Register a codegen handler for a loop modifier. Args: modifier_name: Name of the modifier (e.g., "unroll", "para

Register a loop modifier. Args: name: Name of the modifier (e.g., "unroll", "parallel") modifier: Lo

Register a pass for a specific backend. Args: backend: Backend name (e.g., "cuda") pass_func: Pass f

Rotates half the hidden dims of the input.

Execute command with retries, show full output on failure

Show generated CUDA code.

Not in use; reserved for future path (symmetric buffers are allocated eagerly in ``__init__``).

Test correctness.

Test MoRI SHMEM P2P bandwidth using putmem_nbi Args: src_rank: Source PE rank dst_rank: Destination PE rank num_

Copy contents to a PyTorch tensor for inspection.

Copy to a PyTorch tensor for inspection.

return C = all_gather(A) @ B.T

Reference PyTorch forward pass using AllGather-GEMM.

Reference PyTorch forward pass using GEMM-ReduceScatter.

Unwrap a loop modifier call to get the inner iterator. Args: node: AST Call node representing the modifier call

Cast a value to a specific type. Usage: val_cast(value, ll.uint32) # Cast value to uint32 ll.uint32(value) # Same as above

Process constants (map to C++ syntax).

Handle EnumDef IR node.

Handle SpecialStructDef IR node.

Write a line to builtin buffer (for mixin compatibility).

Shuffle xor (butterfly): each lane reads from (laneid ^ mask).

Add C++ headers for std types.

Analyze return type from annotation (e.g., -> Tuple[bool, int, int]).

Analyze return statements to determine return value structure.