Count kernels per lane (tid).

Create a GPT-2 small model with random weights.

Return a stack of linear->norm->sigmoid layers

Register forward pass hook (which registers a backward hook) to model outputs Returns: - layers: a dict with keys as layer/module and val

Merge annotations into the trace and emit a Perfetto ``.pftrace``.

Profile graph replays and save the trace.

Run training loop and capture memory snapshot. Args: activation_checkpointing: Whether to enable gradient checkpointing. snapshot

Save kernel annotations to a pickle file.

Carry a Chrome event arg over as a typed Perfetto debug annotation.

Split overlapping slices into backing lanes so each lane is nestable. A lane only holds slices that are either disjoint or fully contained, so a

Detect Wikitext article title lines like ' = Some Title = '.

Given a state, choose an epsilon-greedy action

Analyze a memory snapshot using Mosaic's Python API. Returns a structured dictionary with memory breakdown. Args: snapshot_path: Pat

Annotate a collective with the metadata eager NCCL traces expose. Returns a ``mark_kernels`` context manager. Any kernels launched inside (i.

Computes batched dot by reducing to ``bmm``

Computes batched dot by multiplying and summing

Benchmark data loading with a specific batch size.

Create a compute + collective block annotated for profiling.

Create a simple transformer block with parameters.

Add the experience to memory

Check if memory usage exceeds threshold for CI/CD pipelines. Args: report: Memory analysis report from training_pipeline_with_memory_moni