[Deps] Split up of dependencies

[saiarthiraguram]

Ticket

Fixes: Link to Github Issue

Problem description

dev-requirements - contains model based requirements
core-requirements - contains compiler based requirements

Changes

1. Added Compiler only dependencies to core-requirements
2. Added other requirements to dev-requirements - Will split this to model based requirements in future
3. Added sanity tests for each frameworks to be tested on slim docker.
4. Adding a slim docker with only compiler based requirements installed
5. Having a dev docker with all requirements installed

To-Do

• Raise model based requirements separately for each model.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 73.11828% with 25 lines in your changes missing coverage. Please review.
✅ Project coverage is 62.39%. Comparing base (7d35604) to head (a69f5b2).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
forge/csrc/test/passes/gtest_main.cpp	73.11%	25 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2364      +/-   ##
==========================================
+ Coverage   62.33%   62.39%   +0.06%     
==========================================
  Files         157      157              
  Lines       13362    13451      +89     
==========================================
+ Hits         8329     8393      +64     
- Misses       5033     5058      +25     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[saiarthiraguram]

Unit tests, Failure update:

**Current Test Results: 15/31 PASSING **

All consistently passing tests:

1. DecomposeNdReshapeSplitTest.basic_dimension_split_optimization
2. DecomposeNdReshapeSplitTest.invalid_cases_should_be_skipped
3. CommuteBroadcastThroughTranspose.commute_broadcast_through_transpose
4. UpdateReshapeNamedAttrsTest.update_named_attrs
5. UpdateConcatNamedAttrsTest.update_named_attrs
6. UpdateReduceSumAttrsTest.ReduceSumDim
7. UpdateReduceMaxAttrsTest.ReduceMaxDim
8. EraseInverseOpsSqueezeAndUnsqueeze.erase_inv_ops_sq_unsq
9. CommuteTransposeThroughReduce.commute_transpose_through_higher_dim_reduce
10. EraseUnnecessary4DSeqTwoOps.two_operands
11. EraseUnnecessary4DSeqTwoOps.na1
12. EraseUnnecessary4DSeqTwoOps.na2
13. EraseUnnecessary4DSeqThreeOps.three_operands

[saiarthiraguram]

@nvukobratTT ,
MLIR seg fault issue:

**Problem ** :

• TensorFlow links its own LLVM with command-line options for StableHLO/quantization
• TT-MLIR links its own LLVM with command-line options for TT passes
• Both try to register options in overlapping LLVM registries
• Iterator corruption occurs when one library's registry is modified while another is iterating

Technical Details:

• llvm::cl::AddLiteralOption() modifies global LLVM command-line option collections
• SmallPtrSetIteratorImpl becomes invalid when underlying container is modified during iteration
• Multiple LLVM static initializations
• Memory layout corruption leads to invalid pointer dereference (0x100000001)
  planning on finding a way to have split up but with this issue mitigated by installing some problematic requirements in core itself.

Solution added:
Added guard on loads which systematically loads LLVM in a way that seg fault is fixed

To fix the problem, explicitly cast the call to log_error on line 89 to void. This documents that the value is intentionally ignored and the function is called for its side effect. This change should be made only to line 89 in forge/csrc/passes/mlir_passes.cpp. No imports or additional definitions are required.

---

To address the warning and clarify that the value of the expression is intentionally ignored (and that the function is called for its side effect), explicitly cast the call to log_error to (void). This is a common idiom in C/C++ to indicate that the return value is intentionally discarded, and it will silence static analysis tools like CodeQL. Only line 117 in forge/csrc/passes/mlir_passes.cpp needs to be changed:

• Replace:
  log_error(LogMLIRCompiler, "Unknown error during MLIR pass registration");
• With:
  (void)log_error(LogMLIRCompiler, "Unknown error during MLIR pass registration");

No new imports, methods, or definitions are needed.

---

To fix the warning, explicitly cast the call to log_warning to (void). This documents that the return value is intentionally ignored and that the expression is used for its side effects (logging). This is a standard idiom in C/C++ to silence such warnings from static analysis tools.
Change:

• In forge/csrc/passes/mlir_passes.cpp, line 125, change
  log_warning(LogMLIRCompiler, "MLIR pass registration was not successful");
  to
  (void)log_warning(LogMLIRCompiler, "MLIR pass registration was not successful");
  No new imports or definitions are needed.

---

[saiarthiraguram]

9/09/2025
Based on discussion on call, posting current status to get info on it
Context: Segmentation fault (SegFault) occurs in the CI when executing the test-sub workflow.

• Debugged which files cause the seg fault by creating a simple test which doesn't import many modules and checking each imports in test files.
• When commenting out the problematic imports in conftest.py we are able to collect the pytests.

Current Fix in docker env is to

• Uninstall and reinstall torch, tensorflow, paddlepaddle - fixes the segfault
• trying to find a more sensible fix that reinstallation in test-sub workflow which defeats the purpose of docker installation