Check that we found all expected start/end pairs. Raises on missing scopes.

Add proton.record statements to the FMMS kernel TTGIR file. Uses a two-pass approach: first scan for insertion points, then insert from botto

Allocate kernel output buffers (maxs, maxs_idx) in symmetric memory. Returns (maxs, maxs_idx, symm_mem_hdl, storage_offset_maxs_idx). maxs an

Rename FMMS columns/values to FlashSampling equivalents.

Apply top-k and top-p filtering and return probabilities. Single softmax.

Apply top-k and top-p filtering using vLLM's Qitra Triton kernel and return probabilities.

Apply combined top-k and top-p masking using Triton. Top-k is applied first (by logit value), then top-p is applied to the remaining k v

Time kernel execution using CUDA events.

Bootstrap 95% CI for the mean of a binary array. Returns (lo%, hi%).

Read the per-TP batch-scaling CSVs and return a long DataFrame. When ``args.use_reruns`` is set, also load sibling ``{gpu}-rerun*`` folders s

Create a benchmark function for a specific case.

Extract the final numeric answer after '#### ' in the GSM8K answer string.

Find all FMMS kernel TTGIR files in the dump directory and process them.

Find the latest samples_gsm8k*.jsonl under eval_dir.

Fused matmul + per-tile top-k reduction. Returns: topk_vals: [num_tiles_v, H, top_k] float32 logits (temperature-scaled) topk_ids

Stage 1: each (V, H) tile computes its local max and argmax in parallel.

Judge all samples in an eval directory. Returns {doc_id: correct}.

Ask the judge LLM whether the model's response matches the reference.

Return the most recent timestamped subdirectory.

Load and combine NCU + Proton data into a DataFrame. Columns: bsz, method, matmul_us, sampling_us, total_us.

Load already-judged results keyed by doc_id, for resume support.

Load NCU data for a single method across all batch sizes. Kernel names are shortened and durations aggregated by short name, so multiple raw

Load samples from a JSONL file, deduplicated by doc_id. lm-eval writes one line per (doc_id, filter) combination. The response text is the sa

FLOPs for the fused matmul: 2 * V * D * H.

Run under nsys: warmup, sync ranks, then profile timed iterations.

Source: https://cs.stanford.edu/people/mmahoney/cs369m/Lectures/lecture1.pdf

Pad D to 16-byte alignment for TMA on SM 90+ (H100, B200, etc.). TMA requires the innermost tensor dimension to be 16-byte aligned. For bf16

Paired bootstrap test: is there a significant difference between two variants? Both variants answer the same set of questions, so we use a *paire

Compute paired speedup ratios between baseline and FMMS runs. Comparisons are made only within the same trial directory (e.g. Qwen3-8B-trial1

Parse engine stats grouped by benchmark run. Uses "Starting main benchmark run..." as the start delimiter and "Serving Benchmark Result" as t

Stacked bar chart: one bar per batch size, one segment per kernel.

Compute relative performance vs ref_method and save plot + CSV. ``bdf`` may contain multiple rows per ``n_hidden_states`` (one per run); when

Classic roofline plot: achieved TFLOP/s vs arithmetic intensity (FLOP/byte).

Print per-scope runtime breakdown as percentages of kernel time. Raises if any expected scope is missing, since filling 0 would produce a bog

Print per-variant accuracy with bootstrap CIs, and pairwise comparisons.

Read metadata.json dumped by the benchmark runner. Falls back to args.json and logs.txt for older result directories.

Return the set of token indices allowed after top-k then top-p filtering.

Run benchmarks using FlashInfer's CUPTI-based bench_gpu_time.

Run benchmarks using NVBench.

Shift all logits by -offset without touching the existing weights. Appends a bias column so that ``h_new @ W_new^T = h @ W^T - offset``. Sinc