+ j!Jla0t$Rt^RIHt^RIHt^RIt^RIH uH t ]'d~.t R] R&^t R] R&.tR] R &.tR] R &.tR ] R &.tR ] R &.tR] R&.tR] R&.tR] R&.tR] R&.tR] R&RtR] R&RtR] R&.tR] R&.tR] R&.tR] R&.tR] R&RR ltR!tR"tR#t R$t!R%t"R#)&a#Runtime driver for torch.compiler.precompile artifacts, authored as real code. Nothing here is imported or run by torch at runtime, and the generated artifacts never import this module. Instead torch._precompile emits these function bodies VERBATIM (via inspect.getsource, on the to_python_code / emit path only) into the self-contained python_code string, after the calling-convention metadata and the compiled/captured graph. Authoring the driver as real code -- instead of a hand-written string literal -- lets pyrefly / ruff / IDEs type-check and navigate the load-bearing driver logic that would otherwise be invisible inside a string (and drops the wall of ``# noqa: F821``). Keeping python_code self-contained and version-frozen (its behavior is hashed via code_hash, invariant 7) still holds: the artifact carries the driver TEXT, it does not import it, so there is no torch-version skew. The emit path runs getsource in-process where torch source is present; load() never touches this module. The names the emitted bodies read from the artifact's own namespace -- the metadata constants, the ``_torch`` / ``_pytree`` import aliases, and the graph's ``call`` -- are declared under TYPE_CHECKING below so the bodies type-check here; at emit time they resolve against the metadata + graph sections that precede the driver in python_code. INVARIANT: ``_extract_param_buffers`` reproduces ``torch._precompile._intern_param_buffers``'s params-then-buffers, intern-by-identity ordering VERBATIM; the two must stay in sync (see Note [precompile programming model]). ) annotations) TYPE_CHECKINGNz list[int]MODULE_POSITIONSintNUM_POSITIONAL_ARGSz list[str] PARAM_NAMES BUFFER_NAMESzlist[tuple[int, ...]] PARAM_SHAPES BUFFER_SHAPES PARAM_DTYPES BUFFER_DTYPES PARAM_DEVICESBUFFER_DEVICESGRAD_PARAM_INDICESz str | NoneIN_SPECstrOUT_SPECz#list[tuple[int | None, ...] | None]USER_INPUT_SHAPESzlist[str | None]USER_INPUT_DTYPESUSER_INPUT_DEVICESz5list[dict[int, tuple[int | None, int | None]] | None]USER_INPUT_BOUNDSc V^8dQhRRRR/#) flat_inputsz list[object]return)formats"Y/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/_precompile_driver.py __annotate__r:s<<,<<<cR#Nr)rs&rcallr#:sr c<aaa a \V4^8o\4o .o .oVVV V 3Rlp\V4F)wr#VPRR7FwrEV!W$V4K K+ \V4F)wr#VP RR7FwrFV!W$V4K K+ S S3#)aHLift the runtime modules' params then buffers, interning by identity, in the same order as capture, so the list lines up with the compiled/captured graph. Returns (pb, names) where names mirrors PARAM_NAMES + BUFFER_NAMES. This ordering AND the naming must match torch._precompile._intern_param_buffers verbatim (its INVARIANT).c<\V4S9dOSP\V44SPV4SPS'd RV RV 2MT4R#R#)m.N)idaddappend)mintmultinamespbseens&&&rintern&_extract_param_buffers..internGsJ a5  HHRUO IIaL LL1RD!A 6 r F)remove_duplicate)lenset enumeratenamed_parameters named_buffers) modsr2r+r&r,pbr.r/r0r1s & @@@@r_extract_param_buffersr==s IME 5D B E77 4&&&>DA 2! ?!4OOUO;DA 2! <! u9r c^RIHpV!V4h))PrecompileError)torch._precompiler@)msg_PrecompileErrors& r_failrDVsF 3 r c \\4\\4,pW8wd\RV: RV: R24\\4\\ 4,p\\ 4\\4,p\\4\\4,p\WW4V4Fwrgrp \VP4\V48wd3\RV: R\VP4 R\V4 R24\VP4V 8wd!\RV: RVP RV R24\VP4V 8wgK\RV: RVP RV R24K R#) z3precompile: the runtime model's param/buffer names z! do not match the traced model's zU; the runtime model must be structurally identical to the traced model (invariant 2).z%precompile: the runtime param/buffer z has shape z but the traced model's was z has dtype z is on device N)listrrrDr r r r r rziptupleshaperdtypedevice) r0r/expectedexpected_shapesexpected_dtypesexpected_devices_nm_t_shp_dt_devs && r_check_structurerU^soK 4 #55H  A%K""*.; < <(4 +>>O<(4 +>>OM*T.-AA$' ?5E% D ?eDk ) 7wk?##?d }MZ[  rxx=C  7wk"((T**-/?@  ryy>T ! 7wnRYYKX..2V4LM !%r c \V4\8wd \R\ R\V4 R24.p\Fup\ W,\ P P4'g,\RV R\W,4P R24VPW,4Kw \V4UUu.uFwr4V\\49gKVNK ppp\P!\V44wrg\ e+V\P"!\ 48wd \R4\V4\\$48wd \R 4\'V\$\(\*4EF<wrrV e!\ V\ P,4'gK-\VP.4p \V 4\V 48wgM\0;QJd%R \'W44F 'gK R M R M!R \'W444'd\R V R\V 4 R24V e7\3VP44V 8wd\RVP4 RV R24V fEK\3VP64V 8wgEK \RVP6 RV R24EK? \9V4wr\;W4\ P<!4;_uu_4\?\A.V OVO44pRRR4\B'd\\B4pX\V4V, RpVR\V4V, p\'\BV4FAwppV V,pVPDf VVn"K&VPDPGV4KC \PH!X\P"!\J44#uuppi +'giL;i)aRun the captured ATen graph eagerly. Pass the same args the traced fn took -- the module(s) in the same positions plus the runtime inputs. The module(s) must be structurally identical to the ones precompile traced (same param/buffer order and tying); only the weight values may differ. The eager backend runs the graph as captured: inputs (including tensor subclasses) are passed through unchanged (no dense flatten/unflatten), and the graph's flat outputs are reassembled into fn's output structure. If fn ran a backward, the trailing grad outputs (one per GRAD_PARAM_INDICES entry) are parameter grads, scattered (accumulated) onto the params that received one like eager .backward() -- frozen / non-contributing params keep .grad = None.precompile: expected 2 positional args (the same as the traced fn), got  (invariant 2).!precompile: argument at position = must be the nn.Module the traced fn took (invariant 2), got r'Nprecompile: runtime inputs have a different structure than the traced example inputs (invariant 3); they must match in nesting and count.precompile: runtime inputs flattened to a different number of leaves than the traced example (invariant 3); they must match the traced structure.c3."TF wrW8gxK R#5ir"r.0aes& r !_eager_forward..s(LODAOsTF&precompile: a runtime input has shape ( but the artifact was traced with shape zl; the graph is specialized to the static dims (invariant 3). Retrace for this shape, or use backend='eager'.&precompile: a runtime input has dtype ( but the artifact was traced with dtype n; the graph is specialized to the example dtype (invariant 6). Cast the input to the traced dtype, or retrace.)precompile: a runtime input is on device ' but the artifact was traced on device p; the graph is specialized to the example device (invariant 6). Move the input to the traced device, or retrace.)&r5rrDr isinstance_torchnnModuletype__name__r*r7r6_pytree tree_flattenrHrtreespec_loadsrrGrrTensorrIanyrrJrKr=rUno_gradrFr#rgradadd_tree_unflattenr)argsr:_iira user_inputs user_flat_runtime_in_specrQrRrSrT_actr0_namesoutr,gradsidxgr;s* r_eager_forwardrs@ 4y'' #$7#89##&t9+_ >  D$(FII$4$455 3B48..248n.E.E-FaI  DH "+4SASAQ=R4R11KS")"6"6u[7I"JI/73I3I'3RR  R  9~.//  V  #$&79K # .s&) 1@ATM $ $af $s! !TFrerfz (None = a dynamic dim, any size); the graph is specialized to the static dims (invariant 3). Retrace, mark the dim dynamic via mark_unbacked, or use backend='eager'.rgrhrirjrkrlzprecompile: runtime input dim z has size z but mark_unbacked declared min=z (invariant 3).z but mark_unbacked declared max=z expected sizestridezwrong number of dimensionsc3"TFDqRJ;'d5\;QJdRV4F 'gK RM RM !RV44xKF R#5i)Nc3("TFqRJxK R#5ir"r)r`_ds& rrc._inductor_forward...Os";":sTF)rw)r`_ss& rrcrNs< FWdN ; ;ss";";sss";";; ;FWsA AAArz If two inputs share a mark_unbacked shape_id, their marked dims are bound to one symbol and so MUST have equal sizes at runtime; this can also be a shape_id equality violation.aprecompile: a runtime tensor's shape or memory format differs from the traced example; the inductor backend specializes on input shape and memory format (invariants 3 and 6). The mismatch can be a user INPUT or a model PARAMETER/BUFFER whose layout (memory format) differs from the example weight, since the inductor backend also bakes each param/buffer's layout. Pass the model/inputs in the example's shape and layout (.contiguous() to match a contiguous example, or match the example weight's layout), or use backend='eager'.z Underlying: )(r5rrDrrmrnrorprqrrr*r7r6rsrtrHrrurrGrrrrvrIrwrrJrKitemsr=rUrFr#AssertionErrorrryrzr{r)r|r:r}r~rarrrrQrRrSrT_bndrr_lo_hi_szr0rr_e_m _has_dynamic_shape_id_noter,rrrr;s* r_inductor_forwardrs& 4y'' #$7#89##&t9+_ >  D$(FII$4$455 3B48..248n.E.E-FaI  DH "+4SASAQ=R4R11KS")"6"6u[7I"JI/73I3I'3RR  R  9~.//  V &) &!#T S"S9SS)#__conditional_annotations____doc__ __future__rtypingrtorchrntorch.utils._pytreeutilsrsr__annotations__rrrr r r r r rrrrrrrrr#r=rDrUrr)rs@rrs2# %% #%i$  K L) *,L',+-M(- L) !M9!!M9! "NI"$& &GZHc=?:?*,',+-(-OQLQ=2 &RQIhUIr