rocm_jax/jax/_src/cache_key.py
2025-04-07 23:30:31 +00:00

367 lines
13 KiB
Python

# Copyright 2023 The JAX Authors.
#
# 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
#
# https://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.
import copy
import enum
import hashlib
import io
import logging
import os
import sys
from typing import cast as type_cast
from jax._src import config
from jax._src.lib import version_str as jaxlib_version_str
from jax._src.lib import xla_client
from jax._src.lib.mlir import ir
from jax._src.lib.mlir import passmanager as pm
import numpy as np
logger = logging.getLogger(__name__)
_extra_flag_prefixes: list[str] = []
def add_flag_prefixes(flag_prefixes: list[str]) -> None:
"""Add flag prefixes to include in the cache key. Call prior to get().
"""
global _extra_flag_prefixes
_extra_flag_prefixes += flag_prefixes
def clear_flag_prefixes() -> None:
"""Clear flag prefixes added by add_flag_prefixes().
"""
global _extra_flag_prefixes
_extra_flag_prefixes = []
def get_flag_prefixes() -> list[str]:
"""Return flag prefixes added by add_flag_prefixes().
"""
return _extra_flag_prefixes
def custom_hook() -> str:
"""Custom hook for any addition to the cache key.
The custom hook will be called everytime get() is called and can be
defined to return a string that will be hashed into the cache key.
"""
return ""
class IgnoreCallbacks(enum.IntEnum):
# Do not remove any callback pointers from precompiled IR.
NO = enum.auto()
# Remove all callback pointers from precompiled IR.
ALL = enum.auto()
# Remove only custom_partitioning callback pointer from precompiled IR.
CUSTOM_PARTITIONING = enum.auto()
def get(
module: ir.Module,
devices: np.ndarray,
compile_options: xla_client.CompileOptions,
backend: xla_client.Client,
compression_algorithm: str = "zstandard",
ignore_callbacks: IgnoreCallbacks = IgnoreCallbacks.NO,
) -> str:
"""Creates a hashed string to use as a key to the compilation cache.
Creates a cache key that is a hex-encoded string of a unique hash based on
the arguments. The hex-encoded string is 256 characters long.
Args:
module: the input program
devices: an array of accelerator devices that the program will run on
compile_options: options passed to the XLA compiler
backend: description of the platform (e.g., TPU version)
compression_algorithm: a string representing the compression algorithm used
for the executable before persisting in the cache
ignore_callbacks: whether to remove the all callback pointer from the
computation.
Typical return value example:
'jit__psum-14ac577cdb2ef6d986078b4054cc9893a9a14a16dbb0d8f37b89167c1f1aacdf'
"""
entries = [
(
"computation",
lambda hash_obj: _hash_computation(
hash_obj, module, ignore_callbacks
),
),
(
"jax_lib version",
lambda hash_obj: hash_obj.update(
bytes(jaxlib_version_str.encode("utf-8"))
),
),
(
"backend version",
lambda hash_obj: _hash_platform(hash_obj, backend)
),
(
"XLA flags",
lambda hash_obj: _hash_xla_flags(hash_obj, get_flag_prefixes()),
),
(
"compile_options",
lambda hash_obj: _hash_serialized_compile_options(
hash_obj,
compile_options,
# In case of GPU multi-process tasks we need to strip device
# assignment to use cache key as invariant between processes.
strip_device_assignment=(backend.platform == "gpu"),
),
),
(
"accelerator_config",
lambda hash_obj: _hash_accelerator_config(hash_obj, devices),
),
(
"compression",
lambda hash_obj: _hash_string(hash_obj, compression_algorithm),
),
("custom_hook", lambda hash_obj: _hash_string(hash_obj, custom_hook())),
]
hash_obj = hashlib.sha256()
for name, hashfn in entries:
hashfn(hash_obj)
_log_cache_key_hash(hash_obj, name, hashfn)
sym_name = module.operation.attributes['sym_name']
module_name = ir.StringAttr(sym_name).value
return module_name + "-" + hash_obj.digest().hex()
def _log_cache_key_hash(hash_obj, last_serialized: str, hashfn):
if logger.isEnabledFor(logging.DEBUG):
# Log the hash of just this entry
fresh_hash_obj = hashlib.sha256()
hashfn(fresh_hash_obj)
logger.debug(
"get_cache_key hash of serialized %s: %s",
last_serialized,
fresh_hash_obj.digest().hex(),
)
# Log the cumulative hash
logger.debug(
"get_cache_key hash after serializing %s: %s",
last_serialized,
hash_obj.digest().hex(),
)
def _remove_callbacks(m: ir.Module, ignore_callbacks: IgnoreCallbacks):
"""Removes callback pointers from precompiled IR.
Python function pointers are not deterministic across executions.
"""
def _update_bc_attribute(op: ir.Operation) -> ir.WalkResult:
if op.name == "stablehlo.custom_call" and (
(
ignore_callbacks == IgnoreCallbacks.ALL
and op.attributes["call_target_name"].value.endswith("callback")
)
or op.attributes["call_target_name"].value == "CustomSPMDPartitioning"
):
op.attributes["backend_config"] = ir.StringAttr.get("REMOVED")
return ir.WalkResult.ADVANCE
if ignore_callbacks == IgnoreCallbacks.NO:
return m
m.operation.walk(_update_bc_attribute)
return m
def _serialize_ir(m: ir.Module, ignore_callbacks: IgnoreCallbacks) -> bytes:
output = io.BytesIO()
if ignore_callbacks != IgnoreCallbacks.NO:
m = _remove_callbacks(
type_cast(ir.Module, m.operation.clone()), ignore_callbacks
)
m.operation.write_bytecode(file=output)
return output.getvalue()
def _canonicalize_ir(
m_original: ir.Module, ignore_callbacks: IgnoreCallbacks
) -> bytes:
with m_original.context:
m = type_cast(ir.Module, m_original.operation.clone())
passes = pm.PassManager.parse(
"builtin.module(strip-debuginfo)"
)
passes.run(m.operation)
return _serialize_ir(m, ignore_callbacks)
def _hash_computation(hash_obj, module, ignore_callbacks: IgnoreCallbacks):
if config.compilation_cache_include_metadata_in_key.value:
canonical_ir = _serialize_ir(module, ignore_callbacks)
else:
canonical_ir = _canonicalize_ir(module, ignore_callbacks)
hash_obj.update(canonical_ir)
def _hash_devices(hash_obj, devices: np.ndarray) -> None:
for device in devices.flat:
_hash_string(hash_obj, device.device_kind)
def _hash_accelerator_config(hash_obj, accelerators: np.ndarray):
accelerator_devices = []
for accelerator in accelerators.flat:
accelerator_devices.append(accelerator)
try:
hash_obj.update(
xla_client.get_topology_for_devices(accelerator_devices).serialize()
)
except xla_client._xla.XlaRuntimeError as ex:
# Fall back for those backends that do not support serialized
# PjRtTopologyDescription as yet.
logger.info("get (_hash_accelerator_config): unable to hash "
"accelerator config, falling back to hashing "
"devices %s (type %s)", ex, type(ex))
_hash_devices(hash_obj, accelerators)
# LINT.IfChange(xla_flags)
xla_flags_to_exclude_from_cache_key = [
"--xla_dump_compress_protos",
"--xla_dump_module_metadata",
"--xla_dump_max_hlo_modules",
"--xla_dump_include_timestamp",
"--xla_dump_hlo_pass_re",
"--xla_dump_hlo_module_re",
"--xla_dump_hlo_snapshots",
"--xla_dump_fusion_visualization",
"--xla_dump_hlo_as_url",
"--xla_dump_hlo_as_proto",
"--xla_dump_hlo_as_text",
"--xla_dump_hlo_as_long_text",
"--xla_dump_hlo_as_html",
"--xla_dump_hlo_as_dot",
"--xla_dump_to",
"--xla_force_host_platform_device_count",
"--xla_dump_disable_metadata",
"--xla_dump_hlo_pipeline_re",
"--xla_tpu_sdc_checker_streamz_metric",
"--xla_tpu_sdc_checker_enable_sdc_event_callbacks",
"--xla_tpu_sdc_checker_enable_coresweep_ng_callbacks",
"--xla_tpu_sdc_checker_no_logging_if_callbacks_are_present",
"--xla_gpu_cuda_data_dir",
"--xla_gpu_experimental_autotune_cache_mode",
]
env_override_flags_to_exclude_from_cache_key = {
x.strip("-") for x in xla_flags_to_exclude_from_cache_key
}
# LINT.ThenChange(:debug_options)
def _hash_serialized_compile_options(hash_obj, compile_options_obj,
strip_device_assignment=False):
# Do not mess with the original CompileOptions object since it is passed to
# the compiler. Create a deep copy for the purpose of cache key generation.
compile_options_copy = copy.deepcopy(compile_options_obj)
# Certain debug options do not affect the compile result and thus, should not
# be part of the cache key as their inclusion will result in unnecessary cache
# misses. Clear them here by setting bool values to False, ints to 0, and
# strings to empty. The exact values used to clear are not relevant as long
# as the same values are used every time for each field.
debug_options = compile_options_copy.executable_build_options.debug_options
# LINT.IfChange(debug_options)
debug_options.xla_force_host_platform_device_count = 0
debug_options.xla_dump_to = ""
debug_options.xla_dump_hlo_module_re = ""
debug_options.xla_dump_hlo_pass_re = ""
debug_options.xla_dump_hlo_as_text = False
debug_options.xla_dump_hlo_as_proto = False
debug_options.xla_dump_hlo_as_dot = False
debug_options.xla_dump_hlo_as_url = False
debug_options.xla_dump_hlo_as_html = False
debug_options.xla_dump_fusion_visualization = False
debug_options.xla_dump_hlo_snapshots = False
debug_options.xla_dump_max_hlo_modules = False
debug_options.xla_dump_module_metadata = False
debug_options.xla_dump_compress_protos = False
debug_options.xla_dump_hlo_as_long_text = False
debug_options.xla_dump_disable_metadata = False
debug_options.xla_dump_hlo_pipeline_re = ""
debug_options.xla_gpu_experimental_autotune_cache_mode = 0
# Optional way to specify the cuda install path to be used by the compiler.
# This could possibly affect the cuda version compiled with, but this should
# already be included in the platform information (and might not be reflected
# by the cuda path regardless, since this only hashes on the directory name
# and not the contents). It can also cause spurious cache misses if the cuda
# path changes across runs despite being the same version, so we clear it
# here.
debug_options.xla_gpu_cuda_data_dir = ""
# LINT.ThenChange(:xla_flags)
compile_options_copy.env_option_overrides = [
flag_value
for flag_value in compile_options_copy.env_option_overrides
if flag_value[0] not in env_override_flags_to_exclude_from_cache_key
]
if strip_device_assignment and compile_options_copy.device_assignment:
replica_count = compile_options_copy.device_assignment.replica_count()
computation_count = compile_options_copy.device_assignment.computation_count()
compile_options_copy.device_assignment = xla_client.DeviceAssignment.create(
np.arange(replica_count * computation_count).reshape(
[replica_count, computation_count])
)
return hash_obj.update(compile_options_copy.SerializeAsString())
def _hash_platform(hash_obj, backend):
_hash_string(hash_obj, backend.platform)
_hash_string(hash_obj, backend.platform_version)
_hash_string(hash_obj, backend.runtime_type)
def _hash_xla_flags(hash_obj, extra_flag_prefixes: list[str]):
xla_flags = []
xla_flags_env_var = os.getenv("XLA_FLAGS")
if xla_flags_env_var:
xla_flags.extend(xla_flags_env_var.split())
libtpu_init_args_env_var = os.getenv("LIBTPU_INIT_ARGS")
if libtpu_init_args_env_var:
xla_flags.extend(libtpu_init_args_env_var.split())
for arg in sys.argv:
if arg.startswith("--xla") or any(
arg.startswith(p) for p in extra_flag_prefixes
):
xla_flags.append(arg)
# N.B. all XLA flags that take an argument must use '=' and not a space
# (e.g. --xla_force_host_platform_device_count=8) (I think).
for flag in sorted(xla_flags):
if flag.split("=")[0] in xla_flags_to_exclude_from_cache_key:
logger.debug("Not including XLA flag in cache key: %s", flag)
continue
logger.debug("Including XLA flag in cache key: %s", flag)
_hash_string(hash_obj, flag)
def _hash_string(hash_obj, str_var):
hash_obj.update(str_var.encode("utf-8").strip())