File size: 5,462 Bytes
d643072 |
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 |
# Copyright 2024 MIT Han Lab
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0
from typing import Optional
import ipdb
import torch
from torch import nn
from torch.nn import functional as F
from .triton_lite_mla_kernels.linear_relu_fwd import linear_relu_fwd
from .triton_lite_mla_kernels.mm import matmul # for autocast
from .triton_lite_mla_kernels.pad_vk_mm_fwd import pad_vk_mm_fwd
from .triton_lite_mla_kernels.proj_divide_bwd import proj_divide_bwd
from .triton_lite_mla_kernels.vk_mm_relu_bwd import vk_mm_relu_bwd
from .triton_lite_mla_kernels.vk_q_mm_divide_fwd import vk_q_mm_divide_fwd
from .triton_lite_mla_kernels.vk_q_mm_relu_bwd import vk_q_mm_relu_bwd
class TritonLiteMLAFunction(torch.autograd.Function):
@staticmethod
def forward(
ctx,
x: torch.Tensor,
qkv_weight: torch.Tensor,
proj_weight: torch.Tensor,
proj_bias: Optional[torch.Tensor],
num_heads: int,
head_dim: int,
eps: float,
) -> torch.Tensor:
ctx.x_dtype, ctx.qkv_weight_dtype, ctx.proj_dtype = x.dtype, qkv_weight.dtype, proj_weight.dtype
if torch.is_autocast_enabled():
autocast_dtype = torch.get_autocast_gpu_dtype()
x = x.to(autocast_dtype)
qkv_weight = qkv_weight.to(autocast_dtype)
proj_weight = proj_weight.to(autocast_dtype)
if proj_bias is not None:
proj_bias = proj_bias.to(autocast_dtype)
B, N, C = x.shape
qkv, relu_mask = linear_relu_fwd(x, qkv_weight) # B, N, 3*C. autocast is processed here
qkv, relu_mask = qkv.view(B, N, 3, C), relu_mask.view(B, N, 3, C)
q, k, v = qkv.unbind(2) # B, N, C
k = k.reshape(B, N, num_heads, head_dim)
v = v.reshape(B, N, num_heads, head_dim)
q = q.reshape(B, N, num_heads, head_dim)
vk = pad_vk_mm_fwd(v, k, torch.float, torch.float)
proj_input, vk_q = vk_q_mm_divide_fwd(vk, q, eps, torch.float, qkv.dtype)
proj_input = proj_input.view(B, N, C)
y = F.linear(proj_input, proj_weight, proj_bias)
if ctx.needs_input_grad[0] or ctx.needs_input_grad[1] or ctx.needs_input_grad[2] or ctx.needs_input_grad[3]:
ctx.save_for_backward(x, qkv_weight, relu_mask, v, k, vk, q, vk_q, proj_input, proj_weight)
ctx.eps = eps
if torch.get_autocast_gpu_dtype() == torch.float16:
y = y.clip(-65504, 65504)
return y
@staticmethod
def backward(ctx, grad_y: torch.Tensor):
x, qkv_weight, relu_mask, v, k, vk, q, vk_q, proj_input, proj_weight = ctx.saved_tensors
B, N, H, C1 = vk_q.shape
C = C1 - 1
# ipdb.set_trace()
grad_proj_weight = (
(grad_y.reshape(-1, H * C).T @ proj_input.view(-1, H * C)).to(ctx.proj_dtype)
if ctx.needs_input_grad[2]
else None
)
grad_proj_bias = grad_y.sum((0, 1)).to(ctx.proj_dtype) if ctx.needs_input_grad[3] else None
#
grad_vk_q = proj_divide_bwd(grad_y, proj_weight, vk_q, ctx.eps)
del grad_y, vk_q
grad_qkv = torch.empty(B, N, 3, H, C, dtype=q.dtype, device=q.device)
grad_vk = vk_q_mm_relu_bwd(grad_vk_q, vk, q, relu_mask[:, :, 0].view(B, N, H, C), grad_qkv[:, :, 0])
del grad_vk_q, vk
vk_mm_relu_bwd(grad_vk, k, v, relu_mask[:, :, 1].view(B, N, H, C), grad_qkv[:, :, 1], grad_qkv[:, :, 2])
del grad_vk, q, k, v, relu_mask
grad_qkv_weight = (
(grad_qkv.view(B * N, 3 * H * C).T @ x.view(B * N, H * C)).to(ctx.qkv_weight_dtype)
if ctx.needs_input_grad[1]
else None
)
grad_x = (grad_qkv.view(B, N, 3 * H * C) @ qkv_weight).to(ctx.x_dtype) if ctx.needs_input_grad[0] else None
del grad_qkv
return grad_x, grad_qkv_weight, grad_proj_weight, grad_proj_bias, None, None, None
class TritonLiteMLA(nn.Module):
def __init__(
self,
dim: int,
num_heads: int,
eps=1e-15,
use_bias=False,
):
super().__init__()
self.dim, self.num_heads, self.head_dim, self.eps = dim, num_heads, dim // num_heads, eps
if use_bias:
raise NotImplementedError(f"use_bias is not supported for TritonLiteMLA")
self.qkv = nn.Linear(dim, dim * 3, bias=use_bias)
self.proj = nn.Linear(dim, dim)
def forward(self, x: torch.Tensor, mask=None, HW=None, block_id=None) -> torch.Tensor:
return TritonLiteMLAFunction.apply(
x, self.qkv.weight, self.proj.weight, self.proj.bias, self.num_heads, self.head_dim, self.eps
)
@property
def module_str(self) -> str:
_str = type(self).__name__ + "("
eps = f"{self.eps:.1E}"
_str += f"i={self.in_dim},o={self.out_dim},h={self.heads},d={self.dim},eps={eps}"
return _str
def __repr__(self):
return f"EPS{self.eps}-" + super().__repr__()
|