torch.memory_stats
function memory_stats(): voidReturns memory statistics from the WebGPU buffer allocator.
Tracks memory usage including active allocations, pooled memory, and peak memory. Essential for profiling, debugging memory issues, and monitoring GPU resource usage. Useful for:
- Monitoring GPU memory consumption during training
- Detecting memory leaks
- Profiling performance hotspots
- Debugging out-of-memory errors
- Optimizing model architecture for memory constraints
- Pooling: pooled_bytes are cached but not actively in use; use empty_cache() to free
- Peak tracking: peak_bytes includes all memory ever allocated, not just current
- Allocation count: Useful for detecting fragmentation issues
- Byte counts: All sizes in bytes; divide by 1024/1024 for MB
Returns
Object containing memory statistics with fields: - active_bytes: Bytes currently allocated for active tensor data - pooled_bytes: Bytes pooled and ready for reuse (not actively in use) - peak_bytes: Peak memory usage since last reset (maximum ever allocated) - total_allocations: Total count of allocation operations performed
Examples
// Get current memory stats
const stats = torch.webgpu.memory_stats();
console.log(`Active: ${stats.active_bytes / 1024 / 1024} MB`);
console.log(`Pooled: ${stats.pooled_bytes / 1024 / 1024} MB`);
console.log(`Peak: ${stats.peak_bytes / 1024 / 1024} MB`);
console.log(`Total allocations: ${stats.total_allocations}`);
// Monitor memory during training
for (let epoch = 0; epoch < 10; epoch++) {
const stats_before = torch.webgpu.memory_stats();
trainEpoch();
const stats_after = torch.webgpu.memory_stats();
console.log(`Epoch ${epoch} used ${(stats_after.peak_bytes - stats_before.peak_bytes) / 1024 / 1024} MB`);
}See Also
- PyTorch torch.cuda.memory_stats()
- memory_summary - Get formatted memory summary string
- empty_cache - Free pooled memory
- reset_peak_memory_stats - Reset peak counter for profiling