thunder.core.trace.TraceCtx

class thunder.core.trace.TraceCtx(fn=None)

Bases: object

Trace representing fn.

Parameters:

• fn (Optional[Callable]) – Callable to represent.

• fn (None | Callable) –

fn

Callable to represent. It’s either a callable written with pytorch functions or Module.

Type:

Callable | None

args

Arguments of fn. Elements are proxified, e.g., torch.Tensor is converted to TensorProxy.

Type:

Any

kwargs

Keyword arguments of fn. Values are converted to Proxy if possible.

Type:

dict[str, Any]

bound_symbols

Each BoundSymbol represents one line of trace.

Type:

list[BoundSymbol]

scopes

In most cases, same as [self.bound_symbols]. Direct modification of this attribute would provide better flexibility to trace transformation as in insert_inplace() and replace_inplace(). Also [tutorial] How to Implement CPU Offloading as Trace Transform would be a great read.

Type:

list[list[BoundSymbol]]

__init__(fn=None)

Parameters:

fn (Optional[Callable]) –

Methods

__init__([fn])
add_bound_symbol(bsym)	rtype: None
add_name(name)	rtype: None
add_object(x)	rtype: ContextObject
get_provenance()	rtype: Optional[TraceProvenance]
has_name(name)	rtype: bool
is_constant(x)	rtype: bool
make_name([name, prefix])	rtype: str
make_object_name(x)	rtype: str
mark_complete()	rtype: None
peek_scope()	rtype: Optional[list]
pop_scope()	rtype: list
post_unpack(fn)	rtype: None
push_scope(scope)	rtype: None
python(*[, print_depth, include_decorators])	rtype: str
python_callable(*[, global_dicts])	rtype: Callable
python_ctx()	rtype: dict
save_trace(filename)	rtype: None
set_current_source_location(filename, positions)
set_provenance(provenance)	rtype: None
siginfo()	rtype: SigInfo
unpacked()	rtype: None
unpacking()	rtype: None

Attributes

bound_symbols
output
tags