jax.experimental.custom_partitioning 模块#
API#
- jax.experimental.custom_partitioning.custom_partitioning(fun, static_argnums=())[source]#
在 XLA 图中插入一个带有自定义 SPMD 降低(lowering)规则的 CustomCallOp。
@custom_partitioning def f(*args): return ... def propagate_user_sharding(mesh, user_shape): '''Update the sharding of the op from a user's shape.sharding.''' user_sharding = jax.tree.map(lambda x: x.sharding, user_shape) def partition(mesh, arg_shapes, result_shape): def lower_fn(*args): ... builds computation on per-device shapes ... result_shardings = jax.tree.map(lambda x: x.sharding, result_shape) arg_shardings = jax.tree.map(lambda x: x.sharding, arg_shapes) # result_sharding and arg_shardings may optionally be modified and the # partitioner will insert collectives to reshape. return mesh, lower_fn, result_sharding, arg_shardings def infer_sharding_from_operands(mesh, arg_shapes, shape): '''Compute the result sharding from the sharding of the operands.''' arg_shardings = jax.tree.map(lambda x: x.sharding, arg_shapes) f.def_partition(partition, propagate_user_sharding, infer_sharding_from_operands=infer_sharding_from_operands, sharding_rule='i j -> 'i j')
def_partition的参数如下:propagate_user_sharding:一个可调用对象,它接收用户(在 DAG 中)的分片信息,并返回对新的 NamedSharding 的建议。默认值为 None。一个简单的实现就是直接返回输入的切分。partition:一个可调用对象,它接收 SPMD 建议的分区形状和分区规格(partition specs),并返回网格(mesh)、每个分片的降低函数,以及最终的输入和输出分片规格(SPMD 分区器将重新分区输入以进行匹配)。返回网格是为了在未提供网格时允许配置集合通信操作(collectives)的 axis_names。infer_sharding_from_operands:一个可调用对象,它根据为每个参数选择的NamedSharding计算出输出的NamedSharding。decode_shardings:当设置为 True 时,尽可能将输入的GSPMDSharding转换为NamedSharding。如果用户未提供上下文网格,则可能无法实现。sharding_rule:一个 SdyShardingRule 对象、一个描述分片规则的类 Einsum 符号字符串,或者一个可以生成上述任一内容的可调用对象。我们将分片规则中 Einsum 符号的索引标签称为“因子”(factors)。我们借鉴了 einops.rearrange 字符串的概念,在因子之间使用空格分隔,并允许使用多字母因子名称。默认情况下,因子对应于直通/元素级维度。对应于其他维度的因子可以通过下述关键字参数指定。详情和示例请参阅 jax-shardy-guide。reduction_factors:一个字符串元组,指定字符串 sharding_rule 的缩减因子(reduction factors)。缩减因子对应于出现在操作数中但不出现在结果中的维度,例如矩阵乘法中的收缩维度。如果缩减因子被分片,结果将需要在相同的轴上进行 all-reduce。need_replication_factors:一个字符串元组,指定字符串 sharding_rule 的 need_replication 因子。need_replication 因子对应于为了支持实现而不应被分片的维度。permutation_factors:一个字符串元组,指定字符串 sharding_rule 的置换因子。置换因子对应于如果被分片则会触发集合置换(collective permute)的维度。factor_sizes:一个变量关键字参数字典,指定仅在字符串 sharding_rule 的复合因子中使用的因子大小。
当 config.use_shardy_partitioner.value 为 True 时,使用 sharding_rule;否则使用 propagate_user_sharding 和 infer_sharding_from_operands。
可以使用 static_argnums 将位置参数指定为静态参数。JAX 使用
inspect.signature(fun)来解析这些位置参数。示例
例如,假设我们要增强现有的
jax.numpy.fft.fft。该函数计算 N 维输入沿最后一个维度的离散傅里叶变换,并沿前 N-1 个维度进行批处理。然而,默认情况下,它会忽略输入的分片并将输入收集到所有设备上。由于jax.numpy.fft.fft是沿前 N-1 个维度进行批处理的,这样做是不必要的。我们将创建一个新的my_fft操作,它不会改变沿前 N-1 个维度的分片,并且仅在必要时沿最后一个维度收集输入。import jax from jax.sharding import NamedSharding from jax.experimental.custom_partitioning import custom_partitioning from jax.experimental.pjit import pjit from jax.sharding import PartitionSpec as P from jax.sharding import Mesh from jax.numpy.fft import fft import regex as re import numpy as np # Pattern to detect all-gather or dynamic-slice in the generated HLO _PATTERN = '(dynamic-slice|all-gather)' # For an N-D input, keeps sharding along the first N-1 dimensions # but replicate along the last dimension def supported_sharding(sharding, shape): rank = len(shape.shape) max_shared_dims = min(len(sharding.spec), rank-1) names = tuple(sharding.spec[:max_shared_dims]) + tuple(None for _ in range(rank - max_shared_dims)) return NamedSharding(sharding.mesh, P(*names)) def partition(mesh, arg_shapes, result_shape): result_shardings = jax.tree.map(lambda x: x.sharding, result_shape) arg_shardings = jax.tree.map(lambda x: x.sharding, arg_shapes) return mesh, fft, supported_sharding(arg_shardings[0], arg_shapes[0]), (supported_sharding(arg_shardings[0], arg_shapes[0]),) def infer_sharding_from_operands(mesh, arg_shapes, result_shape): arg_shardings = jax.tree.map(lambda x: x.sharding, arg_shapes) return supported_sharding(arg_shardings[0], arg_shapes[0]) @custom_partitioning def my_fft(x): return fft(x) # Use Einsum-like notation to specify the sharding rule. my_fft.def_partition( infer_sharding_from_operands=infer_sharding_from_operands, partition=partition, sharding_rule='...i -> ...i') # Use SdyShardingRule object to specify the sharding rule. my_fft.def_partition( infer_sharding_from_operands=infer_sharding_from_operands, partition=partition, sharding_rule=SdyShardingRule(operand_mappings=((BATCHING, 'i'),), result_mappings=((BATCHING, 'i'),))))
现在创建一个沿第一个轴分片的 2D 数组,将其通过
my_fft,观察它如何保持预期的分片方式,并且与fft的输出相同。然而,检查 HLO(使用lower(x).compile().runtime_executable().hlo_modules())会发现,my_fft不会创建任何 all-gather 或 dynamic-slice 操作,而fft会。with Mesh(np.array(jax.devices()), ('x',)): x = np.asarray(np.random.randn(32*1024, 1024), dtype=np.complex64) y = pjit(lambda x: x, in_shardings=None, out_shardings=P('x'))(x) pjit_my_fft = pjit(my_fft, in_shardings=P('x'), out_shardings=P('x')) pjit_fft = pjit(fft, in_shardings=P('x'), out_shardings=P('x')) print(pjit_my_fft(y)) print(pjit_fft(y)) # dynamic-slice or all-gather are not present in the HLO for my_fft, because x is a 2D array assert(re.search(_PATTERN, pjit_my_fft.lower(x).compile().runtime_executable().hlo_modules()[0].to_string()) is None) # dynamic-slice or all-gather are present in the HLO for fft assert(re.search(_PATTERN, pjit_fft.lower(x).compile().runtime_executable().hlo_modules()[0].to_string()) is not None)
# my_fft [[-38.840824 +0.j -40.649452 +11.845365j ... -1.6937828 +0.8402481j 15.999859 -4.0156755j]] # jax.numpy.fft.fft [[-38.840824 +0.j -40.649452 +11.845365j ... -1.6937828 +0.8402481j 15.999859 -4.0156755j]]
由于
supported_sharding中的逻辑,my_fft也适用于一维数组。但在这种情况下,my_fft的 HLO 确实显示了 dynamic-slice,因为最后一个维度是计算 FFT 的维度,在进行计算之前需要将其复制到所有设备上。with Mesh(np.array(jax.devices()), ('x',)): x = np.asarray(np.random.randn(32*1024*1024), dtype=np.complex64) y = pjit(lambda x: x, in_shardings=None, out_shardings=P('x'))(x) pjit_my_fft = pjit(my_fft, in_shardings=P('x'), out_shardings=P('x')) pjit_fft = pjit(fft, in_shardings=P('x'), out_shardings=P('x')) print(pjit_my_fft(y)) print(pjit_fft(y)) # dynamic-slice or all-gather are present in the HLO for my_fft, because x is a 1D array assert(re.search(_PATTERN, pjit_my_fft.lower(x).compile().runtime_executable().hlo_modules()[0].to_string()) is None) # dynamic-slice or all-gather are present in the HLO for fft assert(re.search(_PATTERN, pjit_fft.lower(x).compile().runtime_executable().hlo_modules()[0].to_string()) is not None)
# my_fft [ 7.217285 +0.j -3012.4937 +4287.635j -405.83594 +3042.984j ... 1422.4502 +7271.4297j -405.84033 -3042.983j -3012.4963 -4287.6343j] # jax.numpy.fft.fft [ 7.217285 +0.j -3012.4937 +4287.635j -405.83594 +3042.984j ... 1422.4502 +7271.4297j -405.84033 -3042.983j -3012.4963 -4287.6343j]