wehub-resource-sync
945 lines
37 KiB
Python
945 lines
37 KiB
Python
# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
|
||
#
|
||
# 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.
|
||
from __future__ import annotations
|
||
|
||
import ast
|
||
import logging
|
||
import re
|
||
from dataclasses import dataclass
|
||
|
||
import numpy as np
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
from ..dcp.sharded_weight import ShardedWeightDesc
|
||
from .lexer import Lexer
|
||
from .parser import Parser
|
||
from .traceback import AOATraceback
|
||
|
||
_ShardInfo = dict[str, list[ShardedWeightDesc]]
|
||
|
||
# SliceRef := (key, src_slice, dst_slice, postprocess_list)
|
||
SliceRef = tuple[str, tuple[slice, ...], tuple[slice, ...], list[str] | None]
|
||
SUPPORTED_DTYPES = ['float16', 'float32', 'bfloat16']
|
||
|
||
|
||
class TensorDesc:
|
||
def __init__(
|
||
self,
|
||
slices: list[SliceRef],
|
||
shape: tuple[int],
|
||
in_degree: int = 0,
|
||
out_degree: int = 0,
|
||
dtype: str | None = None,
|
||
):
|
||
self.slices = slices
|
||
self.shape = shape
|
||
self.in_degree = in_degree
|
||
self.out_degree = out_degree
|
||
self.dtype = dtype
|
||
|
||
def __repr__(self):
|
||
s = []
|
||
for key, sl_src, sl_dst, pp_list in self.slices:
|
||
s.append(
|
||
f"{key}{sl_src} -> self{sl_dst}, postprocess_list={pp_list}"
|
||
)
|
||
return f"Tensor(shape={self.shape}, slices={s}, in_degree={self.in_degree}, out_degree={self.out_degree}, dtype={self.dtype})"
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class ShardMappingEntry:
|
||
target_slice: ShardedWeightDesc
|
||
source_slice: ShardedWeightDesc
|
||
postprocess_list: list[str] | None = None
|
||
|
||
|
||
ShardMapping = list[ShardMappingEntry]
|
||
|
||
OPTIMIZER_STATE_NAME = [
|
||
".w_0",
|
||
".moment1_0",
|
||
".moment2_0",
|
||
".beta1_pow_acc_0",
|
||
".beta2_pow_acc_0",
|
||
]
|
||
|
||
|
||
def split_optimizer_state_key(key: str) -> tuple[str, str]:
|
||
for opt_state_name in OPTIMIZER_STATE_NAME:
|
||
if key.endswith(opt_state_name):
|
||
return key[: -len(opt_state_name)], opt_state_name
|
||
return key, None
|
||
|
||
|
||
class AOAShardInfoContext:
|
||
def __init__(
|
||
self,
|
||
source_state_shard_info: _ShardInfo,
|
||
destination_state_shard_info: _ShardInfo,
|
||
aoa_config_reverse: bool = False,
|
||
) -> None:
|
||
self.source_state_shard_info = source_state_shard_info
|
||
self.destination_state_shard_info = destination_state_shard_info
|
||
self.aoa_config_reverse = aoa_config_reverse
|
||
self.left_var_to_right_var_mapping = {}
|
||
self.right_var_from_left_var_mapping = {}
|
||
self.src_state_keys = set()
|
||
self.dst_state_keys = set()
|
||
self.init_src_state_keys()
|
||
self.init_dst_state_keys()
|
||
|
||
def init_src_state_keys(self):
|
||
for k in self.source_state_shard_info.keys():
|
||
model_state_key, _ = split_optimizer_state_key(k)
|
||
self.src_state_keys.add(model_state_key)
|
||
|
||
def init_dst_state_keys(self):
|
||
if self.destination_state_shard_info is None:
|
||
return
|
||
for k in self.destination_state_shard_info.keys():
|
||
model_state_key, _ = split_optimizer_state_key(k)
|
||
self.dst_state_keys.add(model_state_key)
|
||
|
||
def get_all_dst_state_keys(self):
|
||
return self.dst_state_keys
|
||
|
||
def get_all_src_state_keys(self):
|
||
return self.src_state_keys
|
||
|
||
def get_num_hidden_layers(
|
||
self,
|
||
name_with_layer_id: str,
|
||
layer_id_macro_tag: str,
|
||
) -> int:
|
||
if layer_id_macro_tag not in name_with_layer_id:
|
||
raise ValueError(
|
||
f"layer_id_macro_tag '{layer_id_macro_tag}' not in name_with_layer_id '{name_with_layer_id}'"
|
||
)
|
||
prefix, suffix = name_with_layer_id.split(layer_id_macro_tag, 1)
|
||
pattern = re.compile(rf"{re.escape(prefix)}(\d+){re.escape(suffix)}")
|
||
match_layer_id = set()
|
||
for key in self.get_all_src_state_keys():
|
||
match = pattern.fullmatch(key)
|
||
if match:
|
||
layer_num = int(match.group(1))
|
||
match_layer_id.add(layer_num)
|
||
return match_layer_id
|
||
|
||
def get_src_state_shard_num(self, src_state_key: str) -> int:
|
||
model_state_key, opt_state_name = split_optimizer_state_key(
|
||
src_state_key
|
||
)
|
||
|
||
assert opt_state_name is None, (
|
||
"AOA notions apply only to the model state, but are automatically propagated to the optimizer state.Now the src_state_key is {src_state_key}, which is a optimizer state key."
|
||
)
|
||
reverse = True
|
||
if self.aoa_config_reverse:
|
||
reverse = False
|
||
# Only need to parse the model state key for optimizer state shard num, because the optimizer state slice info is completely consistent with the model state slice info.
|
||
resolved_model_state_key = self.resolve_mapping_chain(
|
||
model_state_key, reverse=reverse
|
||
)
|
||
|
||
state_keys = [
|
||
resolved_model_state_key,
|
||
f"{resolved_model_state_key}.w_0",
|
||
f"{resolved_model_state_key}.moment1_0",
|
||
f"{resolved_model_state_key}.moment2_0",
|
||
]
|
||
|
||
shard_nums = {
|
||
len(
|
||
{
|
||
shard_info.global_offset
|
||
for shard_info in self.source_state_shard_info[key]
|
||
}
|
||
)
|
||
for key in state_keys
|
||
if key in self.source_state_shard_info
|
||
}
|
||
|
||
if not shard_nums:
|
||
logger.warning(
|
||
f"No shard information found for any of the keys: {state_keys}, return 1."
|
||
)
|
||
return 1
|
||
if len(shard_nums) > 1:
|
||
raise AssertionError(
|
||
f"Inconsistent shard numbers among keys in source_sharded_state_dict for the key {src_state_key}: shard_nums={shard_nums}."
|
||
)
|
||
return shard_nums.pop()
|
||
|
||
def get_dst_state_shard_num(self, dst_state_key: str) -> int:
|
||
if self.destination_state_shard_info is None:
|
||
# Default `dst_state_shard_num=1` if `destination_state_shard_info` is missing.
|
||
return 1
|
||
model_state_key, opt_state_name = split_optimizer_state_key(
|
||
dst_state_key
|
||
)
|
||
|
||
assert opt_state_name is None, (
|
||
"AOA notions apply only to the model state, but are automatically propagated to the optimizer state.Now the dst_state_key is {dst_state_key}, which is a optimizer state key."
|
||
)
|
||
reverse = False
|
||
if self.aoa_config_reverse:
|
||
reverse = True
|
||
# Only need to parse the model state key for optimizer state shard num, because the optimizer state slice info is completely consistent with the model state slice info.
|
||
resolved_model_state_key = self.resolve_mapping_chain(
|
||
model_state_key, reverse=reverse
|
||
)
|
||
|
||
state_keys = [
|
||
resolved_model_state_key,
|
||
f"{resolved_model_state_key}.w_0",
|
||
f"{resolved_model_state_key}.moment1_0",
|
||
f"{resolved_model_state_key}.moment2_0",
|
||
]
|
||
|
||
shard_nums = {
|
||
len(
|
||
{
|
||
shard_info.global_offset
|
||
for shard_info in self.destination_state_shard_info[key]
|
||
}
|
||
)
|
||
for key in state_keys
|
||
if key in self.destination_state_shard_info
|
||
}
|
||
|
||
if not shard_nums:
|
||
logger.warning(
|
||
f"No shard information found for any of the keys: {state_keys}, return 1."
|
||
)
|
||
return 1
|
||
if len(shard_nums) > 1:
|
||
raise AssertionError(
|
||
f"Inconsistent shard numbers among keys in destination_state_shard_info for the key {dst_state_key}: shard_nums={shard_nums}."
|
||
)
|
||
return shard_nums.pop()
|
||
|
||
def resolve_mapping_chain(self, key: str, reverse: bool = False) -> str:
|
||
"""
|
||
Recursively resolve the mapping chain, find the final leaf node
|
||
|
||
Args:
|
||
key: The key to be resolved
|
||
reverse: False use left_var_to_right_var_mapping，True use right_var_from_left_var_mapping
|
||
|
||
For example:
|
||
- reverse=False: temp_var -> dst_key
|
||
- reverse=True: temp_var -> src_key
|
||
"""
|
||
visited = set() # avoid infinite loop
|
||
current_key = key
|
||
|
||
if reverse:
|
||
mapping_dict = self.right_var_from_left_var_mapping
|
||
else:
|
||
mapping_dict = self.left_var_to_right_var_mapping
|
||
|
||
while current_key in mapping_dict:
|
||
assert current_key not in visited, (
|
||
f"Infinite loop detected in resolve_mapping_chain, which means the start key is not src_key or the end key is not dst_key, the aoa_config is error. current_key={current_key}, the loop is: {'->'.join(visited)}->{current_key}"
|
||
)
|
||
visited.add(current_key)
|
||
if reverse and current_key in self.get_all_src_state_keys():
|
||
break
|
||
elif not reverse and current_key in self.get_all_dst_state_keys():
|
||
break
|
||
|
||
mapped_vars = mapping_dict[current_key]
|
||
if mapped_vars and len(mapped_vars) > 0:
|
||
assert len(mapped_vars) == 1, (
|
||
f"Reference chain resolution failed: "
|
||
f"Unable to determine which leaf node the intermediate node '{key}' is directly associated with, "
|
||
f"because a many-to-one mapping was found in the mapping relationship. "
|
||
f"The many-to-one mapping is {current_key} : {mapped_vars}."
|
||
)
|
||
current_key = mapped_vars[0]
|
||
else:
|
||
break
|
||
|
||
return current_key
|
||
|
||
|
||
class AOAEngine:
|
||
def __init__(
|
||
self,
|
||
aoa_config: dict[str, list[str]],
|
||
source_state_shard_info: _ShardInfo,
|
||
destination_state_shard_info: _ShardInfo,
|
||
):
|
||
self.aoa_config = aoa_config
|
||
self.source_state_shard_info = source_state_shard_info
|
||
self.destination_state_shard_info = destination_state_shard_info
|
||
self.aoa_config_reverse = self.aoa_config.get(
|
||
"aoa_config_reverse", False
|
||
)
|
||
enable_traceback = self.aoa_config.get("enable_traceback", True)
|
||
self.traceback = AOATraceback() if enable_traceback else None
|
||
self.context = AOAShardInfoContext(
|
||
source_state_shard_info,
|
||
destination_state_shard_info,
|
||
self.aoa_config_reverse,
|
||
)
|
||
self.lexer = Lexer(self.context, traceback=self.traceback)
|
||
tokens = self.lexer.all_tokens(
|
||
self.aoa_config.get("aoa_statements", [])
|
||
)
|
||
self.parser = Parser(tokens)
|
||
self.statements = self.parser.parse_program()
|
||
|
||
if self.traceback and getattr(self.lexer, "final_expressions", None):
|
||
final_exprs = self.lexer.final_expressions
|
||
if len(final_exprs) == len(self.statements):
|
||
for expr, stmt in zip(final_exprs, self.statements):
|
||
self.traceback.record_children(
|
||
expr, [repr(stmt)], macro_name="parser"
|
||
)
|
||
|
||
if self.aoa_config_reverse:
|
||
self.statements = list(reversed(self.statements))
|
||
self.input_vars = self.build_input_vars()
|
||
self.output_vars = {}
|
||
self.intermediate_vars = {}
|
||
self.need_remove_input_vars = set()
|
||
self.need_add_output_vars = set()
|
||
|
||
self.shape_propagation()
|
||
|
||
def make_input_tensor(
|
||
self, key: str, shape: tuple[int], dtype: str
|
||
) -> TensorDesc:
|
||
base_slice = tuple([slice(0, s) for s in shape])
|
||
return TensorDesc(
|
||
[(key, base_slice, base_slice, None)],
|
||
shape,
|
||
in_degree=0,
|
||
out_degree=0,
|
||
dtype=dtype,
|
||
)
|
||
|
||
def build_input_vars(self):
|
||
input_vars = {}
|
||
dtype = None
|
||
for key, shards in sorted(self.source_state_shard_info.items()):
|
||
global_shape = shards[0].global_shape
|
||
model_state_key, opt_state_name = split_optimizer_state_key(key)
|
||
if opt_state_name is None:
|
||
dtype = shards[0].dtype
|
||
if model_state_key in input_vars.keys() or opt_state_name in [
|
||
".beta1_pow_acc_0",
|
||
".beta2_pow_acc_0",
|
||
]:
|
||
continue
|
||
input_vars[model_state_key] = self.make_input_tensor(
|
||
model_state_key, global_shape, dtype
|
||
)
|
||
return input_vars
|
||
|
||
def split(
|
||
self, tensor: TensorDesc, axis: int, sizes: list[int]
|
||
) -> list[TensorDesc]:
|
||
results = []
|
||
start = 0
|
||
tensor.out_degree += len(sizes)
|
||
dtype = tensor.dtype
|
||
for sz in sizes:
|
||
sub_dst_slice = [slice(None)] * len(tensor.shape)
|
||
sub_dst_slice[axis] = slice(0, sz)
|
||
sub_slices = []
|
||
for aidx, src_sl, dst_sl, pp_list in tensor.slices:
|
||
if pp_list is not None:
|
||
src_sl = postprocess_transpose(list(src_sl), pp_list)
|
||
|
||
dst_start = (
|
||
dst_sl[axis].start if dst_sl[axis].start is not None else 0
|
||
)
|
||
dst_stop = (
|
||
dst_sl[axis].stop
|
||
if dst_sl[axis].stop is not None
|
||
else tensor.shape[axis]
|
||
)
|
||
inter_begin = max(start, dst_start)
|
||
inter_end = min(start + sz, dst_stop)
|
||
if inter_begin < inter_end:
|
||
src_axis_start = (
|
||
src_sl[axis].start
|
||
if src_sl[axis].start is not None
|
||
else 0
|
||
)
|
||
sub_src_sl = list(src_sl)
|
||
sub_dst_sl = list(dst_sl)
|
||
offset = inter_begin - dst_start
|
||
length = inter_end - inter_begin
|
||
sub_src_sl[axis] = slice(
|
||
src_axis_start + offset,
|
||
src_axis_start + offset + length,
|
||
)
|
||
sub_dst_sl[axis] = slice(
|
||
inter_begin - start, inter_begin - start + length
|
||
)
|
||
if pp_list is not None:
|
||
sub_src_sl = postprocess_transpose(
|
||
list(sub_src_sl), pp_list, reverse=True
|
||
)
|
||
sub_slices.append(
|
||
(
|
||
aidx,
|
||
tuple(sub_src_sl),
|
||
tuple(sub_dst_sl),
|
||
pp_list.copy(),
|
||
)
|
||
)
|
||
else:
|
||
sub_slices.append(
|
||
(aidx, tuple(sub_src_sl), tuple(sub_dst_sl), None)
|
||
)
|
||
new_shape = list(tensor.shape)
|
||
new_shape[axis] = sz
|
||
results.append(
|
||
TensorDesc(
|
||
sub_slices,
|
||
tuple(new_shape),
|
||
in_degree=1,
|
||
out_degree=0,
|
||
dtype=dtype,
|
||
)
|
||
)
|
||
start += sz
|
||
return results
|
||
|
||
def concat(self, tensors: list[TensorDesc], axis: int) -> TensorDesc:
|
||
slices = []
|
||
assert len(tensors) >= 1, (
|
||
"When concatenating multiple tensors, there should be at least one!"
|
||
)
|
||
shape = list(tensors[0].shape)
|
||
ndim = len(shape)
|
||
assert 0 <= axis < ndim, (
|
||
f"when concat, the axis {axis} is out of range for tensors "
|
||
f"with shape {shape} (valid range: {0} to {ndim - 1})."
|
||
)
|
||
shape[axis] = sum(t.shape[axis] for t in tensors)
|
||
dtype = tensors[0].dtype
|
||
assert all(t.dtype == dtype for t in tensors), (
|
||
f"All tensors must have the same dtype when concatenating multiple tensors!But the tensors {tensors} have different dtypes: {[t.dtype for t in tensors]}."
|
||
)
|
||
curr = 0
|
||
for t in tensors:
|
||
t.out_degree += 1
|
||
for aidx, src_sl, dst_sl, pp_list in t.slices:
|
||
new_dst_sl = list(dst_sl)
|
||
dst_start = (
|
||
dst_sl[axis].start if dst_sl[axis].start is not None else 0
|
||
)
|
||
dst_stop = (
|
||
dst_sl[axis].stop
|
||
if dst_sl[axis].stop is not None
|
||
else t.shape[axis]
|
||
)
|
||
length = dst_stop - dst_start
|
||
new_dst_sl[axis] = slice(
|
||
dst_start + curr, dst_start + curr + length
|
||
)
|
||
if pp_list is not None:
|
||
slices.append(
|
||
(aidx, src_sl, tuple(new_dst_sl), pp_list.copy())
|
||
)
|
||
else:
|
||
slices.append((aidx, src_sl, tuple(new_dst_sl), None))
|
||
curr += t.shape[axis]
|
||
return TensorDesc(
|
||
slices,
|
||
tuple(shape),
|
||
in_degree=len(tensors),
|
||
out_degree=0,
|
||
dtype=dtype,
|
||
)
|
||
|
||
def transpose(self, tensor: TensorDesc, permutation: str) -> TensorDesc:
|
||
slices = []
|
||
tensor.out_degree += 1
|
||
tensor_shape = transpose_list(
|
||
tensor.shape, ast.literal_eval(permutation)
|
||
)
|
||
dtype = tensor.dtype
|
||
for aidx, src_sl, dst_sl, pp_list in tensor.slices:
|
||
trans_dst_sl = transpose_list(dst_sl, ast.literal_eval(permutation))
|
||
if pp_list is not None:
|
||
new_pp_list = pp_list.copy()
|
||
new_pp_list.append(permutation)
|
||
slices.append((aidx, src_sl, trans_dst_sl, new_pp_list))
|
||
else:
|
||
slices.append((aidx, src_sl, trans_dst_sl, [permutation]))
|
||
return TensorDesc(
|
||
slices, tensor_shape, in_degree=1, out_degree=0, dtype=dtype
|
||
)
|
||
|
||
def cast(self, tensor: TensorDesc, dtype: str) -> TensorDesc:
|
||
slices = []
|
||
tensor.out_degree += 1
|
||
for aidx, src_sl, dst_sl, pp_list in tensor.slices:
|
||
if pp_list is not None:
|
||
new_pp_list = pp_list.copy()
|
||
new_pp_list.append(dtype)
|
||
slices.append((aidx, src_sl, dst_sl, new_pp_list))
|
||
else:
|
||
slices.append((aidx, src_sl, dst_sl, [dtype]))
|
||
# For the cast operation, post_process is required. Therefore, the returned
|
||
# Tensor's dtype here is the same as the input tensor's dtype, rather than the casted dtype.
|
||
return TensorDesc(
|
||
slices, tensor.shape, in_degree=1, out_degree=0, dtype=tensor.dtype
|
||
)
|
||
|
||
def identity(self, tensor: TensorDesc) -> TensorDesc:
|
||
tensor.out_degree += 1
|
||
return TensorDesc(
|
||
tensor.slices,
|
||
tensor.shape,
|
||
in_degree=1,
|
||
out_degree=0,
|
||
dtype=tensor.dtype,
|
||
)
|
||
|
||
def shape_propagation(self):
|
||
def _get_var_ref(var):
|
||
if var.name in self.intermediate_vars:
|
||
return self.intermediate_vars[var.name]
|
||
elif var.name in self.input_vars:
|
||
return self.input_vars[var.name]
|
||
else:
|
||
raise ValueError(f"{var.name} should be assigned before!")
|
||
|
||
for stmt in self.statements:
|
||
stmt_repr = repr(stmt)
|
||
left_vars = stmt.left_vars
|
||
right_vars = stmt.right_vars
|
||
if self.aoa_config_reverse:
|
||
left_vars, right_vars = right_vars, left_vars
|
||
attrs = stmt.attrs
|
||
|
||
try:
|
||
if len(left_vars) > 1 or len(right_vars) > 1:
|
||
if not (len(attrs) == 1 and attrs[0].key == "axis"):
|
||
raise ValueError(
|
||
f"When split/concat, only support one attr named `axis`, but got {attrs}."
|
||
)
|
||
axis = attrs[0].value
|
||
|
||
if len(left_vars) == 1:
|
||
in_name = left_vars[0].name
|
||
in_ref = _get_var_ref(left_vars[0])
|
||
ndim = len(in_ref.shape)
|
||
assert 0 <= axis < ndim, (
|
||
f"when split, the axis {axis} is out of range for tensor {in_name} "
|
||
f"with shape {in_ref.shape} (valid range: {0} to {ndim - 1})."
|
||
)
|
||
assert in_ref.shape[axis] % len(right_vars) == 0, (
|
||
f"when split, the shape of the input tensor {in_name} is {in_ref.shape}, the axis is {axis}, the number of right_vars is {len(right_vars)}, but the shape of the input tensor {in_name} is not divisible by the number of right_vars."
|
||
)
|
||
sizes = [
|
||
in_ref.shape[axis] // len(right_vars)
|
||
for var in right_vars
|
||
]
|
||
result = self.split(in_ref, axis, sizes)
|
||
for out_var, out_ref in zip(right_vars, result):
|
||
self.intermediate_vars[out_var.name] = out_ref
|
||
if (
|
||
out_var.name
|
||
in self.context.get_all_dst_state_keys()
|
||
):
|
||
self.output_vars[out_var.name] = out_ref
|
||
|
||
elif len(right_vars) == 1:
|
||
left_refs = [_get_var_ref(var) for var in left_vars]
|
||
result = self.concat(left_refs, axis)
|
||
out_name = right_vars[0].name
|
||
self.intermediate_vars[out_name] = result
|
||
if out_name in self.context.get_all_dst_state_keys():
|
||
self.output_vars[out_name] = result
|
||
|
||
else:
|
||
raise SyntaxError(
|
||
f'Unexpected split/concat statement: {stmt}'
|
||
)
|
||
|
||
elif len(left_vars) == 1 and len(right_vars) == 1:
|
||
lvar, rvar = left_vars[0], right_vars[0]
|
||
if rvar.name == "_":
|
||
self.need_remove_input_vars.add(lvar.name)
|
||
elif lvar.name == "_":
|
||
self.need_add_output_vars.add(rvar.name)
|
||
else:
|
||
if len(attrs) > 0:
|
||
assert len(attrs) == 1 or (
|
||
len(attrs) == 2
|
||
and {attr.key for attr in attrs}
|
||
== {"src_dtype", "dst_dtype"}
|
||
), (
|
||
"Only support:\n"
|
||
" - One operator, OR\n"
|
||
" - Two operators with keys {'src_dtype', 'dst_dtype'}."
|
||
)
|
||
attr = attrs[0]
|
||
in_ref = _get_var_ref(lvar)
|
||
if attr.key == "permute":
|
||
if attr.value == "[]":
|
||
ndim = len(in_ref.shape)
|
||
perm = str(list(range(ndim - 1, -1, -1)))
|
||
else:
|
||
perm = attr.value
|
||
if self.aoa_config_reverse:
|
||
perm = str(
|
||
invert_permutation(
|
||
ast.literal_eval(perm)
|
||
)
|
||
)
|
||
result = self.transpose(in_ref, perm)
|
||
elif attr.key == "dtype":
|
||
assert not self.aoa_config_reverse, (
|
||
"When `aoa_config_reverse=True`, the dtype must be specified as "
|
||
"'src_dtype=...,dst_dtype=...'. Formats like 'dtype=xxx' are not supported."
|
||
)
|
||
assert attr.value in SUPPORTED_DTYPES, (
|
||
f"Unsupported cast dtype: {attr.value}"
|
||
)
|
||
result = self.cast(in_ref, attr.value)
|
||
elif (
|
||
attrs[0].key == "src_dtype"
|
||
and attrs[1].key == "dst_dtype"
|
||
):
|
||
src_dtype, dst_dtype = (
|
||
attrs[0].value,
|
||
attrs[1].value,
|
||
)
|
||
assert src_dtype in SUPPORTED_DTYPES, (
|
||
f"Unsupported cast dtype: {src_dtype}"
|
||
)
|
||
assert dst_dtype in SUPPORTED_DTYPES, (
|
||
f"Unsupported cast dtype: {dst_dtype}"
|
||
)
|
||
if self.aoa_config_reverse:
|
||
src_dtype, dst_dtype = dst_dtype, src_dtype
|
||
result = self.cast(in_ref, dst_dtype)
|
||
elif attr.key == "axis":
|
||
result = in_ref
|
||
else:
|
||
raise ValueError(
|
||
f"Unsupported attribute: {attr}"
|
||
)
|
||
|
||
self.intermediate_vars[rvar.name] = result
|
||
if (
|
||
rvar.name
|
||
in self.context.get_all_dst_state_keys()
|
||
):
|
||
self.output_vars[rvar.name] = result
|
||
else:
|
||
# rename operation
|
||
in_ref = _get_var_ref(lvar)
|
||
result = self.identity(in_ref)
|
||
self.intermediate_vars[rvar.name] = result
|
||
if (
|
||
rvar.name
|
||
in self.context.get_all_dst_state_keys()
|
||
):
|
||
self.output_vars[rvar.name] = result
|
||
else:
|
||
raise SyntaxError(f'Unexpected statement: {stmt}')
|
||
except (
|
||
AssertionError,
|
||
ValueError,
|
||
KeyError,
|
||
SyntaxError,
|
||
RuntimeError,
|
||
) as e:
|
||
if self.traceback:
|
||
chain = self.traceback.build_chain(stmt_repr)
|
||
self.traceback.add_error(
|
||
error_message=str(e),
|
||
stage="shape_propagation",
|
||
chain=chain,
|
||
error_type=type(e).__name__,
|
||
)
|
||
self.traceback.print()
|
||
raise
|
||
if self.destination_state_shard_info is not None:
|
||
for name in self.destination_state_shard_info:
|
||
model_state_key, _ = split_optimizer_state_key(name)
|
||
if model_state_key not in self.output_vars:
|
||
if model_state_key in self.need_add_output_vars:
|
||
self.output_vars[model_state_key] = None
|
||
else:
|
||
assert model_state_key in self.input_vars, (
|
||
f"{model_state_key} needs to be loaded, "
|
||
f"but not found in checkpoint. "
|
||
f"If the key exists in the current model but not in the loaded checkpoint, please use the add primitive in aoa_statements: "
|
||
f"_ -> {model_state_key}, and {model_state_key} will be randomly initialized."
|
||
)
|
||
self.output_vars[model_state_key] = self.input_vars[
|
||
model_state_key
|
||
]
|
||
else:
|
||
# When destination_state_shard_info is not provided, the AOAEngine automatically derives it
|
||
# from source_state_shard_info and aha_statements. In this case, all destination_states
|
||
# remain unsharded (not partitioned).
|
||
for name, ref_t in self.input_vars.items():
|
||
if (
|
||
name not in self.output_vars
|
||
and ref_t.out_degree == 0
|
||
and name not in self.need_remove_input_vars
|
||
):
|
||
self.output_vars[name] = self.identity(ref_t)
|
||
for name, ref_t in self.intermediate_vars.items():
|
||
if name not in self.output_vars and ref_t.out_degree == 0:
|
||
self.output_vars[name] = self.identity(ref_t)
|
||
|
||
def find_source_slices(
|
||
self, key: str, local_slice: tuple[slice, ...]
|
||
) -> list[SliceRef]:
|
||
assert key in self.output_vars, (
|
||
f"The key {key} is not in the output_vars (which is built during load_state_dict)."
|
||
)
|
||
tensor = self.output_vars[key]
|
||
if tensor is None:
|
||
return []
|
||
results = []
|
||
assert len(local_slice) == len(tensor.shape), (
|
||
f"For the key {key}, the target_tensor has {len(local_slice)} dimensions, "
|
||
f"but the tensor in output_vars has {len(tensor.shape)} dimensions (shape={tensor.shape}). "
|
||
)
|
||
ndim = len(tensor.shape)
|
||
|
||
def slice_intersect(a: slice, b: slice):
|
||
start = max(a.start, b.start)
|
||
stop = min(a.stop, b.stop)
|
||
if start >= stop:
|
||
return None
|
||
return slice(start, stop, 1)
|
||
|
||
for src_key, sl_src, sl_dst, pp_list in tensor.slices:
|
||
intersection = []
|
||
for i in range(ndim):
|
||
inter = slice_intersect(local_slice[i], sl_dst[i])
|
||
if inter is None:
|
||
break
|
||
intersection.append(inter)
|
||
else:
|
||
# Compute corresponding src_slice for the intersection
|
||
if pp_list is not None:
|
||
sl_src = postprocess_transpose(list(sl_src), pp_list)
|
||
src_slice = []
|
||
for i in range(ndim):
|
||
dst = sl_dst[i]
|
||
src = sl_src[i]
|
||
dst_start = dst.start
|
||
src_start = src.start
|
||
inter_start, inter_stop = (
|
||
intersection[i].start,
|
||
intersection[i].stop,
|
||
)
|
||
offset = inter_start - dst_start
|
||
src_inter_start = src_start + offset
|
||
src_inter_stop = src_inter_start + (
|
||
inter_stop - inter_start
|
||
)
|
||
src_slice.append(slice(src_inter_start, src_inter_stop, 1))
|
||
if pp_list is not None:
|
||
src_slice = postprocess_transpose(
|
||
list(src_slice), pp_list, reverse=True
|
||
)
|
||
results.append(
|
||
(
|
||
src_key,
|
||
tuple(src_slice),
|
||
tuple(intersection),
|
||
pp_list.copy(),
|
||
),
|
||
)
|
||
else:
|
||
results.append(
|
||
(src_key, tuple(src_slice), tuple(intersection), None)
|
||
)
|
||
return results
|
||
|
||
def find_shard_sources(
|
||
self,
|
||
target: ShardedWeightDesc,
|
||
) -> ShardMapping:
|
||
target_key, opt_state_name = split_optimizer_state_key(target.key)
|
||
target_local_shape = target.local_shape
|
||
target_global_offset = target.global_offset
|
||
target_global_shape = target.global_shape
|
||
|
||
if opt_state_name in [".beta1_pow_acc_0", ".beta2_pow_acc_0"]:
|
||
assert target_key in self.output_vars, (
|
||
f"The key {target_key} is not in the output_vars (which is built during load_state_dict)."
|
||
)
|
||
tensor = self.output_vars[target_key]
|
||
target_local_shape = tensor.shape
|
||
target_global_offset = (0,) * len(target_local_shape)
|
||
target_global_shape = target_local_shape
|
||
|
||
slices = tuple(
|
||
slice(offset, offset + size, 1)
|
||
for offset, size in zip(target_global_offset, target_local_shape)
|
||
)
|
||
|
||
results = self.find_source_slices(target_key, slices)
|
||
|
||
shard_mappings = []
|
||
|
||
target_key = (
|
||
target_key + opt_state_name
|
||
if opt_state_name is not None
|
||
else target_key
|
||
)
|
||
|
||
src_keys = {
|
||
result[0]
|
||
for result in results
|
||
if result[0] not in self.need_remove_input_vars
|
||
}
|
||
if opt_state_name in [".beta1_pow_acc_0", ".beta2_pow_acc_0"]:
|
||
if len(src_keys) == 0:
|
||
return shard_mappings
|
||
elif len(src_keys) > 1:
|
||
logger.warning(
|
||
f"{target_key} has multiple sources: {src_keys} (e.g., .beta1_pow_acc_0). Returning one arbitrarily."
|
||
)
|
||
src_key = next(iter(src_keys))
|
||
else:
|
||
src_key = next(iter(src_keys))
|
||
return [
|
||
ShardMappingEntry(
|
||
target,
|
||
ShardedWeightDesc(
|
||
src_key + opt_state_name,
|
||
target.local_shape,
|
||
target.global_shape,
|
||
target.global_offset,
|
||
target.dtype,
|
||
),
|
||
None,
|
||
)
|
||
]
|
||
|
||
for src_key, src_slices, local_slices, pp_list in results:
|
||
src_var = self.input_vars[src_key]
|
||
target_model_state_key, target_opt_state_name = (
|
||
split_optimizer_state_key(target.key)
|
||
)
|
||
if target_opt_state_name is None:
|
||
if src_var.dtype != target.dtype:
|
||
assert pp_list is not None and target.dtype in str(
|
||
pp_list
|
||
), (
|
||
"Direct assignment of Tensors with different types is prohibited in AOA. "
|
||
f"If you want to achieve this functionality, please use the cast semantics provided by AOA. "
|
||
f"Now the src_var.dtype is {src_var.dtype}, the target.dtype is {target.dtype}, the pp_list is {pp_list}."
|
||
f"The src_key is {src_key}, the target_key is {target.key}."
|
||
)
|
||
else:
|
||
src_var.dtype = target.dtype
|
||
|
||
src_global_shape = src_var.shape
|
||
|
||
src_local_shape = tuple(slc.stop - slc.start for slc in src_slices)
|
||
src_global_offset = tuple(slc.start for slc in src_slices)
|
||
|
||
tgt_local_shape = tuple(
|
||
slc.stop - slc.start for slc in local_slices
|
||
)
|
||
tgt_global_offset = tuple(slc.start for slc in local_slices)
|
||
|
||
new_src_key = (
|
||
src_key + opt_state_name
|
||
if opt_state_name is not None
|
||
else src_key
|
||
)
|
||
|
||
source_sharded_weight = ShardedWeightDesc(
|
||
new_src_key,
|
||
src_local_shape,
|
||
tuple(src_global_shape),
|
||
src_global_offset,
|
||
src_var.dtype,
|
||
)
|
||
target_sharded_weight = ShardedWeightDesc(
|
||
target_key,
|
||
tgt_local_shape,
|
||
tuple(target_global_shape),
|
||
tgt_global_offset,
|
||
target.dtype,
|
||
)
|
||
|
||
if src_key in self.need_remove_input_vars:
|
||
mapping_entry = ShardMappingEntry(
|
||
target_sharded_weight,
|
||
source_sharded_weight,
|
||
[],
|
||
)
|
||
continue
|
||
|
||
shard_mappings.append(
|
||
ShardMappingEntry(
|
||
target_sharded_weight,
|
||
source_sharded_weight,
|
||
pp_list,
|
||
)
|
||
)
|
||
|
||
return shard_mappings
|
||
|
||
|
||
def postprocess_transpose(
|
||
li: list[tuple[slice, ...]] | tuple[tuple[slice, ...]],
|
||
postprocess_list: list[str],
|
||
reverse: bool = False,
|
||
) -> list[tuple[slice, ...]] | tuple[tuple[slice, ...]]:
|
||
result = li
|
||
if reverse:
|
||
for pp in list(reversed(postprocess_list)):
|
||
if pp.startswith("["):
|
||
reversed_transpose = np.argsort(ast.literal_eval(pp)).tolist()
|
||
result = transpose_list(result, reversed_transpose)
|
||
else:
|
||
for pp in postprocess_list:
|
||
if pp.startswith("["):
|
||
result = transpose_list(result, ast.literal_eval(pp))
|
||
return result
|
||
|
||
|
||
def transpose_list(
|
||
li: list[tuple[slice, ...]] | tuple[tuple[slice, ...]],
|
||
permutation: list[int],
|
||
) -> list[tuple[slice, ...]] | tuple[tuple[slice, ...]]:
|
||
trans_list = []
|
||
for idx in permutation:
|
||
trans_list.append(li[idx])
|
||
if isinstance(li, tuple):
|
||
return tuple(trans_list)
|
||
else:
|
||
return trans_list
|
||
|
||
|
||
def invert_permutation(p: list[int]) -> list[int]:
|
||
q = [0] * len(p)
|
||
for i, pi in enumerate(p):
|
||
q[pi] = i
|
||
return q
|