+ jY$a0t$Rt^RIHt^RIt^RIHt^RIH t ^RI t ^RI H t ^RI Ht^RIHtHt^RIHt]P(!]4tR R R R R R/tRRRR/t] P2P4R!RRll4t] !RR7!RR44tRsR]R&] P2P>RRl4t RRlt!RR lt"R#)"zSonicMoE integration: fused MoE using CuteDSL kernels from `kernels-community/sonic-moe`. Provides `sonicmoe_experts_forward` registered as "sonicmoe" in the ExpertsInterface. Requirements: CUDA, `kernels`, `nvidia-cutlass-dsl`, has_gate=True. ) annotationsN)Callable) dataclass)version)logging)is_kernels_availablemaybe_import_error)lazy_load_kernelsiluswiglugelugeglureluregluznvidia-cutlass-dslz4.6.0zapache-tvm-ffiz0.1.10c V^8dQhRRRR/#) raise_errorboolreturn)formats"c/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/integrations/sonicmoe.py __annotate__r.s,,d,t,c \4'g\RVR7#\PP 4'g\RVR7#\PP 4^,pV^ 8d\RV R2VR7#\ P4Fmwr#\PPV4p\P!T4\P!T48gKV\RT RT R T R 2TR7u# R # \PPd\RT R2TR7uu#i;i) u Whether the sonic-moe kernel can be loaded in this environment: `kernels` installed, a CUDA GPU on Hopper (SM90) or newer, and `nvidia-cutlass-dsl` / `apache-tvm-ffi` no newer than the versions the kernel was validated against (`SONICMOE_DEPENDENCIES`) — newer releases may break its CuteDSL / TVM FFI dispatch. A one-glance gate for callers, including external stacks. With `raise_error=True` (used by the loader) it re-raises the specific `ImportError` instead of returning `False`. `@assume_constant_result` makes dynamo evaluate this once at trace time and inline the bool (its probe is untraceable — an `lru_cache`'d import check, the `get_device_capability` pybind, and `importlib.metadata` version lookups), so it stays compile-safe if called from a compiled region. zSsonic-moe requires the `kernels` package. Use a different `experts_implementation`.rzdsonic-moe kernel requires CUDA, but CUDA is not available. Use a different `experts_implementation`.z`sonic-moe requires a Hopper (SM90+) or newer GPU, but the current device has compute capability z-.x. Use a different `experts_implementation`.zsonic-moe requires `z`, but it is not installed.z` >= z, but z is installed.T) rrtorchcuda is_availableget_device_capabilitySONICMOE_DEPENDENCIESitems importlibmetadatarPackageNotFoundErrorparse)rmajor distribution min_version installeds& ris_sonicmoe_loadabler*-sC ! !! a#   :: " " $ $! r#   JJ , , .q 1E qy! M O#  &;%@%@%B!  !**22<@I == #gmmK&@ @%&|nE+fYKWef' &C !!66 %&|n4OP^i  s&D0EET)frozenc0]tRt^]t$RtR]R&R]R&RtR#)SonicMoEzAEntry points exposed by the `kernels-community/sonic-moe` kernel.typeactivation_type_enumrmoe_general_routing_inputsrN)__name__ __module__ __qualname____firstlineno____doc____annotations____static_attributes__rrrr-r-]sK ((rr-zSonicMoE | None _SONICMOEcV^8dQhRR/#)rrNoner)rs"rrrls..t.rcd\eR#\RR7\R4pVf \R4h\ \ VRR4RR4p\ VRR4pR V3RV33UUu.uF wr4VeK VNK pppV'd\R R P V4 R 24h\ VVR 7sR#uuppi)a Load sonic-moe once into the `_SONICMOE` module global. `@allow_in_graph` makes `torch.compile` treat the untraceable hub download + dynamic import as a single opaque node instead of tracing into it; it returns `None` (proxyable) and populates the global, which `load_sonicmoe_kernel` then returns. Under NO circumstances may this function return a value: an `@allow_in_graph` fx node's return must be proxyable, and returning the bundle (e.g. from the warm-cache short-circuit) breaks torch.compile with `Unsupported: torch.* op returned non-Tensor`. Raises `ImportError` if CUDA/hardware requirements are not met, or if the kernel or required symbols are not found. NTrz sonic-moeu}Failed to load the sonic-moe kernel — check that `kernels-community/sonic-moe` has a build matching the current torch/CUDA.enumsActivationTyper0zenums.ActivationTypez.sonic-moe kernel is missing required symbols: z, zN. Make sure you have the `kernels` package and `nvidia-cutlass-dsl` installed.)r/r0)r8r*r ImportErrorgetattrjoinr-)kernelr/r0nameattrmissings r_load_sonicmoe_kernelrEks"T* k *F ~ ;  #767D#ACSUYZ!(1Mt!T $%9 : )+E F  JD    rs #$! I)   H % 8VWfg >.w8H(S%%,&,^ $)))" ?!..b 9r