`core.resharding.nvshmem_copy_service.validation`#

Validation utilities for GPU-to-GPU communication.

Provides deterministic data generation and validation for verifying

correctness of communication operations.

Module Contents#

Classes#

`ValidationResult`	Result of validating a single task.
`ValidationSummary`	Summary of validation across all tasks.

Functions#

`generate_deterministic_data`	Generate deterministic data pattern for a task.
`validate_received_data`	Validate received data against expected deterministic pattern.
`log_validation_summary`	Log validation summary.

API#

class core.resharding.nvshmem_copy_service.validation.ValidationResult#

Result of validating a single task.

task_id: int#: None

size: int#: None

passed: bool#: None

src_pe: int#: None

mismatches: int#: 0

first_mismatch_idx: int#: None

first_mismatch_expected: int#: 0

first_mismatch_actual: int#: 0

batch_index: int#: None

iteration: int#: None

class core.resharding.nvshmem_copy_service.validation.ValidationSummary#

Summary of validation across all tasks.

total_tasks: int#: None

passed_tasks: int#: None

failed_tasks: int#: None

total_bytes: int#: None

results: List[core.resharding.nvshmem_copy_service.validation.ValidationResult]#: None

property all_passed: bool#: Check if all validated tasks passed.

core.resharding.nvshmem_copy_service.validation.generate_deterministic_data( task_id: int, size: int, device: str = 'cuda', ) → torch.Tensor#

Generate deterministic data pattern for a task.

Pattern: Each byte = (task_id * 31 + position) % 256 This creates a unique pattern per task that varies along the data.

Parameters:

task_id – Unique task identifier
size – Number of bytes to generate
device – Device to create tensor on (‘cuda’ or ‘cpu’)

Returns:

torch.Tensor of uint8 with deterministic pattern

core.resharding.nvshmem_copy_service.validation.validate_received_data( task_id: int, tensor: torch.Tensor, size: int, src_pe: int = -1, ) → core.resharding.nvshmem_copy_service.validation.ValidationResult#

Validate received data against expected deterministic pattern.