kernbench2/benches/ccl_allreduce.py

"""CCL all-reduce bench — single unified entry point.

Driven entirely by ``ccl.yaml`` + ``topology.yaml``:

- ``defaults.algorithm`` in ``ccl.yaml`` picks which kernel to run
  (``ring_allreduce_{tcm,hbm,sram}`` / ``mesh_allreduce_4`` /
  ``tree_allreduce_7``).
- ``world_size`` is derived from the algorithm entry's override or from
  the topology spec (``sips × cubes_per_sip × pes_per_cube``).
- The host code uses only real PyTorch ``torch.distributed`` names:
  ``init_process_group``, ``get_world_size``, ``get_rank``, ``all_reduce``.

The bench is split into ``worker(rank, world_size, torch)`` — the
per-rank business logic, designed to look like a real PyTorch DDP
training worker so future model benches can reuse the same skeleton —
and ``run(torch)`` — the kernbench-specific launcher that initializes
the process group and invokes the worker.
"""
from __future__ import annotations

import numpy as np

from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
from kernbench.policy.placement.dp import DPPolicy

# Default per-rank tile size if ccl.yaml doesn't override it. Real
# pytorch benches hardcode batch/feature dims similarly.
DEFAULT_N_ELEM = 32


def _derive_dp(spec: dict, world_size: int) -> DPPolicy:
    """Pick a DPPolicy that fans the tensor across exactly ``world_size`` PEs.

    Mirrors what a real PyTorch DDP user does manually with
    ``tensor.to(f"cuda:{rank}")``: the host code chooses the placement so
    that the collective sees the right number of participating ranks.
    """
    sips = int(spec["system"]["sips"]["count"])
    cm = spec["sip"]["cube_mesh"]
    pl = spec["cube"]["pe_layout"]
    pes_per_cube = int(pl["pe_per_corner"]) * len(pl["corners"])
    cubes_per_sip = int(cm["w"]) * int(cm["h"])
    total = sips * cubes_per_sip * pes_per_cube
    if world_size == total:
        return DPPolicy(sip="column_wise", cube="column_wise", pe="column_wise")
    if world_size <= pes_per_cube:
        return DPPolicy(
            sip="replicate", cube="replicate", pe="column_wise",
            num_sips=1, num_cubes=1, num_pes=world_size,
        )
    if world_size <= cubes_per_sip * pes_per_cube:
        return DPPolicy(
            sip="replicate", cube="column_wise", pe="column_wise",
            num_sips=1, num_cubes=world_size // pes_per_cube,
        )
    return DPPolicy(sip="column_wise", cube="column_wise", pe="column_wise")


def worker(rank: int, world_size: int, torch) -> None:
    """Per-rank business logic. Mirrors a real PyTorch DDP worker.

    In real PyTorch DDP, this function runs in N separate processes,
    each with its own ``rank``. In kernbench (single-process multi-device)
    it is invoked once with ``rank=0`` on the single host driver; the
    actual per-PE parallelism is handled by ``torch.launch`` fanning out
    the kernel across all participating PEs via the tensor's DPPolicy.
    The ``rank`` parameter is therefore always 0 today, and is kept as
    an explicit argument for parity with real DDP workers (``if rank ==
    0`` logging guards, future multi-host extensions).
    """
    cfg = resolve_algorithm_config(load_ccl_config())
    algo_name = cfg["algorithm"]
    n_elem = int(cfg.get("n_elem", DEFAULT_N_ELEM))

    # Pick a DP that produces exactly ``world_size`` shards on this topology.
    dp = _derive_dp(torch.spec, world_size)
    tensor = torch.zeros(
        (1, world_size * n_elem), dtype="f16", dp=dp, name="ccl_in",
    )

    # Initialize: CCL rank r's slice gets value (r + 1). Real PyTorch idiom:
    #     target.copy_(torch.from_numpy(source))
    init = np.zeros((1, world_size * n_elem), dtype=np.float16)
    for r in range(world_size):
        init[0, r * n_elem : (r + 1) * n_elem] = float(r + 1)
    tensor.copy_(torch.from_numpy(init))

    # The main act: one all_reduce call — the backend installs IPCQ at
    # init_process_group time and here only dispatches the kernel.
    torch.distributed.all_reduce(tensor, op="sum")

    # Verify: each shard should hold sum(1..world_size) after all-reduce.
    result = tensor.numpy()
    expected = float(sum(range(1, world_size + 1)))
    all_ok = bool(np.allclose(result, expected, rtol=1e-1, atol=1e-1))

    # Print only on rank 0 — real PyTorch DDP idiom for single-source logs.
    if rank == 0:
        if all_ok:
            print(f"  {algo_name} (ws={world_size}): {world_size} OK")
        else:
            flat = result.reshape(-1)
            n_fail = 0
            for r in range(world_size):
                slice_r = flat[r * n_elem : (r + 1) * n_elem]
                if not np.allclose(slice_r, expected, rtol=1e-1, atol=1e-1):
                    n_fail += 1
                    if n_fail <= 5:
                        print(
                            f"  [FAIL] rank {r} "
                            f"(ws={world_size}, algo={algo_name}): "
                            f"got mean={float(slice_r.mean()):.3f}, "
                            f"expected={expected:.3f}"
                        )
            print(
                f"  {algo_name} (ws={world_size}): "
                f"{world_size - n_fail} OK / {n_fail} FAIL"
            )


def run(torch) -> None:
    """CLI entry point: initialize the process group, invoke worker."""
    dist = torch.distributed
    dist.init_process_group(backend="ahbm")
    worker(
        rank=dist.get_rank(),
        world_size=dist.get_world_size(),
        torch=torch,
    )