eThQ FSSKrSSKrSSKJr SSKrSr Sr SrSrS rS rSS jr\"\R&R(5\"S 5:aS r\R,R.r\rO\ r\ r\\\\\\\\R,R2\R,R4\R,R6\R,R6\R,R8S. rSrg!\\4a0 SSKr\"\R5RS:a\ "S5eGN f=f)N)parsezYour currently installed version of Keras is Keras 3, but this is not yet supported in Transformers. Please install the backwards-compatible tf-keras package with `pip install tf-keras`.c [R"U5nSS[RRU[R"[R "S5UR 5- 5--nX-$)ab Gaussian Error Linear Unit. Original Implementation of the gelu activation function in Google Bert repo when initially created. For information: OpenAI GPT's gelu is slightly different (and gives slightly different results): 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) Also see https://arxiv.org/abs/1606.08415 ??@)tfconvert_to_tensormatherfcastsqrtdtype)xcdfs S/var/www/auris/envauris/lib/python3.13/site-packages/transformers/activations_tf.py_gelur"sS QA rww{{1rwwrwws|QWW'E#EFF GC 7Nc v[R"U5n[R"[RUR 5n[R"SUR 5nSS[R "[R"SU- 5X[R"US5---5--nX-$)z Gaussian Error Linear Unit. This is a smoother version of the GELU. Original paper: https://arxiv.org/abs/1606.0841 Args: x: float Tensor to perform activation Returns: `x` with the GELU activation applied. Hm?rrr) r r r r pirtanhrpow)rrcoeffrs r _gelu_newr/s QA !'' "B GGHagg &E rwwrwwsRx0Aq! 8L4LMNN OC 7Nrc[R"U5nU[R"[RR U55-$)N)r r rr softplusrs rmishr As5 QA rwwrww''*+ ++rc [R"U5n[R"SUR5n[R"SUR5nSU-S[R"X-SX-U---5--$)Nrg3E?rr)r r r rr)rcoeff1coeff2s r gelu_fastr$Gsi QA WWXqww 'F WW\177 +F 7cBGGAJ# Q2F$GHH IIrc[R"U5n[R"SUR5nU[RR X-5-$)NgZd;?)r r r rr sigmoid)rrs r quick_gelur'Os@ QA GGE177 #E rwwuy) ))rcD[R"[U5SS5$)ax Clip the range of possible GeLU outputs between [-10, 10]. This is especially useful for quantization purpose, as it allows mapping 2 negatives values in the GeLU spectrum. For more information on this trick, please refer to https://arxiv.org/abs/2004.09602 Gaussian Error Linear Unit. Original Implementation of the gelu activation function in Google Bert repo when initially created. For information: OpenAI GPT's gelu is slightly different (and gives slightly different results): 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) Also see https://arxiv.org/abs/1606.08415 :param x: :return: i )r clip_by_valuerrs rgelu_10r+Us   E!Hc2 ..rcv[R"USUS9up#U[RRU5-$)a Gated Linear Unit. Implementation as defined in the original paper (see https://arxiv.org/abs/1612.08083), where the input `x` is split in two halves across a dimension (`axis`), A and B, returning A * sigmoid(B). Args: `x`: float Tensor to perform activation `axis`: dimension across which `x` be split in half Returns: `x` with the GLU activation applied (with its size halved across the dimension `axis`). r)axis)r splitr r&)rr-abs rglur1cs0 88Aqt $DA rwwq! !!rz2.4c>[RRUSS9$)NT) approximate)keras activationsgelurs rapproximate_gelu_wrapr7us  %%aT%::r) r6r+r$gelu_newr1r r'relur&siluswishrc U[;a [U$[SUS[[R5535e)Nz function z not found in ACT2FN mapping )ACT2FNKeyErrorlistkeys)activation_strings rget_tf_activationrBsBF"'((#4"55RSWX^XcXcXeSfRghiir))r tensorflowr packaging.versionrtf_kerasr4ModuleNotFoundError ImportError __version__major ValueErrorrrr r$r'r+r1versionVERSIONr7r5r6r8r9r&r;rr=rBrrrOs9 #   $, J* / "    e ,;    ! !D$H DH       " "  ((    # #    $ $    " "  jq [)  U  %%) &  * sC((4D D