File size: 10,340 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 | # SPDX-FileCopyrightText: © 2025 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import pytest
import torch
# This is the Conv3d used by the mochi models in huggingface's diffusers module.
from diffusers.models.autoencoders.autoencoder_kl_cogvideox import CogVideoXCausalConv3d as RefContextParallelConv3d
from loguru import logger
import ttnn
from ...layers.conv3d import ContextParallelConv3d as TtContextParallelConv3d
from ...parallel.config import MochiVAEParallelConfig, ParallelFactor
from ...parallel.manager import CCLManager
from ...utils.check import assert_quality
from ...utils.tensor import print_tensor_mem_info
# Common test configurations
PCC_REQUIRED = 0.999
conv3d_args = {
"context_parallel": True,
"causal": True,
"padding_mode": "replicate",
"bias": True,
"groups": 1,
}
def create_random_models(
mesh_device,
**model_args,
):
"""Initialize both reference and TT models."""
# CogVideoXCausalConv3d takes a narrower kwarg set than genmo's class: no
# `causal`/`context_parallel`/`bias`/`groups` knobs (causality is always on;
# bias=True and groups=1 are hardcoded). Rename `padding_mode` -> `pad_mode`.
ref_args = {
"in_channels": model_args["in_channels"],
"out_channels": model_args["out_channels"],
"kernel_size": model_args["kernel_size"],
"stride": model_args["stride"],
"pad_mode": model_args["padding_mode"],
}
reference_model = RefContextParallelConv3d(**ref_args)
# Create TT model
tt_model = TtContextParallelConv3d(mesh_device=mesh_device, **model_args)
# Diffusers nests the conv at `self.conv`; flatten to TT's `weight`/`bias`.
ref_state = {k.removeprefix("conv."): v for k, v in reference_model.state_dict().items()}
tt_model.load_torch_state_dict(ref_state)
return reference_model, tt_model
def validate_outputs(tt_output, ref_output, test_name):
"""Validate and compare model outputs."""
pcc, mse, mae = compute_metrics(ref_output, tt_output)
logger.info(f"Output - PCC: {pcc}, MSE: {mse}, MAE: {mae}")
passing = pcc >= PCC_REQUIRED
if passing:
logger.info(f"{test_name} Passed!")
else:
logger.warning(f"{test_name} Failed!")
logger.error(f"Output failed with PCC: {pcc}, MSE: {mse}, MAE: {mae}")
assert passing, f"{test_name} output does not meet PCC requirement {PCC_REQUIRED}"
# TODO: Merge this with the conv3d tests in test_vae_mochi.py
@torch.no_grad()
@pytest.mark.parametrize(
"input_shape, out_channels, kernel_size, stride",
[
[(1, 768, 28, 60, 106), 768, (3, 3, 3), (1, 1, 1)],
# [(1, 512, 82, 120, 212), 512, (3, 3, 3), (1, 1, 1)],
],
ids=["768"], # , "512"],
)
@pytest.mark.parametrize(
"mesh_device",
[(1, 1)],
indirect=True,
)
def test_context_parallel_conv3d_forward_noshard(
mesh_device, input_shape, out_channels, kernel_size, stride, reset_seeds
):
input_channels = input_shape[1]
model_args = conv3d_args.copy()
model_args.update(
{
"in_channels": input_channels,
"out_channels": out_channels,
"kernel_size": kernel_size,
"stride": stride,
"parallel_config": MochiVAEParallelConfig(
time_parallel=ParallelFactor(factor=1, mesh_axis=1),
h_parallel=ParallelFactor(factor=1, mesh_axis=0),
w_parallel=ParallelFactor(factor=1, mesh_axis=1),
),
"ccl_manager": CCLManager(
mesh_device=mesh_device,
topology=ttnn.Topology.Linear,
),
}
)
reference_model, tt_model = create_random_models(mesh_device, **model_args)
# Create input tensor (NCTHW format for PyTorch)
torch_input = torch.randn(input_shape)
# Convert to NTHWC format for TT
tt_input_NTHWC = torch_input.permute(0, 2, 3, 4, 1)
logger.info(f"Creating tensor: {tt_input_NTHWC.shape}")
tt_input_NTHWC = ttnn.from_torch(
tt_input_NTHWC,
device=mesh_device,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
mesh_mapper=None,
)
tt_input_NTHWC = ttnn.to_layout(tt_input_NTHWC, ttnn.ROW_MAJOR_LAYOUT)
logger.info("Run TtContextParallelConv3d forward")
tt_output = tt_model(tt_input_NTHWC)
# Convert TT output to torch tensor (from NTHWC to NCHW format)
tt_output_torch = ttnn.to_torch(ttnn.get_device_tensors(tt_output)[0])
tt_output_torch = tt_output_torch.permute(0, 4, 1, 2, 3) # NTHWC -> NCHW
# Get reference output
logger.info("Run reference model forward")
with torch.no_grad():
ref_output, _ = reference_model(torch_input) # returns (output, new_conv_cache)
# Compare shapes
logger.info(f"TT output shape: {tt_output_torch.shape}, Ref output shape: {ref_output.shape}")
assert tt_output_torch.shape == ref_output.shape, "Output shapes don't match"
assert_quality(ref_output, tt_output_torch, pcc=PCC_REQUIRED)
# Original test iterated over the following arches/device combos. Unclear how to
# iterate in this exact fashion within the current test framework.
# {
# "N150": (1, 1),
# "N300": (1, 2),
# "T3K": (1, 1),
# "TG": (8, 4)
# }
@torch.no_grad()
@pytest.mark.parametrize(
"input_shape, out_channels, kernel_size, stride",
[
[(1, 768, 28, 60, 106), 768, (3, 3, 3), (1, 1, 1)],
[(1, 512, 82, 120, 212), 512, (3, 3, 3), (1, 1, 1)],
[(1, 256, 163, 240, 424), 256, (3, 3, 3), (1, 1, 1)],
[(1, 128, 163, 480, 848), 128, (3, 3, 3), (1, 1, 1)],
],
ids=["768", "512", "256", "128"],
)
@pytest.mark.parametrize(
"mesh_device",
[(1, 1)], # TODO: make this the full suite of meshes want to test.
indirect=True,
)
@pytest.mark.skip(
reason=(
"This test is currently busted. "
"Some tt tensors hang on calls to from_torch,"
"The sharding is also not quite right as written. "
"Fix this test when we bring up the mochi model tests, as "
"test_vae_mochi.py has a lot of overlap and we can consolidate them."
)
)
def test_context_parallel_conv3d_forward(mesh_device, input_shape, out_channels, kernel_size, stride, reset_seeds):
"""Test complete forward pass of TtContextParallelConv3d."""
input_channels = input_shape[1]
model_args = conv3d_args.copy()
model_args.update(
{
"in_channels": input_channels,
"out_channels": out_channels,
"kernel_size": kernel_size,
"stride": stride,
# TODO: see test_vae_mochi.py for how to call these with various mesh device shapes
# "parallel_config": MochiVAEParallelConfig(
# time_parallel=ParallelFactor(factor=4, mesh_axis=1),
# h_parallel=ParallelFactor(factor=1, mesh_axis=0),
# w_parallel=ParallelFactor(factor=1, mesh_axis=1),
# ),
# "ccl_manager": CCLManager(
# mesh_device=mesh_device,
# topology=ttnn.Topology.Linear,
# ),
}
)
# Create the models
reference_model, tt_model = create_random_models(mesh_device, **model_args)
# Create input tensor (NCTHW format for PyTorch)
torch_input = torch.randn(input_shape)
# Convert to NTHWC format for TT
tt_input_NTHWC = torch_input.permute(0, 2, 3, 4, 1)
# TODO: see test_vae_mochi.py for how to set this up.
dims = [None, 1]
mesh_mapper = ttnn.create_mesh_mapper(
mesh_device,
ttnn.MeshMapperConfig(
[
ttnn.PlacementReplicate() if dims[0] is None else ttnn.PlacementShard(dims[0]),
ttnn.PlacementReplicate() if dims[1] is None else ttnn.PlacementShard(dims[1]),
],
ttnn.MeshShape(mesh_device.shape[0], mesh_device.shape[1]),
),
)
tt_input_NTHWC = ttnn.from_torch(
tt_input_NTHWC,
device=mesh_device,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
mesh_mapper=mesh_mapper,
)
tt_input_NTHWC = ttnn.to_layout(tt_input_NTHWC, ttnn.ROW_MAJOR_LAYOUT)
logger.info("Run TtContextParallelConv3d forward")
tt_output = tt_model(tt_input_NTHWC)
# Convert TT output to torch tensor (from NTHWC to NCHW format)
tt_output_torch = ttnn.to_torch(ttnn.get_device_tensors(tt_output)[0])
tt_output_torch = tt_output_torch.permute(0, 4, 1, 2, 3) # NTHWC -> NCHW
# Get reference output
logger.info("Run reference model forward")
with torch.no_grad():
ref_output, _ = reference_model(torch_input) # returns (output, new_conv_cache)
# Compare shapes
logger.info(f"TT output shape: {tt_output_torch.shape}, Ref output shape: {ref_output.shape}")
assert tt_output_torch.shape == ref_output.shape, "Output shapes don't match"
assert_quality(ref_output, tt_output_torch, pcc=PCC_REQUIRED)
@torch.no_grad()
@pytest.mark.parametrize(
"input_shape, out_channels, kernel_size, stride",
[
[(1, 512, 82, 120, 212), 512, (3, 3, 3), (1, 1, 1)],
],
ids=["512"],
)
@pytest.mark.parametrize(
"mesh_device",
[(1, 1)],
indirect=True,
)
@pytest.mark.skip(reason="This allocation hangs on blackhole. TODO Debug and fix in the future.")
def test_sharding(mesh_device, input_shape, out_channels, kernel_size, stride):
# Create input tensor (NCTHW format for PyTorch)
torch_input = torch.randn(input_shape)
logger.info(f"initial shape {torch_input.shape}")
# Convert to NTHWC format for TT
tt_input_NTHWC = torch_input.permute(0, 2, 3, 4, 1).contiguous()
logger.info(f"initial reshape {tt_input_NTHWC.shape}")
logger.info(f"contiguous? : {tt_input_NTHWC.is_contiguous()}")
tt_input_NTHWC = ttnn.from_torch(
tt_input_NTHWC,
device=mesh_device,
dtype=ttnn.bfloat16,
layout=ttnn.ROW_MAJOR_LAYOUT,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
mesh_mapper=None,
)
logger.info("Create DONE!")
print_tensor_mem_info(tt_input_NTHWC)
logger.info("reorder")
tt_input_NTHWC_rm = ttnn.to_layout(tt_input_NTHWC, ttnn.ROW_MAJOR_LAYOUT)
print_tensor_mem_info(tt_input_NTHWC_rm)
|