Aiter Op Test

aiter op_test standard

How every aiter op test in op_tests/test_*.py must be built. The canonical reference in-tree is op_tests/test_quant.py — match its shape. A test is both a correctness check (vs a torch reference) and a perf sweep that ends in a markdown summary table.

Follow this whenever you create a new test_*.py, rewrite an old one, or add shapes/candidates to an existing one.

The hard rules

1. Mirror test_quant.py. Same imports, same decorator, same table-at-the-end flow. Don't invent a different structure.
2. @benchmark() on the test fn. It logs the function's call args (the shape params) as table columns automatically and merges the dict you return. So the test fn signature is the table's left-hand columns — name params accordingly.
3. Candidates live in a dict; build ret in a loop. Per candidate record raw us, plus TFLOPS and TB/s, plus err — ret[f"{name} us"], ret[f"{name} TFLOPS"], ret[f"{name} TB/s"], ret[f"{name} err"]. Never hand-write ratio columns. (TFLOPS/TB-s section below.)
4. torch is the reference only — compute it, compare against it, but do not time it and do not put it in the table. (A pure-torch candidate is allowed only when torch is one of the kernels under test, e.g. torch.einsum.)
5. Time with run_perftest, check with checkAllclose — both, for every candidate. Compare in fp32 (.to(dtypes.fp32)).
6. End with a markdown summary table — one per test function. Sweep the shape lists with itertools.product, collect per-shape dicts into a pd.DataFrame, print via aiter.logger.info("... :\n%s", df.to_markdown(index=False)). A file with several test fns of different arg signatures emits one table each — never force-merge them (it scatters NaN columns). Mandatory — a test with no summary table is incomplete.
7. __main__ guard. All argparse + the sweep loop go inside main(), called under if __name__ == "__main__": main(). The reference (run_torch) and the @benchmark test fn stay at module top level so other scripts can import them for combination testing.
8. Standard argparse only. Use -d/--dtype, -b/--batch, -s/--mnk plus op-specific sweep axes as needed (e.g. --layout, --modes, --mtp). Those are legitimate data lists. Do not add bespoke behavior-toggle flags (no --dsv4, no --only-*) — every flag is a list the sweep iterates.
9. Run clean on every supported card. Gate on get_gfx() in main() so the test passes on all supported archs; arch-unsupported ops/candidates are filtered out before launch. Prefer the kernel's arch-dispatching wrapper over a file-per-arch (full section below).

Canonical template

# SPDX-License-Identifier: MIT
# Copyright (C) 2024-2026, Advanced Micro Devices, Inc. All rights reserved.

import argparse
import itertools

import aiter
import pandas as pd
import torch
import torch.nn.functional as F
from aiter import dtypes
from aiter.test_common import (
    benchmark,
    checkAllclose,
    run_perftest,
)
from aiter.jit.utils.chip_info import get_gfx  # "gfx942", "gfx950", "gfx1250", ...

torch.set_default_device("cuda")

SUPPORTED_GFX = ["gfx942", "gfx950"]  # every card this op is built/validated for


def run_torch(x, weight, dtype=dtypes.bf16):
    # Reference only: fp32 math, cast back. Not timed, not in the table.
    ...
    return out.to(dtype)


@benchmark()  # call args (b, m, n, k, dtype, ...) become the table's left columns
def test_op(b, m, n, k, dtype, layout):
    # build inputs/outputs in the layout the MODEL actually uses (see below);
    # ref = run_torch(...)
    candidates = {
        "triton": lambda: ...,        # the path the model really runs
        "torch_einsum": lambda: ...,  # optional torch kernel under test
    }
    if <kernel supported on this arch + config>:   # e.g. get_gfx() != "gfx1250"
        candidates["ck"] = lambda: ...   # else skip it (see rules below)

    flops = 2 * b * m * n * k                                   # roofline numerator
    nbytes = (b * m * k + b * n * k + b * m * n) * x.element_size()

    ret = {"gfx": get_gfx()}             # record the card in the table
    for name, fn in candidates.items():
        out, us = run_perftest(fn)
        err = checkAllclose(ref.to(dtypes.fp32), out.to(dtypes.fp32),
                            rtol=1e-2, atol=1e-2, msg=f"{name}: <op>")
        ret[f"{name} us"] = us
        ret[f"{name} TFLOPS"] = flops / us / 1e6
        ret[f"{name} TB/s"] = nbytes / us / 1e6
        ret[f"{name} err"] = err
    return ret


def main():
    # Whole-op arch gate goes HERE, not inside test_op: @benchmark always returns
    # the call-args dict, so an in-fn `return` still emits an args-only row.
    if get_gfx() not in SUPPORTED_GFX:
        aiter.logger.warning("<op> unsupported on %s; skipping", get_gfx())
        return

    parser = argparse.ArgumentParser(
        formatter_class=argparse.RawTextHelpFormatter, description="config input of test"
    )
    parser.add_argument("-d", "--dtype", type=dtypes.str2Dtype, nargs="*", default="bf16,", ...)
    parser.add_argument("-b", "--batch", type=int, nargs="*", default=[...], ...)
    parser.add_argument("-s", "--mnk", type=dtypes.str2tuple, nargs="*", default=[...], ...)
    # add -l/--layout ONLY if the op has real layout variants
    args = parser.parse_args()

    for dtype in args.dtype:          # one table per outer-config element
        df = []
        for layout, b, (m, n, k) in itertools.product(   # sweep via itertools.product
            args.layout, args.batch, args.mnk
        ):
            df.append(test_op(b, m, n, k, dtype, layout))
        df = pd.DataFrame(df)
        aiter.logger.info("<op> summary (markdown):\n%s", df.to_markdown(index=False))


if __name__ == "__main__":
    main()

op_tests/test_batched_gemm_bf16.py is a complete worked instance of this template (GEMM with triton/CK/einsum candidates, bmn/mbn layout, V4 shapes).

Faithful to the real model call (do not test an idealized op)

The point of the test is the kernel as the model invokes it, not a clean textbook version. Reproduce exactly:

• Preallocated output buffers. If the model passes YQ=/out= a buffer it allocated, allocate and pass the same — don't let the kernel allocate its own.
• Real tensor layout, including non-contiguous views. If the model feeds a transposed view (e.g. o.transpose(0, 1)), build the input as a transposed view of a contiguous tensor — not a fresh contiguous tensor of that shape.
• Couple input and output layout. They are linked in the model. If you sweep an output layout (bmn vs mbn), the input must follow: the mbn (model) case is a transposed view of [m, b, k] (physically mbk) and a transposed view of [m, b, n] — not contiguous [b, m, k].
• For a torch.einsum candidate, switch the layout by editing the subscript string, not by transposing afterwards: ->sgr is physically [m,b,n] (mbn), ->gsr is physically [b,m,n] (bmn). Feed it the model's natural contiguous operand (e.g. o is contiguous [s,g,d], so x.transpose(0,1).contiguous()).

A test that quietly uses contiguous inputs when the model uses a transposed view gives the wrong perf and hides correctness bugs (see next rule).

Report TFLOPS and TB/s, not just us

A bare us doesn't say whether a kernel is compute- or memory-bound. Always add both roofline metrics per candidate, derived from the same us:

flops = 2 * b * m * n * k                                   # GEMM: 2*M*N*K mul-add
nbytes = (b * m * k + b * n * k + b * m * n) * x.element_size()  # in + weight + out
ret[f"{name} TFLOPS"] = flops / us / 1e6   # us -> s is 1e-6, FLOP -> T is 1e-12
ret[f"{name} TB/s"]   = nbytes / us / 1e6

Count the FLOPs and bytes the op actually does (adjust the formula per op: a quant/norm/attention kernel has its own element-traffic and arithmetic). Use tensor.element_size() for dtype width so fp8/bf16/fp16 are handled. Reading the table: small m (decode) is memory-bound (high TB/s, low TFLOPS); large m is compute-bound (TFLOPS approaches peak).

Multiple test functions → multiple tables; reference patterns

• One @benchmark fn per distinct arg signature, one table each. A file may hold several (e.g. a main bf16/fp8 sweep + an fp8 nm-asm cross-check with different columns). Give each its own pd.DataFrame + aiter.logger.info(...) via a tiny summarize(name, rows) helper. Forcing them into one table scatters NaN columns and is unreadable.
• Two correctness shapes are common:
  • Multi-candidate (this template): several kernels vs one torch reference, all timed, dict-loop.
  • Single-kernel-vs-reference, in-place output (e.g. cache-writing kernels): clone the output buffer, run the kernel into one copy and the reference into another, then checkAllclose. Still record us/TFLOPS/TB-s/err.
• Prefer a shared reference from aiter.ops.torch_ref when one exists; only hand-write run_torch when there is none.
• Do all the asserts a case needs; record one representative err. A quant path checks the dequantized output and bit-exact scales: use tol_err_ratio= for the fraction of allowed element mismatches (fp8/bf16 rounding), and rtol=0, atol=0 for values that must match to the bit (scales).

Skip a candidate in configs it does not support

Some kernels are only correct for some layouts/dtypes. Running them anyway pollutes the table with wrong-but-fast numbers. Conditionally add such a candidate and leave its cells nan elsewhere — e.g. batched_gemm_bf16_CK returns garbage (err ≈ 0.99) on a non-contiguous mbk input, so it is only added for bmn:

if layout == "bmn":
    candidates["ck"] = lambda: aiter.batched_gemm_bf16_CK(x, weight)

When you skip something, say so in a code comment with the reason. The non-zero err column is exactly how you discover these — never silently drop a candidate because its error is high; first confirm whether it's a real bug or an unsupported config, then skip with a comment.

Run on every supported card (arch gating)

A test must run clean on every currently-supported card — today gfx942 (MI300) and gfx950 (MI35x), plus any arch the kernel specifically targets. Never assume one GPU. Detect arch at runtime and filter before launching:

from aiter.jit.utils.chip_info import get_gfx   # "gfx942" / "gfx950" / "gfx1250" / ...

• One test, all archs — drive the arch-dispatching wrapper; do NOT write a file per arch. Most aiter kernels expose a public wrapper that routes to the wave64/wave32 (or gfx-specific) implementation internally by get_gfx() — call that wrapper and the single test covers every arch. Only the behavioral arch differences need handling in the test (e.g. gfx1250 uses a linear FP8 layout, so force preshuffle=False there; keep the kernel call and the reference in sync). Never import an *_gfx1250/arch-suffixed kernel directly to make a parallel test file.
• Op not built/supported on this arch → skip in main() with an allow-list early return (skips the whole sweep cleanly — no rows):

  SUPPORTED_GFX = ["gfx942", "gfx950", "gfx1250"]
  if get_gfx() not in SUPPORTED_GFX:
      aiter.logger.warning("<op> unsupported on %s; skipping", get_gfx())
      return
  

  Do not gate by returning from inside the @benchmark fn: that wrapper always returns the call-args dict, so an in-fn return still emits an args-only NaN row. Prefer a positive allow-list (not in [...]) over a deny-list so an unknown new card doesn't silently run an unbuilt kernel and crash.
• One candidate / one sub-check unsupported on this arch → drop just it, with a warning naming arch + reason — e.g. the fp8 nm-asm cross-checks are wave64-only:

  if get_gfx() != "gfx1250":
      summarize("hca_fp8", [test_hca_fp8(bs) for bs in args.fp8_bs])
  else:
      aiter.logger.warning("gfx1250: skipping wave64-only fp8 cross-checks")
  

• Record the card: put "gfx": get_gfx() in the returned dict so one table is self-describing across cards.

In-tree precedent: the flydsl fused_compress_attn wrappers dispatch wave64/wave32 internally (one test_flydsl_compress_attn.py covers all archs); test_deepgemm.py / test_gemm_a4w4.py allow-list a single arch; test_gemm_a8w8.py does per-arch candidate/dtype/shape gating with warnings.

Deriving model shapes (don't guess)

When adding "test op X for model Y" shapes:

1. Read the real config.json for the actual dims (grep/python -json), don't assume the dataclass defaults.
2. Map model semantics → the kernel's (b, m, n, k) and write the mapping in a comment. Worked example — DeepSeek-V4 grouped output LoRA (atom/models/deepseek_v4.py, batched_gemm_bf16(o.transpose(0,1), wo_a, YQ=y)):
   • b (batch) = n_local_groups = o_groups // tp
   • m = num_tokens (the swept dim)
   • n = o_lora_rank
   • k = n_heads * head_dim // o_groups
3. Cover the real parallelism configs, because they change b:
   • V4-Flash: o_groups=8, tp8 → b=1, tp2 → b=4
   • V4-Pro: o_groups=16, tp8 → b=2, dp (tp1 attn, full groups) → b=16, tp1 → b=16 Put the candidate b values in -b defaults and the (m,n,k) rows in -s; the sweep's cross product covers each config, identifiable by the b/n/k columns. Ask the user for the tp/dp set if it isn't given.

Workflow when asked to write/extend a test

1. Make the edit (smallest change that follows the rules; for layout/equation tweaks, edit the one operand or subscript string in place).
2. python3 -c "import ast; ast.parse(open(path).read())" then a tiny subset run to confirm it executes and err == 0.
3. Run the requested sweep and paste the markdown table verbatim — the table is the deliverable. When the user asks for "the result", give the raw df.to_markdown block, not a re-summary.
4. Keep import <module> side-effect-free (the __main__ guard) so the user can compose tests.

Anti-patterns (all previously rejected)

• ❌ Custom behavior-toggle flags (--dsv4, --only-*) — every flag is a swept list.
• ❌ Hand-written ratio columns (triton/ck) — table holds raw us/TFLOPS/TB-s/err per candidate; compute ratios outside if needed.
• ❌ Reporting only us without TFLOPS and TB/s.
• ❌ torch reference timed into the table.
• ❌ Module-level argparse/sweep that runs on import.
• ❌ Gating arch by returning from inside @benchmark (emits an args-only NaN row); gate in main() instead.
• ❌ A separate test_*_gfx<arch>.py per arch when one dispatching wrapper covers all — merge into one test.
• ❌ Force-merging test fns with different arg signatures into one table (NaN scatter).
• ❌ Idealized contiguous inputs when the model uses a transposed view.
• ❌ Leaving a wrong-result kernel in the table instead of skipping its config.
• ❌ Hardcoding one arch / no get_gfx() gate — crashes on other supported cards.
• ❌ Deny-list arch checks that let an unknown new card run an unbuilt kernel.
• ❌ Deep manual nested loops for the sweep instead of itertools.product.
• ❌ No final summary table.