uki :dZddlmZddlmZmZddlmZddlZddl Z ddl Z ddl Z ddl m Z mZddlZddlmZddlZddlZddlmZdd lmZdd lmZdd lmZdd lmZdd lmZddlmZddlmZddlmZddlmZddl m!Z!ddl m"Z"ddl#m$Z$ddl%Z&ddl'm(Z)ddl*m+Z+ ddl,m-Z-ddl0m1Z2ejfZ3e"jhZ4e Z5ejlZ6ejnZ8ejrZ:e Z;eGddZ?e?Z@daAGdde jZCeCZDejdZF dd)ZKeejd*"dd+ d?d,ZL d@d-ZM dAd.ZN dBd/ZO dCd0ZPdDd1ZQd2ZRdEd3ZSejZUd4ZVd5ZW dF dGd6ZXeeYe= e2` or `e1 <= e2`. See more details at https://github.com/jax-ml/jax/blob/main/jax/experimental/jax2tf/README.md#user-specified-symbolic-constraints. with_gradient: if set (default), add a tf.custom_gradient to the lowered function, by converting the ``jax.vjp(fun)``. This means that reverse-mode TensorFlow AD is supported for the output TensorFlow function, and the value of the gradient will be JAX-accurate. native_serialization_platforms: Specifies the platform(s) for which to lower the code. Must be a tuple of strings, including a subset of: 'cpu', 'cuda', 'rocm', 'tpu'. The default (`None``), specifies the JAX default backend on the machine where the lowering is done. native_serialization_disabled_checks: Disables the specified safety checks. See docstring of `DisabledSafetyCheck`. Returns: A version of `fun_jax` that expects TfVals as arguments (or tuple/lists/dicts thereof), and returns TfVals as outputs, and uses only TensorFlow ops and thus can be called from a TensorFlow program. zbThe `native_serialization` parameter is deprecated and will be removed in a future version of JAX.) stacklevelzXThe `enable_xla` parameter is deprecated and will be removed in a future version of JAX.c3$K|]}|dv yw))cpucudarocmtpuNr+.0ps r$ zconvert..s9339sznnative_serialization_platforms must be a sequence containing a subset of {'cpu', 'cuda', 'rocm', 'tpu'}. Got: c ^tjs/tjst ddtj zt sBtjtjjtjdad d}tj||}t!j"|tj||}t!j"|d||f}tj$|\}}t't)t*t-t/||}t1 j3dr$tj4d fd }||} nEj7|\} } d} | D cgc]#} tj8j;| | %} } j=| Dcgc]}tj>|d } }tj@| }|Scc} w#j=wxYwcc}w) Nz5convert must be used outside all JAX transformations.z Trace state: Tctj|}td|D}t|\}}t j ||S)Nc3K|]@}t|tjjjr |j n|Bywr:)rrcompatv1 DimensionvaluerOds r$rQzRconvert..converted_fun_tf..jax_arg_spec_from_tf..s?2 )BIILL,B,BC77IJK2sAA)npshapetuple_tfval_to_tensor_jax_dtypejaxShapeDtypeStruct)a tf_arg_shape_ a_jax_dtypes r$jax_arg_spec_from_tfz?convert..converted_fun_tf..jax_arg_spec_from_tfsKXXa[l2$022l2!4na ! !, <.converted_fun_tf..converted_fun_flat_with_custom_gradient_tf(sJ,0??<+H (':y.-$1!+,*!+%'( (r,zkThe jax2tf-converted function does not support gradients. Use `with_gradient` parameter to enable gradients)inputmessage jax2tf_out)raTfValreturnzjax.ShapeDtypeStruct)rprwrxrw)!rtrace_state_cleanr7r0 ValueError trace_ctx_has_registered_tf_source_pathrregister_exclusionospathdirnamer__file__r tree_mapr symbolic_args_specs tree_flattenr]mappreprocess_arg_tfrangelenNativeSerializationImplbefore_conversioncustom_gradientrnraw_opsPreventGradientafter_conversionidentitytree_unflatten)rk kwargs_tfreargs_jax_specskwargs_jax_specscombined_args_tfrpargs_kwargs_treers outs_flat_tfrmrcruoxout_tfrgrjrhrrrqrErDr@r?rAs @@@@r$converted_fun_tfz!convert..converted_fun_tfs  ! ! #,?,N,N  A $..) * + ,, *))"''//"++*FG'+$ =''( J--$ L+%.%;%;F M  Gs%AH(HHH*HH')rkrwrrwrxrw) DEFAULT_NATIVE_SERIALIZATIONwarningswarnDeprecationWarningrlistr]allrzr check_callable) rqr?r@rArBrCrDrErs ```` `` r$convertrs^!== MM 6q*33 MM 6q*# 5e} E 979 9  01 3 44&++I%J"WQQf r,c@eZdZ ddZdZdZ ddZd dZy) rct|||_t|dr|}ntj|}||_||_||_||_||_ y)N)rDrEtrace) dictconvert_kwargshasattrr_jitrqrgrhrErD)r4rqrgrhrDrEfun_jits r$r5z NativeSerializationImpl.__init__Ks_>\DhjDw g  gDL DO$D0TD-*HD'r,c(tjgt_fd}||_dg}tj|j |j |j||ji|j|_ |dJ|d|_ y)Nct_yr:r;)_prev_func_listsr$_restore_contextzCNativeSerializationImpl.before_conversion.._restore_contextds ;J8r,) platformsdisabled_checks/_device_assignment_for_internal_jax2tf_use_onlyr) r7r2rr_export_internalrqrDrErgrhexporteddevice_assignment)r4r_exported_device_assignmentrs @r$rz)NativeSerializationImpl.before_conversion`s)HHO9;6K-D#'&G,, 55AA8S    - ++ -DM 'q ) 56 58;Dr,c$|jyr:)rr3s r$rz(NativeSerializationImpl.after_conversionrsr,ct||j}|t|jj|jjfSr:)_run_exported_as_tfrr] out_avalsout_tree)r4rpresultss r$rnz"NativeSerializationImpl.run_fun_tfus:", >G E$--112DMM4J4J JJr,c ,tj|j|jj|jj |jj |jj|jj|jdS)NT)in_treein_avalsin_shardings_hlorout_shardings_hlor apply_jit) r _get_vjp_funrqrrrrrrrr3s r$ get_vjp_funz#NativeSerializationImpl.get_vjp_fun{sh    (, (=(=)-)?)?151O1O*.--*A*A26--2Q2Q262H2H*. 00r,N)rDSequence[str] | NonerESequence[DisabledSafetyCheck])rpSequence[TfVal]rxzGtuple[Sequence[TfVal], Sequence[core.ShapedArray], tree_util.PyTreeDef])rxz-tuple[Callable, Sequence[core.AbstractValue]])r(r)r*r5rrrnrr+r,r$rrJsDI/CI6SI*<$K.K]K 0r,rc6t|\}}|jS)aXComputes the TensorFlow dtype using JAX's typing rules. If the value is a tf.Tensor, it starts with its dtype. If the value is a constant it uses JAX to infer its dtype. The resulting dtype follows the JAX type inference rules, and depends on the value of the JAX_ENABLE_X64 flag. See README.md for how 64-bit values are treated. )r^dtype)valtvalrcs r$ dtype_of_valrs 's +'$ r,jax2tf_eval_polymorphic_shapes)r?cdfd }|S)aEvaluates the output shape in presence of shape polymorphism. This is done without lowering or executing the function, same as for `jax.eval_shape`. Args: fun_jax: target JAX function to be called. Its arguments and return value should be JAX arrays, or nested standard Python containers (tuple/list/dict) thereof (pytrees). polymorphic_shapes: Specifies input shapes to be treated polymorphically during shape evaluation. See discussion for `jax2tf.convert`. .. warning:: The shape-polymorphic lowering is an experimental feature. Returns: a function that takes `jax.ShapeDtypeStruct`s (or any values with `.shape` and `.dtype` attributes) corresponding to the inputs for `fun_jax`, and returns a tuple with: * the jax.ShapeDtypeStruct corresponding to the result, as for `jax.eval_shape`. The shape may contain symbolic dimension expressions. * the value that can be passed to `polymorphic_shapes` for a subsequent call to `jax2tf.eval_polymorphic_shape`, or `jax2tf.convert`. For example: >>> import jax >>> from jax.experimental import jax2tf >>> from jax import numpy as jnp >>> >>> f = lambda A, x: jnp.sin(jnp.dot(A, x)) >>> A = jax.ShapeDtypeStruct((2000, 3000), jnp.float32) >>> x = jax.ShapeDtypeStruct((3000, 1000), jnp.float32) >>> out_spec, out_poly_shape = jax2tf.eval_polymorphic_shape(f, polymorphic_shapes=["a, b", "b, c"])(A, x) >>> print(out_spec.shape) ("a", "c") >>> print(out_poly_shape) (a, c) >>> res_spec, res_poly_shape = jax2tf.eval_polymorphic_shape(lambda x: x.T, polymorphic_shapes=[out_poly_shape])(out_spec) >>> print(res_poly_shape) (c, a) ctj|}tjg|}t j d|}||fS)Nc,t|jSr:strr\)rs r$zKeval_polymorphic_shape..do_eval_polymorphic_shape..sQWWr,)r rr_ eval_shaper r)rgargs_poly_specs res_poly_specres_polymorphic_shaperqr?s r$do_eval_polymorphic_shapez9eval_polymorphic_shape..do_eval_polymorphic_shapesL00&(ONN7=_=M%../E}U / //r,)rxrr+)rqr?rs`` r$eval_polymorphic_shapersZ0 #"r,c(dfd}|fdfS)aWraps the function to take a (flat) list of positional args. jax2tf works better and is simpler when the JAX function takes and returns just a tuple of values (no pytrees, no kwargs). This is in part because jax.vjp does not support kwargs and we can only set tf.custom_gradient on functions with flat arguments and results Returns: * the wrapped JAX function taking and returning a flat list of arguments * a thunk that can be called after the wrapped function has been called to return the output pytree. Nctj|\}}|i|}tj|\}}|k(sJ||Sr:)r rr) args_flat_jax tree_args tree_kwargstree_res res_flat_jaxrrqr out_tree_refs r$ fun_flat_jaxz%flatten_fun_jax..fun_flat_jaxs\&55g}MI{ 1[1H&33H=L(  <8#;; ;L r,cSr:r+)rsr$rz!flatten_fun_jax..s|r,r+)rqrrrs`` @r$flatten_fun_jaxrs, + ++r,ct|sd|dt|d}t|t|\}}t j |d|}|S)zwPre-processes the TF args. Returns: a tuple with the pre-processed TF arg, the TF shape, and the JAX dtype. z Argument z of type zN of jax2tf.convert(f) should be NumPy array, scalar, tf.Variable, or tf.Tensor jax2tf_arg_)r%type TypeErrorr^rr)arg_idxarg_tfmsgrcs r$rrs_ 6  vhiV ~6? ?C C.)0)&! ;;vWI6 7& -r,c*dddfd}|S)aPrepares the TF function to be used with tf.custom_gradient. Args: impl: the serialization implementation details with_gradient: whether to include a tf.custom_gradient args_specs, kwargs_specs: the jax.ShapeDtypeArrays for the args and kwargs args_tf: the flattened TF arguments of the primal function outs_avals: the flattened output JAX abstract values of the primal function outs_tf: the flattened TF outputs of the primal function N) variablesc R|rtd|tjd5dd}tt || }t |z} j \}}td|D}t |f |d j|}ddd|S#1swYSxYw)NzmUnexpected variables used in forward pass. This should not happen for first-order differentiation. variables= jax2tf_vjpc||Stj|jtjk(s Jd|t j |tS)Nz out_ct_tf=r)rprimal_dtype_to_tangent_dtyperrfloat0r zeros_like_tf_np_dtype_for_float0) out_ct_tf out_ct_avalrs r$ fix_out_ctzC_make_custom_gradient_fn_tf..grad_fn_tf..fix_out_ctsV   11+2C2CD UfZdZcYeWffU}}V+BCCr,c3FK|]}t|jywr:r)rOras r$rQzB_make_custom_gradient_fn_tf..grad_fn_tf..*s %  AGG %!)rAr?)rcore.ShapedArrayrrw)rzr name_scoper]rrrr)rout_cts_flat_tfrout_cts_fixed_flat_tfvjp_args_flat_tf fun_vjp_jax vjp_in_avalsvjp_polymorphic_shapes in_cts_flatrkrjrlrmrAs r$ grad_fn_tfz/_make_custom_gradient_fn_tf..grad_fn_tfs   L   | $2 D$C OZQX$YZw*??"&"2"2"4k<$%% G#1    !1 2k+2: ;2: s A+BB&)rrwr+) rqrjrArgrhrkrlrmrs `` ``` r$roros& &&P r,c |j}d}tt|||}td|jD}td|jD}t |Dcgc]\}}||j vs|}}}|j } tj} | r| } nd} | | kDrtd| d| dt| ||tj_run_exported_as_tf.._convert_value..Fs.0 65&&v.0s'))rzipr\ _to_tf_dtyperrcast)ravalconversion_dtypes r$_convert_valuez+_run_exported_as_tf.._convert_valueDsp 0$' CII$>0 0=26= 0$DJJ/4::% WWS* ++ jr,c3TK|] }td|jD"yw)c3NK|]}tj|r|ndywr:rrYs r$rQz0_run_exported_as_tf...Rs* %%%a(Ad 2 %s#%N)r]r\rOout_avals r$rQz&_run_exported_as_tf..Qs0*   %^^ %%*s&(c3FK|]}t|jywr:)r rrs r$rQz&_run_exported_as_tf..VsTXL0TrNzUXlaCallModule from your TensorFlow installation supports up to serialization version z) but the serialized module needs version z5. You should upgrade TensorFlow, e.g., to tf_nightly.F)versionToutSout function_listhas_token_input_outputc3<K|]}|jywr:)upperrNs r$rQz&_run_exported_as_tf..{s(Oq(Osrc32K|]}t|ywr:)r)rOdcs r$rQz&_run_exported_as_tf..}s13  B13rr+ use_shardy_partitionerzXlaCallModule %smodule)skip_replicated_sharding)+rr]rr enumeratemodule_kept_var_idxcalling_convention_versiontfxla%call_module_maximum_supported_versionAttributeErrorNotImplementedErrorrr7r2 function_def signaturenamerdisabled_safety_checksDisabledSafetyCheckplatformrr$rXr vlog_is_onvlogrmlir_module_serializedappendrr _shard_valuerexecuting_eagerly call_modulerUrVget_default_graph_add_function_recursive_inference_functionrr)rpr args_avalsr out_shapes_tf out_typesiatfkept_args_flat_tfrget_max_supported_versionmax_supported_version concrete_fncall_module_attrskept_in_shardingsress r$rr:sY  * s><DE,*((**- TASASTT))2<)@fvq#AIeIeDesff  / /'% % K K57 $$ 456%Y'<<     = = I1OO  " " , , 1 1PR# $)(OH? ( ? ?H  ' ';aX66q9:; )+)=)=)? A,./, ,<*;<#77C*II iill$$&>>  ) ) S242F2F2HJh002 3# c.#x'9'9:;# *kg % $%(s$,MMM4'M MMctj|}t|jt|jk\r/|jdt|j|jk(s J||f|S)aConverts JAX avals from logical to physical, if relevant. JAX might have avals whose logical vs physical shape/dtype may differ, and only the physical view is expected to possibly relate to TF. TF impl rules should operate on the physical form. A JAX logical aval might even correspond, in principle, to several physical avals, but we don't support those here. Instead we assert there is only one and return it. N)r physical_avalrr\)rrKs r$_jax_physical_avalrLsq$$T*- m!! "c$**o 5   .s4:: /4:: =VANPT@UV > r,cTttjd|jS)Nr+)rLr ShapedArrayrrs r$_jax_physical_dtyperOs D,,R7 8 > >>r,cXt|}ttd|jS)zIGenerate a TF shape, possibly containing None for polymorphic dimensions.c4tj|rdS|Sr:)r is_symbolic_dim)rZs r$rz#_aval_to_tf_shape..sV%;%;A%>TAr,)rLr]rr\)rs r$_aval_to_tf_shaperSs, D !$ sE::  r,c t|}|tjk(rt}t jj |S#t$rYCwxYwr:)rOrrrrras_dtype) jax_dtypes r$r r sO #I.I &--'I   I &&    s A AActj|j}|tjvrt d|d|S)Nzdtype zR is not a valid JAX array type. Only arrays of numeric types are supported by JAX.)rcanonicalize_dtypeas_numpy_dtype_jax_dtype_setr)tf_dtypedts r$ _to_jax_dtyper]sN  !8!89"v$$$ fRD!OO PP )r,ct|tjtjfrO|xst |j }t |}||j k7rtj|||fS||fS|xstj|j }t||f}|xr,tj|dkDxrtjdu}|r$tjj|d\}}nd}|t |}|t j"k(r3tj$tj&||j(}t+|drSt!j,|j t j.r%|j j0j3|}tj4||}|r||ftj|<||fS)a<Converts a scalar, ndarray, or tf.Tensor to a tf.Tensor with proper type. If `jax_dtype` is missing, uses JAX typing rules. See README.md for details regarding 64-bit values. Args: val: a scalar, ndarray, tf.Tensor, or tf.Variable jax_dtype: an optional dtype to use. If missing, uses JAX type inference rules for constants. memoize_constants: whether to memoize TF constants. We can't do this everywhere, we may be outside of a conversion scope. Returns: a tuple with a tf.Tensor with the type as needed by JAX, and the JAX type. N)NNrr)rrrr r]rr rr abstractifyidr[sizer7constant_cachegetrrzerosr\rYr issubdtypeextended_rulesphysical_constconvert_to_tensor)rrVmemoize_constantsr const_key do_memoizerctf_vals r$r^r^s$bii-.5]3995I#I.399$ WWS* +Y 66 )^8T--c288IC)$I$k q(8k=P=_=_gk=kJ%4488LQiaf ~%i0 fmm #hhrxx}&6&E&EF g 6#4#4SYY#Pii--c2##C/?@f 9.vsCSVCr!.rG)rG)protoTrp)rqsharding_v2_proto)"to_protoris_hlo_sharding_replicatedrrrr[rursrtr~rv transposerflatr OpShardingrrrrrr$rX tf_contextr9rzr]r __version__splitrShardingapply_to_tensor)rsdr&sharding_prototadxla_sharding_v1_protoxla_sharding_v2_proto tf_versions r$r8r8AsZ J;;=.--b1 J  0 0  * *  $))NEEFG 11 2 >55 6   C  0 0C4@3J3J ~"" #!/!J!J!!/!J!J#22 4 ""((5A5L5L $$ %#1#L#L . F F(::*>>#1#L#L%446!!!#  B CCCR^^%9%9#%>r%BCC*'  '< = M M T N    %: ; K K 43H L r,ctj}tjtjf|_tjtjg|_dtji|_t |j}t |j}t |j|t usJ|tusJtusJtjj|ddtjj|ddtjjdfdy) z/Registers TF custom container types as pytrees.ract|dfSr:r]xss r$rz._register_checkpoint_pytrees..s&+Bi%6r,ct|Sr:rrcrs r$rz._register_checkpoint_pytrees..s eBir,ct|dfSr:rrs r$rz._register_checkpoint_pytrees..suRy$>Or,ct|Sr:)rrs r$rz._register_checkpoint_pytrees..s 48r,cft|jt|jfSr:)r]valueskeys)ss r$rz._register_checkpoint_pytrees..s qxxz"E!&&(O4r,c(t||Sr:)r )kr dict_wrappers r$rz._register_checkpoint_pytrees..sLQ,r,N) rModulerabcrr]rrr_r register_pytree_node)m tuple_wrapper list_wrapperrs @r$_register_checkpoint_pytreesrsiik! biik"!# biik"!# biik!#qss)-acc,acc, e ## # T !! ! T !! !--$$]578OQ--$$\3O%;=--$$4,.r,)r#rwrxbool)rxzlist[Any] | None)rqrr?z9str | PolyShape | None | Sequence[str | PolyShape | None]r@z Sequence[str]rArrBrrCz"bool | _DefaultNativeSerializationrDrrErrxr)rrwrxDType)rqrrxr)rqrrxztuple[Callable, Callable])rrrrwrxrw) rjrrArrkrrlzSequence[core.ShapedArray]rmr)rprrzexport.Exportedrxr)rrrxr)rrrxztuple[int | None, ...])NF)rrwrVz DType | Nonerxztuple[TfVal, DType])rxrwryPartitionsOrReplicated)rrwrzxla_client.HloSharding | Noner&rrxrw)^__doc__ __future__rcollections.abcrr functoolsr contextlibrsr~ threadingtypingrrrabslr numpyr[r_r r jax._srcr r rrrrrrjax._src.exportrr jax._src.libr tensorflowr!tensorflow.compiler.tf2xla.pythonrr*tensorflow.compiler.xlar+tensorflow.python.compiler.xla.experimentalrModuleNotFoundError1tensorflow.compiler.xla.experimental.xla_shardingtensorflow.python.eagerrr NameStack PolyShaperr2safe_maprsafe_zipr rwr PrecisionTyper%r'rr|localr.r7contextmanagerr0r<api_hookrrrrrrrorrLrOrSbool_rr r]r^r]rr|r8rr+r,r$rs7".  !%#&#;0MF:  & &   00 mm mm   :<""' A  A() ..<<     01]a57"&;GcCGRTz Yz&3z z  z #E z-Az3Pzz2zz:0:0z     @A#'4#+34#B4#r,2,6#(-(:&=:04: *9 : -G :*9:zs "1s ,s l ?  (( ' :>167*67;N7tuS#X45J3IJ:>2>+/>49>B.:SMLLMs H HH