ȅiDfSSKJr SSKJrJr SSKrSSKJr \(aSSKJ r J r "SS5r g) ) annotations)Any TYPE_CHECKINGN)DimEntry)DimTensorcl\rSrSr%SrS\S'S\S'SSjrSSjrSS jrSS jr SS jr S r g )EnableAllLayersa, RAII-style context manager for enabling functorch vmap layers. It manages the creation and cleanup of functorch dynamic layers. This is probably one of the more algorithmically important parts of first class dims. Intuitively, FCD can be thought of as another way of using vmap, where you don't actually have to vmap at the top level, instead the vmaps are implicitly determined by inspecting the bound dimensions on the FCD tensors involved in a compute (this is similar to our concept of non-lexical modes that we spent a long time talking about years ago). But under the hood you still need to actually enable the vmap mode. So once FCD has determined all of the dims we are batching over, it needs to enable all those layers so functorch can actually apply the batching rules. Therefore enable all layers! int levels_startz list[Dim] levels_to_dimcSSKJn SUl/UlUHWnUR 5(aMUR 5n[ XB5(deURRU5 MY URRSS9 g)z Initialize and push dynamic layers for all first-class dimensions. Args: levels: List of dimension entries to create layers for r)rrcUR$N)_level)ds Z/home/james-whalen/.local/lib/python3.13/site-packages/functorch/dim/_enable_all_layers.py*EnableAllLayers.__init__..6sahh)keyN) rrr is_positionaldim isinstanceappendsort)selflevelsrlrs r__init__EnableAllLayers.__init__"sx A??$$EEG!!))))""))!,  $67rc[UR5HIupURn[RR R US5nUS:XdMCX@lMK U$)N differentr) enumeratersizetorch_C _functorch_vmap_increment_nestingr)r ir batch_sizelevels r __enter__EnableAllLayers.__enter__8sR 2 23FAJHH''?? KXEAv$)! 4  rcUR[UR5-S- n[[UR55HBn[R R R5nXdU- :XaM5SXE- SU35e g)z)Clean up dynamic layers in reverse order.rzExpected layer z, got N)rlenrranger)r*r+_vmap_decrement_nesting)r exc_typeexc_valexc_tb to_remover-poppeds r__exit__EnableAllLayers.__exit__As}%%D,>,>(??!C s4--./AXX((@@BF]* !)-vh? *0rcp/n[UR5*S5HnUR[U55 M Un[R R RU5(Ga![R R RU5nUceX`R:a%X`R[UR5-:de[URX`R- 5n[R R RU5nUceURX5 [R R RU5n[R R RU5(aGM!SSKJn U "5n XZlXlX*lX:lU $)z Create a Tensor from a batched tensor by unwrapping functorch layers. Args: batchedtensor: Batched tensor from functorch operation has_device: Whether tensor has device info Returns: Tensor with appropriate levels rr)r )r4rrrr)r*r+is_batchedtensormaybe_get_levelrr3rmaybe_get_bdiminsert get_unwrappedrr _tensor _batchtensor _has_device_levels) r batchedtensor has_devicer!r-tensorr/rbdimr results r from_batchedEnableAllLayers.from_batchedJsg"$ ))++Q/A MM(1+ &0hh!!226::HH''77?E$ $$---%:K:Kc""O;3 4--e6G6G.GHIC88&&55f=D# ## MM$ $XX((66v>Fhh!!226:: +' rc^^[RRRU5(dgUn[ [ [ TR555HmUc g[UU4SjU55(dM%[RRRUTRT-5 [RRRU5nM g)z Update the levels of a batched tensor in place. This requires the _maybe_unsafe_set_level binding that we'll add to functorch. Args: batchtensor: Batched tensor to update levels: New levels to set Nc3X># UHo[TRT5:Hv M! g7fr)rr).0r"r-r s r 8EnableAllLayers.inplace_update_layers..s$HA!3!3A!677s'*) r)r*r+r>reversedr4r3rany_maybe_unsafe_set_levelrrB)r batchtensorr!implr-s` @rinplace_update_layers%EnableAllLayers.inplace_update_layersqsxx""33K@@ %D$6$6 789A|HHHH##;;D$BSBSVWBWXxx**88>:r)rrN)r!list[DimEntry])returnr )r6rr7rr8rr[None)rG torch.TensorrHboolr[r )rVr]r!rZr[r\) __name__ __module__ __qualname____firstlineno____doc____annotations__r#r0r;rLrX__static_attributes__rrr r sG 8,%N?'?1?? ?rr ) __future__rtypingrrr) _dim_entryrrrr r rfrrrjs%"% }?}?r