aquin.compute.vram_guard Namespace Reference

Functions
tuple[float, float]|None	accelerator_vram_gib ()
tuple[float, float]|None	cuda_vram_gib ()
None	log_accelerator_vram (Callable[[str], None]|None log=None)
None	log_cuda_vram (Callable[[str], None]|None log=None)
None	release_inspection_models ()
None	cleanup_heavy_job_vram ()
None	restore_session_model (str|None model_id)
str	oom_message (*, str job, str model_id)
str	cuda_oom_message (*, str job, str model_id)
None	raise_if_cuda_oom (BaseException exc, *, str job, str model_id)
str	backend_summary ()

Function Documentation

accelerator_vram_gib()

tuple[float, float] | None accelerator_vram_gib

(

)

Free/total GiB for the active accelerator (CUDA/ROCm only; MPS has no public API).

Definition at line 21 of file vram_guard.py.

Referenced by cuda_vram_gib(), log_accelerator_vram(), and oom_message().

backend_summary()

str backend_summary

(

)

Definition at line 114 of file vram_guard.py.

cleanup_heavy_job_vram()

None cleanup_heavy_job_vram

(

)

Drop any models loaded during a heavy job (simulate TL cache, HF weights).

Definition at line 66 of file vram_guard.py.

References release_inspection_models().

cuda_oom_message()

str cuda_oom_message	(	*	,
		str	job,
		str	model_id )

Definition at line 104 of file vram_guard.py.

References oom_message().

cuda_vram_gib()

tuple[float, float] | None cuda_vram_gib

(

)

Backward-compatible alias.

Definition at line 31 of file vram_guard.py.

References accelerator_vram_gib().

log_accelerator_vram()

None log_accelerator_vram

(

Callable[[str], None] | None

log = None

)

Definition at line 36 of file vram_guard.py.

References accelerator_vram_gib().

Referenced by log_cuda_vram().

log_cuda_vram()

None log_cuda_vram

(

Callable[[str], None] | None

log = None

)

Definition at line 53 of file vram_guard.py.

References log_accelerator_vram().

oom_message()

str oom_message	(	*	,
		str	job,
		str	model_id )

Definition at line 87 of file vram_guard.py.

References accelerator_vram_gib().

Referenced by cuda_oom_message(), and raise_if_cuda_oom().

raise_if_cuda_oom()

None raise_if_cuda_oom	(	BaseException	exc,
		*	,
		str	job,
		str	model_id )

Definition at line 108 of file vram_guard.py.

References oom_message().

release_inspection_models()

None release_inspection_models

(

)

Unload HookedTransformer / embedding caches so simulate can own the GPU.

Definition at line 57 of file vram_guard.py.

Referenced by cleanup_heavy_job_vram().

restore_session_model()

None restore_session_model

(

str | None

model_id

)

Definition at line 71 of file vram_guard.py.