File size: 8,589 Bytes
a62b6c9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 | # SPDX-FileCopyrightText: © 2025 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
"""
Tests for distributing a torch.Tensor using the topology of an existing TTNN tensor.
Follows the same style as test_auto_compose.py, but in the opposite
direction: given a reference distributed tensor, verify that
`from_torch_dist_as(torch_tensor, ref_tensor)` produces a TTNN tensor
with the same topology and data that composes back to the original
torch input.
"""
import pytest
import torch
import ttnn
from models.common.auto_compose import extract_tensor_topology_info, to_torch_auto_compose
from models.common.distribute_as import from_torch_dist_as
# ======================================================================================
# Test Parameters
# ======================================================================================
pytestmark = [
pytest.mark.parametrize(
"ttnn_mesh_device",
[
(1, 1), # single device
(1, 2), # 1D mesh, 2 devices
(1, 8), # 1D mesh, 8 devices
(2, 4), # 2D mesh, 8 devices
],
ids=[
"1x1",
"1x2",
"1x8",
"2x4",
],
indirect=True,
),
pytest.mark.parametrize(
"layout,dtype",
[
(ttnn.ROW_MAJOR_LAYOUT, ttnn.bfloat16),
(ttnn.TILE_LAYOUT, ttnn.bfloat16),
(ttnn.TILE_LAYOUT, ttnn.bfloat8_b), # bfloat8_b only works with TILE_LAYOUT
(ttnn.TILE_LAYOUT, ttnn.bfloat4_b), # bfloat4_b only works with TILE_LAYOUT
],
ids=["row_major_bf16", "tile_bf16", "tile_bf8b", "tile_bf4b"],
),
pytest.mark.parametrize(
"memory_config",
[ttnn.DRAM_MEMORY_CONFIG, ttnn.L1_MEMORY_CONFIG],
ids=["dram", "l1"],
),
]
# ======================================================================================
# Helpers
# ======================================================================================
def _make_arange_dtype(
shape: tuple[int, ...], dtype: torch.dtype = torch.bfloat16, min_value: float = 0, max_value: float = 100
) -> torch.Tensor:
"""Create a deterministic tensor with arange data and specified dtype."""
numel = 1
for s in shape:
numel *= s
# Generate values from min_value to max_value with step of 1
values = torch.arange(min_value, max_value + 1, dtype=dtype)
# Randomly sample indices (with replacement) to fill the tensor
indices = torch.randint(0, len(values), size=(numel,))
data = values[indices].reshape(shape)
return data
def _pos_dim(dim: int, rank: int) -> int:
return dim % rank
def _topology_signature(tt: ttnn.Tensor) -> tuple[tuple[str, tuple[int, ...]], tuple[int, ...]]:
placements, dist_shape = extract_tensor_topology_info(tt)
sig_p = []
for p in placements:
if isinstance(p, ttnn.PlacementShard):
sig_p.append(("shard", (p.dim,)))
else:
assert isinstance(p, ttnn.PlacementReplicate)
sig_p.append(("replicate", ()))
return tuple(sig_p), tuple(dist_shape)
# ======================================================================================
# Tests
# ======================================================================================
@pytest.mark.parametrize("storage", ["host", "device"])
def test_sharded_1d(
ttnn_mesh_device: ttnn.MeshDevice, layout, dtype, memory_config: ttnn.MemoryConfig, storage: str
) -> None:
is_host_ref = storage == "host"
num_devices = ttnn_mesh_device.get_num_devices()
# Reference topology: host-sharded along dim 0
shape = (num_devices, 1, 3, 1)
# get range of values for bfloat4_b quantization which has 4 bits for the mantissa and shared 8-bit exponent
max_value = 6 # allows for range of value to grow to 7 for new_input
ref_input = _make_arange_dtype(shape, dtype=torch.float32, min_value=-7, max_value=max_value)
ref_tt = ttnn.from_torch(
ref_input,
device=None if is_host_ref else ttnn_mesh_device,
dtype=dtype,
layout=layout,
memory_config=memory_config,
mesh_mapper=ttnn.ShardTensorToMesh(ttnn_mesh_device, dim=0),
)
# Distribute a different tensor using the same topology
new_input = ref_input + 1 # change values to avoid accidental equality
if is_host_ref:
tt_as = from_torch_dist_as(new_input, ref_tt, device=ttnn_mesh_device)
else:
tt_as = from_torch_dist_as(new_input, ref_tt)
# Verify topology matches reference
expected_topology = _topology_signature(ref_tt)
assert _topology_signature(tt_as) == expected_topology
assert tt_as.dtype == dtype
assert tt_as.memory_config() == memory_config
# Verify round-trip via auto-composition matches the source
if is_host_ref:
torch_auto = to_torch_auto_compose(tt_as, device=ttnn_mesh_device)
else:
torch_auto = to_torch_auto_compose(tt_as)
assert torch.equal(torch_auto, new_input)
@pytest.mark.parametrize("storage", ["host", "device"])
@pytest.mark.parametrize("dim", [0, 1, 2, -1])
def test_sharded_various_dims(
ttnn_mesh_device: ttnn.MeshDevice,
layout,
dtype,
memory_config: ttnn.MemoryConfig,
dim: int,
storage: str,
) -> None:
is_host_ref = storage == "host"
num_devices = ttnn_mesh_device.get_num_devices()
rank = 4
axis = _pos_dim(dim, rank)
shape = [2, 3, 4, 1]
shape[axis] = num_devices
# get range of values for bfloat4_b quantization which has 4 bits for the mantissa and shared 8-bit exponent
max_value = 6 # allows for range of value to grow to 7 for new_input
ref_input = _make_arange_dtype(shape, dtype=torch.float32, min_value=-7, max_value=max_value)
ref_tt = ttnn.from_torch(
ref_input,
device=None if is_host_ref else ttnn_mesh_device,
dtype=dtype,
layout=layout,
memory_config=memory_config,
mesh_mapper=ttnn.ShardTensorToMesh(ttnn_mesh_device, dim=dim),
)
new_input = ref_input + 1 # change values to avoid accidental equality
if is_host_ref:
tt_as = from_torch_dist_as(new_input, ref_tt, device=ttnn_mesh_device)
else:
tt_as = from_torch_dist_as(new_input, ref_tt)
assert _topology_signature(tt_as) == _topology_signature(ref_tt)
assert tt_as.dtype == dtype
assert tt_as.memory_config() == memory_config
if is_host_ref:
torch_auto = to_torch_auto_compose(tt_as, device=ttnn_mesh_device)
else:
torch_auto = to_torch_auto_compose(tt_as)
assert torch.equal(torch_auto, new_input)
@pytest.mark.parametrize("storage", ["host", "device"])
@pytest.mark.parametrize("dims_pair", [(0, 1), (0, -1), (1, -1)])
def test_sharded_2d(
ttnn_mesh_device: ttnn.MeshDevice,
layout,
dtype,
memory_config: ttnn.MemoryConfig,
dims_pair: tuple[int, int],
storage: str,
) -> None:
is_host_ref = storage == "host"
mesh_shape = tuple(ttnn_mesh_device.shape)
if len(mesh_shape) != 2 and torch.prod(torch.tensor(mesh_shape)).item() <= 1:
pytest.skip("Requires a 2D mesh with both dims > 1")
rank = 4
d0 = _pos_dim(dims_pair[0], rank)
d1 = _pos_dim(dims_pair[1], rank)
assert d0 != d1
shape = [2, 3, 4, 5]
shape[d0] = mesh_shape[0]
shape[d1] = mesh_shape[1]
# get range of values for bfloat4_b quantization which has 4 bits for the mantissa and shared 8-bit exponent
max_value = 6 # allows for range of value to grow to 7 for new_input
ref_input = _make_arange_dtype(shape, dtype=torch.float32, min_value=-7, max_value=max_value)
mapper = ttnn.ShardTensor2dMesh(ttnn_mesh_device, mesh_shape=mesh_shape, dims=(dims_pair[0], dims_pair[1]))
ref_tt = ttnn.from_torch(
ref_input,
device=None if is_host_ref else ttnn_mesh_device,
dtype=dtype,
layout=layout,
memory_config=memory_config,
mesh_mapper=mapper,
)
new_input = ref_input + 1 # change values to avoid accidental equality
if is_host_ref:
tt_as = from_torch_dist_as(new_input, ref_tt, device=ttnn_mesh_device)
else:
tt_as = from_torch_dist_as(new_input, ref_tt)
assert _topology_signature(tt_as) == _topology_signature(ref_tt)
assert tt_as.dtype == dtype
assert tt_as.memory_config() == memory_config
if is_host_ref:
torch_auto = to_torch_auto_compose(tt_as, device=ttnn_mesh_device)
else:
torch_auto = to_torch_auto_compose(tt_as)
assert torch.equal(torch_auto, new_input)
|