File size: 7,360 Bytes
33d23e1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 | # SPDX-FileCopyrightText: © 2025 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import pytest
import torch
import ttnn
from ...layers.feedforward import FeedForward, ParallelFeedForward
from ...parallel.manager import CCLManager
from ...utils.check import assert_quality
from ...utils.tensor import bf16_tensor
from ...utils.test import mesh_device_config_to_string
class TorchFeedForward(torch.nn.Module):
def __init__(self, dim, dim_out, bias=True, activation_fn=None, inner_dim=None):
super().__init__()
self.dim = dim
self.dim_out = dim_out
self.bias = bias
self.activation_fn = activation_fn
if activation_fn == "swiglu":
ff1_inner_dim = inner_dim * 2
else:
ff1_inner_dim = inner_dim
self.ff1 = torch.nn.Linear(dim, ff1_inner_dim, bias=bias)
self.ff2 = torch.nn.Linear(inner_dim, dim_out, bias=bias)
def forward(self, x):
x = self.ff1(x)
if self.activation_fn == "gelu":
x = torch.nn.functional.gelu(x)
elif self.activation_fn == "swiglu":
x, gate = torch.chunk(x, 2, -1)
x = x * torch.nn.functional.silu(gate)
return self.ff2(x)
@pytest.mark.parametrize(
"mesh_device",
[(1, 1), (1, 2), (2, 1)],
indirect=True,
)
@pytest.mark.parametrize(
("B, seq, dim, inner_dim, dim_out"),
[
(1, 1, 256, 2432, 2432), # SD3.5 timestep_embedder
(1, 333, 2048, 2432, 2432), # SD3.5 text_embedder
(1, 4096, 2432, 9728, 2432), # SD3.5 spatial FF
(1, 333, 2432, 9728, 2432), # SD3.5 text FF
],
)
@pytest.mark.parametrize("activation_fn", [None, "gelu"])
@pytest.mark.parametrize(
("bias"),
[
True,
# False,
],
)
def test_feedforward(
mesh_device: ttnn.MeshDevice,
B: int,
seq: int,
dim: int,
inner_dim: int,
dim_out: int,
bias: bool,
activation_fn: str,
) -> None:
torch_dtype = torch.bfloat16
torch_model = TorchFeedForward(dim, dim_out, bias=bias, activation_fn=activation_fn, inner_dim=inner_dim).to(
dtype=torch_dtype
)
torch_model.eval()
tt_model = FeedForward(
dim, dim_out, inner_dim=inner_dim, bias=bias, activation_fn=activation_fn, mesh_device=mesh_device
)
tt_model.load_torch_state_dict(torch_model.state_dict())
torch_input_tensor = torch.randn((1, B, seq, dim), dtype=torch_dtype)
tt_input_tensor = bf16_tensor(torch_input_tensor, device=mesh_device)
with torch.no_grad():
torch_output = torch_model(torch_input_tensor)
tt_output = tt_model(tt_input_tensor)
for t in ttnn.get_device_tensors(tt_output):
t = ttnn.to_torch(t)
assert_quality(torch_output, t, pcc=0.999_400)
@pytest.mark.parametrize(
"mesh_device",
[(1, 1)],
indirect=True,
)
@pytest.mark.parametrize(
("B, seq, dim, inner_dim, dim_out, bias"),
[
(1, 256, 2048, 4096, 2048, False), # square-ish swiglu FF, no bias
(1, 512, 1536, 4096, 1536, True), # swiglu FF with bias
(1, 128, 3072, 8192, 3072, False), # wide swiglu FF
],
ids=["swiglu_2048_4096", "swiglu_bias_1536_4096", "swiglu_3072_8192"],
)
def test_feedforward_swiglu_fused(
mesh_device: ttnn.MeshDevice,
B: int,
seq: int,
dim: int,
inner_dim: int,
dim_out: int,
bias: bool,
) -> None:
"""FUSE_SWIGLU minimal_matmul path: silu(gate)*up fused into the matmul output stage.
Compares the fused FeedForward against the torch golden (a*silu(b)).
"""
torch_dtype = torch.bfloat16
torch_model = TorchFeedForward(dim, dim_out, bias=bias, activation_fn="swiglu", inner_dim=inner_dim).to(
dtype=torch_dtype
)
torch_model.eval()
tt_model = FeedForward(
dim,
dim_out,
inner_dim=inner_dim,
bias=bias,
activation_fn="swiglu",
mesh_device=mesh_device,
)
tt_model.load_torch_state_dict(torch_model.state_dict())
torch_input_tensor = torch.randn((1, B, seq, dim), dtype=torch_dtype)
tt_input_tensor = bf16_tensor(torch_input_tensor, device=mesh_device)
with torch.no_grad():
torch_output = torch_model(torch_input_tensor)
tt_output = tt_model(tt_input_tensor)
for t in ttnn.get_device_tensors(tt_output):
t = ttnn.to_torch(t)
assert_quality(torch_output, t, pcc=0.999_900)
@pytest.mark.parametrize(
"mesh_device, device_params",
[
(
shape,
{
"fabric_config": None if shape == (1, 1) else ttnn.FabricConfig.FABRIC_1D,
"require_exact_physical_num_devices": False if shape == (1, 1) else True,
},
)
for shape in [(1, 1), (1, 2), (2, 1), (2, 2), (2, 4), (4, 2)]
],
ids=mesh_device_config_to_string,
indirect=True,
)
@pytest.mark.parametrize(
("tp_mesh_axis"),
[
0,
1,
],
)
@pytest.mark.parametrize("is_fsdp", [True, False], ids=["yes_fsdp", "no_fsdp"])
@pytest.mark.parametrize(
("B, seq, dim, inner_dim, dim_out, bias, activation_fn"),
[
(1, 1, 256, 2432, 2432, True, "gelu"), # SD3.5 timestep_embedder
(1, 333, 2048, 2432, 2432, True, "gelu"), # SD3.5 text_embedder
(1, 4096, 2432, 9728, 2432, True, "gelu"), # SD3.5 spatial FF
(1, 333, 2432, 9728, 2432, True, "gelu"), # SD3.5 text FF
(1, 44520, 3072, 8192, 3072, False, "swiglu"), # Mochi spatial FF
(1, 118, 1536, 4096, 1536, False, "swiglu"), # Mochi prompt FF
],
ids=[
"sd35_timestep_embedder",
"sd35_text_embedder",
"sd35_spatial_ff",
"sd35_text_ff",
"mochi_spatial_ff",
"mochi_prompt_ff",
],
)
def test_parallel_feedforward(
mesh_device: ttnn.MeshDevice,
B: int,
seq: int,
dim: int,
inner_dim: int,
dim_out: int,
bias: bool,
activation_fn: str,
tp_mesh_axis: int,
is_fsdp: bool,
) -> None:
torch_dtype = torch.bfloat16
torch_model = TorchFeedForward(dim, dim_out, bias=bias, activation_fn=activation_fn, inner_dim=inner_dim).to(
dtype=torch_dtype
)
torch_model.eval()
fsdp_mesh_axis = 1 - tp_mesh_axis if is_fsdp else None
ccl_manager = CCLManager(mesh_device, topology=ttnn.Topology.Linear)
tt_model = ParallelFeedForward(
dim,
dim_out,
inner_dim=inner_dim,
bias=bias,
activation_fn=activation_fn,
mesh_device=mesh_device,
mesh_axis=tp_mesh_axis,
fsdp_mesh_axis=fsdp_mesh_axis,
ccl_manager=ccl_manager,
)
tt_model.load_torch_state_dict(torch_model.state_dict())
torch_input_tensor = torch.randn((1, B, seq, dim), dtype=torch_dtype)
tt_input_tensor = bf16_tensor(torch_input_tensor, device=mesh_device)
with torch.no_grad():
torch_output = torch_model(torch_input_tensor)
tt_output = tt_model(tt_input_tensor)
shard_dims = [None, None]
shard_dims[tp_mesh_axis] = -1
shard_dims[1 - tp_mesh_axis] = 0
tt_output = ttnn.to_torch(
tt_output,
mesh_composer=ttnn.ConcatMesh2dToTensor(mesh_device, dims=shard_dims, mesh_shape=tuple(mesh_device.shape)),
)
for i in range(tt_output.shape[0]):
assert_quality(torch_output.squeeze(), tt_output[i].squeeze(), pcc=0.999_400)
|