[NPUW] Add subgraph composition dump (.sg) support #33876
+163
−14
Conversation
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
github-actions
bot
added
category: NPU
dylanneve1
force-pushed
the
EISW-196509
branch
3 times, most recently
from
e065500 to
c221f5e
Compare
dmatveev
reviewed
Jan 30, 2026
Contributor
dmatveev
left a comment
dmatveev
left a comment
There was a problem hiding this comment.
Looks good for the first PR! Here's some comments though.
- Is the PR description up to date? I belive there's no mentioned option in place now?
- I believe you're familiar now with the attention mechanisms and the way we handle it, would your schema reflect such structure with e.g. a PYRAMID hint? How would it look like?
| if (subgraph_counts.find(name) == subgraph_counts.end()) { | ||
| subgraph_order.push_back(name); | ||
| } | ||
| subgraph_counts[name]++; |
Contributor
There was a problem hiding this comment.
So I believe you're counting by name(s) which you synthesize yourself based on real_id.. Are you sure these names match the subgraph names as we dump them? What I see here you reconstruct the same format but if it changes during the dump procedure well'get a discrepancy
Don't we have this name specified as the subgraph's ov::Model's friendly_name? I don't remember.
Probably it makes sense to:
- Introduce a tiny helper method here to format you a subgraph name - the way it'll be dumped - and use it in both places
- Simplify the counting method - you can just traverse through orderedSubgraphs and collect their
.replaced_by()to a set. This set will give you a list of uniqe subgraphs.. so then you iterate over it again and poll their names
dylanneve1
force-pushed
the
EISW-196509
branch
3 times, most recently
from
ab4a46e to
84bb68f
Compare
dylanneve1
force-pushed
the
EISW-196509
branch
3 times, most recently
from
047b116 to
29f2f2a
Compare
Introduces the NPUW_DUMP_SUBS_COMPOSITION configuration option. When enabled, generates a JSON file (.sg) containing the pipeline structure, including unique subgraphs and their repetitions. Also adds NPUW_DUMP_SUBS_DIR option to specify output directory.
Remove dedicated NPUW_DUMP_SUBS_COMPOSITION option. The .sg file is now automatically generated whenever subgraphs are dumped via NPUW_DUMP_SUBS. Remove unnecessary "./" prefix from subgraph paths in the JSON output.
Refactor path construction in dump_subgraph_model and dump_subgraph_composition to use ov::util::join_path instead of manual string with dir_prefix.
- Reorder includes alphabetically (fstream before iostream) - Fix indentation alignment for multi-line statements
- Introduce format_subgraph_name() helper to ensure consistent naming - Simplify dump_subgraph_composition() to use replaced_by() for unique subgraphs - Update all dump paths to use the helper method
…FA logic - Use format_subgraph_name helper for consistent naming - Always use first occurrence's id for subgraph names (matches dump) - Optimize counting of subgraph repetitions (single-pass) - Fix MoE early-return logic: pyramid/HFA files only listed when MoE is NOT present - Add clarifying comments about naming, ordering, and dump logic - Ensure composition file matches actual dumped files
dylanneve1
force-pushed
the
EISW-196509
branch
from
29f2f2a to
f71964b
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Introduces the NPUW_DUMP_SUBS_DIR configuration option for specifying NPUW dump directory. When NPUW_DUMP_SUBS is enabled, the plugin will now generate JSON files (.sg format) containing the pipeline structure, including unique subgraphs and their repetitions.
Details:
Tickets: