25641) ### Details: - *using register_new_node in the matches to allow for seamless matcher application for the FQ node* - *added optional Convert node, so that pattern would work on different quantized weights forms*  ### Tickets: - *CVS-39818*
Details:
Tickets: