§
    9²¸j€‚  ã                  óÐ  — d Z ddlmZ ddlZddlZddlZddlZddlZddlm	Z	 ddl
mZ ddlmZ ddlmZ dd	lmZmZmZ d
dlmZ d
dlmZ d
dlmZmZmZmZ  e¦   «         rddlZddlmZ ddlm Z  ddl!m"Z"  ej#        e$¦  «        Z% G d„ de¦  «        Z&edid„¦   «         Z'edjd„¦   «         Z( edd¦  «        d„ ¦   «         Z) edd¦  «        d„ ¦   «         Z* eddd d!d"d#d$d%¦  «        d&„ ¦   «         Z+	 	 	 dkdld1„Z, edd2d3d4d5d6d7d8d9d:d;d<d=d>d?d@dAdBdCdDdEdFdG¦  «        dH„ ¦   «         Z-dmdM„Z.dndO„Z/dodR„Z0dpdV„Z1dqdX„Z2drdZ„Z3dsd[„Z4dtd]„Z5dud_„Z6dvd`„Z7dwda„Z8dxdd„Z9dyde„Z:dfZ;edzdh„¦   «         Z<dS ){u  Dynamo exporter.

Wraps `torch.export.export(strict=False)` with helpers that make Transformers
models exportable. The export pipeline uses five sections, in execution order:

1. **Model signature patch** (`patch_forward_signature`): replaces `model.forward`
   with a flat explicit signature derived from `sample_inputs` so `torch.export` does
   not expand `**kwargs` into a `combined_args` bundle that mismatches `dynamic_shapes`.
   This is the entry contract `torch.export` reads before tracing.
2. **Model patches** (`_PATCHES["dynamo"]` via `apply_patches("dynamo")`): reversible
   class-attribute swaps applied during tracing to replace non-exportable model patterns
   (data-dependent loops, in-place ops, mask checks) with export-safe equivalents.
   Modeling code itself is not updated because these patches are too model-specific.
3. **Pytree registration** (`register_cache_pytrees_for_model`): flatten/unflatten
   hooks (via `torch.utils._pytree.register_pytree_node`) for Cache subclasses and
   custom containers so `torch.export` can trace through them.
4. **Dynamic shapes** (`get_auto_dynamic_shapes`): automatic `Dim.AUTO` inference
   for all tensor and cache inputs when `DynamoConfig.dynamic=True`.
5. **Model state cleanup** (`reset_model_state`): non-Cache stateful module attributes
   (`_STATEFUL_CACHE_ATTRS`) are saved on entry, set to `None` during the trace, and
   restored on exit â€” so a previous eager forward doesn't leak into the trace and any
   FakeTensors the trace planted are discarded before the next eager forward.
é    )ÚannotationsN)ÚMutableMapping)Úcontextmanager)ÚAnyé   )Úlogging)Úis_detectron2_availableÚis_torch_availableÚtorch_compilable_checké   )Ú
HfExporter)ÚDynamoConfig)Úapply_patchesÚpatch_attributesÚprepare_for_exportÚregister_patch)ÚExportedProgram)ÚCache)ÚPreTrainedModelc                  ó0   — e Zd ZdZdgZddiZddiZdd„ZdS )ÚDynamoExporterao  Exporter that converts a [`PreTrainedModel`] to an `ExportedProgram`.

    Example:

    ```python
    >>> from transformers.exporters.exporter_dynamo import DynamoExporter, DynamoConfig

    >>> exporter = DynamoExporter()
    >>> exported = exporter.export(model, inputs, config=DynamoConfig(dynamic=True))
    >>> outputs = exported.module()(**inputs)
    ```
    Útorchz2.11.0z2.12.0Úmodelr   Úsample_inputsúMutableMapping[str, Any]ÚconfigúDynamoConfig | dict[str, Any]Úreturnr   c                óB  — t          |t          ¦  «        rt          di |¤Ž}n4t          |t          ¦  «        st          dt	          |¦  «        › �¦  «        ‚t          ||¦  «        \  }}}|j        }|j        r+|€)t           	                    d¦  «         t          |¦  «        }t          |¦  «         t          d¦  «        5  t          |¦  «        5  t          ||¦  «        5  t          ||¦  «        5  t           j                             |dt%          j        t          |¦  «        ¦  «        |j        ||j        ¬¦  «        }d d d ¦  «         n# 1 swxY w Y   d d d ¦  «         n# 1 swxY w Y   d d d ¦  «         n# 1 swxY w Y   d d d ¦  «         n# 1 swxY w Y   |S )Nz2Expected config to be a DynamoConfig or dict, got uy  `dynamic=True` with no explicit `dynamic_shapes` marks every input axis `Dim.AUTO`, so torch.export resolves symbolic shapes for all of them â€” including axes that are actually fixed (batch, a size-1 decode step, num_heads/head_dim). Passing explicit `dynamic_shapes` that mark only the axes which vary bypasses that symbolic-shape resolution and exports significantly faster.Údynamo© )ÚargsÚkwargsÚstrictÚdynamic_shapesÚ+prefer_deferred_runtime_asserts_over_guards)Ú
isinstanceÚdictr   Ú	TypeErrorÚtyper   r%   ÚdynamicÚloggerÚwarning_onceÚget_auto_dynamic_shapesÚ register_cache_pytrees_for_modelr   Úreset_model_stateÚpatch_model_configÚpatch_forward_signaturer   ÚexportÚcopyÚdeepcopyr$   r&   )Úselfr   r   r   Úoutput_flagsr%   Úexported_programs          úv/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/transformers/exporters/exporter_dynamo.pyr3   zDynamoExporter.exportW   s¯  € õ �f�dÑ#Ô#ð 	aÝ!Ð+Ð+ FÐ+Ð+ˆFˆFÝ˜F¥LÑ1Ô1ð 	aÝÐ_ÕQUÐV\ÑQ]ÔQ]Ð_Ð_Ñ`Ô`Ð`å-?ÀÀ}Ñ-UÔ-UÑ*ˆˆ}˜làÔ.ˆØŒ>ð 	D˜nÐ4Ý×Òð(ñô ð õ 5°]ÑCÔCˆNå(¨Ñ/Ô/Ð/õ ˜(Ñ#Ô#ð	ð 	å˜eÑ$Ô$ð	ð 	õ ˜u lÑ3Ô3ð	ð 	õ $ E¨=Ñ9Ô9ð		ð 	õ 16´×0CÒ0CØØÝ”}¥T¨-Ñ%8Ô%8Ñ9Ô9Ø”}Ø-Ø<BÔ<nð 1Dñ 1ô 1Ðð	ð 	ð 	ñ 	ô 	ð 	ð 	ð 	ð 	ð 	ð 	øøøð 	ð 	ð 	ð 	ð 	ð 	ð 	ñ 	ô 	ð 	ð 	ð 	ð 	ð 	ð 	øøøð 	ð 	ð 	ð 	ð 	ð 	ð 	ñ 	ô 	ð 	ð 	ð 	ð 	ð 	ð 	øøøð 	ð 	ð 	ð 	ð 	ð 	ð 	ñ 	ô 	ð 	ð 	ð 	ð 	ð 	ð 	øøøð 	ð 	ð 	ð 	ð  Ðs   ÃFÃE=Ã#E&Ã4AE	ÅE&ÅEÅE&ÅEÅE&ÅE=Å&E*Å*E=Å-E*Å.E=Å1FÅ=F	ÆFÆF	ÆFÆFÆFN)r   r   r   r   r   r   r   r   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__Úrequired_packagesÚmin_versionsÚtested_versionsr3   r!   ó    r9   r   r   E   sP   € € € € € ðð ð !˜	ÐØ˜XÐ&€LØ Ð)€Oð) ð ) ð ) ð ) ð ) ð ) rA   r   r   r   r7   údict[str, Any]c              #  óÞ  K  — g }|                      ¦   «         D ]M\  }}|�%t          | d¦  «        rt          | j        |¦  «        sŒ-|                     | j        ||fd„f¦  «         ŒN|                      ¦   «         D ]E}t          |d¦  «        r3t          |j        d¦  «        r|                     |j        dd„ f¦  «         ŒFt          |¦  «        5  dV — ddd¦  «         dS # 1 swxY w Y   dS )uK  Reversibly tweak `model.config` for the trace:

    - Applies `output_flags` (popped from inputs by `prepare_for_export`) onto
      `model.config.<flag>` so the model picks them up via its usual `<flag> if <flag> is
      not None else self.config.<flag>` fallback.
    - Disables `use_mamba_kernels` on every submodel's config that declares it (mamba/jamba
      kernels are not exportable).

    Originals are restored on exit. Flags whose value is `None`, or that the config doesn't
    declare, are silently skipped â€” useful for submodels that don't accept every parent flag.
    Nr   c                ó   — |S ©Nr!   )Ú	_originalÚvs     r9   ú<lambda>z$patch_model_config.<locals>.<lambda>œ   s   € Èa€ rA   Úuse_mamba_kernelsc                ó   — dS ©NFr!   ©rF   s    r9   rH   z$patch_model_config.<locals>.<lambda>Ÿ   s   € ÐY^€ rA   )ÚitemsÚhasattrr   ÚappendÚmodulesr   )r   r7   Úconfig_patchesÚflagÚvalueÚmodules         r9   r1   r1   ‹   s]  è è € ð €NØ#×)Ò)Ñ+Ô+ð Rð R‰ˆˆeØˆ=¥¨¨xÑ 8Ô 8ˆ=ÅÈÌÐVZÑ@[Ô@[ˆ=ØØ×Ò˜uœ|¨TÀuÐ3OÐ3OÐ3OÐPÑQÔQÐQÐQØ—-’-‘/”/ð að aˆÝ�6˜8Ñ$Ô$ð 	a­°´Ð@SÑ)TÔ)Tð 	aØ×!Ò! 6¤=Ð2EÐG^ÐG^Ð"_Ñ`Ô`Ð`øÝ	˜.Ñ	)Ô	)ð ð Øˆˆˆðð ð ñ ô ð ð ð ð ð ð ð øøøð ð ð ð ð ð s   ÃC"Ã"C&Ã)C&Úinputsc              #  ó¨   ‡K  — | j         Šˆfd„}t          j        d„ |D ¦   «         ¦  «        |_        	 || _         dV — ‰| _         dS # ‰| _         w xY w)aü  Temporarily replace `model.forward` with a flat explicit signature derived from `inputs`.

    `torch.export` infers the exported function signature from `model.forward.__signature__`.
    Most transformers models use `**kwargs: Unpack[TransformersKwargs]`, which causes
    `torch.export` to expand the signature into a large `combined_args` bundle that
    mismatches the `dynamic_shapes` dict. This patch replaces the forward with a
    minimal signature containing only the keys present in `inputs`.
    c                 ó   •—  ‰di | ¤ŽS )Nr!   r!   )r#   Úoriginal_forwards    €r9   Ú_flat_forwardz.patch_forward_signature.<locals>._flat_forward°   s   ø€ ØÐÐ)Ð) &Ð)Ð)Ð)rA   c                óZ   — g | ](}t          j        |t           j        j        d ¬¦  «        ‘Œ)S )N)Údefault)ÚinspectÚ	ParameterÚPOSITIONAL_OR_KEYWORD)Ú.0Úks     r9   ú
<listcomp>z+patch_forward_signature.<locals>.<listcomp>´   s1   € ÐeÐeÐeÐYZ�Ô	˜1�gÔ/ÔEÈtÐ	TÑ	TÔ	TÐeÐeÐerA   N)Úforwardr\   Ú	SignatureÚ__signature__)r   rU   rY   rX   s      @r9   r2   r2   ¤   s‡   øè è € ð ”}Ðð*ð *ð *ð *ð *õ #*Ô"3ØeÐeÐ^dÐeÑeÔeñ#ô #€MÔð)Ø%ˆŒØˆˆˆà(ˆŒˆˆøÐ(ˆŒÐ(Ð(Ð(Ð(s   ´A Á	Ar    zQtransformers.models.nllb_moe.modeling_nllb_moe.NllbMoeTop2Router._cast_classifierc                ó   — d„ S )z:Disable classifier dtype cast in nllb-moe (not traceable).c                ó   — d S rE   r!   )r6   r"   r#   s      r9   rH   z(_patch_classifier_cast.<locals>.<lambda>Ñ   s   € ¨€ rA   r!   rL   s    r9   Ú_patch_classifier_castrg   Î   s   € ð .Ð-Ð-rA   z0torch.nn.functional.scaled_dot_product_attentionc                ó*   ‡ ‡‡— ddl mŠmŠ ˆˆ ˆfd„}|S )u4  Route SDPA through the MATH backend on CPU during tracing â€” CPU SDPA's flash/efficient
    paths guard on ``Eq(batch, 1)`` (upstream https://github.com/pytorch/pytorch/issues/180202),
    which trips ``GuardOnDataDependentSymNode`` whenever the batch dim comes from a data-dependent
    op like ``pixel_values[bool_mask]`` (Idefics2/3 and most VLMs). The MATH decomposition has no
    batch-1 dispatch, so the guard never fires. CUDA exports are left alone â€” the GPU kernels
    don't have this guard, and we want the flash/efficient decompositions there.
    r   )Ú
SDPBackendÚsdpa_kernelc                ó¤   •— | j         j        dk    r4 ‰‰j        ¦  «        5   ‰| g|¢R i |¤Žcd d d ¦  «         S # 1 swxY w Y    ‰| g|¢R i |¤ŽS )NÚcpu)Údevicer*   ÚMATH)Úqueryr"   r#   ri   Úoriginalrj   s      €€€r9   Úpatchz_patch_sdpa.<locals>.patchß   s¼   ø€ ØŒ<Ô Ò%Ð%Ø�˜Zœ_Ñ-Ô-ð 8ð 8Ø�x Ð7¨Ð7Ð7Ð7°Ð7Ð7ð8ð 8ð 8ð 8ñ 8ô 8ð 8ð 8ð 8ð 8ð 8ð 8øøøð 8ð 8ð 8ð 8àˆx˜Ð/ Ð/Ð/Ð/¨Ð/Ð/Ð/s   ¢;»?Á?)Útorch.nn.attentionri   rj   )rp   rq   ri   rj   s   ` @@r9   Ú_patch_sdpars   Ô   sI   øøø€ ð ;Ð:Ð:Ð:Ð:Ð:Ð:Ð:ð0ð 0ð 0ð 0ð 0ð 0ð 0ð €LrA   z4transformers.utils.import_utils.is_kernels_availablez'transformers.utils.is_kernels_availablez0transformers.modeling_utils.is_kernels_availablezGtransformers.models.sam3_video.modeling_sam3_video.is_kernels_availablez9transformers.models.mra.modeling_mra.is_kernels_availablez;transformers.models.rwkv.modeling_rwkv.is_kernels_availablez;transformers.models.yoso.modeling_yoso.is_kernels_availablec                ó   — d„ S )uÑ   Force-disable the optional ``kernels`` library during export â€” its kernels
    call into native code that ``torch.export`` cannot trace, and the pure-PyTorch
    fallbacks in each model are always traceable.c                 ó   — dS rK   r!   )r"   r#   s     r9   rH   z-_patch_is_kernels_available.<locals>.<lambda>ù   s   €  5€ rA   r!   rL   s    r9   Ú_patch_is_kernels_availablerv   è   s   € ð" )Ð(Ð(rA   FÚhidden_statesútorch.TensorÚ
cu_seqlensÚrotary_pos_embútorch.Tensor | NoneÚposition_embeddingsú(tuple[torch.Tensor, torch.Tensor] | NoneÚreturns_tupleÚboolc                ó`	  ‡‡— |j         dk    }|r|                     d¦  «        }|j        d         }t          |dk    d¦  «         |j        d         dz
  Š|dd…         |dd…         z
  }	t          |	|	d         k                         ¦   «         d¦  «         t          | d¦  «        �r1t          | d	¦  «        rÅt          | d
¦  «        rµ| j        | j        k    r¥|                      |¦  «         	                    | j        | j        | j        gd¬¦  «        \  }
}}|
 
                    || j        | j        ¦  «        }
| 
                    || j        | j        ¦  «        }| 
                    || j        | j        ¦  «        }�n:|                      |¦  «                             |d| j        d¦  «                             dd¦  «                             d¦  «        \  }
}}nÞt#          | dt#          | dd¦  «        ¦  «        }t#          | dt#          | dd¦  «        ¦  «        }t#          | dt#          | dd¦  «        ¦  «        } ||¦  «         
                    || j        | j        ¦  «        }
 ||¦  «         
                    || j        | j        ¦  «        } ||¦  «         
                    || j        | j        ¦  «        }|�Àt$          j        t)          | ¦  «        j                 j        }t/          |t0          t2          f¦  «        r|\  }} ||
|||¦  «        \  }
}nd ||
                     d¦  «        |¦  «                             d¦  «        }
 ||                     d¦  «        |¦  «                             d¦  «        }|‰z  Šˆˆfd„} ||
¦  «        }
 ||¦  «        } ||¦  «        }t          |
j        d         dk    d¦  «         t          |
j        d         dk    d¦  «         t6          j        j                             |
||d| j        | j         sdn| j!        t#          | d| j        ¦  «        | j        k    ¬¦  «        }|                     dd¦  «                             |d¦  «        }t          | d¦  «        r| j"        n| j#        } ||¦  «        }|r|                     d¦  «        }|r|dfn|S )z¬Export-safe chunked vision/audio attention: reshape segments into a batch dim,
    apply rotary if provided, run one SDPA call, project, and re-emit in the original layout.é   r   z1Chunked vision attention received an empty input.r   Néÿÿÿÿz¹Chunked vision attention requires uniform segment lengths during export. Ensure all images have the same resolution (use do_resize=True in the processor) or pad inputs to a common size.ÚqkvÚq_dimÚkv_dim)ÚdimÚq_projÚqÚk_projr`   Úv_projrG   c                ó\   •— |                       d‰‰f¦  «                             dd¦  «        S )Nr   r   r   )Ú	unflattenÚ	transpose)ÚtÚnum_segmentsÚseg_lens    €€r9   Ú_to_batchedz7_reshaped_vision_attention_forward.<locals>._to_batchedY  s,   ø€ Ø�{Š{˜1˜|¨WÐ5Ñ6Ô6×@Ò@ÀÀAÑFÔFÐFrA   z1Reshaped chunked-vision attention got zero batch.r   z/Reshaped chunked-vision attention got zero seq.Fg        Únum_key_value_heads)Ú	is_causalÚscaleÚ	dropout_pÚ
enable_gqaÚproj)$ÚndimÚsqueezeÚshaper   ÚallrN   r„   r…   rƒ   ÚsplitÚviewÚ	num_headsÚhead_dimr’   Úreshaper�   ÚunbindÚgetattrÚsysrP   r*   r;   Úapply_rotary_pos_emb_visionr'   ÚtupleÚlistÚ	unsqueezer   ÚnnÚ
functionalÚscaled_dot_product_attentionÚscalingÚtrainingÚattention_dropoutr—   Úout_proj)r6   rw   ry   rz   r|   r~   r#   Úneeds_batch_restoreÚ
seq_lengthÚsegment_lengthsÚquery_statesÚ
key_statesÚvalue_statesr‡   r‰   rŠ   r¤   ÚcosÚsinr‘   Úattn_outputr®   r�   r�   s                         @@r9   Ú"_reshaped_vision_attention_forwardr¸     sî  øø€ ð (Ô,°Ò1ÐØð 1Ø%×-Ò-¨aÑ0Ô0ˆàÔ$ QÔ'€JÝØ�aŠØ;ñô ð ð Ô# AÔ&¨Ñ*€Là    ”n z°#°2°#¤Ñ6€OÝØ	˜O¨AÔ.Ò	.×3Ò3Ñ5Ô5ð	*ñô ð õ ˆt�UÑÔñ ]õ �4˜Ñ!Ô!ð 
	¥g¨d°HÑ&=Ô&=ð 
	À$Ä*ÐPTÔP[ÒB[ÐB[Ø59·X²X¸mÑ5LÔ5L×5RÒ5RØ”˜Tœ[¨$¬+Ð6¸Bð 6Sñ 6ô 6Ñ2ˆL˜* lð (×,Ò,¨Z¸¼ÈÌÑWÔWˆLØ#Ÿš¨°TÔ5MÈtÌ}Ñ]Ô]ˆJØ'×,Ò,¨Z¸Ô9QÐSWÔS`ÑaÔaˆL‰Lð —’˜Ñ'Ô'×/Ò/°
¸A¸t¼~ÈrÑRÔR×\Ò\Ð]^Ð`aÑbÔb×iÒiÐjkÑlÔlñ 3ˆL˜* l lõ ˜˜x­°°s¸DÑ)AÔ)AÑBÔBˆÝ˜˜x­°°s¸DÑ)AÔ)AÑBÔBˆÝ˜˜x­°°s¸DÑ)AÔ)AÑBÔBˆØ�v˜mÑ,Ô,×1Ò1°*¸d¼nÈdÌmÑ\Ô\ˆØ�V˜MÑ*Ô*×/Ò/°
¸D¼NÈDÌMÑZÔZˆ
Ø�v˜mÑ,Ô,×1Ò1°*¸d¼nÈdÌmÑ\Ô\ˆàÐ&õ
 '*¤kµ$°t±*´*Ô2GÔ&HÔ&dÐ#ÝÐ)­Eµ4¨=Ñ9Ô9ð 	nà*‰HˆC�Ø'BÐ'BÀ<ÐQ[Ð]`ÐbeÑ'fÔ'fÑ$ˆL˜*˜*ð 7Ð6°|×7MÒ7MÈaÑ7PÔ7PÐReÑfÔf×nÒnÐopÑqÔqˆLØ4Ð4°Z×5IÒ5IÈ!Ñ5LÔ5LÐNaÑbÔb×jÒjÐklÑmÔmˆJà˜LÑ(€GðGð Gð Gð Gð Gð Gð �;˜|Ñ,Ô,€LØ�˜ZÑ(Ô(€JØ�;˜|Ñ,Ô,€Lå˜<Ô-¨aÔ0°AÒ5Ð7jÑkÔkÐkÝ˜<Ô-¨aÔ0°AÒ5Ð7hÑiÔiÐiÝ”(Ô%×BÒBØØØØØŒlØ!œ]ÐF�#�#°Ô0FÝ˜4Ð!6¸¼ÑGÔGÈ4Ì>ÒYð Cñ ô €Kð ×'Ò'¨¨1Ñ-Ô-×5Ò5°jÀ"ÑEÔE€KÝ# D¨&Ñ1Ô1ÐDˆtŒyˆy°t´}€HØ�(˜;Ñ'Ô'€Kàð /Ø!×+Ò+¨AÑ.Ô.ˆà"/Ð@ˆK˜ÐÐ°[Ð@rA   zFtransformers.models.qwen2_vl.modeling_qwen2_vl.VisionAttention.forwardzTtransformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionAttention.forwardzMtransformers.models.qwen3_vl.modeling_qwen3_vl.Qwen3VLVisionAttention.forwardzXtransformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe.Qwen3VLMoeVisionAttention.forwardzKtransformers.models.qwen3_5.modeling_qwen3_5.Qwen3_5VisionAttention.forwardzVtransformers.models.qwen3_5_moe.modeling_qwen3_5_moe.Qwen3_5MoeVisionAttention.forwardz^transformers.models.qwen3_omni_moe.modeling_qwen3_omni_moe.Qwen3OmniMoeVisionAttention.forwardzEtransformers.models.glm4v.modeling_glm4v.Glm4vVisionAttention.forwardzPtransformers.models.glm4v_moe.modeling_glm4v_moe.Glm4vMoeVisionAttention.forwardzJtransformers.models.glm_ocr.modeling_glm_ocr.GlmOcrVisionAttention.forwardzbtransformers.models.ernie4_5_vl_moe.modeling_ernie4_5_vl_moe.Ernie4_5_VLMoeVisionAttention.forwardz_transformers.models.cohere_compass.modeling_cohere_compass.CohereCompassVisionAttention.forwardzRtransformers.models.exaone4_5.modeling_exaone4_5.Exaone4_5_VisionAttention.forwardzYtransformers.models.qwen2_5_omni.modeling_qwen2_5_omni.Qwen2_5OmniVisionAttention.forwardzNtransformers.models.kimi_k25.modeling_kimi_k25.Kimi_K25VisionAttention.forwardzYtransformers.models.muse_glimmer.modeling_muse_glimmer.MuseGlimmerVisionAttention.forwardz[transformers.models.video_llama_3.modeling_video_llama_3.VideoLlama3VisionAttention.forwardzWtransformers.models.paddleocr_vl.modeling_paddleocr_vl.PaddleOCRVisionAttention.forwardzWtransformers.models.minicpmv4_6.modeling_minicpmv4_6.MiniCPMV4_6VisionAttention.forwardzXtransformers.models.qwen2_5_omni.modeling_qwen2_5_omni.Qwen2_5OmniAudioAttention.forwardz]transformers.models.qwen3_omni_moe.modeling_qwen3_omni_moe.Qwen3OmniMoeAudioAttention.forwardzOtransformers.models.qwen3_asr.modeling_qwen3_asr.Qwen3ASRAudioAttention.forwardc                óJ   ‡— t          j        | ¦  «        }d|v pd|v Šˆfd„}|S )zPBind `returns_tuple` once per class by inspecting the original forward's source.zreturn attn_output, attn_weightzreturn attn_output, Nonec                ó(   •— t          | g|¢R d‰i|¤ŽS )Nr~   )r¸   )r6   r"   r#   r~   s      €r9   rb   z0_patch_chunked_vision_attention.<locals>.forward�  s'   ø€ Ý1°$Ðe¸ÐeÐeÐeÈ]ÐeÐ^dÐeÐeÐerA   )r\   Ú	getsource)rp   Úsrcrb   r~   s      @r9   Ú_patch_chunked_vision_attentionr½   w  sQ   ø€ õF Ô
˜HÑ
%Ô
%€CØ5¸Ð<ÐaÐ@ZÐ^aÐ@a€Mðfð fð fð fð fð €NrA   Úclsr*   r   Ústrc                ó$   — | j         › d| j        › �S )Nú:)r;   r<   )r¾   s    r9   Ú_class_to_pathrÂ   ­  s   € ØŒnÐ1Ð1˜sÔ/Ð1Ð1Ð1rA   Úpathc                ó°   — |                       dd¦  «        \  }}t          j        |¦  «        }|                      d¦  «        D ]}t          ||¦  «        }Œ|S )NrÁ   r   ú.)rœ   Ú	importlibÚimport_moduler¢   )rÃ   Úmodule_nameÚqualnameÚobjÚparts        r9   Ú_path_to_classrÌ   ±  s[   € Ø ŸJšJ s¨AÑ.Ô.Ñ€K�Ý
Ô
! +Ñ
.Ô
.€CØ—’˜sÑ#Ô#ð !ð !ˆÝ�c˜4Ñ Ô ˆˆØ€JrA   Úsymr   c                óð   — | j         }t          | t          j        ¦  «        r|                     ¦   «         S t          | t          j        ¦  «        r|                     ¦   «         S |                     ¦   «         S )zÞThe concrete value a ``Sym*`` has already specialized to, or ``None`` if it's still dynamic.
    Each ``Sym*`` type has its own accessor (``maybe_as_bool``/``maybe_as_float``/``maybe_as_int``);
    there is no generic one.)Únoder'   r   ÚSymBoolÚmaybe_as_boolÚSymFloatÚmaybe_as_floatÚmaybe_as_int)rÍ   rÏ   s     r9   Ú_maybe_sym_constantrÕ   ¹  si   € ð Œ8€DÝ�#•u”}Ñ%Ô%ð $Ø×!Ò!Ñ#Ô#Ð#Ý�#•u”~Ñ&Ô&ð %Ø×"Ò"Ñ$Ô$Ð$Ø×ÒÑÔÐrA   rÊ   Útensorsr¦   c                óà  ‡— | �)t          | ¦  «        t          t          t          t          fv r| S t          | ¦  «        t
          u rˆfd„| D ¦   «         S t          | ¦  «        t          u r ˆfd„|                      ¦   «         D ¦   «         S t          | t          j
        ¦  «        r)t          ‰¦  «        }‰                     | ¦  «         d|dœS t          | t          j        ¦  «        rdt          | ¦  «        dœS t          | t          j        ¦  «        rdt	          | ¦  «        d	œS t          | t          j        ¦  «        r%d
t	          | ¦  «                             d¦  «        dœS t          | t          j        ¦  «        r%dt	          | ¦  «                             d¦  «        dœS t          | t          j        t          j        t          j        f¦  «        r<t+          | ¦  «        }|�|S t          ‰¦  «        }‰                     | ¦  «         d|dœS t          | t           ¦  «        rdt-          | ¦  «        dœS t          | ¦  «        }t          | t          ¦  «        r1dt-          |¦  «        ˆfd„|                      ¦   «         D ¦   «         dœS t          | t.          t
          t0          t2          f¦  «        rdt-          |¦  «        ˆfd„| D ¦   «         dœS t          | t4          j        ¦  «        rt9          d¦  «        ‚t;          | d¦  «        r@ˆfd„t=          | ¦  «                             ¦   «         D ¦   «         }dt-          |¦  «        |dœS t9          dt          | ¦  «        j        › d�¦  «        ‚)z]Single-pass: recursively build a JSON-native context while collecting tensors into `tensors`.Nc                ó0   •— g | ]}t          |‰¦  «        ‘ŒS r!   ©Ú_flatten_to_context©r_   ÚirÖ   s     €r9   ra   z'_flatten_to_context.<locals>.<listcomp>Ë  s$   ø€ Ð=Ð=Ð=°AÕ# A wÑ/Ô/Ð=Ð=Ð=rA   c                ó8   •— i | ]\  }}|t          |‰¦  «        “ŒS r!   rÙ   ©r_   r`   rG   rÖ   s      €r9   ú
<dictcomp>z'_flatten_to_context.<locals>.<dictcomp>Í  s*   ø€ ÐKÐKÐK±t°q¸!�Õ& q¨'Ñ2Ô2ÐKÐKÐKrA   Útensor)Ú_trÜ   Úsize)rá   rG   rm   )rá   ÚsÚdtypeztorch.)rá   ÚnÚlayoutrÍ   r*   )rá   ÚpÚmapc                ó8   •— i | ]\  }}|t          |‰¦  «        “ŒS r!   rÙ   rÞ   s      €r9   rß   z'_flatten_to_context.<locals>.<dictcomp>ö  s*   ø€ ÐMÐMÐM¹¸¸A�!Õ(¨¨GÑ4Ô4ÐMÐMÐMrA   )rá   rç   rG   Úseqc                ó0   •— g | ]}t          |‰¦  «        ‘ŒS r!   rÙ   rÛ   s     €r9   ra   z'_flatten_to_context.<locals>.<listcomp>ü  s$   ø€ Ð?Ð?Ð?°aÕ% a¨Ñ1Ô1Ð?Ð?Ð?rA   z0Cannot flatten a bound method for pytree contextÚ__dict__c                ó8   •— i | ]\  }}|t          |‰¦  «        “ŒS r!   rÙ   rÞ   s      €r9   rß   z'_flatten_to_context.<locals>.<dictcomp>  s*   ø€ ÐRÐRÐR¹¸¸1�Õ'¨¨7Ñ3Ô3ÐRÐRÐRrA   rÊ   )rá   rç   rã   zCannot flatten z for pytree context) r*   r   ÚintÚfloatr¿   r¦   r(   rM   r'   r   ÚTensorÚlenrO   ÚSizerm   rä   Úremoveprefixræ   ÚSymIntrÒ   rÐ   rÕ   rÂ   r¥   ÚsetÚ	frozensetÚtypesÚ
MethodTyper)   rN   Úvarsr:   )rÊ   rÖ   ÚidxÚconstr¾   Ústates    `    r9   rÚ   rÚ   Å  s_  ø€ ð €{•d˜3‘i”i¥D­#­uµcÐ#:Ð:Ð:Øˆ
ÝˆC�y„y•DÐÐØ=Ð=Ð=Ð=¸Ð=Ñ=Ô=Ð=ÝˆC�y„y•DÐÐØKÐKÐKÐK¸s¿yºy¹{¼{ÐKÑKÔKÐKõ �#•u”|Ñ$Ô$ð *Ý�'‰lŒlˆØ�Š�sÑÔÐØ SÐ)Ð)Ð)Ý�#•u”zÑ"Ô"ð .Ø¥4¨¡9¤9Ð-Ð-Ð-Ý�#•u”|Ñ$Ô$ð /Ø¥S¨¡X¤XÐ.Ð.Ð.Ý�#•u”{Ñ#Ô#ð EØ¥C¨¡H¤H×$9Ò$9¸(Ñ$CÔ$CÐDÐDÐDÝ�#•u”|Ñ$Ô$ð FØ¥S¨¡X¤X×%:Ò%:¸8Ñ%DÔ%DÐEÐEÐEÝ�#�œ¥e¤nµe´mÐDÑEÔEð 'õ $ CÑ(Ô(ˆØÐØˆLÝ�'‰lŒlˆØ�Š�sÑÔÐØ #Ð&Ð&Ð&õ �#•tÑÔð 8Ø¥>°#Ñ#6Ô#6Ð7Ð7Ð7õ ˆs‰)Œ)€CÝ�#•tÑÔð 
àÝ Ñ$Ô$ØMÐMÐMÐMÀÇÂÁÄÐMÑMÔMð
ð 
ð 	
õ
 �#��t¥S­)Ð4Ñ5Ô5ð 
àÝ Ñ$Ô$Ø?Ð?Ð?Ð?¸3Ð?Ñ?Ô?ð
ð 
ð 	
õ
 �#•uÔ'Ñ(Ô(ð Lõ
 ÐJÑKÔKÐKÝˆs�JÑÔð CØRÐRÐRÐRÅÀSÁ	Ä	ÇÂÑ@QÔ@QÐRÑRÔRˆØ¥.°Ñ"5Ô"5¸EÐBÐBÐBå
ÐM¥d¨3¡i¤iÔ&8ÐMÐMÐMÑ
NÔ
NÐNrA   Úctxc                óâ  ‡— | �)t          | ¦  «        t          t          t          t          fv r| S t          | ¦  «        t
          u rˆfd„| D ¦   «         S t          | ¦  «        t          u r$d| vr ˆfd„|                      ¦   «         D ¦   «         S | d         }|dk    r‰| d                  S |dk    rt          t          | d         ¦  «        S |d	k    rt          t          | d         ¦  «        S |d
k    rt          j
        | d         ¦  «        S |dk    rt          j        | d         ¦  «        S |dk    r‰| d                  S |dk    rt          | d         ¦  «        S |dk    rDt          | d         ¦  «        } |ˆfd„| d                              ¦   «         D ¦   «         ¦  «        S |dk    rJt          | d         ¦  «        }ˆfd„| d         D ¦   «         }	  ||¦  «        S # t          $ r  ||Ž cY S w xY w|dk    rdt          | d         ¦  «        }|                     |¦  «        }| d                              ¦   «         D ]\  }}t          |‰¦  «        |j        |<   Œ|S t          d|›d�¦  «        ‚)zVReconstruct an object from its JSON-native context, substituting tensor index markers.Nc                ó0   •— g | ]}t          |‰¦  «        ‘ŒS r!   ©Ú_unflatten_from_contextrÛ   s     €r9   ra   z+_unflatten_from_context.<locals>.<listcomp>  s$   ø€ ÐAÐAÐA¸Õ'¨¨7Ñ3Ô3ÐAÐAÐArA   rá   c                ó8   •— i | ]\  }}|t          |‰¦  «        “ŒS r!   r   rÞ   s      €r9   rß   z+_unflatten_from_context.<locals>.<dictcomp>  s*   ø€ ÐOÐOÐO¹4¸1¸a�Õ*¨1¨gÑ6Ô6ÐOÐOÐOrA   rà   rÜ   ræ   rå   rä   rm   rã   râ   rG   rÍ   r*   rç   rè   c                ó8   •— i | ]\  }}|t          |‰¦  «        “ŒS r!   r   rÞ   s      €r9   rß   z+_unflatten_from_context.<locals>.<dictcomp>+  s*   ø€ ÐXÐXÐX¹t¸qÀ!�AÕ.¨q°'Ñ:Ô:ÐXÐXÐXrA   rê   c                ó0   •— g | ]}t          |‰¦  «        ‘ŒS r!   r   rÛ   s     €r9   ra   z+_unflatten_from_context.<locals>.<listcomp>.  s$   ø€ ÐGÐGÐG¸Õ(¨¨GÑ4Ô4ÐGÐGÐGrA   rÊ   zUnknown tag z in pytree context)r*   r   rî   rï   r¿   r¦   r(   rM   r¢   r   rm   rò   rÌ   r)   Ú__new__r  rì   )rý   rÖ   rŽ   r¾   rM   Úinstancer`   rG   s    `      r9   r  r    s§  ø€ ð €{•d˜3‘i”i¥D­#­uµcÐ#:Ð:Ð:Øˆ
ÝˆC�y„y•DÐÐØAÐAÐAÐA¸SÐAÑAÔAÐAÝˆC�y„y•DÐÐ˜T¨˜_˜_ØOÐOÐOÐOÀ3Ç9Â9Á;Ä;ÐOÑOÔOÐOð 	ˆDŒ	€AØˆH‚}€}Ø�s˜3”xÔ Ð ØˆH‚}€}Ý•u˜c #œhÑ'Ô'Ð'ØˆG‚|€|Ý•u˜c #œhÑ'Ô'Ð'ØˆH‚}€}ÝŒ|˜C œHÑ%Ô%Ð%ØˆF‚{€{ÝŒz˜#˜cœ(Ñ#Ô#Ð#ØˆE‚z€zØ�s˜3”xÔ Ð ð 	ˆF‚{€{Ý˜c #œhÑ'Ô'Ð'ð 	ˆE‚z€zÝ˜S œXÑ&Ô&ˆØˆsÐXÐXÐXÐXÀsÈ3ÄxÇ~Â~ÑGWÔGWÐXÑXÔXÑYÔYÐYØˆE‚z€zÝ˜S œXÑ&Ô&ˆØGÐGÐGÐG¸cÀ#¼hÐGÑGÔGˆð	Ø�3�u‘:”:ÐøÝð 	ð 	ð 	Ø�3˜�;ÐÐÐð	øøøàˆE‚z€zÝ˜S œXÑ&Ô&ˆØ—;’;˜sÑ#Ô#ˆØ˜”H—N’NÑ$Ô$ð 	Gð 	G‰DˆAˆqÝ#:¸1¸gÑ#FÔ#FˆHÔ˜aÑ Ð Øˆå
Ð: 1Ð:Ð:Ð:Ñ
;Ô
;Ð;s   Ç
G ÇG1Ç0G1útuple[list, Any]c                ó.   — g }t          | |¦  «        }||fS rE   rÙ   )rÊ   rÖ   Úcontexts      r9   Ú_pytree_flattenr
  =  s!   € Ø€GÝ! # wÑ/Ô/€GØ�GÐÐrA   c                ó\   — t          | ¦  «        \  }}d„ t          |¦  «        D ¦   «         |fS )Nc                ó`   — g | ]+\  }}t           j        j                             |¦  «        |f‘Œ,S r!   )r   ÚutilsÚ_pytreeÚSequenceKey)r_   rÜ   Úleafs      r9   ra   z-_pytree_flatten_with_keys.<locals>.<listcomp>E  s4   € ÐXÐXÐX¹7¸1¸d�UŒ[Ô ×,Ò,¨QÑ/Ô/°Ð6ÐXÐXÐXrA   )r
  Ú	enumerate)rÊ   Úleavesr	  s      r9   Ú_pytree_flatten_with_keysr  C  s6   € Ý% cÑ*Ô*�O€FˆGØXÐXÅiÐPVÑFWÔFWÐXÑXÔXÐZaÐaÐarA   r	  c                ó<   — t          |t          | ¦  «        ¦  «        S rE   )r  r¦   )Úvaluesr	  s     r9   Ú_pytree_unflattenr  H  s   € Ý" 7­D°©L¬LÑ9Ô9Ð9rA   Ú
object_clsc                óæ   — 	 t           j        j                             | t          t
          t          | ¦  «        t          ¬¦  «         dS # t          $ r}dt          |¦  «        vr‚ Y d}~dS d}~ww xY w)zÄRegister a single class (e.g. a `Cache` subclass like `StaticCache`) as a torch.export pytree
    node, so `torch.export.load` can unflatten it as a graph input without needing the original model.)Úserialized_type_nameÚflatten_with_keys_fnz!already registered as pytree nodeN)
r   r  r  Úregister_pytree_noder
  r  rÂ   r  Ú
ValueErrorr¿   )r  Úes     r9   r  r  L  s“   € ð
ÝŒÔ×0Ò0ØÝÝÝ!/°
Ñ!;Ô!;Ý!:ð 	1ñ 	
ô 	
ð 	
ð 	
ð 	
øõ ð ð ð Ø.µc¸!±f´fÐ<Ð<Øð =Ð<Ð<Ð<Ð<Ð<øøøøðøøøs   ‚AA	 Á	
A0ÁA+Á+A0c              #  ój   K  — |                       ¦   «         D ]}|V — t          |¦  «        E d {V —† Œd S rE   )Ú__subclasses__Ú_iter_subclasses)r¾   Úsubclasss     r9   r   r   \  sW   è è € Ø×&Ò&Ñ(Ô(ð .ð .ˆØˆˆˆÝ# HÑ-Ô-Ð-Ð-Ð-Ð-Ð-Ð-Ð-Ð-ð.ð .rA   c                ó  — t          t          ¦  «        D ]}t          |¦  «         Œt          j        t          j        | ¦  «        ¦  «        D ]l\  }}t          j        |¦  «        rS|j        | j        j        k    r>|j	         
                    d¦  «        r$t          |t          ¦  «        st          |¦  «         Œmt          ¦   «         r<t          | t          ¦  «        r)| j        j        dk    rddlm} t          |¦  «         dS dS dS dS )zCRegister all relevant cache types as pytree nodes for torch.export.r   Ú
layoutlmv2r   )Ú	ImageListN)r   r   r  r\   Ú
getmembersÚ	getmoduleÚisclassr;   Ú	__class__r:   ÚendswithÚ
issubclassr	   r'   r   r   Ú
model_typeÚ detectron2.structures.image_listr$  )r   Ú
cache_typeÚ_rÊ   r$  s        r9   r/   r/   b  s'  € õ '¥uÑ-Ô-ð )ð )ˆ
Ý˜ZÑ(Ô(Ð(Ð(õ Ô$¥WÔ%6°uÑ%=Ô%=Ñ>Ô>ð &ð &‰ˆˆ3åŒO˜CÑ Ô ð	&à” %¤/Ô"<Ò<Ð<Ø”×%Ò% gÑ.Ô.ð =å˜s¥EÑ*Ô*ð =õ ! Ñ%Ô%Ð%øõ Ñ Ô ð (¥Z°µÑ%GÔ%Gð (ÈEÌLÔLcÐgsÒLsÐLsØ>Ð>Ð>Ð>Ð>Ð>å˜YÑ'Ô'Ð'Ð'Ð'ð(ð (ð (ð (ÐLsÐLsrA   rà   údict[int, torch.export.Dim]c                óž   — t                                t          |                      ¦   «         ¦  «        t          j        j        j        ¦  «        S )zPGenerate a dynamic shape with all dimensions set to Dim.AUTO for a given tensor.)r(   ÚfromkeysÚranger†   r   r3   ÚDimÚAUTO)rà   s    r9   Ú_auto_dynamic_shaper5  ~  s.   € å�=Š=�˜vŸzšz™|œ|Ñ,Ô,­e¬lÔ.>Ô.CÑDÔDÐDrA   c                ó  — t          | t          j        ¦  «        rt          | ¦  «        S | �(t          | t          t
          t          t          f¦  «        rdS t          | d¦  «        r!t          | ¦  «        \  }}t          |¦  «        S t          | ¦  «        t          t          t          t          fv r" t          | ¦  «        d„ | D ¦   «         ¦  «        S t          | ¦  «        t           u rd„ |                      ¦   «         D ¦   «         S dS )uÜ  Recursively build dynamic shapes for any input value.

    - Tensors â†’ per-dimension Dim.AUTO spec.
    - Scalars / None â†’ None (no dynamic dims).
    - Objects with ``__dict__`` (ModelOutput, Cache, â€¦) â†’ flat list of leaf specs,
      matching the ``TreeSpec(list, â€¦)`` that torch.export produces for these types.
    - Lists / tuples â†’ same container type, recursed element-wise.
    - Plain dicts â†’ recursed dict of specs.
    - Everything else â†’ None.
    Nrì   c              3  ó4   K  — | ]}t          |¦  «        V — Œd S rE   ©r.   )r_   rG   s     r9   ú	<genexpr>z*get_auto_dynamic_shapes.<locals>.<genexpr>–  s+   è è € ÐGÐG¸1Õ3°AÑ6Ô6ÐGÐGÐGÐGÐGÐGrA   c                ó4   — i | ]\  }}|t          |¦  «        “ŒS r!   r8  )r_   r`   rG   s      r9   rß   z+get_auto_dynamic_shapes.<locals>.<dictcomp>˜  s'   € ÐIÐIÐI±$°!°Q�Õ*¨1Ñ-Ô-ÐIÐIÐIrA   )r'   r   rð   r5  rî   rï   r   r¿   rN   r
  r.   r*   r¦   r¥   rõ   rö   r(   rM   )rU   r  r.  s      r9   r.   r.   ƒ  sé   € õ �&�%œ,Ñ'Ô'ð +Ý" 6Ñ*Ô*Ð*Ø€~� F­Sµ%½½sÐ,CÑDÔD€~ØˆtÝˆv�zÑ"Ô"ð /Ý# FÑ+Ô+‰	ˆ�Ý& vÑ.Ô.Ð.ÝˆF�|„|��e¥S­)Ð4Ð4Ð4Ø�t�F‰|Œ|ÐGÐGÀÐGÑGÔGÑGÔGÐGÝˆF�|„|•tÐÐØIÐI¸&¿,º,¹.¼.ÐIÑIÔIÐIØˆ4rA   )Ú"cached_rotary_positional_embeddingÚcached_sequence_lengthútorch.nn.Modulec              #  óö   K  — d„ |                       ¦   «         D ¦   «         }|D ]\  }}}t          ||d¦  «         Œ	 dV — |D ]\  }}}t          |||¦  «         ŒdS # |D ]\  }}}t          |||¦  «         Œw xY w)zÊSave each `_STATEFUL_CACHE_ATTRS` value, null it for the trace, restore on exit.

    FakeTensors that `torch.export` plants into these attributes during the trace are
    discarded by the restore.
    c           	     óh   — g | ]/}t           D ]%}t          ||¦  «        ¯||t          ||¦  «        f‘Œ&Œ0S r!   )Ú_STATEFUL_CACHE_ATTRSrN   r¢   )r_   rT   Úattrs      r9   ra   z%reset_model_state.<locals>.<listcomp>´  s`   € ð ð ð àÝ)ðð ð Ý�6˜4Ñ Ô ð	Ø	�•w˜v tÑ,Ô,Ð-ðð ð ð rA   N)rP   Úsetattr)r   Ú	originalsrT   rA  r.  rp   s         r9   r0   r0   ­  sÎ   è è € ðð à—m’m‘o”oðñ ô €Ið %ð $ð $‰ˆ��aÝ�˜˜dÑ#Ô#Ð#Ð#ð,Øˆˆˆà&/ð 	,ð 	,Ñ"ˆF�D˜(Ý�F˜D (Ñ+Ô+Ð+Ð+ð	,ð 	,ø ið 	,ð 	,Ñ"ˆF�D˜(Ý�F˜D (Ñ+Ô+Ð+Ð+ð	,øøøs   ¼A ÁA8)r   r   r7   rB   )r   r   rU   rB   )NNF)
rw   rx   ry   rx   rz   r{   r|   r}   r~   r   )r¾   r*   r   r¿   )rÃ   r¿   r   r*   )rÍ   r   r   r   )rÊ   r   rÖ   r¦   r   r   )rý   r   rÖ   r¦   r   r   )rÊ   r   r   r  )rÊ   r   )r	  r   r   r   )r  r*   )r¾   r*   )r   r   )rà   rx   r   r/  )rU   r   r   r   )r   r=  )=r=   Ú
__future__r   r4   rÆ   r\   r£   r÷   Úcollections.abcr   Ú
contextlibr   Útypingr   r  r   Úutils.import_utilsr	   r
   r   Úbaser   Úconfigsr   r   r   r   r   r   Útorch.exportr   Úcache_utilsr   Úmodeling_utilsr   Ú
get_loggerÚ__file__r,   r   r1   r2   rg   rs   rv   r¸   r½   rÂ   rÌ   rÕ   rÚ   r  r
  r  r  r  r   r/   r5  r.   r@  r0   r!   rA   r9   ú<module>rP     sU  ðð ð ð0 #Ð "Ð "Ð "Ð "Ð "à €€€Ø Ð Ð Ð Ø €€€Ø 
€
€
€
Ø €€€Ø *Ð *Ð *Ð *Ð *Ð *Ø %Ð %Ð %Ð %Ð %Ð %Ø Ð Ð Ð Ð Ð à Ð Ð Ð Ð Ð Ø dÐ dÐ dÐ dÐ dÐ dÐ dÐ dÐ dÐ dØ Ð Ð Ð Ð Ð Ø !Ð !Ð !Ð !Ð !Ð !Ø VÐ VÐ VÐ VÐ VÐ VÐ VÐ VÐ VÐ VÐ VÐ Vð ÐÑÔð 1Ø€L€L€LØ,Ð,Ð,Ð,Ð,Ð,à#Ð#Ð#Ð#Ð#Ð#Ø0Ð0Ð0Ð0Ð0Ð0ð 
ˆÔ	˜HÑ	%Ô	%€ð; ð ; ð ; ð ; ð ; �Zñ ; ô ; ð ; ðL ðð ð ñ „ðð0 ð)ð )ð )ñ „ð)ðR €�ÐmÑnÔnð.ð .ñ oÔnð.ð
 €�ÐLÑMÔMðð ñ NÔMðð& €Øà:Ø-ð 7ØMØ?ØAØAñô ð)ð )ñô ð)ð< +/ØDHØðeAð eAð eAð eAð eAðP €ØàLØZØSØ^ØQØ\ØdØKØVØPØhàeàXà_àTØ_àaØ]à]à^ØcØUñ? ô  ðBð ñC ô  ðBð*2ð 2ð 2ð 2ðð ð ð ð	ð 	ð 	ð 	ðCOð COð COð COðL/<ð /<ð /<ð /<ðdð ð ð ðbð bð bð bð
:ð :ð :ð :ðð ð ð ð .ð .ð .ð .ð(ð (ð (ð (ð8Eð Eð Eð Eð
ð ð ð ðHÐ ð ð,ð ,ð ,ñ „ð,ð ,ð ,rA   