File size: 3,664 Bytes
b025706 | 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 | # SPDX-FileCopyrightText: © 2023 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import torch
import ttnn
from models.common.lightweightmodule import LightweightModule
from models.common.utility_functions import nearest_32
class TtLlamaConv2dPatch(LightweightModule):
"""Conv2D Patching layer.
Column parallel over unfolded input.
Arguments:
in_channels: Input channels.
out_channels: Output channels.
kernel_size: Size of convolution kernel.
stride (default 1): Stride for convolution.
bias (default False): Use bias in Conv2d.
Input: (bsz, in_channels, width, height)
Output: (bsz, num_tokens, out_channels)
"""
def __init__(
self,
mesh_device,
state_dict,
state_dict_prefix,
dtype,
in_channels: int,
out_channels: int,
kernel_size: int,
stride: int,
bias,
):
super().__init__()
self.mesh_device = mesh_device
self.num_devices = self.mesh_device.get_num_devices()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.bias = (
ttnn.as_tensor(
torch.reshape(state_dict[f"{state_dict_prefix}_linear.bias"], (1, -1)),
dtype=dtype,
layout=ttnn.TILE_LAYOUT,
device=self.mesh_device,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
mesh_mapper=ttnn.ReplicateTensorToMesh(self.mesh_device),
)
if bias
else None
)
self._unfold = torch.nn.Unfold(kernel_size=self.kernel_size, stride=self.stride)
weight = state_dict[f"{state_dict_prefix}_linear.weight"]
pad_len = nearest_32(weight.shape[-1]) - weight.shape[-1]
padding = torch.zeros(self.out_channels, pad_len, dtype=weight.dtype)
padded_weight = torch.cat([weight, padding], dim=-1)
padded_weight = padded_weight.permute(1, 0).reshape(1, 1, -1, self.out_channels)
self._linear_weight = ttnn.as_tensor(
padded_weight,
dtype=dtype,
layout=ttnn.TILE_LAYOUT,
device=self.mesh_device,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
mesh_mapper=ttnn.ReplicateTensorToMesh(self.mesh_device),
)
self.compute_kernel_config = ttnn.init_device_compute_kernel_config(
mesh_device.arch(),
math_fidelity=ttnn.MathFidelity.HiFi2,
math_approx_mode=True,
fp32_dest_acc_en=True,
packer_l1_acc=True,
)
def forward(self, x: torch.Tensor):
x = self._unfold(x)
x = x.permute(0, 2, 1)
# Need to pad the last dimension of x to be a multiple of a tile
pad_len = nearest_32(x.shape[-1]) - x.shape[-1]
padding = torch.zeros((x.shape[0], x.shape[1], pad_len), dtype=x.dtype, device=x.device)
x = torch.cat([x, padding], dim=-1)
x = ttnn.as_tensor(
x,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
device=self.mesh_device,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
mesh_mapper=ttnn.ReplicateTensorToMesh(self.mesh_device),
)
out = ttnn.linear(
x,
self._linear_weight,
bias=self.bias,
dtype=ttnn.bfloat16,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
compute_kernel_config=self.compute_kernel_config,
core_grid=ttnn.CoreGrid(y=8, x=8),
)
return out
|