globals

globals

Object

Bases: Tensor

save

save(_=0)

Save an object to be accessable after the trace context is exited.

Examples:

model = LanguageModel("gpt2", device_map='auto', dispatch=True) with model.trace("Hello World"): ... attn_0 = model.transformer.h[0].attn.output.save() print(attn_0)

__getattr__

__getattr__(name: str) -> Self

__getitem__

__getitem__(key: Any) -> Self

__call__

__call__(*args: Any, **kwargs: Any) -> Self

TracingCache

TracingCache()

cache instance-attribute

cache = {}

code_cache instance-attribute

code_cache = {}

get

get(cache_key: Tuple)

Check if the given filename and lineno is in the cache.

add

add(cache_key: Tuple, value: Any)

Add the given value to the cache.

get_code

get_code(cache_key)

Get a cached compiled code object.

add_code

add_code(cache_key, code_obj)

Cache a compiled code object.

clear

clear()

Clear all cached source, AST, and code objects.

Globals

Process-wide tracing state.

Holds two pieces of true global state: - saves: set of id() for objects marked via .save(). The root tracer's push() filters its frame locals against this set so only saved values propagate out of the trace. - cache: source/AST/code-object memoization across traces.

Root-vs-inner detection lives on the tracer itself — see Tracer.push — by checking whether the target frame is an nnsight-generated frame (i.e., another trace's compiled body).

saves class-attribute instance-attribute

saves = set()

cache class-attribute instance-attribute

cache = TracingCache()

clear staticmethod

clear()

save

save(object: Any)