kernbench2/tests/test_runtime_api_tensor.py

"""Tests for the pytorch-compat Tensor API extensions.

Covers the new ``torch.from_numpy`` factory and ``Tensor.numpy``,
``Tensor.copy_`` methods used by the unified ``ccl_allreduce`` bench.
"""
from __future__ import annotations

import numpy as np
import pytest

from kernbench.policy.placement.dp import DPPolicy
from kernbench.runtime_api.bench_runner import run_bench
from kernbench.runtime_api.types import resolve_device
from kernbench.sim_engine.engine import GraphEngine
from kernbench.topology.builder import resolve_topology


def _engine_factory(topology, device):
    return GraphEngine(getattr(topology, "topology_obj", topology), enable_data=True)


def _run_with(bench_body):
    topo = resolve_topology("topology.yaml")
    return run_bench(
        topology=topo,
        bench_fn=bench_body,
        device=resolve_device("all"),
        engine_factory=_engine_factory,
    )


# ── from_numpy ──────────────────────────────────────────────────────


def test_from_numpy_creates_host_tensor():
    """torch.from_numpy returns a kernbench Tensor with the array stored
    in its host buffer (not deployed to any PE)."""

    def body(torch):
        arr = np.arange(8, dtype=np.float16).reshape(1, 8)
        h = torch.from_numpy(arr)
        # Host tensor has shape/dtype matching the array.
        assert h.shape == (1, 8)
        assert h.dtype == "f16"
        # numpy() round-trips the host buffer.
        assert np.array_equal(h.numpy(), arr)
        # No deploy → no real shards.
        assert h._handle is None
        # Submit a no-op so run_bench has at least one handle.
        torch.zeros((1, 8), dtype="f16",
                    dp=DPPolicy(sip="replicate", cube="replicate", pe="replicate",
                                num_sips=1, num_cubes=1, num_pes=1),
                    name="dummy")

    _run_with(body)


# ── single-PE replicated tensor ─────────────────────────────────────


def test_copy_and_numpy_single_pe():
    """copy_ from a numpy array, then numpy() round-trips correctly on
    a single-PE (no real sharding) tensor."""

    def body(torch):
        dp = DPPolicy(sip="replicate", cube="replicate", pe="replicate",
                      num_sips=1, num_cubes=1, num_pes=1)
        t = torch.zeros((1, 16), dtype="f16", dp=dp, name="t")
        src = np.arange(16, dtype=np.float16).reshape(1, 16)
        t.copy_(torch.from_numpy(src))
        gathered = t.numpy()
        assert gathered.shape == (1, 16)
        assert np.array_equal(gathered, src)

    _run_with(body)


# ── multi-PE column-wise sharding (1 cube) ──────────────────────────


def test_copy_and_numpy_multi_pe_column_wise():
    """copy_ splits across 8 PEs in one cube, numpy() reassembles."""

    def body(torch):
        n_pe = 8
        dp = DPPolicy(sip="replicate", cube="replicate", pe="column_wise",
                      num_sips=1, num_cubes=1, num_pes=n_pe)
        t = torch.zeros((1, n_pe * 4), dtype="f16", dp=dp, name="t")
        src = np.arange(n_pe * 4, dtype=np.float16).reshape(1, n_pe * 4)
        t.copy_(torch.from_numpy(src))
        gathered = t.numpy()
        assert gathered.shape == (1, n_pe * 4)
        assert np.array_equal(gathered, src)
        # Sanity: there really were 8 shards.
        assert len(t._handle.shards) == n_pe

    _run_with(body)


# ── multi-cube sharding ─────────────────────────────────────────────


def test_copy_and_numpy_multi_cube():
    """copy_ across 2 cubes (16 PEs total), numpy() reassembles."""

    def body(torch):
        n_pe_per_cube = 8
        n_cubes = 2
        total = n_cubes * n_pe_per_cube  # 16
        dp = DPPolicy(sip="replicate", cube="column_wise", pe="column_wise",
                      num_sips=1, num_cubes=n_cubes)
        t = torch.zeros((1, total * 4), dtype="f16", dp=dp, name="t")
        src = np.arange(total * 4, dtype=np.float16).reshape(1, total * 4)
        t.copy_(torch.from_numpy(src))
        gathered = t.numpy()
        assert np.array_equal(gathered, src)
        assert len(t._handle.shards) == total

    _run_with(body)


# ── shape mismatch raises ───────────────────────────────────────────


def test_copy_shape_mismatch_raises():
    """copy_ with mismatched shapes raises ValueError."""

    def body(torch):
        dp = DPPolicy(sip="replicate", cube="replicate", pe="replicate",
                      num_sips=1, num_cubes=1, num_pes=1)
        t = torch.zeros((1, 8), dtype="f16", dp=dp, name="t")
        src = np.zeros((1, 16), dtype=np.float16)
        with pytest.raises(ValueError, match="copy_ shape mismatch"):
            t.copy_(torch.from_numpy(src))

    _run_with(body)


# ── __setitem__ / __getitem__ (shard-aligned) ───────────────────────


def test_setitem_getitem_single_pe():
    """Scalar and slice assignment on a single-PE tensor round-trips."""

    def body(torch):
        dp = DPPolicy(sip="replicate", cube="replicate", pe="replicate",
                      num_sips=1, num_cubes=1, num_pes=1)
        t = torch.zeros((1, 8), dtype="f16", dp=dp, name="t")

        # Scalar broadcast
        t[0, 0:4] = 3.0
        assert np.allclose(t[0, 0:4], 3.0)
        assert np.allclose(t[0, 4:8], 0.0)  # untouched

        # Array assignment
        t[0, 4:8] = np.array([10, 20, 30, 40], dtype=np.float16)
        assert np.array_equal(t[0, 4:8], [10, 20, 30, 40])

        # Full read-back via numpy()
        expected = np.array([[3, 3, 3, 3, 10, 20, 30, 40]], dtype=np.float16)
        assert np.array_equal(t.numpy(), expected)

    _run_with(body)


def test_setitem_getitem_multi_pe_shard_aligned():
    """Shard-aligned slice assignment on an 8-PE column-wise tensor."""

    def body(torch):
        n_pe = 8
        n_elem = 4  # per shard
        dp = DPPolicy(sip="replicate", cube="replicate", pe="column_wise",
                      num_sips=1, num_cubes=1, num_pes=n_pe)
        t = torch.zeros((1, n_pe * n_elem), dtype="f16", dp=dp, name="t")

        # Write each shard with its rank value
        for r in range(n_pe):
            t[0, r * n_elem : (r + 1) * n_elem] = float(r + 1)

        # Read back each shard
        for r in range(n_pe):
            expected = float(r + 1)
            arr = t[0, r * n_elem : (r + 1) * n_elem]
            assert np.allclose(arr, expected), f"shard {r}: {arr} != {expected}"

        # Full gather
        full = t.numpy().reshape(-1)
        for r in range(n_pe):
            assert np.allclose(full[r * n_elem : (r + 1) * n_elem], float(r + 1))

    _run_with(body)


def test_setitem_cross_shard_raises():
    """Slice spanning two shards raises NotImplementedError."""

    def body(torch):
        n_pe = 4
        n_elem = 4
        dp = DPPolicy(sip="replicate", cube="replicate", pe="column_wise",
                      num_sips=1, num_cubes=1, num_pes=n_pe)
        t = torch.zeros((1, n_pe * n_elem), dtype="f16", dp=dp, name="t")
        with pytest.raises(NotImplementedError, match="spans multiple shards"):
            t[0, 2:6] = 1.0  # crosses shard 0 (0:4) and shard 1 (4:8)

    _run_with(body)