+ "ݳj^Rt^RIHtHtHt^RIt^RIHt^RIH t ^RI H t ^RI H t ^RIHt] !4tRRlt!R R 4t!R R ]4t!R R]4t!RR]4tR#)z MoE (Mixture of Experts) lifecycle hooks system. This module provides a base class for model-specific MoE lifecycle hooks that allow customization of MoE forward passes and routing logic during quantization. )AnyDictOptionalN) StopForward) get_device) setup_logger)move_toczV^8dQhR\PR\R\\P,/#)parent relative_pathreturn)nnModulestrr)formats"_/home/luca/screen-checking/.venv/lib/python3.14/site-packages/gptqmodel/models/moe_lifecycle.py __annotate__rs1!!!3!8TVT]T]K^!c vV'gV#VPR4pTp\V4FwrERPVRV^,4p\W54'd\ W54pK>\VR4'd,VP 4'dV\ V4,pK{\V4Uu.uFqPR4'dKVNK upR,p \RV R V R \V4P R V 24h V# \\3dp\RT RT 24hRp?ii;iuupi) a Get a submodule from a parent module by relative path. Args: parent: The parent module (e.g., a layer replica) relative_path: Dot-separated path to the submodule (e.g., 'mlp.experts.0.gate_proj') Returns: The submodule if found, None otherwise .N __getitem__z[MoE PATH] Failed indexing 'z': _:NNz[MoE PATH] Failed at 'z' (z ), current=z, attrs=)split enumeratejoinhasattrgetattrisdigitint IndexErrorKeyError ValueErrordir startswithtype__name__) r r partscurrentipart path_so_fareaattrss && r_get_module_by_relative_pathr0s+     $EGU#hhuVa!e}- 7 ! !g,G Wm , , N!#d), !$G F 1LL4EQQ FsKE(c+k$w-J`J`Iaaijoipqs s$ N) N #?vS!LMM NGs$D 1D6D6 D3D..D3ca]tRt^?toRtR.tRR.tV3RlRltRV3RlR lltV3R lR lt V3R lR lt V3RlRlt V3RlRlt V3RlRlt RV3RlRlltRtVtR#)MoELifecycleHooksaA Base class for model-specific MoE lifecycle hooks. This class provides customization points for MoE-specific operations during quantization. Subclasses should override these methods to provide model-specific implementations. Main use case: Forward whole calibration dataset to all experts instead of only routed ones. expertsshared_experts shared_expertcd<V^8dQhRS[PRS[RS[S[P,/#)r layer_module model_classr rrr&r)r __classdict__s"rrMoELifecycleHooks.__annotate__Os2"))$8TVT]T]K^rc VP4pVf%\PRVP 24R#\ W^,R4pV#)a Extract the MoE block from a layer module using the :moe flag from module_tree. Args: layer_module: The layer module (e.g., DecoderLayer) model_class: The model class (to access module_tree) Returns: The MoE block module, or None if not found Example: For GLM-4, module_tree has "mlp:moe", so this returns layer_module.mlp For MiniMax-M2, module_tree has "block_sparse_moe:moe", so this returns layer_module.block_sparse_moe Nz&No :moe flag found in module_tree for )get_moe_module_namelogerrorr'r)selfr7r8moe_module_name moe_blocks&&& r get_moe_blockMoELifecycleHooks.get_moe_blockOsO&99;  " II>{?S?S>TU VL!*!>5(LL*} !>!>E)LL+( rc<V^8dQhRS[PRS[PRS[RS[S[S[3,RS[S[S[,,RS[ RS[ RS[R S[S[,R S[P/ #) r rB hidden_states processorsubsetrdoriginal_forwardr8 module_looperrer rrtorchTensorrrrrrhcallabler&)rr:s"rrr;s$2$2yy$2!<<$2 $2 cN $2 #+49"5 $2'$2$2$2'sm$2 $2rc  l\PR\V4P R24V!V3/V B#)a_ Forward hidden states to all experts, bypassing routing. Subclasses should implement this with model-specific logic. Args: moe_block: The MoE block module hidden_states: Input tensor processor: The quantization processor subset: Dict[str, NamedModule] being calibrated original_forward: Original forward function model_class: The model class moe_block_prefix: Optional prefix for MoE block modules (optimized parameter) module_looper: ModuleLooper instance for TLS-based hooks pausing **kwargs: Additional arguments Returns: Output tensor from routed forward pass z+forward_to_all_experts not implemented for zY. Falling back to normal forward. Override this method for model-specific implementation.)r>r?r&r') r@rBrsrtrurdrvr8rwrekwargss &&&&&&&&&&,rforward_to_all_experts(MoELifecycleHooks.forward_to_all_expertss?> 9$y/:R:R9ST1 2 1&11r)N)r' __module__ __qualname____firstlineno____doc__rVr[rCrHrPrMr\rarpr~__static_attributes____classdictcell__r:s@rr2r2?sw$"2? C4 = =,**.  D$2$2$2rr2ca]tRtRtoRtRtRtRtR V3RlRlltRt V3RlRlt RV3R lR llt V3R lt R t VtR#)!ExpertProjectionMoELifecycleHooksia Base MoE lifecycle hooks for expert architectures with 3 projections. This class extracts the common logic and allows subclasses to configure the attribute names used for the three projections. Expert forward pattern: output = down_proj(act_fn(gate_proj(x)) * up_proj(x)) Subclasses should set: - gate_proj_name: Name of gate/w1 projection - up_proj_name: Name of up/w3 projection - down_proj_name: Name of down/w2 projection Nc,<V^8dQhRS[RS[RS[/#r gate_proj_name up_proj_namedown_proj_namer)rr:s"rr.ExpertProjectionMoELifecycleHooks.__annotate__+s#s]`rc VeWnVeW nVeW0n\VPVPVP.4'g3\ RVP RVP RVP 24hR#)a Initialize with custom projection names. Args: gate_proj_name: Name of gate projection (e.g., "gate_proj" or "w1") up_proj_name: Name of up projection (e.g., "up_proj" or "w3") down_proj_name: Name of down projection (e.g., "down_proj" or "w2") Nz]Projection names must be set either as class attributes or constructor parameters. Got: gate=z, up=z, down=)rrrallr#)r@rrrs&&&&r__init__*ExpertProjectionMoELifecycleHooks.__init__+s  %"0   # ,   %"0 D''):):DRV R2V9gKKVPRV R24^,u# R#)z Extract moe_block_prefix from subset keys. Args: subset: Dict[str, NamedModule] being calibrated moe_block: The MoE block module Returns: The moe_block_prefix string, or None if not found Nr)rMr\keysr)r@rurBrfrgros&&& r_extract_moe_block_prefix;ExpertProjectionMoELifecycleHooks._extract_moe_block_prefixLs 88C"&"E"Ei"P;;=C q):(;1%=%Dyy1%6$7q!9:1==&&Q/F.Gq+IS+Pyy1%<$=Q!?@CC ! rc<V^8dQhRS[PRS[PRS[RS[S[S[3,RS[S[S[,,RS[ RS[ RS[R S[S[,R S[S[P,R S[P/ #) r rBrsrtrurdrvr8rwrereplica_moduler rx)rr:s"rrresVVyyV!<<V V cN V #+49"5 V'VVV'smV%RYY/V Vrc aaaaa a aaaaaaaa V'd V'g#Rp \PV 4\V 4hS f V!S3/V B#\VRR4p \ V 4'd V !S S4^oRo SP SP SP.pRV V3RlloSPSV4oSPSV4pSPS4oSPS4oRoVe/S'd'S 'dVFpS RS RV 2pVS9gKRoM RoSe\SR 4'd\S4^8dS'dS 'dzS RS R2pSP4F]pVPV4'gKV\V4RP!R4p\V4^8gKKV^,V9gK[RoM VVVVVV 3R lpVVVVVVV VV V3 R lpSP#T;'g.S SSR 7pS'dR V9dVP%R 4S'dRV9dVP%R4VF#pVR 8Xd V!4KVR8XgKV!4K% S 'd \'4hS^8d1VP)VR4V!S3/V BpVP)VR4V#V!S3/V BpV# TP)TR4i;i)a Forward to all experts using configurable projection names. Args: moe_block: The MoE block module hidden_states: Input tensor processor: The quantization processor (needed for hook pausing) subset: Dict[str, NamedModule] containing modules currently being calibrated original_forward: Original forward function to call for final output model_class: The model class (to access module_tree) replica_module: Optional replica of the layer module. When provided, modules are looked up from the replica instead of using subset directly. This is needed for multi-GPU parallel execution where the replica is on a different device than the original modules in subset. **kwargs: Additional arguments (attention_mask, etc.) This implementation replays shared-expert and routed-expert paths in the order they appear in the subset/module tree, then calls the original routed forward for the final output. z%Missing processor or original_forwardNrecord_moe_root_input_featureFc$V^8dQhR\/#)r ror)rs"rrNExpertProjectionMoELifecycleHooks.forward_to_all_experts..__annotate__s ! !S !rcr<Se!\SV4pVeV#\RV 24hSPV4pV#)a5 Get the callable module for a given subset key. When replica_module is provided, resolves the module from the replica using the key as a relative path. This ensures forward passes happen on the correct device for multi-GPU execution. Falls back to subset[key] when replica is not provided or lookup fails. z.[MoE DEBUG] replica_submodule is None for key=)r0r#get)roreplica_submodule subset_modulerrus& rget_callable_moduleUExpertProjectionMoELifecycleHooks.forward_to_all_experts..get_callable_modulesO)$@QT$U!$0,,$'UVYUZ%[\\#JJsOM rrT__iter__c<S'd S'gR#\SS4p\V4pV!\SV44S^, oR# \dRoR#i;i)NT)rrrr)shared_expert_moduleshared_expert_device expert_counthas_shared_expertsrsrBrgstop_forward_raiseds rrun_shared_expertsTExpertProjectionMoELifecycleHooks.forward_to_all_experts..run_shared_expertssZ%-D +'.y:Q'R$'12F'G$$W].run_routed_expertss%#~'=W^]gEhEhqB  "a'#0#8#8]=P=PQS=T#U #0 &/&?" ./q1B0C1ZLPQRVReReQfg,-Q/@.A:,aPTPaPaObc./q1B0C1ZLPQRVReReQfg6)fF.BxW]G]")&$2E2Et"L?N?Z ? ;`jkq`r &'7G  %6)&-fd6I6I&J $+FD4E4E$F #.|#<!*r?r#rr{rrrrPrar\rMrlenrr%rrprkr_set_processor_hooks_paused)!r@rBrsrtrurdrvr8rwrerr} error_msg root_recorder proj_namesr`rNrormr(rrexecution_order group_nameresultrrfrOrrrrgrs!fff&f&&&&ff, @@@@@@@@rr~8ExpertProjectionMoELifecycleHooks.forward_to_all_expertsesD 0?I IIi Y' '  #$MQ>QR  ! !000KH $ > >y+ V"&"E"E # 88C# ,1HM]")*!,C+DAdVL&=)-& #!  %'.**M*MRUS "#S$(9>N/02C1DAFM{{}>>-00M 2 34::3?E5zQ58z+A+/( % + +6 %6 %p99!5!;!;-/$; :  (/"A  " "8 ,   @  " "9 -)JX%"$y("$ * -  !   5 5i F L)-B6B99)UK  &m>v>F 99)UKs  KKc><V^8dQh/S[;R&S[;R&S[;R&#rr)rr:s"rrrs, !"#$%r)rrr)NNN)NN)r'rrrrrrrrrrr~__annotate_func__rrrs@rrrsH NLN.;2VV]rrc&]tRtRtRtRtRtRtRtR#)GateUpDownMoELifecycleHooksi>z MoE lifecycle hooks for models using gate_proj/up_proj/down_proj naming. Used by: GLM-4, Qwen2-MoE, Mixtral, Phi-3 MoE, most Llama-based MoE models gate_projup_proj down_projrN r'rrrrrrrrrrrrr>s !NL Nrrc&]tRtRtRtRtRtRtRtR#)W1W3W2MoELifecycleHooksiIz MoE lifecycle hooks for models using w1/w3/w2 naming. Used by: Some architectures that explicitly name layers w1, w3, w2 (e.g., MiniMax-M2) w1w3w2rNrrrrrrIs NLNrr)rtypingrrrrytorch.nnrnn_modules.hooked_linearr utils.devicer utils.loggerr utils.modelrr>r0r2rrrrrrrsj '& 2%'!n!HU2U2pd(9dN !"C!?r