Repository navigation
[PT FE] Fix timm model filter selecting huge models, update timm to 1.0.30, support aten.meshgrid in FX - #38443
Merged
Merged
Conversation
mvafin
requested review from
ValentinaKats,
mmikolajcz and
rkazants
and removed request for
a team
September 28, 2026 11:11
github-actions
Bot
added
category: CI
….meshgrid in FX - Fix filter_timm size ordering (was alphabetical) and recognize parameter-count (7b, so400m), width-multiplier (50x3, 32x8d) and f/e size tokens, so the smallest variant of each architecture is selected. This removes multi-GB models that made the nightly runner run out of memory. - Update timm to 1.0.30 and regenerate timm_models. - Add FX translator for aten.meshgrid and enable meshgrid layer tests in export mode. - Document how to run model hub tests and reuse the model cache.
Contributor
There was a problem hiding this comment.
Copilot review overview
🟡 Changes recommended
The filter still retains large one-letter iFormer size variants instead of selecting only the smallest model.
Review effort: Balanced
Findings: 1
What changed in this PR
Reduces PyTorch timm nightly memory usage, updates timm, and adds FX support for aten.meshgrid.
Changes:
- Improves timm model filtering and regenerates the model scope.
- Adds shared meshgrid conversion with TorchScript/FX tests.
- Documents model-hub testing and caching.
| File | Description |
|---|---|
tests/model_hub_tests/pytorch/timm_models |
Regenerates the timm model scope. |
tests/model_hub_tests/pytorch/test_timm.py |
Improves model-size filtering. |
tests/model_hub_tests/pytorch/envs/timm.txt |
Updates timm to 1.0.30. |
tests/layer_tests/pytorch_tests/test_listunpack.py |
Adds FX meshgrid coverage. |
src/frontends/pytorch/src/utils.hpp |
Declares shared meshgrid builder. |
src/frontends/pytorch/src/utils.cpp |
Implements meshgrid decomposition. |
src/frontends/pytorch/src/transforms/prim_list_unpack_replacer.cpp |
Reuses shared decomposition. |
src/frontends/pytorch/src/op/meshgrid.cpp |
Converts known lists and supports FX. |
src/frontends/pytorch/src/op_table.cpp |
Registers FX meshgrid overloads. |
src/frontends/pytorch/README.md |
Documents model-hub testing. |
.github/agents-prototype/skills/conversion-issues/pytorch.md |
Links model-hub guidance. |
.claude/skills/ov-update-pytorch-version/SKILL.md |
Adds model-test and cache instructions. |
💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Use one translator for TorchScript and FX. When the input list is known at conversion, decompose meshgrid directly and return a list, otherwise keep deferring to PrimListUnpackReplacer. Share the decomposition in build_meshgrid.
Recognize t and h as size tokens when the name has no other size token, so iformer_h/iformer_s fold into iformer_t and efficientnetv2_rw_s into efficientnetv2_rw_t, while efficientnet_h_b5 and regnetz_040_h keep h as a variant marker.
DunnBC22/efficientformer-l3-300-Brain_Tumors_Image_Classification was removed from HF Hub, which fails the nightly scope2 HF Transformers step and skips the LLM, Detectron2 and Torchbench steps after it. Use the original snap-research/efficientformer-l1-300 checkpoint instead.
github-actions
Bot
added
the
github_actions
Nightly PyTorch Models Scope1/Scope2 jobs ran only on schedule. Also run them for PRs that change PyTorch model hub tests (checked in the Smart_CI job) and on workflow_dispatch with the pytorch-models-nightly input.
cyankiwi/Qwen3.5-4B-AWQ-4bit is a multimodal checkpoint. AutoModelForCausalLM loads it with text_config, which has no quantization_config, so the model is loaded unquantized and patch_quantized fails with 'Unknown quantization type: None'.
The multimodal checkpoint can't be loaded correctly via AutoModelForCausalLM: transformers<5.8 drops its quantization_config, and newer versions still don't match its ignore list to the text-only module names.
…erged Reverts the wrap_with_autocast translator (030af3e); it is added in openvinotoolkit#38396.
Build legacy setup.py packages in isolation to avoid the kernels egg_info plugin, and allow numpy<2 pickled inputs of maml models with numpy 2.
Contributor
There was a problem hiding this comment.
Warning
Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.
Copilot review overview
Review effort: Lite
Findings: 2
Open (5)
pytorch-models-nightly), so… · Newpytorch-models-nightly), so… · Newindexingis typically not passed as a positional input. This code calls… · Newtorch.meshgridonly accepts 0D/1D inputs (and errors for higher-rank tensors), but this… · Newsetup_classand never restored, which can… · New
| NodeVector reshapes; | ||
| auto const_neg_1 = v0::Constant::create(element::i32, Shape{1}, {-1}); | ||
| auto const_1 = v0::Constant::create(element::i32, Shape{1}, {1}); | ||
| for (size_t i = 0; i < inputs.size(); ++i) { |
Contributor
Author
There was a problem hiding this comment.
torch.meshgrid itself rejects inputs with rank > 1, so a valid model can't reach this decomposition with such inputs. Keeping it as is.
Comment on lines
+45
to
+49
| os.environ["PIP_BUILD_CONSTRAINT"] = build_constraints | ||
| os.environ["PIP_CONSTRAINT"] = build_constraints | ||
| # build legacy setup.py packages in isolation, so setuptools plugins from the test env | ||
| # (e.g. kernels egg_info writer) are not loaded | ||
| os.environ["PIP_USE_PEP517"] = "1" |
Contributor
Author
There was a problem hiding this comment.
Torchbench tests run in a separate pytest invocation, which is the last step of the job, so these variables don't affect other tests. Keeping it simple.
evkotov
reviewed
Sep 30, 2026
evkotov
approved these changes
Oct 1, 2026
bumbosiepsak
approved these changes
Oct 1, 2026
| # parameter count, e.g. "300m", "7b", "so400m" | ||
| param_count_pattern = re.compile(r"^(?:so)?(\d+)([mb])$") | ||
| # depth/cardinality x width multiplier, e.g. "50x1", "152x4", "32x8d" | ||
| width_multiplier_pattern = re.compile(r"^(\d+)x(\d+)d?$") |
Contributor
There was a problem hiding this comment.
[nitpick] This is brittle. width_multiplier_pattern = r"^(\d+)x(\d+)d?$" would misclassify a hypothetical NxN resolution token (e.g. 224x224) as a large width multiplier and skew size ranking. No such tokens exist in the current timm name set and test_models_list_complete validates the regenerated list, so impact is limited to test selection.
| with open(deps_file) as f: | ||
| deps = f.read() | ||
| with open(deps_file, "w") as f: | ||
| f.write(deps.replace('"torchvision", "torchaudio"]', '"torchvision"]')) |
Contributor
There was a problem hiding this comment.
[nitpick] This is brittle. deps.replace('"torchvision", "torchaudio"]', '"torchvision"]') is mostly acceptable thanks to pinning of the torchbench commit.
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.
Details:
filter_timmsorted size tokens alphabetically (enormous<giant< ... <small) and didn't recognize7b/so400m, so it picked the largest variants (e.g.vit_7b_dinov3, 26.9 GB). Several of these were loaded at the same time under-n 4.filter_timmnow keeps the smallest variant of each architecture. Size tokens are ordered by size, parameter counts and baret/hsizes are recognized, and width-multiplier variants (resnetv2_50x3,resnext101_32x8d) are folded into their base architecture (they use the same ops).timm_modelsis regenerated: 517 models, down from 532. The largest model is now 4.05 GB, and total weights drop from ~248 GB to ~108 GB. New failing models are marked xfail:iformer_*(numpy<2 load error) andgemma4_vit_167m*in export mode.aten::meshgridis now converted during translation for both TorchScript and FX.PrimListUnpackReplaceris still the fallback when the input list is unknown at that point. This fixes export mode for 9 new timm models and meshgrid outputs used directly bystack/cat/[i].DunnBC22/efficientformer-l3-300-Brain_Tumors_Image_Classificationwas removed from HF Hub. That failure also skips the LLM, Detectron2 and Torchbench steps. The model is replaced withsnap-research/efficientformer-l1-300.qwen3_compressed_tensors(cyankiwi/Qwen3.5-4B-AWQ-4bit, a multimodal checkpoint) is marked xfail. With the pinned transformers 5.5.3,AutoModelForCausalLMdrops itsquantization_config(fixed in 5.8, Fix: propagate quantization_config to text sub-config for composite models in AutoModelForCausalLM huggingface/transformers#45494). Newer versions still don't match the checkpoint's ignore list to the text-only module names, solinear_attn.in_proj_a/bare initialized randomly.install.py, which requires torchaudio. torchaudio was dropped from the requirements in [PT FE] Upgrade tests to torch 2.12 #36098 and doesn't support torch 2.14. The test removes torchaudio from torchbench'sTORCH_DEPS, pinssetuptools<81for sdist builds (visdomneedspkg_resources), and setsPIP_USE_PEP517=1. With pip 24, legacysetup.pypackages are built in the test venv, where theegg_infohook ofkernels(installed for LLM tests) crashes against the olderhuggingface_hubinstalled by torchbench. The maml models' inputs, pickled with numpy<2, now load with numpy 2.convit_tiny(trace): the model fills a cache on its first forward, so the test runs one warm-up forward first.ExportedProgramare taken fromout_spec. A key is skipped if it's ambiguous (one tensor under several keys, or the same key in several nested dicts) or matches the name of another graph node or input.efficientvit_b0(export): fixed by [PyTorch] Support exported grad-mode and autocast regions without decompositions #38396, now passes.PyTorch Models Nightly Scope1/Scope2 testsjobs inubuntu_22.ymlnow also run for PRs that changetests/model_hub_tests/pytorch,tests/model_hub_tests/models_hub_commonorjob_pytorch_models_tests.yml. TheSmart_CIjob checks the PR's changed files withgit diffon the merge commit, without extra API calls. The jobs also run on manual dispatch with the newpytorch-models-nightlyinput. They use the same runner and settings as the nightly, and aren't part ofOverall_Status, so a failure doesn't block merge. This PR runs them itself. Moved from [CI] Run nightly PyTorch model hub scopes on related PRs and on dispatch #38457.Tickets:
AI Assistance:
test_models_list_complete: passes.