mstar.utils

Modules

adarms_norm	Fused AdaRMS normalisation + scale/shift/gate Triton kernel.
cache_home	The user's cache root, which M*'s on-disk caches (compiled kernels, downloaded checkpoints) live under.
client_errors	Client-visible text for an exception raised while serving a request.
containers	Small generic containers with no engine or worker knowledge
cpus	The CPUs this process may use, and the plan every process of a deployment reads.
device
exitcode	Words for a child process's exit status, for log lines and error messages.
fast_tokenize	Tokenize a long prompt as parallel chunks, byte-identical to one encode.
fp8_block	FP8 block-scale checkpoints: the quantization_config they declare, and their weights dequantized.
gc_freeze	Move a process's set-up heap out of the garbage collector's reach.
generation_kwargs	The client-facing generation knobs, checked once where every endpoint submits.
h2d	Copy-only host-to-device staging, for work that runs beside a live CUDA graph.
hf_snapshot	Resolving a Hugging Face repo id to a local snapshot directory.
host_copy	Large host copies split over threads of their own.
ipc_format
knobs	Engine knobs read by more than one package.
logging_config
orphan	Watchdog that leaves once the process that spawned this one is gone.
profiler	Utilities for NVTX range annotations for profiling with nsys.
socket_prefix	Default IPC socket prefixes for the server entry points.
streams	CUDA stream state a worker has to manage.
threads	Starting threads, and moving the threads a step passes through onto their CPUs.
utf8_stream	Decode a model's text chunks without splitting characters.