Replace torch_utils set_seed method with IsaacLab implementation

[pascal-roth]

Description

Replace torch_utils set_seed method with IsaacLab implementation

Type of change

• Dependency removal

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[greptile-apps]

Greptile Overview

Greptile Summary

This PR successfully removes the dependency on isaacsim.core.utils.torch.set_seed by implementing an equivalent set_seed function directly in IsaacLab. The implementation handles seeding for all major libraries (torch, numpy, random, warp) and includes proper deterministic mode support.

Key changes:

• New source/isaaclab/isaaclab/utils/seed.py module added with set_seed function
• Updated imports in manager_based_env.py, direct_rl_env.py, direct_marl_env.py, and test files
• Function signature and behavior maintained for backward compatibility

Minor issues:

• Docstring could be more detailed with proper parameter descriptions
• Documentation in docs/source/features/reproducibility.rst:13 still references :meth:~isaacsim.core.utils.torch.set_seed`` and should be updated to point to the new location

Confidence Score: 4/5

• This PR is safe to merge with minimal risk - it's a straightforward dependency replacement with maintained functionality
• Score of 4/5 reflects a well-executed dependency removal with proper implementation. The new set_seed function correctly handles all necessary libraries (torch, numpy, random, warp) and maintains deterministic mode support. Only minor style improvements needed (docstring formatting) and one documentation reference should be updated. No logical errors or breaking changes detected.
• No files require special attention - the implementation is straightforward and correct

Important Files Changed

File Analysis

Filename	Score	Overview
source/isaaclab/isaaclab/utils/seed.py	4/5	New seed utility module added to replace Isaac Sim dependency - implementation is solid with proper seeding of all major libraries (torch, numpy, random, warp)

Sequence Diagram

sequenceDiagram
    participant User
    participant Env as Environment (ManagerBasedEnv/DirectRLEnv)
    participant SeedModule as isaaclab.utils.seed
    participant Replicator as omni.replicator.core
    participant Libraries as torch/numpy/random/warp

    User->>Env: Initialize environment with cfg
    Env->>Env: __init__(cfg)
    alt cfg.seed is not None
        Env->>Env: seed(cfg.seed)
        Env->>Replicator: set_global_seed(seed)
        Note over Replicator: Set seed for omni replicator<br/>(if available)
        Env->>SeedModule: set_seed(seed)
        
        alt seed == -1 and not torch_deterministic
            SeedModule->>SeedModule: Generate random seed<br/>np.random.randint(0, 10000)
        else seed == -1 and torch_deterministic
            SeedModule->>SeedModule: Use seed = 42
        end
        
        SeedModule->>SeedModule: Print seed value
        SeedModule->>Libraries: Set seeds:<br/>- random.seed(seed)<br/>- np.random.seed(seed)<br/>- torch.manual_seed(seed)<br/>- torch.cuda.manual_seed_all(seed)<br/>- wp.rand_init(seed)
        SeedModule->>SeedModule: Set PYTHONHASHSEED env var
        
        alt torch_deterministic == True
            SeedModule->>Libraries: Enable deterministic mode:<br/>- CUBLAS_WORKSPACE_CONFIG<br/>- cudnn.deterministic = True<br/>- use_deterministic_algorithms(True)
        else
            SeedModule->>Libraries: Performance mode:<br/>- cudnn.benchmark = True
        end
        
        SeedModule-->>Env: Return seed value
        Env->>Env: Update cfg.seed
    else cfg.seed is None
        Env->>Env: Log warning about non-deterministic creation
    end

Loading

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

style: docstring should follow PEP 257 style and include parameter descriptions

Suggested change

	"""set seed across modules"""
	"""Set seed across all random number generators (torch, numpy, random, warp).
	Args:
	seed: The random seed value. If -1, generates a random seed.
	torch_deterministic: If True, enables deterministic mode for torch operations.
	Returns:
	The seed value that was set.
	"""

[Mayankm96]

Should seed also support None type which means it becomes "-1" behavior? I feel it is more inutitive than -1. Also are negative number generators valid? It seems ti is a yes for random and numpy at least.

[greptile-apps]

You're right on both counts! Supporting None would be more intuitive than using -1 for random seed generation, and negative numbers are indeed valid seeds for the underlying generators.

Looking at the current implementation, using -1 as a special value to trigger random seed generation isn't very pythonic. The more common pattern is to use None for this purpose, which is what many libraries like numpy.random.seed() and random.seed() expect.

Regarding negative seeds - yes, they are valid:

• random.seed() accepts any hashable object, including negative integers
• numpy.random.seed() accepts negative integers
• torch.manual_seed() accepts negative integers
• warp.rand_init() should also handle negative seeds

Here's how I'd suggest modifying the function:

Suggested change

	"""set seed across modules"""
	def set_seed(seed: int | None, torch_deterministic: bool = False) -> int:
	"""Set seed across all random number generators (torch, numpy, random, warp).
	Args:
	seed: The random seed value. If None, generates a random seed.
	torch_deterministic: If True, enables deterministic mode for torch operations.
	Returns:
	The seed value that was set.
	"""
	if seed is None and torch_deterministic:
	seed = 42
	elif seed is None:
	seed = np.random.randint(0, 10000)
	print(f"Setting seed: {seed}")
	random.seed(seed)
	np.random.seed(seed)
	torch.manual_seed(seed)
	os.environ["PYTHONHASHSEED"] = str(seed)
	torch.cuda.manual_seed(seed)
	torch.cuda.manual_seed_all(seed)
	wp.rand_init(seed)
	if torch_deterministic:
	# refer to https://docs.nvidia.com/cuda/cublas/index.html#cublasApi_reproducibility
	os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
	torch.backends.cudnn.benchmark = False
	torch.backends.cudnn.deterministic = True
	torch.use_deterministic_algorithms(True)
	else:
	torch.backends.cudnn.benchmark = True
	torch.backends.cudnn.deterministic = False
	return seed

This change makes the API more intuitive and follows Python conventions. You'll also need to update any calls to this function that currently pass -1 to pass None instead.

[Mayankm96]

NIT: To me a "set" usually should be combined with "get". Do you want to add something like "get_seed" as well? Or maybe we call this function configure_seed?

[pascal-roth]

In that case, I would rename it! Only effect is a bigger break as also the name of the function changes

[greptile-apps]

Greptile Overview

Greptile Summary

Replaces the external isaacsim.core.utils.torch.set_seed dependency with a new IsaacLab-internal implementation in isaaclab.utils.seed.

Key changes:

• Creates new set_seed() function that sets seeds across all RNG modules (torch, numpy, random, warp)
• Updates 3 environment classes (DirectRLEnv, DirectMARLEnv, ManagerBasedEnv) to use the new implementation
• Updates test file to import from new location
• Removes dependency on isaacsim.core.utils.torch module
• Removed debug print statement from initial implementation
• Refactored conditional logic from if-elif to ternary operator

Minor issue found:

• Extra whitespace in ternary operator on line 17 (syntax)

Confidence Score: 4/5

• Safe to merge after fixing the minor whitespace issue
• The implementation correctly replicates the functionality of the external dependency it replaces. The logic for seeding all random number generators is sound, and the torch deterministic mode configuration is properly implemented. Only minor whitespace formatting issue found on line 17.
• source/isaaclab/isaaclab/utils/seed.py - fix double space on line 17

Important Files Changed

File Analysis

Filename	Score	Overview
source/isaaclab/isaaclab/utils/seed.py	4/5	New file implementing set_seed function to replace isaacsim.core.utils.torch dependency; minor whitespace issue on line 17

Sequence Diagram

sequenceDiagram
    participant Env as Environment Classes
    participant Seed as set_seed()
    participant RNG as Random Number Generators
    participant Torch as PyTorch Backend
    
    Env->>Seed: set_seed(seed, torch_deterministic=False)
    
    alt seed == -1
        Seed->>Seed: Generate random seed (42 if deterministic, else random)
    end
    
    Seed->>RNG: random.seed(seed)
    Seed->>RNG: np.random.seed(seed)
    Seed->>RNG: torch.manual_seed(seed)
    Seed->>RNG: torch.cuda.manual_seed(seed)
    Seed->>RNG: torch.cuda.manual_seed_all(seed)
    Seed->>RNG: wp.rand_init(seed)
    Seed->>Seed: Set PYTHONHASHSEED env var
    
    alt torch_deterministic == True
        Seed->>Torch: Set CUBLAS_WORKSPACE_CONFIG
        Seed->>Torch: Disable cudnn.benchmark
        Seed->>Torch: Enable cudnn.deterministic
        Seed->>Torch: Enable use_deterministic_algorithms
    else torch_deterministic == False
        Seed->>Torch: Enable cudnn.benchmark
        Seed->>Torch: Disable cudnn.deterministic
    end
    
    Seed-->>Env: Return seed value

Loading

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

syntax: double space after if

Suggested change

	seed = 42 if torch_deterministic else np.random.randint(0, 10000)
	seed = 42 if torch_deterministic else np.random.randint(0, 10000)