guides/content/static_pages/jax-xla-options/index.md at main · Findus23/guides

title	All XLA Options/Flags
slug	all-xla-options
date	2025-04-08
toc	end
description	A list of all XLA options extracted from the latest JAX version

Unfortunately the JAX documentation only seems to list a few common XLA flags. The rest of them is not documented at all outside of the OpenXLA source code. Here I am listing all of them as of JAX/jaxlib 0.10.2 (XLA 5a9e73cb). Keep in mind that most of them are experimental and don't depend on their behaviour to be stable between JAX/XLA versions.

If you are interested in the XLA flags of earlier JAX versions, check out the older versions of this page.

import subprocess
import sys

use_jax = False

if use_jax:
    out = subprocess.run(
        [
            sys.executable,
            "-c",
            "import jax;jax.numpy.zeros(10)"
        ],
        env={"XLA_FLAGS": "--help"},
        capture_output=True
    )
    text = out.stderr.decode().split("Flags:")[2].rstrip()
else:
    # read from output of `bazel run //xla/tools:hlo-opt -- --help`
    with open("out.txt") as f:
        text = f.read().rstrip()

for line in text.splitlines()[1:]:
    print("")
    if not line.lstrip().startswith("--"):
        print(line)
        continue
    _, arg, type, description = line.split("\t", maxsplit=3)
    arg_name, default = arg.strip().split("=")
    print(f"## {arg_name}")
    print(f"- default: `{default}`")
    print(f"- type: **{type}**")
    if description.startswith("[Stable]"):
        print(f"- **[Stable]**")
        description = description.replace("[Stable]", "").lstrip()
    print("")
    print(description)

All content below is extracted from https://github.com/openxla/xla and licensed under the Apache License 2.0.

--xla_hlo_evaluator_use_fast_path

default: false
type: bool

Enable fast evaluation of dots in the HloEvaluator

--xla_cpu_enable_fast_math

default: false
type: bool

Enable unsafe fast-math optimizations in the CPU compiler; this may produce faster code at the expense of some accuracy.

--xla_cpu_enable_platform_dependent_math

default: true
type: bool

Enable platform dependent math in the CPU compiler; this may produce faster code at the expense of consistent results across CPUs.

--xla_cpu_experimental_enable_tiling_propagation

default: false
type: bool

If true, enable experimental tiling propagation for CPU.

--xla_cpu_fast_math_honor_nans

default: true
type: bool

When xla_cpu_enable_fast_math is true then this controls whether we allow operations to produce NaNs. Ignored when xla_cpu_enable_fast_math is false.

--xla_cpu_fast_math_honor_infs

default: true
type: bool

When xla_cpu_enable_fast_math is true then this controls whether we allow operations to produce infinites. Ignored when xla_cpu_enable_fast_math is false.

--xla_cpu_fast_math_honor_division

default: true
type: bool

When xla_cpu_enable_fast_math is true then this controls whether we forbid to use multiplication by the reciprocal instead of division. Ignored when xla_cpu_enable_fast_math is false.

--xla_cpu_fast_math_honor_functions

default: true
type: bool

When xla_cpu_enable_fast_math is true then this controls whether we forbid to approximate calculations for functions. Ignored when xla_cpu_enable_fast_math is false.

--xla_cpu_ftz

default: true
type: bool

If true, flush-to-zero semantics are enabled in the code generated for CPUs. Default is true.

--xla_cpu_enable_fast_min_max

default: true
type: bool

Enable fast floating point min/max lowering that might not propagate NaNs.

--xla_gpu_enable_fast_min_max

default: false
type: bool

Enable fast floating point min/max lowering that does not propagate NaNs.

--xla_llvm_enable_alias_scope_metadata

default: true
type: bool

In LLVM-based backends, enable the emission of !alias.scope metadata in the generated IR.

--xla_llvm_enable_noalias_metadata

default: true
type: bool

In LLVM-based backends, enable the emission of !noalias metadata in the generated IR.

--xla_llvm_enable_invariant_load_metadata

default: true
type: bool

In LLVM-based backends, enable the emission of !invariant.load metadata in the generated IR.

--xla_llvm_disable_expensive_passes

default: false
type: bool

In LLVM-based backends, disable a custom set of expensive optimization passes.

--xla_backend_optimization_level

default: 3
type: int32

Numerical optimization level for the XLA compiler backend.

--xla_disable_hlo_passes

default: ""
type: string

Comma-separated list of hlo passes to be disabled. These names must exactly match the passes' names; no whitespace around commas.

--xla_enable_hlo_passes_only

default: ""
type: string

Comma-separated list of hlo passes to be enabled. These names must exactly match the passes' names; no whitespace around commas. The unspecified passes are all disabled.

--xla_disable_all_hlo_passes

default: false
type: bool

Disables all HLO passes. Notes that some passes are necessary for correctness and the invariants that must be satisfied by 'fully optimized' HLO are different for different devices and may change over time. The only 'guarantee', such as it is, is that if you compile XLA and dump the optimized HLO for some graph, you should be able to run it again on the same device with the same build of XLA.

--xla_embed_ir_in_executable

default: false
type: bool

Embed the compiler IR as a string in the executable.

--xla_eliminate_hlo_implicit_broadcast

default: true
type: bool

Eliminate implicit broadcasts when lowering user computations to HLO instructions; use explicit broadcast instead.

--xla_cpu_multi_thread_eigen

default: true
type: bool

When generating calls to Eigen in the CPU backend, use multi-threaded Eigen mode.

--xla_gpu_cuda_data_dir

default: "./cuda_sdk_lib"
type: string

If non-empty, specifies a local directory containing ptxas and nvvm libdevice files; otherwise we use those from runfile directories.

--xla_gpu_ftz

default: false
type: bool

If true, flush-to-zero semantics are enabled in the code generated for GPUs.

--xla_gpu_ptx_file

default: ""
type: string

If non-empty, specifies a file containing ptx to use. The filename prefix must have the same pattern as PTX dumped by XLA. This allows to match one specific module. General workflow. Get the generated module ptx from XLA, modify it, then pass it back via this option.

--xla_gpu_llvm_ir_file

default: ""
type: string

If non-empty, specifies a file containing textual LLVM IR to use. The filename prefix must have the same pattern as LLVM dumped by XLA (i.e. module_0001.ir-no-opt.ll -> module_0001.MY_NEW_FILE.ll). This allows to match one specific module. General workflow. Get the not optimized LLVM IR from XLA, modify it, then pass it back via this option.

--xla_test_all_output_layouts

default: false
type: bool

Let ClientLibraryTestBase::ComputeAndCompare* test all permutations of output layouts. For example, with a 3D shape, all permutations of the set {0, 1, 2} are tried.

--xla_test_all_input_layouts

default: false
type: bool

Let ClientLibraryTestBase::ComputeAndCompare* test all permutations of input layouts. For example, for 2 input arguments with 2D shape and 4D shape, the computation will run 2! * 4! times for every possible layouts

--xla_test_add_command_buffer_mode

default: false
type: bool

If true, the test launched with ClientLibraryTestBase will use command buffer to execute the computation.

--xla_hlo_profile

default: false
type: bool

Instrument the computation to collect per-HLO cycle counts

--xla_backend_extra_options

default: ""
type: string

Extra options to pass to a backend; comma-separated list of 'key=val' strings (=val may be omitted); no whitespace around commas.

--xla_cpu_use_onednn

default: false
type: bool

Call oneDNN thunks for matmul and convolution fusions in the CPU backend.

--xla_cpu_experimental_onednn_custom_call

default: false
type: bool

Call oneDNN custom call thunks in the CPU backend.

--xla_cpu_experimental_onednn_fusion_type

default: ""
type: string

Comma-separated list of oneDNN fusion types to be enabled; no whitespace around commas. Two ways to pass values:

Exact type names. This overwrites the default setting.
'+' or '-' prefix: This adds or removes a fusion type from the default list. Cannot be mixed with the overwrite mode. Every item must have the sign prefix.

Available fusion types: dot, eltwise, and reduce.

The default list is currently empty.

--xla_cpu_use_acl

default: false
type: bool

Generate calls to ACL (Arm Compute Library) in the CPU backend.

--xla_cpu_opt_preset

default: "CPU_OPT_PRESET_DEFAULT"
type: string

Set CPU optimization preset (FAST_RUNTIME, FAST_COMPILE)

--xla_cpu_use_fusion_emitters

default: true
type: bool

Use fusion emitters for code generation in the CPU backend.

--xla_cpu_use_thunk_runtime

default: true
type: bool

Deprecated.

--xla_cpu_use_xnnpack

default: true
type: bool

Use XNNPACK for supported operations.

--xla_cpu_experimental_xnn_fusion_type

default: ""
type: string

Comma-separated list of XNN fusion types to be enabled; no whitespace around commas. Two ways to pass values:

Exact type names. This overwrites the default setting.
'+' or '-' prefix: This adds or removes a fusion type from the default list. Cannot be mixed with the overwrite mode. Every item must have the sign prefix.

Available fusion types: dot, eltwise, and reduce.

The default list is currently empty.

--xla_cpu_experimental_ynn_fusion_type

default: ""
type: string

Comma-separated list of YNN fusion types to be enabled; no whitespace around commas. Two ways to pass values:

Exact type names. This overwrites the default setting.
'+' or '-' prefix: This adds or removes a fusion type from the default list. Cannot be mixed with the overwrite mode. Every item must have the sign prefix.

The default list is currently empty.

--xla_cpu_experimental_xnn_graph_fusion_mode

default: "XNN_GRAPH_FUSION_MODE_DISABLED"
type: string

Controls XnnGraphFusion pass. XNN_GRAPH_FUSION_MODE_DISABLED - default value,

XNN_GRAPH_FUSION_MODE_GREEDY - greedy extraction of XNNPACK-compatible subgraphs starting from root instructions,

XNN_GRAPH_FUSION_MODE_GREEDY_SLINKY - same as GREEDY plus operations that are only supported with slinky, XNN_GRAPH_FUSION_MODE_BYPASS_COST_MODEL - test-only value for disabling XNNPACK cost models.

--xla_cpu_parallel_codegen_split_count

default: 32
type: int32

Split LLVM module into at most this many parts before codegen to enable parallel compilation for the CPU backend.

--xla_cpu_copy_insertion_use_region_analysis

default: false
type: bool

Use region based analysis in copy insertion pass.

--xla_cpu_enable_concurrency_optimized_scheduler

default: false
type: bool

[Deprecated, do not use].

--xla_cpu_scheduler_type

default: "CPU_SCHEDULER_TYPE_DEFAULT"
type: string

XLA:CPU's scheduler type.

--xla_cpu_prefer_vector_width

default: 256
type: int32

Preferred vector width for the XLA:CPU LLVM backend.

--xla_cpu_max_isa

default: ""
type: string

Maximum ISA that XLA:CPU LLVM backend will codegen, i.e., it will not use newer instructions. Available values: SSE4_2, AVX, AVX2, AVX512, AVX512_VNNI, AVX512_BF16, AMX, and AMX_FP16. (AMX will enable both AMX_BF16 and AMX_INT8 instructions.)

--xla_cpu_emitter_verification_level

default: 0
type: int32

Sets how often we verify the emitted modules. Higher levels mean more frequent verification. Currently supported: 0, 1.

--xla_gpu_crash_on_verification_failures

default: false
type: bool

Crashes the program on extra verification failures, e.g. cuDNN cross checking failures

--xla_gpu_strict_conv_algorithm_picker

default: false
type: bool

[Deprecated, do not use].

--xla_gpu_autotune_level

default: 4
type: int32
[Stable]

Set GEMM and Convolution auto-tuning level. 0 = off; 1 = on; 2 = on+init; 3 = on+init+reinit; 4 = on+init+reinit+check; 5 = on+init+reinit+check and skip WRONG_RESULT solutions. See also the related flag xla_gpu_autotune_gemm_rtol. Remark that, setting the level to 5 only makes sense if you are sure that the reference (first in the list) solution is numerically CORRECT. Otherwise, the autotuner might discard many other correct solutions based on the failed BufferComparator test.

--xla_gpu_autotune_max_solutions

default: 0
type: int64

Maximal number of GEMM solutions to consider for autotuning: 0 means consider all solutions returned by the GEMM library.

--xla_gpu_fusion_autotune_top_k_configs

default: 1
type: int32

Maximum number of candidate configs considered during autotuning non-gemm fusions.

--xla_gpu_autotune_gemm_rtol

default: 0.100000
type: float

Relative precision for comparing GEMM solutions vs the reference one

--xla_force_host_platform_device_count

default: 1
type: int32

Force the host platform to pretend that there are these many host "devices". All of these host devices are backed by the same threadpool. Setting this to anything other than 1 can increase overhead from context switching but we let the user override this behavior to help run tests on the host that run models in parallel across multiple devices.

--xla_gpu_disable_gpuasm_optimizations

default: false
type: bool

In XLA:GPU run ptxas in -O0 (default is -O3).

--xla_gpu_generate_debug_info

default: false
type: bool

Generate debug info for codegened CUDA kernels.

--xla_gpu_generate_line_info

default: false
type: bool

Generate line info for codegened CUDA kernels.

--xla_fuel

default: ""
type: string

Sets compiler fuel, useful for bisecting bugs in passes. Format --xla_fuel=PASS1=NUM1,PASS2=NUM2,...

--xla_dump_to

default: ""
type: string

Directory into which debugging data is written. If not specified but another dumping flag is passed, data will be written to stdout. To explicitly write to stdout, set this to "-". The values "sponge" and "test_undeclared_outputs_dir" have a special meaning: They cause us to dump into the directory specified by the environment variable TEST_UNDECLARED_OUTPUTS_DIR. One or more --xla_dump_hlo_as_* flags can be set to specify the formats of the dumps. For example, if both --xla_dump_hlo_as_text and --xla_dump_hlo_as_proto are set, then the HLO modules will be dumped as text and as protos.

--xla_flags_reset

default: false
type: bool

Whether to reset XLA_FLAGS next time to parse.

--xla_dump_hlo_as_text

default: false
type: bool

Dumps HLO modules as text before and after optimizations. debug_options are written to the --xla_dump_to dir, or, if no dir is specified, to stdout.

--xla_dump_hlo_as_long_text

default: true
type: bool

Dumps HLO modules as long text before and after optimizations. debug_options are written to the --xla_dump_to dir, or, if no dir is specified, to stdout. Ignored unless xla_dump_hlo_as_text is true.

--xla_hlo_print_inline_stack_frames

default: false
type: bool

If true, when printing HLO text, resolve each instruction's stack_frame_id into inline source_file/source_line/source_column fields on its metadata, and append the StackFrameIndex resolution tables (FileNames, FunctionNames, FileLocations, StackFrames) at the end of the module text.

--xla_dump_large_constants

default: false
type: bool

Dumps HLO modules including large constants before and after optimizations. debug_options are written to the --xla_dump_to dir, or, if no dir is specified, to stdout. Ignored unless xla_dump_hlo_as_text is true.

--xla_dump_hlo_as_proto

default: false
type: bool

Dumps HLO modules as HloProtos to the directory specified by --xla_dump_to.

--xla_gpu_experimental_dump_fdo_profiles

default: false
type: bool

Dumps FDO profiles as text to the directory specified by --xla_dump_to.

--xla_gpu_experimental_dump_gpu_executable

default: false
type: bool

Dump the serialized GPU executables to 'gpu_executable_proto' suffixed files, in the directory specified by xla_dump_to. No-op if xla_dump_to isn't set, or during autotuning compilations.

--xla_dump_hlo_as_dot

default: false
type: bool

Dumps HLO modules rendered as dot files to the directory specified by --xla_dump_to.

--xla_dump_hlo_as_html

default: false
type: bool

Dumps HLO modules rendered as HTML files to the directory specified by --xla_dump_to.

--xla_dump_hlo_as_url

default: false
type: bool

Tries to dump HLO modules rendered as URLs to stdout (and also to the directory specified by --xla_dump_to). This is not implemented by default; you need to add a plugin which calls RegisterGraphToURLRenderer().

--xla_dump_fusion_visualization

default: false
type: bool

Tries to generate HLO fusion visualization as an HTML page to the directory specified by --xla_dump_to). This is not implemented by default; you need to add a plugin which calls RegisterGraphToURLRenderer(). Generates a file per computation. Currently only implemented for the GPU backend.

--xla_dump_hlo_snapshots

default: false
type: bool

Every time an HLO module is run, dumps an HloSnapshot to the directory specified by --xla_dump_to.

--xla_dump_hlo_module_re

default: ""
type: string

Limits dumping only to modules which match this regular expression. Default is to dump all modules.

--xla_dump_hlo_pass_re

default: ""
type: string

If specified, dumps HLO before and after optimization passes which match this regular expression, in addition to dumping at the very beginning and end of compilation.

--xla_dump_emitter_re

default: ""
type: string

If specified, dumps debug logs (e.g. IR like LLVM or MLIR) before and after emitters which match this regular expression, in addition to dumping at the very beginning and end of compilation.

--xla_dump_include_timestamp

default: false
type: bool

If specified, includes a timestamp in the dumped filenames.

--xla_dump_hlo_to_subfolder

default: false
type: bool

If true, dumps HLO modules in a subfolder titled the HLO module name in the directory specified by --xla_dump_to.

--xla_dump_max_hlo_modules

default: -1
type: int32

Max number of hlo module dumps in a directory. Set to < 0 for unbounded.

--xla_dump_module_metadata

default: false
type: bool

Dumps HloModuleMetadata as text protos to the directory specified by --xla_dump_to.

--xla_dump_compress_protos

default: false
type: bool

Gzip-compress protos dumped by --xla_dump_hlo_as_proto.

--xla_hlo_graph_addresses

default: false
type: bool

When rendering graphs (--xla_dump_hlo_as_{dot,html,url}), displays the address in memory of each HloInstruction object.

--xla_hlo_graph_sharding_color

default: false
type: bool

Assign colors based on sharding assignments when generating the HLO graphs.

--xla_allow_excess_precision

default: true
type: bool

Allow xla to increase the output precision of an instruction.

--xla_gpu_force_conv_nchw

default: false
type: bool

For cuDNN convolutions, always use NCHW layouts.

--xla_gpu_force_conv_nhwc

default: false
type: bool

For cuDNN convolutions, always use NHWC layouts.

--xla_gpu_algorithm_denylist_path

default: ""
type: string

An AlgorithmDenylist text proto file as a denylist of convolutions to avoid to use.

--xla_gpu_use_runtime_fusion

default: false
type: bool

For using cuDNN runtime compiled fusion kernels.

--xla_tpu_detect_nan

default: false
type: bool

Trigger error on execution on TPU if a NAN value is detected

--xla_tpu_detect_inf

default: false
type: bool

Trigger error on execution on TPU if a INF value is detected

--xla_cpu_enable_xprof_traceme

default: false
type: bool

If true, XLA CPU generates code to call TraceMe::Activity{Start|End} around HLO operations.

--xla_gpu_unsafe_fallback_to_driver_on_ptxas_not_found

default: false
type: bool

If true, XLA GPU falls back to the driver if ptxas is not found. Note that falling back to the driver can have drawbacks like using more memory and/or other bugs during compilation, so we recommend setting this flag to false.

--xla_multiheap_size_constraint_per_heap

default: -1
type: int32

Generates multiple heaps (i.e., temp buffers) with a size constraint on each heap to avoid Out-of-Memory due to memory fragmentation. The constraint is soft, so it works with tensors larger than the given constraint size. -1 corresponds to no constraints.

--xla_gpu_force_compilation_parallelism

default: 0
type: int32

Overrides normal multi-threaded compilation setting to use this many threads. Setting to 0 (the default value) means no enforcement.

--xla_gpu_enable_llvm_module_compilation_parallelism

default: false
type: bool

Decides whether we can do LLVM module compilation in a parallelised way. If set to false, then it will be single threaded, otherwise the number of threads depends on the --xla_gpu_force_compilation_parallelism flag and the thread pool supplied to GpuCompiler.

--xla_gpu_default_to_alg_dot_bf16_bf16_f32

default: false
type: bool

Use the dot precision algorithm `ALG_DOT_BF16_BF16_F32 by default for f32 dots.

--xla_gpu_deterministic_ops

default: false
type: bool

Guarantees run-to-run determinism on GPU.

--xla_gpu_exclude_nondeterministic_ops

default: false
type: bool

Excludes non-deterministic ops from compiled executables.

--xla_gpu_disable_async_collectives

default: ""
type: string

This disables a certain set of async collectives and turn them into synchronous ones. By default, this is empty which indicates enabling async execution for all collectives. A sample usage is: --xla_gpu_disable_async_collectives=ALLREDUCE,REDUCESCATTER

--xla_gpu_enable_while_loop_unrolling

default: "WHILE_LOOP_UNROLLING_AUTO_UNROLL"
type: string

Enables while loop unrolling features. WHILE_LOOP_UNROLLING_DOUBLE_BUFFER unrolls the loop by factor of 2, WHILE_LOOP_UNROLLING_FULL_UNROLL will unroll the entire loop WHILE_LOOP_UNROLLING_AUTO_UNROLL unrolls by a factor of 2, if there is any collective present within a while loop.WHILE_LOOP_UNROLLING_MANUAL_UNROLL will unroll loops annotated with the _xla_loop_unroll_strategy attribute.

--xla_gpu_all_reduce_combine_threshold_bytes

default: 31457287
type: int64
[Stable]

Size threshold (in bytes) for the GPU all-reduce combiner.

--xla_gpu_all_gather_combine_threshold_bytes

default: 31457287
type: int64

Size threshold (in bytes) for the GPU all-gather combiner.

--xla_gpu_reduce_scatter_combine_threshold_bytes

default: 31457287
type: int64
[Stable]

Size threshold (in bytes) for the GPU reduce-scatter combiner.

--xla_gpu_collective_permute_combine_threshold_bytes

default: 31457287
type: int64

Size threshold (in bytes) for the GPU collective-permute combiner.

--xla_gpu_collective_combine_threshold_count

default: 2048
type: int64

Maximum number of instructions to be combined in collective combiners.

--xla_gpu_enable_all_gather_combine_by_dim

default: false
type: bool

Combine all-gather ops with the same gather dimension or irrespective of their dimension.

--xla_gpu_enable_reduce_scatter_combine_by_dim

default: false
type: bool

Combine reduce-scatter ops with the same dimension or irrespective of their dimension.

--xla_gpu_enable_approx_costly_collectives

default: false
type: bool

Enables more accurate latency approximation of collectives. Used in ApproximateLatencyEstimator scheduler.

--xla_gpu_all_reduce_blueconnect_num_devices_per_host

default: 0
type: int32

Number of devices per host for first stage of BlueConnect decomposition pass. The pass will attempt to decompose all-reduces ops into a ReduceScatter-AllReduce-AllGather sequence, with the initial ReduceScatter being performed over all of the devices in the same host. Set to < 1 to disable all-reduce decomposition.

--xla_gpu_enable_while_loop_reduce_scatter_code_motion

default: false
type: bool

Enable hoisting of reduce-scatter outside while loops.

--xla_gpu_collective_inflation_factor

default: 1
type: int32

Inflation factor for collectives. If set to > 1, each XLA/GPU collective will execute multiple times (will yield incorrect results)

--xla_llvm_force_inline_before_split

default: true
type: bool

Decide whether to force inline before llvm module split to get a more balanced splits for parallel compilation

--xla_gpu_enable_reassociation_for_converted_ar

default: true
type: bool

Enable allreduce reassociation on allreduces that are converted to a wider type. The reassociated allreduce will be promoted to a wider-typed allreduce.

--xla_gpu_dump_llvmir

default: false
type: bool

Dump LLVM IR.

--xla_dump_hlo_unoptimized_snapshots

default: false
type: bool

Every time an HLO module is run, dumps an HloUnoptimizedSnapshot to the directory specified by --xla_dump_to.

--xla_gpu_enable_cudnn_fmha

default: false
type: bool

[Deprecated, do not use]

--xla_gpu_fused_attention_use_cudnn_rng

default: false
type: bool

Use cudnn random number generator for fused attention kernel.

--xla_gpu_enable_cudnn_layer_norm

default: false
type: bool

Rewrite layer norm patterns into cuDNN library call.

--xla_gpu_enable_cublaslt

default: true
type: bool

[Deprecated] Use cuBLASLt for GEMMs when possible.

--xla_gpu_enable_command_buffer

default: "CONDITIONAL, CUBLAS, CUBLASLT, CUDNN, CUSTOM_CALL, DYNAMIC_SLICE_FUSION, FUSION"
type: string

The types of the commands that are recorded into command buffers. It can either be a list of command types or a list of command types with + and - as prefix, which indicate adding or removing a command type to/from the default list.

--xla_gpu_graph_min_graph_size

default: 5
type: int32

Capture a region as a function to be launched as cuda graph if the number of moved instructions reaches this threshold.

--xla_gpu_graph_enable_concurrent_region

default: false
type: bool

[Deprecated, do not use]

--xla_gpu_command_buffer_scheduling_mode

default: "LHS"
type: string

The command buffer scheduling mode for XLA:GPU.

--xla_cmd_buffer_trace_cache_size

default: 16
type: int64

Set the command buffer trace cache size, increasing the cache size may sometimes reduces the chances of doing command buffer tracing for updating command buffer instance.

--xla_dump_disable_metadata

default: false
type: bool

Disable dumping HLO metadata in HLO dumps.

--xla_dump_hlo_pipeline_re

default: ""
type: string

If specified, dumps HLO before and after optimization passes in the pass pipelines that match this regular expression.

--xla_dump_enable_mlir_pretty_form

default: true
type: bool

Enable dumping MLIR using pretty print form. If set to false, the dumped MLIR will be in the llvm-parsable format and can be processed by mlir-opt tools. Pretty print form is not legal MLIR.

--xla_dump_full_hlo_config

default: true
type: bool

Enable dumping the full HloModuleConfig proto.

--xla_gpu_enable_dynamic_slice_fusion

default: false
type: bool
[Stable]

Whether to enable dynamic slice fusion to optimize dynamic-slice and dynamic-update-slice operations.

--xla_gpu_enable_dus_accumulator_zero_init_elimination

default: false
type: bool

Replaces broadcast(0) scan-accumulator init with AllocateBuffer when the body's DUS covers every slot.

--xla_gpu_experimental_dynamic_slice_fusion_verify_offsets

default: false
type: bool

When true, DynamicSliceFusionV2Thunk verifies at runtime that annotated DynamicSliceConfig offsets match actual XLA-computed DS/DUS offsets. Adds D2H sync overhead — for debugging only.

--xla_gpu_nccl_termination_timeout_seconds

default: -1
type: int64

Timeout in seconds before terminating jobs stuck in NCCL Rendezvous.

--xla_gpu_enable_nccl_user_buffers

default: false
type: bool

Enables NCCL User Buffer Registration. collective_memory_size in the allocator config must also be set to a non-zero value that is large enough to meet peak collective memory usage.

--xla_gpu_experimental_enable_nccl_symmetric_buffers

default: false
type: bool

Enables NCCL symmetric buffer registration.

--xla_gpu_experimental_aot_compiled_thunks

default: true
type: bool

Enables an Ahead-of-Time (AOT) compilation flow where the compiled binary includes the generated Thunks. In contrast, the legacy flow only compiles up to the HLO optimization stage, before Thunk generation.

--xla_gpu_experimental_enable_nvshmem

default: false
type: bool

Enables NVSHMEM.

--xla_gpu_temp_buffer_use_separate_color

default: false
type: bool

Enables temp User Buffer Registration. Enable this flag will use a separate cuda async memory allocator to allocate temp buffer, this will allocate temp buffer to the fixed address on every iteration

--xla_gpu_require_exclusive_lock

default: false
type: bool

if true, running gpu executable will require exclusive lock on gpu, so there is no multi thread conlicts on gpu. this can enable some optimizations that reduce the cost of resource management, e.g, command buffer update to ensure correctness when running in multi thread mode.

--xla_gpu_enable_nccl_comm_splitting

default: true
type: bool

Enables NCCL communicator splitting which allows sharing NCCL resources between different NCCL cliques.

--xla_gpu_nccl_init_max_rank_per_root_ratio

default: 0
type: int64

Maximum number of ranks associated with a root rank to initialize a NCCL communicator via ncclCommInitRankScalable. A value of zero will lead to a single root.

--xla_gpu_redzone_padding_bytes

default: 8388608
type: int64

Amount of padding the redzone allocator will put on one side of each buffer it allocates. (So the buffer's total size will be increased by 2x this value.)

--xla_gpu_shape_checks

default: "RUNTIME"
type: string

When to perform shape checks in XLA:GPU.

--xla_cpu_strict_dot_conv_math

default: false
type: bool

By default, XLA:CPU will run fp16 dot/conv as fp32, as this is generally (much) faster on our hardware. Set this flag to true to disable this behavior.

--xla_dump_latency_hiding_schedule

default: false
type: bool

Dump the schedule from the latency-hiding scheduler.

--xla_gpu_enable_latency_hiding_scheduler

default: false
type: bool
[Stable]

Enable latency-hiding scheduler for XLA:GPU

--xla_gpu_enable_analytical_latency_estimator

default: false
type: bool

Enable analytical latency estimator for latency-hiding scheduler for XLA:GPU

--xla_gpu_enable_analytical_sol_latency_estimator

default: true
type: bool

Enable analytical Speed-of-Light latency estimator for latency-hiding scheduler for XLA:GPU, must be used without xla_gpu_enable_analytical_latency_estimator. It can also benefit from user-passed options in xla_gpu_analytical_latency_estimator_options

--xla_gpu_analytical_latency_estimator_options

default: ""
type: string

Extra platform-specific options to improve analytical latency estimator precision; comma-separated list of 'key=val' strings (=val may be omitted); no whitespace around commas.Available options: --xla_gpu_analytical_latency_estimator_options='nccl_op_launch_us=55,nic_speed_gbps=40,chunk_prep_us=1,rtt_us=2,gpus_per_node=4,chunk_size_bytes=1024'

--xla_gpu_pgle_profile_file_or_directory_path

default: ""
type: string

Directory or file for PGLE profiles in XLA:GPU

--xla_gpu_memory_limit_slop_factor

default: 95
type: int32

Slop factor for memory limits in XLA:GPU. This flag serves as a multiplier applied to the total available memory, creating a threshold that guides the Latency Hiding Scheduler (LHS) in balancing memory reduction and latency hiding optimizations. This factor effectively establishes a memory limit for compiler passes, determining when the scheduler should prioritize: 1. Memory reduction: When memory usage approaches or exceeds the calculated threshold. 2. Latency hiding: When memory usage is below the threshold, allowing for more aggressive optimizations that may temporarily increase memory usage but improve overall performance. By adjusting this factor, users can fine-tune the trade-off between memory efficiency and performance optimizations. The default value is 95.

--xla_gpu_enable_highest_priority_async_stream

default: true
type: bool

Enable async stream to have the highest priority.

--xla_gpu_enable_pipelined_all_reduce

default: false
type: bool
[Stable]

Enable pipelinling of all-reduce instructions.

--xla_gpu_enable_pipelined_all_gather

default: false
type: bool
[Stable]

Enable pipelinling of all-gather instructions.

--xla_gpu_enable_pipelined_reduce_scatter

default: true
type: bool
[Stable]

Enable pipelinling of reduce-scatter instructions.

--xla_gpu_enable_pipelined_host_offloading

default: false
type: bool

Enable pipelining of host offloading instructions.

--xla_gpu_enable_pipelined_p2p

default: false
type: bool

Enable pipelinling of P2P instructions.

--xla_gpu_collective_permute_decomposer_threshold

default: 9223372036854775807
type: int64
[Stable]

Collective permute decomposer threshold.

--xla_gpu_collectives_implementation

default: ""
type: string

Name of the GPU collectives implementation to use (e.g. "nccl", "loopback"). When empty (the default), the highest-priority registered implementation is used.

--xla_gpu_experimental_pipeline_parallelism_opt_level

default: "PIPELINE_PARALLELISM_OPT_LEVEL_DISABLE"
type: string

Experimental optimizations for SPMD-based pipeline parallelism on GPU.

--xla_enable_enzyme_comms_opt

default: false
type: bool

Enable communication optimization patterns specified in Enzyme. More details in http://shortn/_jXJ2VFoyMN.

--xla_recognize_reduction_optimization_level

default: 0
type: int32

Optimization level for reduction recognition pass.

--xla_partitioning_algorithm

default: "PARTITIONING_ALGORITHM_NOOP"
type: string

The partitioning algorithm to be used in the PartitionAssignment pass

--xla_gpu_enable_triton_gemm

default: true
type: bool
[Stable]

Whether to use Triton-based matrix multiplication.

--xla_gpu_unsupported_enable_triton_multi_output_fusion

default: false
type: bool

Enable Triton multi-output fusions.

--xla_gpu_verify_triton_fusion_numerics

default: false
type: bool

Whether to verify that the numeric results of Triton fusions match the results of regular emitters.

--xla_gpu_enable_cudnn_int8x32_convolution_reordering

default: true
type: bool

Enable cuDNN frontend for int8x32 convolutions with reordered filter.

--xla_gpu_experimental_gemm_fusion_v2

default: false
type: bool

Enable experimental rewrite of GEMM fusion pass for Triton.

--xla_gpu_triton_gemm_any

default: true
type: bool

Use Triton-based matrix multiplication for any GEMM it supports without filtering only faster ones. To make sure only triton gemm is chosen by the autotuner run with xla_gpu_cublas_fallback set to false.

--xla_gpu_exhaustive_tiling_search

default: false
type: bool
[Stable]

Search for Triton GEMM tilings exhaustively during autotuning. This increases the compile time.

--xla_gpu_experimental_enable_subchannel_dequantisation_fusion

default: false
type: bool

Enable fusion for the subchannel dequantisation sequences like [x,z]param -> [x,y,z]broadcast -> [x*y,z]bitcast -> multiply -> dot. Performance can be worse, because some block sizes / split-k > 1 is not considered for subchannel dequant fusions.

--xla_gpu_experimental_enable_triton_heroless_priority_fusion

default: false
type: bool

Enable heroless Triton fusions in the PriorityFusion pass. The pass will try to make Triton fusions first and foremost where it is possible.

--xla_gpu_dump_autotune_results_to

default: ""
type: string

File to write autotune results to. It will be a binary file unless the name ends with .txt or .textproto. Warning: The results are written at every compilation, possibly multiple times per process. This only works on CUDA. In tests, the TEST_UNDECLARED_OUTPUTS_DIR prefix can be used to write to their output directory.

--xla_gpu_load_autotune_results_from

default: ""
type: string

File to load autotune results from. It will be considered a binary file unless the name ends with .txt or .textproto. It will be loaded at most once per process. This only works on CUDA. In tests, the TEST_WORKSPACE prefix can be used to load files from their data dependencies.

--xla_gpu_require_complete_aot_autotune_results

default: false
type: bool

Whether to require complete AOT autotuning results.

--xla_gpu_auto_spmd_partitioning_memory_budget_gb

default: 0
type: int32

Memory budget in GB per device for AutoSharding.

--xla_gpu_auto_spmd_partitioning_memory_budget_ratio

default: 1.100000
type: float

Enabled when xla_gpu_auto_spmd_partitioning_memory_budget_gb is 0. The memory budget is set to xla_gpu_auto_spmd_partitioning_memory_budget_ratio times the estimated memory usage lower bound.

--xla_gpu_dump_autotuned_gemm_fusions

default: false
type: bool

Dumps autotuned GEMM fusions to the directory specified by xla_dump_to or stdout. Each fusion is dumped only once, as an optimized HLO.

--xla_gpu_dump_autotuned_instructions

default: false
type: bool

Dumps autotuned instructions to the directory specified by xla_dump_to or stdout. Each instruction is dumped only once, as an optimized HLO.

--xla_gpu_override_gemm_autotuner

default: ""
type: string

Overrides the GEMM autotuner to use the specified (AutotuneResult::TritonGemmKey) textproto configuration for all Triton GEMM fusions. (You can get such textprotos from the debug logs of the GEMM autotuner.)

--xla_gpu_command_buffer_unroll_loops

default: false
type: bool

During command buffer lowering, unroll the loop command if loop has known loop count.

--xla_gpu_copy_insertion_use_region_analysis

default: false
type: bool

If true, use the new fine-grain region-based live range interference analysis in the copy insertion optimization pass.

--xla_gpu_collect_cost_model_stats

default: false
type: bool

If true, each fusion instruction will have a cost model runtime estimate in backend config after compilation.

--xla_gpu_enable_nccl_clique_optimization

default: false
type: bool

[Deprecated, do not use].

--xla_gpu_cublas_fallback

default: true
type: bool
[Stable]

Whether to allow GEMM fusion autotuning to fall back to cuBLAS when it is faster than Triton.

--xla_gpu_cudnn_gemm_fusion_level

default: 0
type: int32

cuDNN GEMM fusion level; higher level corresponds to more kinds of fused operations.

--xla_gpu_mock_custom_calls

default: false
type: bool

Replace custom calls with noop operations.

--xla_gpu_enable_while_loop_double_buffering

default: false
type: bool
[Stable]

Enable double buffering for while loop

--xla_gpu_filter_kernels_spilling_registers_on_autotuning

default: true
type: bool

Filter out kernels that spill registers during autotuning. Default is true.

--xla_gpu_fail_ptx_compilation_on_register_spilling

default: false
type: bool

Fails the PTX compilation if a kernel spills registers.

--xla_debug_buffer_assignment_show_max

default: 15
type: int64

Number of buffers to display when debugging the buffer assignment

--xla_gpu_llvm_verification_level

default: 0
type: int32

Sets how often we verify the generated llvm modules. Higher levels mean more frequent verification. Currently supported: 0, 1.

--xla_gpu_target_config_filename

default: ""
type: string

Filename for GPU TargetConfig. Triggers devicless compilation: attached device is ignored, and the proto is queried instead

--xla_gpu_enable_cub_radix_sort

default: true
type: bool

Enable radix sort using CUB for simple shapes

--xla_gpu_deviceless_cub_mode

default: "DEVICELESS_CUB_WITH_FALLBACK"
type: string

Control CUB behavior during deviceless compilation. Available options: DEVICELESS_CUB_DISABLED, DEVICELESS_CUB_WITH_FALLBACK, DEVICELESS_CUB_NO_FALLBACK, DEVICELESS_CUB_FORCE_ON_NO_FALLBACK.

--xla_gpu_threshold_for_windowed_einsum_mib

default: 100000
type: int64

Threshold to enable windowed einsum (collective matmul) in MB.Einsums that have partitioned operand(can be either LHS or RHS) that's larger than this threshold will be transformed to use windowed einsums.Default is 100000

--xla_gpu_operand_bytes_threshold_for_windowed_einsum

default: -1
type: int64

This controls whether to enable windowed einsum (collective matmul) based on the sum of sizes of 2 operands if set >= 0.If set >= 0, xla_gpu_threshold_for_windowed_einsum_mib is ignored.Default is -1

--xla_gpu_experimental_enable_fusion_block_level_rewriter

default: false
type: bool

Enabling this flag will attempt to redirect every fusion possible to the Triton emitter

--xla_gpu_enable_libnvptxcompiler

default: false
type: bool

Use libnvptxcompiler for PTX-to-GPU-assembly compilation instead of calling ptxas.

--xla_gpu_enable_libnvjitlink

default: false
type: bool

Use libnvjitlink for PTX-to-GPU-assembly compilation instead of calling ptxas.

--xla_gpu_nccl_async_execution

default: false
type: bool

Whether to use asynchronous execution for NCCL communicators

--xla_gpu_nccl_blocking_communicators

default: true
type: bool

Whether to use non-blocking NCCL communicators

--xla_gpu_nccl_collective_max_nchannels

default: 0
type: int64

Specify the maximum number of channels(SMs) NCCL will use for collective operations. Default is 0 which is to let NCCL decide.

--xla_gpu_nccl_p2p_max_nchannels

default: 0
type: int64

Specify the maximum number of channels(SMs) NCCL will use for p2p operations. Default is 0 which is to let NCCL decide.

--xla_gpu_multi_streamed_windowed_einsum

default: true
type: bool

Whether to run windowed einsum using multiple compute streams.

--xla_gpu_gemm_rewrite_size_threshold

default: 100
type: int64

Threshold until which elemental dot emitter is preferred for GEMMs (minimum combined number of elements of both matrices in non-batch dimensions to be considered for a rewrite).

--xla_gpu_use_embeded_device_lib

default: true
type: bool

Whether to use embeded bitcode library in codegen.

--xla_gpu_use_memcpy_local_p2p

default: false
type: bool

Whether to use memcpy for local p2p communication.

--xla_gpu_collective_permute_connected_components

default: false
type: bool

Split collective-permute into connected-component replica groups instead of one giant clique.

--xla_gpu_collective_permute_mode

default: "private"
type: string

Memory mode for collective-permute: private, symmetric, peer. See CollectivesMode for details.

--xla_gpu_all_gather_mode

default: "private"
type: string

Memory mode for all-gather: private, symmetric, peer. See CollectivesMode for details.

--xla_gpu_use_inprocess_lld

default: true
type: bool

Whether to use lld as a library for the linking.

--xla_gpu_dump_autotune_logs_to

default: ""
type: string

File to write autotune logs to. It will be a binary file unless the name ends with .txt or .textproto.

--xla_reduce_window_rewrite_base_length

default: 16
type: int64

Base length to rewrite the reduce window to, no rewrite if set to 0.

--xla_gpu_enable_host_memory_offloading

default: false
type: bool

Whether to trigger host memory offloading on a device.

--xla_gpu_nccl_terminate_on_error

default: false
type: bool

If set, then NCCL errors will terminate the process.

--xla_gpu_shard_autotuning

default: true
type: bool

Shard autotuning between participating compiler processes (typically in multi-host setups) and join the results when it's done.

--xla_syntax_sugar_async_ops

default: false
type: bool

Enable syntax sugar for async ops in HLO dumps.

--xla_gpu_kernel_cache_file

default: ""
type: string

Path to a file to cache compiled kernels. Cached kernels get reused in further compilations; not yet cached kernels are compiled as usual and get appended to the cache file whenever possible.

--xla_gpu_per_fusion_autotune_cache_dir

default: ""
type: string

Experimental: Maintain a per-fusion autotune cache in the given directory. XLA will try to read existing results when they are needed and write new results when they are determined. The directory must exist. Cache invalidation has to be handled by the user (e.g. please use an empty directory if you want to start with an empty cache). XLA version checks must be done by the user (e.g. if you want to use separate caches for different versions of XLA, please use different directories). Default: no cache.

--xla_gpu_experimental_autotune_cache_mode

default: "AUTOTUNE_CACHE_MODE_UPDATE"
type: string

Experimental: Specify the behavior of per kernel autotuning cache. Supported modes: read (provides readonly access to the cache), update (loads if the cache exists, runs autotuning and dumps the result otherwise). Default: update.

--xla_gpu_experimental_autotuner_cache_dir

default: ""
type: string

Experimental: Specify the directory to read/write autotuner cache to.

--xla_gpu_experimental_autotune_backends

default: ""
type: string

Backends to enable for autotuning. Comma-separated (no spaces). Examples:

'cudnn,triton' (overwrites defaults)

'+cudnn,-cublas' (adds/removes from defaults)

Available: cudnn, triton, cublas, cublaslt etc, check xla.autotuner.Backend for the full list.

--xla_gpu_experimental_all_fusions_with_triton

default: false
type: bool

Experimental: If true, autotune all fusions with block level emitter.

--xla_gpu_gemm_autotuner_override_file

default: ""
type: string

A textproto file to override autotune results. See also xla_gpu_override_gemm_autotuner to override with a single config.

--xla_enable_command_buffers_during_profiling

default: false
type: bool

Experimental: Enable command buffers while a profiling active. By default, enabling profiling switches from command buffers to op-by-op mode.

--xla_gpu_cudnn_gemm_max_plans

default: 5
type: int32

Limit for the number of kernel configurations (plans) to use during autotuning of cuDNN GEMM fusions.

--xla_gpu_enable_triton_gemm_int4

default: true
type: bool

[Deprecated, do not use]

--xla_gpu_async_dot

default: false
type: bool

Wrap dot operations into async computations in an effort to parallelize matrix operations.

--xla_step_marker_location

default: "STEP_MARK_AT_ENTRY"
type: string

Option to emit a target-specific marker to indicate the start of a training. The location of the marker (if any) is determined by the option value of type DebugOptions::StepMarkerLocation.

--xla_gpu_pgle_accuracy_checker

default: "PGLE_STRICTNESS_LEVEL_WARN"
type: string

If an FDO profile is specified and latency hiding scheduler encounters missing instructions in the profile, then the compilation will halt (ERROR), or a warning will be emitted (WARN), or the checker is disabled (OFF)

--xla_gpu_executable_embed_debug_info

default: true
type: bool

Add debug information to the executable such as HLO module, asm_text etc.

--xla_gpu_executable_num_compute_streams

default: 0
type: int32

Number of additional compute streams to allocate for a GPU executable.

--xla_gpu_executable_num_communication_streams

default: 0
type: int32

Number of additional communication streams to allocate for a GPU executable.

--xla_gpu_executable_terminate_timeout

default: 30
type: int32

Set timeout for Rendezvous termination

--xla_gpu_executable_warn_stuck_timeout

default: 10
type: int32

Set timeout for Rendezvous stuck warning

--xla_gpu_execution_terminate_timeout

default: "inf"
type: string

Set timeout for XLA:GPU execution to prevent undetected deadlocks

--xla_gpu_execution_progress_tracking

default: 0
type: int32

Number of thunks to report in progress tracking on execution timeout (0 to disable)

--xla_gpu_first_collective_call_warn_stuck_timeout_seconds

default: 20
type: int32

Set timeout for First Collective Call Rendezvous stuck warning

--xla_gpu_first_collective_call_terminate_timeout_seconds

default: 40
type: int32

Set timeout for First Collective Call Rendezvous termination

--xla_gpu_experimental_disable_binary_libraries

default: false
type: bool

Disable XLA GPU passes that depend on non-open source binary libraries

--xla_gpu_experimental_enable_conv_fusion

default: false
type: bool

enable experimental XLA GPU passes that rewrite conv as hlo fusion instead of custom call.

--xla_gpu_dot_merger_threshold_mb

default: 64
type: int32
[Stable]

Dot merger pass threshold to be set in MB.

--xla_enable_fast_math

default: false
type: bool

Enable optimizations that assume finite math, i.e., no NaN.

--xla_gpu_experimental_stream_annotation

default: true
type: bool

Enable the experimental explicit stream annotation support. If false, the annotations are ignored.

--xla_gpu_experimental_parallel_collective_overlap_limit

default: 1
type: int32

This controls how many in-flight collectives latency hiding scheduler can schedule.

--xla_gpu_experimental_collective_start_as_early_as_possible

default: false
type: bool

This controls whether collectives should start as early as possible.

--xla_gpu_experimental_parallel_async_compute_limit

default: 2
type: int32

This controls how many in-flight asynchronous computations latency hiding scheduler can schedule.

--xla_pjrt_allow_auto_layout_in_hlo

default: false
type: bool

Experimental: Make unset entry computation layout mean auto layout instead of default layout in HLO when run through PjRT. In other cases (StableHLO or non-PjRT) the auto layout is already used.

--xla_gpu_enable_scatter_determinism_expander

default: false
type: bool

Makes scatter ops deterministic and enables the use of the scatter determinism expander. This is an optimized pass that rewrites scatter operations to ensure deterministic behavior with high performance. If the optimization pass does not support a particular scater op, it will be made deterministic using a slower implementation. Note that even when this flag is disabled, scatter operations may still be deterministic. This is the case when 'xla_gpu_exclude_nondeterministic_ops' or 'xla_gpu_deterministic_ops' is enabled.

--xla_gpu_unsupported_enable_all_reduce_decomposer

default: false
type: bool

Internal: Enable the AllReduceDecomposer, an unsupported pass that rewrites small all-reduce operations as a sequence of all-gather and reduce operations.

--xla_gpu_unsupported_enable_ragged_all_to_all_decomposer

default: false
type: bool

Internal: Enable the RaggedAllToAllDecomposer, an experimental pass that rewrites ragged-all-to-all as a dense all-to-all operation.

--xla_gpu_unsupported_enable_ragged_all_to_all_multi_host_decomposer

default: false
type: bool

Internal: Enable the RaggedAllToAllMultiHostDecomposer, an experimental pass to decompose ragged-all-to-all operation in intra-host and inter-host parts.

--xla_gpu_unsupported_override_fast_interconnect_slice_size

default: 0
type: int64

Internal: Override the number of devices in the fast interconnect domain. Default is 0, which means the number of devices is not overridden.

--xla_gpu_unsupported_use_all_reduce_one_shot_kernel

default: true
type: bool

Internal: Enable the one-shot kernel for single-host all-reduce operations.

--xla_gpu_unsupported_use_ragged_all_to_all_one_shot_kernel

default: true
type: bool

Internal: Enable the one-shot kernel for single-host ragged-all-to-all operations.

--xla_gpu_experimental_enable_alltoall_windowed_einsum

default: false
type: bool

Enable windowed einsum rewrite for all-to-all+gemm pattern, This optimization slices the all-to-all into smaller all-to-alls.It is an experimental feature.

--xla_gpu_experimental_pack_dot_operands_along_k_dimension

default: true
type: bool

For sub-byte dot operands, layout them along contracting dimensions.

--xla_unsupported_crash_on_hlo_pass_fix_max_iterations

default: false
type: bool

Crash if HloPassFix can not converge after a fixed number of iterations.

--xla_hlo_pass_fix_detect_cycles

default: false
type: bool

Perform hash-based cycle detection in fixed-point loops.

--xla_gpu_experimental_enable_heuristic_collective_combining

default: false
type: bool

Enable heuristic based collective combining.

--xla_gpu_experimental_collective_cse_distance_threshold

default: 0
type: int64

Set distance threshold for Collective CSE.

--xla_gpu_experimental_collective_perf_table_path

default: ""
type: string

If non empty will interpret this variable as a path for performance tables for collectives. Expects xla.gpu.DeviceHloInstructionProfiles proto.

--xla_unsupported_crash_on_hlo_pass_noop_change

default: false
type: bool

Crash if a pass reports that it did change the HLO but in fact it did not.

--xla_unsupported_crash_on_hlo_pass_silent_hlo_change

default: false
type: bool

Crash if a pass reports that it did not change the HLO but in fact it did.

--xla_disable_automatic_host_compute_offload

default: false
type: bool

Return an error if HostOffloader would have automatically offloaded some compute to the host.

--xla_allow_h2h_copy_when_automatic_host_compute_offload_disabled

default: false
type: bool

Allow host-to-host copy when automatic host compute offload is disabled, i.e. when xla_disable_automatic_host_compute_offload is set.

--xla_enable_scoped_logging_timers

default: true
type: bool

Do not run scoped logging timers (only supported in some places).

--xla_gpu_experimental_matmul_perf_table_path

default: ""
type: string

If non empty will interpret this variable as a path for performance tables for matmuls. Expects xla.gpu.DeviceHloInstructionProfiles proto.

--xla_gpu_experimental_force_split_k

default: 0
type: int32

Force a specific split_k value. Zero (default) means do not force split_k and use the heuristic.

--xla_gpu_experimental_enable_triton_warp_specialization

default: false
type: bool

Enable Triton's auto warp specialization feature where applicable.

--xla_gpu_experimental_enable_collective_multi_streaming

default: false
type: bool

Enable multi-stream runtime for collectives.

--xla_gpu_experimental_max_unroll_factor

default: 32
type: int32

Controls max unroll factor on Blackwell architectures. Should be at least 1.

--xla_detect_unstable_reductions

default: "DETECTION_MODE_NONE"
type: string

Controls the behavior of the unstable reduction detector pass that checks for unstable reductions in HLO computations. Acceptable values are: 'none', 'log', and 'crash'. 'none' is the default.

--xla_detect_unstable_reductions_post_optimizations

default: "DETECTION_MODE_NONE"
type: string

Controls the behavior of the unstable reduction detector pass that checks for unstable reductions in HLO computations after optimizations. Acceptable values are: 'none', 'log', and 'crash'. 'none' is the default.

--xla_gpu_experimental_use_raft_select_k

default: false
type: bool

If true, use the raft::matrix::select_k implementation of TopK.

--xla_gpu_experimental_ragged_all_to_all_use_barrier

default: true
type: bool

If true, use the MultiGpuBarrierKernel in one-shot RaggedAllToAll thunk.

--xla_gpu_experimental_ragged_all_to_all_use_barrier_with_nccl

default: false
type: bool

If true, use the MultiGpuBarrierWithNcclKernel in one-shot RaggedAllToAll thunk.

--xla_gpu_ragged_all_to_all_mode

default: "private"
type: string

Memory mode for ragged-all-to-all: private, symmetric, peer. In symmetric mode, the put/signal path is used. See CollectivesMode for details.

--xla_gpu_experimental_ragged_all_to_all_zero_copy

default: true
type: bool

If true, use the Symmetric Memory mode for MultiGpuBarrierWithNcclKernel

--xla_gpu_experimental_use_ragged_dot_grouped_gemm

default: true
type: bool

If true, use grouped GEMM (hipBLASLt) for ragged dot operations.

--xla_gpu_native_emitter_tune_unroll_factor_for_loops

default: false
type: bool

Experimental: If set, autotune the unroll factor for loop fusions in native emitter.

--xla_gpu_experimental_use_ragged_dot_fusion

default: false
type: bool

If true, use cuDNN fusion for ragged dot operations.

--xla_gpu_experimental_scaled_dot_with_triton

default: true
type: bool

If true, use the Triton emitter for scaled dot.

--xla_cpu_collective_call_warn_stuck_timeout_seconds

default: 20
type: int32

Set timeout for Collective Call Rendezvous stuck warning

--xla_cpu_collective_call_terminate_timeout_seconds

default: 40
type: int32

Set timeout for Collective Call Rendezvous termination

--xla_cpu_collective_timeout_seconds

default: 1800
type: int32

Set timeout for CPU collectives

--xla_keep_shardings_after_spmd

default: false
type: bool

If true, keep shardings after SPMD.

--xla_enable_hlo_sharding_v3

default: false
type: bool

If true, use HloShardingV3 which is a mesh and axis based sharding representation.

--xla_enable_rgv3_materialization

default: true
type: bool

If true, opportunistically materialize MeshAxesReplicaGroupList (RGV3) in SPMD partitioner. If false, fallback to legacy V1/V2 representations.

--xla_sdy_export_all_reduce_scatter

default: false
type: bool

Whether to export all sdy.reduce_scatter operations as native StableHLO collectives wrapped in manual computation blocks. If false (default), only operations with unreduced axes are exported this way.

--xla_gpu_experimental_enable_checksum_tracing_on_thunks

default: false
type: bool

Enables an experimental feature to record checksums of selected thunk inputs/outputs.

--xla_gpu_experimental_enable_buffer_saver_on_thunks

default: false
type: bool

When provided, enables an experimental feature to save results of selected thunks.

--xla_gpu_experimental_thunk_buffer_debug_module_outputs

default: false
type: bool

Whether to enable runtime instrumentation (float check, checksum, buffer saver) on all output buffers of an XLA module after all computation is finished. Setting this flag disables instrumentation of thunks not explicitly enabled by other filter flags.

--xla_gpu_experimental_thunk_buffer_debug_filter_by_thunk_id_ranges

default: "(none)"
type: string

Limits the thunk buffer debug instrumentation to thunks with IDs matching one or more ranges defined as a single integer, min:max (inclusive), or half-open min:/:max.

--xla_gpu_experimental_thunk_buffer_debug_filter_by_profile_annotation_re

default: "(none)"
type: string

Limits the thunk buffer debug instrumentation to thunks with profile annotations matching one or more regexes passed as comma-separated string.

--xla_gpu_experimental_enable_fusion_autotuner

default: true
type: bool

Enable autotuning between the native & triton fusion emitters.

--xla_gpu_rocm_max_trace_events

default: 4194304
type: int64

Maximum number of ROCm trace events (applies to callback/activity/annotation). Set as high as memory allows; up to 1e9.

--xla_gpu_experimental_enable_tiling_propagation

default: false
type: bool

If true, enable experimental tiling propagation.

--xla_gpu_detect_nan

default: "DETECTION_MODE_NONE"
type: string

Controls the behavior of the NaN detector pass that checks for presence of NaN values in kernel outputs. Acceptable values are: 'none', 'warning', and 'fail'. 'none' is the default. If other than 'none' value is provided, additional thunks will be added to detect and warn or fail the execution if NaNs are detected.

--xla_gpu_detect_inf

default: "DETECTION_MODE_NONE"
type: string

Controls the behavior of the Inf detector pass that checks for presence of Inf values in kernel outputs. Acceptable values are: 'none', 'warning', and 'fail'. 'none' is the default. If other than 'none' value is provided, additional thunks will be added to detect and warn or fail the execution if Infs are detected.

--xla_gpu_log_minmax

default: false
type: bool

If true, log min/max values from kernel outputs.

--xla_early_exit_with_layouts

default: false
type: bool

If true, exit early from the layout assignment pass after assigning layouts to entry computations.

--xla_gpu_print_compilation_stats

default: false
type: bool

Prints statistics about the HLO passes: how many times each pass was run, and how long it took.

--xla_gpu_enable_pdl

default: true
type: bool

Enable PDL (Programmatic Dependent Launch).

--xla_dump_buffer_assignment_analysis

default: true
type: bool

Dump BufferAssignment analysis.

--xla_gpu_command_buffer_update_mode

default: "ALWAYS_UPDATE"
type: string

Controls the VA remapping update strategy for command buffer thunks. See CommandBufferUpdateMode for details.

--xla_gpu_experimental_cost_model_gemm_tiling_options

default: ""
type: string

Experimental options for adjusting cost-model guided GEMM tiling selection; comma-separated list of 'key=val' strings (=val may be omitted); no whitespace around commas.

--xla_gpu_ptx_compiler_extra_flags

default: ""
type: string

Whitespace-separated extra flags to pass to the ptxas compiler, e.g. '--maxregcount=32'.

FilesExpand file tree

index.md

Latest commit

History

index.md

File metadata and controls

--xla_hlo_evaluator_use_fast_path

--xla_cpu_enable_fast_math

--xla_cpu_enable_platform_dependent_math

--xla_cpu_experimental_enable_tiling_propagation

--xla_cpu_fast_math_honor_nans

--xla_cpu_fast_math_honor_infs

--xla_cpu_fast_math_honor_division

--xla_cpu_fast_math_honor_functions

--xla_cpu_ftz

--xla_cpu_enable_fast_min_max

--xla_gpu_enable_fast_min_max

--xla_llvm_enable_alias_scope_metadata

--xla_llvm_enable_noalias_metadata

--xla_llvm_enable_invariant_load_metadata

--xla_llvm_disable_expensive_passes

--xla_backend_optimization_level

--xla_disable_hlo_passes

--xla_enable_hlo_passes_only

--xla_disable_all_hlo_passes

--xla_embed_ir_in_executable

--xla_eliminate_hlo_implicit_broadcast

--xla_cpu_multi_thread_eigen

--xla_gpu_cuda_data_dir

--xla_gpu_ftz

--xla_gpu_ptx_file

--xla_gpu_llvm_ir_file

--xla_test_all_output_layouts

--xla_test_all_input_layouts

--xla_test_add_command_buffer_mode

--xla_hlo_profile

--xla_backend_extra_options

--xla_cpu_use_onednn

--xla_cpu_experimental_onednn_custom_call

--xla_cpu_experimental_onednn_fusion_type

--xla_cpu_use_acl

--xla_cpu_opt_preset

--xla_cpu_use_fusion_emitters

--xla_cpu_use_thunk_runtime

--xla_cpu_use_xnnpack

--xla_cpu_experimental_xnn_fusion_type

--xla_cpu_experimental_ynn_fusion_type

--xla_cpu_experimental_xnn_graph_fusion_mode

--xla_cpu_parallel_codegen_split_count

--xla_cpu_copy_insertion_use_region_analysis

--xla_cpu_enable_concurrency_optimized_scheduler

--xla_cpu_scheduler_type

--xla_cpu_prefer_vector_width

--xla_cpu_max_isa

--xla_cpu_emitter_verification_level

--xla_gpu_crash_on_verification_failures

--xla_gpu_strict_conv_algorithm_picker

--xla_gpu_autotune_level

--xla_gpu_autotune_max_solutions

--xla_gpu_fusion_autotune_top_k_configs

--xla_gpu_autotune_gemm_rtol

--xla_force_host_platform_device_count

--xla_gpu_disable_gpuasm_optimizations

--xla_gpu_generate_debug_info

--xla_gpu_generate_line_info

--xla_fuel

--xla_dump_to

--xla_flags_reset

--xla_dump_hlo_as_text

--xla_dump_hlo_as_long_text

--xla_hlo_print_inline_stack_frames

--xla_dump_large_constants

--xla_dump_hlo_as_proto

--xla_gpu_experimental_dump_fdo_profiles

--xla_gpu_experimental_dump_gpu_executable

--xla_dump_hlo_as_dot

--xla_dump_hlo_as_html

--xla_dump_hlo_as_url

--xla_dump_fusion_visualization

--xla_dump_hlo_snapshots