2023-03-08 08:22:30 +00:00
|
|
|
from typing import Any, Callable, Dict, Iterable, List, Tuple
|
2023-01-06 06:14:45 +00:00
|
|
|
|
|
|
|
import torch
|
|
|
|
|
|
|
|
import colossalai
|
2023-01-31 08:00:06 +00:00
|
|
|
from colossalai.fx._compatibility import is_compatible_with_meta
|
2023-01-10 02:56:39 +00:00
|
|
|
from colossalai.fx.codegen.activation_checkpoint_codegen import CODEGEN_AVAILABLE
|
|
|
|
|
2023-01-31 08:00:06 +00:00
|
|
|
AUTOCHUNK_AVAILABLE = CODEGEN_AVAILABLE and is_compatible_with_meta()
|
|
|
|
|
|
|
|
if AUTOCHUNK_AVAILABLE:
|
2023-09-19 06:20:26 +00:00
|
|
|
from torch.fx.graph import (
|
|
|
|
CodeGen,
|
|
|
|
PythonCode,
|
|
|
|
_custom_builtins,
|
|
|
|
_CustomBuiltin,
|
|
|
|
_format_target,
|
|
|
|
_is_from_torch,
|
|
|
|
_Namespace,
|
|
|
|
_origin_type_map,
|
|
|
|
inplace_methods,
|
|
|
|
magic_methods,
|
|
|
|
)
|
2023-01-10 03:10:45 +00:00
|
|
|
|
|
|
|
from torch.fx.node import Argument, Node, _get_qualified_name, _type_repr, map_arg
|
2023-01-10 02:56:39 +00:00
|
|
|
|
2023-01-06 09:09:37 +00:00
|
|
|
from .search_chunk import SearchChunk
|
2023-02-01 05:18:51 +00:00
|
|
|
from .utils import delete_free_var_from_last_use, get_logger, get_node_name, get_node_shape
|
2023-01-06 06:14:45 +00:00
|
|
|
|
|
|
|
|
2023-01-10 01:59:47 +00:00
|
|
|
def _gen_chunk_slice_dim(chunk_dim: int, chunk_indice_name: str, shape: List) -> str:
|
|
|
|
"""
|
|
|
|
Generate chunk slice string, eg. [:, :, chunk_idx_name:chunk_idx_name + chunk_size, :]
|
|
|
|
|
|
|
|
Args:
|
|
|
|
chunk_dim (int)
|
|
|
|
chunk_indice_name (str): chunk indice name
|
|
|
|
shape (List): node shape
|
|
|
|
|
|
|
|
Returns:
|
|
|
|
new_shape (str): return slice
|
|
|
|
"""
|
2023-01-06 06:14:45 +00:00
|
|
|
new_shape = "["
|
2023-01-10 01:59:47 +00:00
|
|
|
for idx, _ in enumerate(shape):
|
2023-01-06 06:14:45 +00:00
|
|
|
if idx == chunk_dim:
|
2023-01-10 01:59:47 +00:00
|
|
|
new_shape += "%s:%s + chunk_size" % (chunk_indice_name, chunk_indice_name)
|
2023-01-06 06:14:45 +00:00
|
|
|
else:
|
|
|
|
new_shape += ":"
|
|
|
|
new_shape += ", "
|
|
|
|
new_shape = new_shape[:-2] + "]"
|
|
|
|
return new_shape
|
|
|
|
|
|
|
|
|
2023-05-15 03:46:25 +00:00
|
|
|
def _gen_loop_start(chunk_input: List[Node], chunk_output: List[Node], chunk_output_dim: int, chunk_size=2) -> str:
|
2023-01-10 01:59:47 +00:00
|
|
|
"""
|
|
|
|
Generate chunk loop start
|
|
|
|
|
|
|
|
eg. chunk_result = torch.empty([100, 100], dtype=input_node.dtype, device=input_node.device)
|
|
|
|
chunk_size = 32
|
|
|
|
for chunk_idx in range(0, 100, 32):
|
|
|
|
......
|
|
|
|
|
|
|
|
Args:
|
|
|
|
chunk_input (List[Node]): chunk input node
|
|
|
|
chunk_output (Node): chunk output node
|
2023-05-15 03:46:25 +00:00
|
|
|
chunk_output_dim (int): chunk output node chunk dim
|
2023-01-10 01:59:47 +00:00
|
|
|
chunk_size (int): chunk size. Defaults to 2.
|
|
|
|
|
|
|
|
Returns:
|
|
|
|
context (str): generated str
|
|
|
|
"""
|
2023-01-06 06:14:45 +00:00
|
|
|
input_node = chunk_input[0]
|
2023-02-01 05:18:51 +00:00
|
|
|
|
|
|
|
context = ""
|
|
|
|
for i in range(len(chunk_output)):
|
|
|
|
shape_str = str(list(get_node_shape(chunk_output[i])))
|
2023-03-10 02:23:26 +00:00
|
|
|
if get_node_name(chunk_output[i]) in ["split", "unbind"]:
|
2023-09-19 06:20:26 +00:00
|
|
|
tensor_str = "torch.empty(%s, dtype=%s.dtype, device=%s.device), " % (
|
|
|
|
shape_str,
|
|
|
|
input_node.name,
|
|
|
|
input_node.name,
|
|
|
|
)
|
|
|
|
tensor_str = tensor_str * len(chunk_output[i].meta["tensor_meta"])
|
2023-02-01 05:18:51 +00:00
|
|
|
tensor_str = "[" + tensor_str[:-2] + "]"
|
|
|
|
context += "%s = %s; " % (chunk_output[i].name, tensor_str)
|
|
|
|
else:
|
2023-09-19 06:20:26 +00:00
|
|
|
context += "%s = torch.empty(%s, dtype=%s.dtype, device=%s.device); " % (
|
|
|
|
chunk_output[i].name,
|
|
|
|
shape_str,
|
|
|
|
input_node.name,
|
|
|
|
input_node.name,
|
|
|
|
)
|
2023-02-01 05:18:51 +00:00
|
|
|
|
|
|
|
out_shape = get_node_shape(chunk_output[0])
|
2023-05-15 03:46:25 +00:00
|
|
|
chunk_shape = out_shape[chunk_output_dim[0]]
|
2023-02-01 05:18:51 +00:00
|
|
|
context += "chunk_size = %d\nfor chunk_idx in range(0, %d, chunk_size):\n" % (chunk_size, chunk_shape)
|
2023-01-06 06:14:45 +00:00
|
|
|
return context
|
|
|
|
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
def _gen_loop_end(
|
|
|
|
chunk_inputs: List[Node],
|
|
|
|
chunk_non_compute_inputs: List[Node],
|
|
|
|
node_list: List[Node],
|
|
|
|
chunk_outputs_idx: int,
|
|
|
|
chunk_outputs_non_tensor: List[Node],
|
|
|
|
search_chunk: SearchChunk,
|
|
|
|
) -> str:
|
2023-01-10 01:59:47 +00:00
|
|
|
"""
|
|
|
|
Generate chunk loop end
|
|
|
|
|
|
|
|
eg. chunk_result[chunk_idx:chunk_idx + chunk_size] = output_node
|
|
|
|
output_node = chunk_result; xx = None; xx = None
|
|
|
|
|
|
|
|
Args:
|
|
|
|
chunk_inputs (List[Node]): chunk input node
|
|
|
|
chunk_non_compute_inputs (List[Node]): input node without chunk
|
|
|
|
chunk_outputs (Node): chunk output node
|
|
|
|
chunk_outputs_dim (int): chunk output node chunk dim
|
|
|
|
node_list (List)
|
|
|
|
|
|
|
|
Returns:
|
|
|
|
context (str): generated str
|
2023-01-10 02:56:39 +00:00
|
|
|
"""
|
2023-02-01 05:18:51 +00:00
|
|
|
context = "chunk_size = None"
|
2023-01-06 06:14:45 +00:00
|
|
|
# determine if its the last use for chunk input
|
|
|
|
for chunk_input in chunk_inputs + chunk_non_compute_inputs:
|
2023-02-01 05:18:51 +00:00
|
|
|
if all([search_chunk.node_mgr.find_node_idx(user) <= chunk_outputs_idx for user in chunk_input.users.keys()]):
|
2023-01-06 06:14:45 +00:00
|
|
|
context += "; %s = None" % chunk_input.name
|
2023-02-01 05:18:51 +00:00
|
|
|
for chunk_output_non_tensor, chunk_output_non_tensor_val in chunk_outputs_non_tensor.items():
|
|
|
|
context += "; %s = %s" % (chunk_output_non_tensor.name, chunk_output_non_tensor_val)
|
2023-01-06 06:14:45 +00:00
|
|
|
context += "\n"
|
|
|
|
return context
|
|
|
|
|
|
|
|
|
2023-01-10 01:59:47 +00:00
|
|
|
def _replace_name(context: str, name_from: str, name_to: str) -> str:
|
|
|
|
"""
|
|
|
|
replace node name
|
|
|
|
"""
|
2023-01-19 03:41:00 +00:00
|
|
|
patterns = [(" ", " "), (" ", "."), (" ", ","), ("(", ")"), ("(", ","), (" ", ")"), (" ", ""), ("", " ")]
|
2023-01-06 06:14:45 +00:00
|
|
|
for p in patterns:
|
|
|
|
source = p[0] + name_from + p[1]
|
|
|
|
target = p[0] + name_to + p[1]
|
|
|
|
if source in context:
|
|
|
|
context = context.replace(source, target)
|
2023-01-19 03:41:00 +00:00
|
|
|
break
|
2023-01-06 06:14:45 +00:00
|
|
|
return context
|
|
|
|
|
|
|
|
|
2023-01-10 01:59:47 +00:00
|
|
|
def _replace_reshape_size(context: str, node_name: str, reshape_size_dict: Dict) -> str:
|
|
|
|
"""
|
|
|
|
replace reshape size, some may have changed due to chunk
|
|
|
|
"""
|
2023-01-06 06:14:45 +00:00
|
|
|
if node_name not in reshape_size_dict:
|
|
|
|
return context
|
2023-01-19 03:41:00 +00:00
|
|
|
context = context.replace(reshape_size_dict[node_name][0], reshape_size_dict[node_name][1])
|
2023-01-06 06:14:45 +00:00
|
|
|
return context
|
|
|
|
|
|
|
|
|
2023-02-07 08:32:45 +00:00
|
|
|
def _replace_new_tensor_like_shape(
|
2023-01-10 01:59:47 +00:00
|
|
|
search_chunk: SearchChunk,
|
|
|
|
chunk_infos: List[Dict],
|
|
|
|
region_idx: int,
|
|
|
|
node_idx: int,
|
|
|
|
node: Node,
|
|
|
|
body: List[str],
|
|
|
|
) -> List[str]:
|
|
|
|
"""
|
|
|
|
add chunk slice for new tensor op such as ones like
|
|
|
|
"""
|
2023-02-07 08:32:45 +00:00
|
|
|
if get_node_name(node) in ["ones_like", "zeros_like", "empty_like"]:
|
2023-02-01 05:18:51 +00:00
|
|
|
meta_node = search_chunk.node_mgr.get_node_by_idx(node_idx)
|
2023-01-06 09:55:22 +00:00
|
|
|
chunk_dim = chunk_infos[region_idx]["node_chunk_dim"][meta_node]["chunk_dim"]
|
|
|
|
if get_node_shape(meta_node)[chunk_dim] != 1:
|
|
|
|
source_node = meta_node.args[0].args[0]
|
2023-09-19 06:20:26 +00:00
|
|
|
if (
|
|
|
|
source_node not in chunk_infos[region_idx]["node_chunk_dim"]
|
|
|
|
or chunk_infos[region_idx]["node_chunk_dim"][source_node]["chunk_dim"] is None
|
|
|
|
):
|
2023-01-16 11:25:05 +00:00
|
|
|
chunk_slice = _gen_chunk_slice_dim(chunk_dim, "chunk_idx", get_node_shape(node))
|
|
|
|
body[-1] = _replace_name(body[-1], node.args[0].name, node.args[0].name + chunk_slice)
|
2023-01-06 09:55:22 +00:00
|
|
|
return body
|
|
|
|
|
|
|
|
|
2023-02-07 08:32:45 +00:00
|
|
|
def _replace_new_tensor_shape(
|
|
|
|
search_chunk: SearchChunk,
|
|
|
|
chunk_infos: List[Dict],
|
|
|
|
region_idx: int,
|
|
|
|
node_idx: int,
|
|
|
|
node: Node,
|
|
|
|
body: List[str],
|
|
|
|
) -> List[str]:
|
|
|
|
"""
|
|
|
|
add chunk slice for new tensor op such as ones
|
|
|
|
"""
|
|
|
|
if get_node_name(node) in ["ones", "zeros", "empty"]:
|
|
|
|
meta_node = search_chunk.node_mgr.get_node_by_idx(node_idx)
|
|
|
|
chunk_dim = chunk_infos[region_idx]["node_chunk_dim"][meta_node]["chunk_dim"]
|
|
|
|
if chunk_dim is None:
|
|
|
|
return
|
|
|
|
if get_node_shape(meta_node)[chunk_dim] == 1:
|
|
|
|
return
|
|
|
|
origin_shape = str(node.args)
|
|
|
|
new_shape = list(node.args)
|
|
|
|
new_shape[chunk_dim] = "min(chunk_size, %d - chunk_idx)" % get_node_shape(meta_node)[chunk_dim]
|
|
|
|
new_shape = str(new_shape)
|
|
|
|
new_shape = new_shape.replace("'", "")
|
|
|
|
body[-1] = _replace_name(body[-1], origin_shape[1:-1], new_shape[1:-1])
|
|
|
|
return body
|
|
|
|
|
|
|
|
|
2023-02-01 05:18:51 +00:00
|
|
|
def _add_node_slice(
|
|
|
|
chunk_nodes: List[Node],
|
2023-01-10 01:59:47 +00:00
|
|
|
region_idx: int,
|
2023-02-01 05:18:51 +00:00
|
|
|
chunk_nodes_dim: Dict,
|
2023-01-10 01:59:47 +00:00
|
|
|
node_idx: int,
|
|
|
|
body: List[str],
|
2023-02-01 05:18:51 +00:00
|
|
|
node: Node,
|
2023-01-10 01:59:47 +00:00
|
|
|
) -> List[str]:
|
|
|
|
"""
|
|
|
|
add chunk slice for input nodes
|
|
|
|
"""
|
2023-02-01 05:18:51 +00:00
|
|
|
for chunk_node_idx, chunk_node in enumerate(chunk_nodes[region_idx]):
|
|
|
|
# inputs node
|
|
|
|
if isinstance(chunk_nodes_dim[region_idx][chunk_node_idx], dict):
|
|
|
|
for idx, dim in chunk_nodes_dim[region_idx][chunk_node_idx].items():
|
|
|
|
if idx == node_idx:
|
|
|
|
chunk_slice = _gen_chunk_slice_dim(dim[0], "chunk_idx", get_node_shape(chunk_node))
|
|
|
|
body[-1] = _replace_name(body[-1], chunk_node.name, chunk_node.name + chunk_slice)
|
|
|
|
# outputs node
|
|
|
|
else:
|
|
|
|
if chunk_node.name == node.name or (chunk_node.name in [i.name for i in node.all_input_nodes]):
|
2023-09-19 06:20:26 +00:00
|
|
|
chunk_slice = _gen_chunk_slice_dim(
|
|
|
|
chunk_nodes_dim[region_idx][chunk_node_idx], "chunk_idx", get_node_shape(chunk_node)
|
|
|
|
)
|
2023-03-10 02:23:26 +00:00
|
|
|
if get_node_name(chunk_node) in ["split", "unbind"]:
|
2023-02-01 05:18:51 +00:00
|
|
|
split_chunk_slice = ""
|
2023-09-19 06:20:26 +00:00
|
|
|
for i in range(len(chunk_node.meta["tensor_meta"])):
|
2023-02-01 05:18:51 +00:00
|
|
|
split_chunk_slice += "%s[%d]%s, " % (chunk_node.name, i, chunk_slice)
|
|
|
|
split_chunk_slice = split_chunk_slice[:-2]
|
|
|
|
body[-1] = _replace_name(body[-1], chunk_node.name, split_chunk_slice)
|
|
|
|
else:
|
|
|
|
body[-1] = _replace_name(body[-1], chunk_node.name, chunk_node.name + chunk_slice)
|
2023-01-06 09:55:22 +00:00
|
|
|
return body
|
|
|
|
|
|
|
|
|
2023-09-19 06:20:26 +00:00
|
|
|
def emit_code_with_chunk(
|
|
|
|
body: List[str],
|
|
|
|
nodes: Iterable[Node],
|
|
|
|
emit_node_func: Callable,
|
|
|
|
delete_unused_value_func: Callable,
|
|
|
|
search_chunk: SearchChunk,
|
|
|
|
chunk_infos: List,
|
|
|
|
eval_mem: bool = False,
|
|
|
|
):
|
2023-01-09 08:29:33 +00:00
|
|
|
"""
|
|
|
|
Emit code with chunk according to chunk_infos.
|
2023-01-09 08:30:16 +00:00
|
|
|
|
2023-01-10 01:59:47 +00:00
|
|
|
It will generate a for loop in chunk regions, and
|
2023-01-09 08:30:16 +00:00
|
|
|
replace inputs and outputs of regions with chunked variables.
|
2023-01-06 06:14:45 +00:00
|
|
|
|
|
|
|
Args:
|
|
|
|
body: forward code
|
|
|
|
nodes: graph.nodes
|
|
|
|
emit_node_func: function to emit node
|
|
|
|
delete_unused_value_func: function to remove the unused value
|
2023-01-09 08:29:33 +00:00
|
|
|
search_chunk: the class to search all chunks
|
|
|
|
chunk_infos: store all information about all chunks.
|
2023-01-06 06:14:45 +00:00
|
|
|
"""
|
|
|
|
node_list = list(nodes)
|
|
|
|
|
2023-01-09 08:29:33 +00:00
|
|
|
# chunk region
|
|
|
|
chunk_starts = [i["region"][0] for i in chunk_infos]
|
|
|
|
chunk_ends = [i["region"][1] for i in chunk_infos]
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-09 08:29:33 +00:00
|
|
|
# chunk inputs
|
2023-09-19 06:20:26 +00:00
|
|
|
chunk_inputs = [i["inputs"] for i in chunk_infos] # input with chunk
|
|
|
|
chunk_inputs_non_chunk = [i["inputs_non_chunk"] for i in chunk_infos] # input without chunk
|
|
|
|
chunk_inputs_dim = [i["inputs_dim"] for i in chunk_infos] # input chunk dim
|
2023-01-16 11:25:05 +00:00
|
|
|
chunk_inputs_names = [j.name for i in chunk_inputs for j in i] + [j.name for i in chunk_inputs_non_chunk for j in i]
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-09 08:29:33 +00:00
|
|
|
# chunk outputs
|
2023-02-01 05:18:51 +00:00
|
|
|
chunk_outputs = [i["outputs"] for i in chunk_infos]
|
|
|
|
chunk_outputs_non_tensor = [i["outputs_non_tensor"] for i in chunk_infos]
|
2023-01-06 06:14:45 +00:00
|
|
|
chunk_outputs_dim = [i["outputs_dim"] for i in chunk_infos]
|
|
|
|
|
2023-01-06 09:31:59 +00:00
|
|
|
node_list = search_chunk.reorder_graph.reorder_node_list(node_list)
|
2023-01-06 06:14:45 +00:00
|
|
|
node_idx = 0
|
|
|
|
region_idx = 0
|
|
|
|
within_chunk_region = False
|
|
|
|
|
2023-03-08 08:22:30 +00:00
|
|
|
if eval_mem:
|
|
|
|
body.append("init_memory = torch.cuda.memory_allocated() / 1024**2\n")
|
|
|
|
|
2023-01-06 06:14:45 +00:00
|
|
|
while node_idx < len(node_list):
|
|
|
|
node = node_list[node_idx]
|
|
|
|
|
2023-01-09 08:29:33 +00:00
|
|
|
# if is chunk start, generate for loop start
|
2023-01-06 06:14:45 +00:00
|
|
|
if node_idx in chunk_starts:
|
|
|
|
within_chunk_region = True
|
|
|
|
region_idx = chunk_starts.index(node_idx)
|
|
|
|
body.append(
|
|
|
|
_gen_loop_start(
|
|
|
|
chunk_inputs[region_idx],
|
|
|
|
chunk_outputs[region_idx],
|
|
|
|
chunk_outputs_dim[region_idx],
|
|
|
|
chunk_infos[region_idx]["chunk_size"],
|
2023-09-19 06:20:26 +00:00
|
|
|
)
|
|
|
|
)
|
2023-01-06 06:14:45 +00:00
|
|
|
|
|
|
|
if within_chunk_region:
|
|
|
|
emit_node_func(node, body)
|
|
|
|
# replace input var with chunk var
|
2023-02-01 05:18:51 +00:00
|
|
|
body = _add_node_slice(chunk_inputs, region_idx, chunk_inputs_dim, node_idx, body, node)
|
|
|
|
# replace output var with chunk var
|
|
|
|
body = _add_node_slice(chunk_outputs, region_idx, chunk_outputs_dim, node_idx, body, node)
|
2023-02-07 08:32:45 +00:00
|
|
|
# new tensor like
|
|
|
|
body = _replace_new_tensor_like_shape(search_chunk, chunk_infos, region_idx, node_idx, node, body)
|
|
|
|
# new tensor
|
|
|
|
body = _replace_new_tensor_shape(search_chunk, chunk_infos, region_idx, node_idx, node, body)
|
2023-04-26 03:38:43 +00:00
|
|
|
# reassign reshape size
|
2023-01-16 11:25:05 +00:00
|
|
|
body[-1] = _replace_reshape_size(body[-1], node.name, chunk_infos[region_idx]["reshape_size"])
|
2023-01-06 06:14:45 +00:00
|
|
|
body[-1] = " " + body[-1]
|
|
|
|
delete_unused_value_func(node, body, chunk_inputs_names)
|
2023-03-08 08:22:30 +00:00
|
|
|
if eval_mem:
|
|
|
|
body.append(
|
|
|
|
" if chunk_idx == 0:\n print('%s', torch.cuda.max_memory_allocated() / 1024**2 - init_memory); torch.cuda.reset_peak_memory_stats()\n"
|
2023-09-19 06:20:26 +00:00
|
|
|
% (node.name)
|
|
|
|
)
|
2023-01-06 06:14:45 +00:00
|
|
|
else:
|
|
|
|
emit_node_func(node, body)
|
|
|
|
if node_idx not in chunk_inputs:
|
|
|
|
delete_unused_value_func(node, body, chunk_inputs_names)
|
2023-03-08 08:22:30 +00:00
|
|
|
if eval_mem:
|
|
|
|
body.append(
|
|
|
|
"print('%s', torch.cuda.max_memory_allocated() / 1024**2 - init_memory); torch.cuda.reset_peak_memory_stats()\n"
|
2023-09-19 06:20:26 +00:00
|
|
|
% (node.name)
|
|
|
|
)
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-09 08:29:33 +00:00
|
|
|
# generate chunk region end
|
2023-01-06 06:14:45 +00:00
|
|
|
if node_idx in chunk_ends:
|
|
|
|
body.append(
|
2023-09-19 06:20:26 +00:00
|
|
|
_gen_loop_end(
|
|
|
|
chunk_inputs[region_idx],
|
|
|
|
chunk_inputs_non_chunk[region_idx],
|
|
|
|
node_list,
|
|
|
|
chunk_ends[region_idx],
|
|
|
|
chunk_outputs_non_tensor[region_idx],
|
|
|
|
search_chunk,
|
|
|
|
)
|
|
|
|
)
|
2023-01-06 06:14:45 +00:00
|
|
|
within_chunk_region = False
|
|
|
|
|
|
|
|
node_idx += 1
|
|
|
|
|
|
|
|
|
2023-01-31 08:00:06 +00:00
|
|
|
if AUTOCHUNK_AVAILABLE:
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
class AutoChunkCodeGen(CodeGen):
|
2023-09-19 06:20:26 +00:00
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
meta_graph,
|
|
|
|
max_memory: int = None,
|
|
|
|
print_mem: bool = False,
|
|
|
|
print_progress: bool = False,
|
|
|
|
eval_mem: bool = False,
|
|
|
|
) -> None:
|
2023-01-10 02:56:39 +00:00
|
|
|
super().__init__()
|
2023-03-08 08:22:30 +00:00
|
|
|
self.eval_mem = eval_mem
|
2023-01-10 02:56:39 +00:00
|
|
|
# find the chunk regions
|
2023-01-20 03:18:17 +00:00
|
|
|
self.search_chunk = SearchChunk(meta_graph, max_memory, print_mem, print_progress)
|
2023-01-10 02:56:39 +00:00
|
|
|
self.chunk_infos = self.search_chunk.search_region()
|
2023-01-20 03:18:17 +00:00
|
|
|
if print_progress:
|
|
|
|
get_logger().info("AutoChunk start codegen")
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2024-04-18 10:15:50 +00:00
|
|
|
def _gen_python_code(self, nodes, root_module: str, namespace: _Namespace, verbose=None) -> PythonCode:
|
2023-01-10 02:56:39 +00:00
|
|
|
free_vars: List[str] = []
|
|
|
|
body: List[str] = []
|
|
|
|
globals_: Dict[str, Any] = {}
|
|
|
|
wrapped_fns: Dict[str, None] = {}
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
# Wrap string in list to pass by reference
|
|
|
|
maybe_return_annotation: List[str] = [""]
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
def add_global(name_hint: str, obj: Any):
|
|
|
|
"""Add an obj to be tracked as a global.
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
We call this for names that reference objects external to the
|
|
|
|
Graph, like functions or types.
|
|
|
|
|
|
|
|
Returns: the global name that should be used to reference 'obj' in generated source.
|
|
|
|
"""
|
2023-09-19 06:20:26 +00:00
|
|
|
if _is_from_torch(obj) and obj != torch.device: # to support registering torch.device
|
2023-01-10 02:56:39 +00:00
|
|
|
# HACK: workaround for how torch custom ops are registered. We
|
|
|
|
# can't import them like normal modules so they must retain their
|
|
|
|
# fully qualified name.
|
|
|
|
return _get_qualified_name(obj)
|
|
|
|
|
|
|
|
# normalize the name hint to get a proper identifier
|
|
|
|
global_name = namespace.create_name(name_hint, obj)
|
|
|
|
|
|
|
|
if global_name in globals_:
|
|
|
|
assert globals_[global_name] is obj
|
|
|
|
return global_name
|
|
|
|
globals_[global_name] = obj
|
2023-01-06 06:14:45 +00:00
|
|
|
return global_name
|
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
# set _custom_builtins here so that we needn't import colossalai in forward
|
2023-01-16 11:25:05 +00:00
|
|
|
_custom_builtins["colossalai"] = _CustomBuiltin("import colossalai", colossalai)
|
2023-01-10 02:56:39 +00:00
|
|
|
|
|
|
|
# Pre-fill the globals table with registered builtins.
|
|
|
|
for name, (_, obj) in _custom_builtins.items():
|
|
|
|
add_global(name, obj)
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
def type_repr(o: Any):
|
|
|
|
if o == ():
|
|
|
|
# Empty tuple is used for empty tuple type annotation Tuple[()]
|
|
|
|
return "()"
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
typename = _type_repr(o)
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
if hasattr(o, "__origin__"):
|
|
|
|
# This is a generic type, e.g. typing.List[torch.Tensor]
|
|
|
|
origin_type = _origin_type_map.get(o.__origin__, o.__origin__)
|
|
|
|
origin_typename = add_global(_type_repr(origin_type), origin_type)
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
if hasattr(o, "__args__"):
|
|
|
|
# Assign global names for each of the inner type variables.
|
|
|
|
args = [type_repr(arg) for arg in o.__args__]
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
if len(args) == 0:
|
|
|
|
# Bare type, such as `typing.Tuple` with no subscript
|
|
|
|
# This code-path used in Python < 3.9
|
|
|
|
return origin_typename
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
return f'{origin_typename}[{",".join(args)}]'
|
|
|
|
else:
|
2023-01-06 06:14:45 +00:00
|
|
|
# Bare type, such as `typing.Tuple` with no subscript
|
2023-01-10 02:56:39 +00:00
|
|
|
# This code-path used in Python 3.9+
|
2023-01-06 06:14:45 +00:00
|
|
|
return origin_typename
|
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
# Common case: this is a regular module name like 'foo.bar.baz'
|
|
|
|
return add_global(typename, o)
|
|
|
|
|
2023-01-16 11:25:05 +00:00
|
|
|
def _format_args(args: Tuple[Argument, ...], kwargs: Dict[str, Argument]) -> str:
|
2023-01-10 02:56:39 +00:00
|
|
|
def _get_repr(arg):
|
|
|
|
# Handle NamedTuples (if it has `_fields`) via add_global.
|
|
|
|
if isinstance(arg, tuple) and hasattr(arg, "_fields"):
|
|
|
|
qualified_name = _get_qualified_name(type(arg))
|
|
|
|
global_name = add_global(qualified_name, type(arg))
|
|
|
|
return f"{global_name}{repr(tuple(arg))}"
|
|
|
|
return repr(arg)
|
|
|
|
|
|
|
|
args_s = ", ".join(_get_repr(a) for a in args)
|
|
|
|
kwargs_s = ", ".join(f"{k} = {_get_repr(v)}" for k, v in kwargs.items())
|
|
|
|
if args_s and kwargs_s:
|
|
|
|
return f"{args_s}, {kwargs_s}"
|
|
|
|
return args_s or kwargs_s
|
|
|
|
|
|
|
|
# Run through reverse nodes and record the first instance of a use
|
|
|
|
# of a given node. This represents the *last* use of the node in the
|
|
|
|
# execution order of the program, which we will use to free unused
|
|
|
|
# values
|
|
|
|
node_to_last_use: Dict[Node, Node] = {}
|
|
|
|
user_to_last_uses: Dict[Node, List[Node]] = {}
|
|
|
|
|
|
|
|
def register_last_uses(n: Node, user: Node):
|
|
|
|
if n not in node_to_last_use:
|
|
|
|
node_to_last_use[n] = user
|
|
|
|
user_to_last_uses.setdefault(user, []).append(n)
|
|
|
|
|
|
|
|
for node in reversed(nodes):
|
|
|
|
map_arg(node.args, lambda n: register_last_uses(n, node))
|
|
|
|
map_arg(node.kwargs, lambda n: register_last_uses(n, node))
|
|
|
|
|
|
|
|
delete_free_var_from_last_use(user_to_last_uses)
|
|
|
|
|
|
|
|
# NOTE: we add a variable to distinguish body and ckpt_func
|
|
|
|
def delete_unused_values(user: Node, body, to_keep=[]):
|
|
|
|
"""
|
|
|
|
Delete values after their last use. This ensures that values that are
|
|
|
|
not used in the remainder of the code are freed and the memory usage
|
|
|
|
of the code is optimal.
|
|
|
|
"""
|
|
|
|
if user.op == "placeholder":
|
|
|
|
return
|
|
|
|
if user.op == "output":
|
|
|
|
body.append("\n")
|
|
|
|
return
|
|
|
|
nodes_to_delete = user_to_last_uses.get(user, [])
|
|
|
|
nodes_to_delete = [i for i in nodes_to_delete if i.name not in to_keep]
|
|
|
|
if len(nodes_to_delete):
|
2023-01-16 11:25:05 +00:00
|
|
|
to_delete_str = " = ".join([repr(n) for n in nodes_to_delete] + ["None"])
|
2023-01-10 02:56:39 +00:00
|
|
|
body.append(f"; {to_delete_str}\n")
|
2023-01-06 06:14:45 +00:00
|
|
|
else:
|
2023-01-10 02:56:39 +00:00
|
|
|
body.append("\n")
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
# NOTE: we add a variable to distinguish body and ckpt_func
|
|
|
|
def emit_node(node: Node, body):
|
2023-09-19 06:20:26 +00:00
|
|
|
maybe_type_annotation = "" if node.type is None else f" : {type_repr(node.type)}"
|
2023-01-10 02:56:39 +00:00
|
|
|
if node.op == "placeholder":
|
|
|
|
assert isinstance(node.target, str)
|
2023-09-19 06:20:26 +00:00
|
|
|
maybe_default_arg = "" if not node.args else f" = {repr(node.args[0])}"
|
2023-01-16 11:25:05 +00:00
|
|
|
free_vars.append(f"{node.target}{maybe_type_annotation}{maybe_default_arg}")
|
2023-01-10 02:56:39 +00:00
|
|
|
raw_name = node.target.replace("*", "")
|
|
|
|
if raw_name != repr(node):
|
|
|
|
body.append(f"{repr(node)} = {raw_name}\n")
|
|
|
|
return
|
|
|
|
elif node.op == "call_method":
|
|
|
|
assert isinstance(node.target, str)
|
2023-01-06 06:14:45 +00:00
|
|
|
body.append(
|
2023-01-10 02:56:39 +00:00
|
|
|
f"{repr(node)}{maybe_type_annotation} = {_format_target(repr(node.args[0]), node.target)}"
|
2023-09-19 06:20:26 +00:00
|
|
|
f"({_format_args(node.args[1:], node.kwargs)})"
|
|
|
|
)
|
2023-01-06 06:14:45 +00:00
|
|
|
return
|
2023-01-10 02:56:39 +00:00
|
|
|
elif node.op == "call_function":
|
|
|
|
assert callable(node.target)
|
|
|
|
# pretty print operators
|
2023-09-19 06:20:26 +00:00
|
|
|
if node.target.__module__ == "_operator" and node.target.__name__ in magic_methods:
|
2023-01-10 02:56:39 +00:00
|
|
|
assert isinstance(node.args, tuple)
|
2023-09-19 06:20:26 +00:00
|
|
|
body.append(
|
|
|
|
f"{repr(node)}{maybe_type_annotation} = "
|
|
|
|
f"{magic_methods[node.target.__name__].format(*(repr(a) for a in node.args))}"
|
|
|
|
)
|
2023-01-10 02:56:39 +00:00
|
|
|
return
|
|
|
|
|
|
|
|
# pretty print inplace operators; required for jit.script to work properly
|
|
|
|
# not currently supported in normal FX graphs, but generated by torchdynamo
|
2023-09-19 06:20:26 +00:00
|
|
|
if node.target.__module__ == "_operator" and node.target.__name__ in inplace_methods:
|
|
|
|
body.append(
|
|
|
|
f"{inplace_methods[node.target.__name__].format(*(repr(a) for a in node.args))}; "
|
|
|
|
f"{repr(node)}{maybe_type_annotation} = {repr(node.args[0])}"
|
|
|
|
)
|
2023-01-10 02:56:39 +00:00
|
|
|
return
|
|
|
|
|
|
|
|
qualified_name = _get_qualified_name(node.target)
|
|
|
|
global_name = add_global(qualified_name, node.target)
|
|
|
|
# special case for getattr: node.args could be 2-argument or 3-argument
|
|
|
|
# 2-argument: attribute access; 3-argument: fall through to attrib function call with default value
|
2023-09-19 06:20:26 +00:00
|
|
|
if (
|
|
|
|
global_name == "getattr"
|
|
|
|
and isinstance(node.args, tuple)
|
|
|
|
and isinstance(node.args[1], str)
|
|
|
|
and node.args[1].isidentifier()
|
|
|
|
and len(node.args) == 2
|
|
|
|
):
|
2023-01-10 02:56:39 +00:00
|
|
|
body.append(
|
2023-09-19 06:20:26 +00:00
|
|
|
f"{repr(node)}{maybe_type_annotation} = {_format_target(repr(node.args[0]), node.args[1])}"
|
|
|
|
)
|
2023-01-10 02:56:39 +00:00
|
|
|
return
|
2023-01-06 06:14:45 +00:00
|
|
|
body.append(
|
2023-09-19 06:20:26 +00:00
|
|
|
f"{repr(node)}{maybe_type_annotation} = {global_name}({_format_args(node.args, node.kwargs)})"
|
|
|
|
)
|
2023-01-10 02:56:39 +00:00
|
|
|
if node.meta.get("is_wrapped", False):
|
|
|
|
wrapped_fns.setdefault(global_name)
|
2023-01-06 06:14:45 +00:00
|
|
|
return
|
2023-01-10 02:56:39 +00:00
|
|
|
elif node.op == "call_module":
|
|
|
|
assert isinstance(node.target, str)
|
2023-09-19 06:20:26 +00:00
|
|
|
body.append(
|
|
|
|
f"{repr(node)}{maybe_type_annotation} = "
|
|
|
|
f"{_format_target(root_module, node.target)}({_format_args(node.args, node.kwargs)})"
|
|
|
|
)
|
2023-01-06 06:14:45 +00:00
|
|
|
return
|
2023-01-10 02:56:39 +00:00
|
|
|
elif node.op == "get_attr":
|
|
|
|
assert isinstance(node.target, str)
|
2023-01-16 11:25:05 +00:00
|
|
|
body.append(f"{repr(node)}{maybe_type_annotation} = {_format_target(root_module, node.target)}")
|
2023-01-10 02:56:39 +00:00
|
|
|
return
|
|
|
|
elif node.op == "output":
|
|
|
|
if node.type is not None:
|
|
|
|
maybe_return_annotation[0] = f" -> {type_repr(node.type)}"
|
|
|
|
body.append(self.generate_output(node.args[0]))
|
|
|
|
return
|
|
|
|
raise NotImplementedError(f"node: {node.op} {node.target}")
|
|
|
|
|
|
|
|
# Modified for activation checkpointing
|
|
|
|
ckpt_func = []
|
|
|
|
|
|
|
|
# if any node has a list of labels for activation_checkpoint, we
|
|
|
|
# will use nested type of activation checkpoint codegen
|
2023-09-19 06:20:26 +00:00
|
|
|
emit_code_with_chunk(
|
|
|
|
body, nodes, emit_node, delete_unused_values, self.search_chunk, self.chunk_infos, self.eval_mem
|
|
|
|
)
|
2023-01-10 02:56:39 +00:00
|
|
|
|
|
|
|
if len(body) == 0:
|
|
|
|
# If the Graph has no non-placeholder nodes, no lines for the body
|
|
|
|
# have been emitted. To continue to have valid Python code, emit a
|
|
|
|
# single pass statement
|
|
|
|
body.append("pass\n")
|
|
|
|
|
|
|
|
if len(wrapped_fns) > 0:
|
|
|
|
wrap_name = add_global("wrap", torch.fx.wrap)
|
2023-01-16 11:25:05 +00:00
|
|
|
wrap_stmts = "\n".join([f'{wrap_name}("{name}")' for name in wrapped_fns])
|
2023-01-10 02:56:39 +00:00
|
|
|
else:
|
|
|
|
wrap_stmts = ""
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
if self._body_transformer:
|
|
|
|
body = self._body_transformer(body)
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
for name, value in self.additional_globals():
|
|
|
|
add_global(name, value)
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
# as we need colossalai.utils.checkpoint, we need to import colossalai
|
|
|
|
# in forward function
|
|
|
|
prologue = self.gen_fn_def(free_vars, maybe_return_annotation[0])
|
|
|
|
prologue = "".join(ckpt_func) + prologue
|
|
|
|
prologue = prologue
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 02:56:39 +00:00
|
|
|
code = "".join(body)
|
|
|
|
code = "\n".join(" " + line for line in code.split("\n"))
|
|
|
|
fn_code = f"""
|
2023-01-10 03:56:00 +00:00
|
|
|
{wrap_stmts}
|
2023-01-06 06:14:45 +00:00
|
|
|
|
2023-01-10 03:56:00 +00:00
|
|
|
{prologue}
|
|
|
|
{code}"""
|
2023-01-10 02:56:39 +00:00
|
|
|
# print(fn_code)
|
|
|
|
return PythonCode(fn_code, globals_)
|