+ PjRt^RIt^RIt^RIt^RIHt]P !]4tRR0t Rt RRlt Rt RR lt R R ltR R ltR#)as Dynamic int8 quantization for pocket-tts. Uses torchao if available (torch 2.10+ with C++ extensions), otherwise falls back to torch.ao.quantization (deprecated but functional on torch 2.5-2.9). Quantizes attention (Q/K/V/output projections) and FFN (linear1/linear2) layers in the FlowLM transformer. The flow matching network and Mimi VAE decoder remain in float32. N attentionffnc^RIpVPPR4fR#^RIp\ VR4'g\ VRR4'gR#R# \ dR#i;i)zDetect the best available quantization backend. Returns "torchao" if torchao is installed with working C++ extensions, otherwise returns "torch.ao". Ntorchaoztorch.ao_C_SKIPPED_CPP_EXTENSIONSF)importlib.utilutil find_specrhasattrgetattr ImportError) importlibrs p/Users/ahmed/devFolder/Ultron/claude-voice/gateway/.venv/lib/python3.14/site-packages/pocket_tts/quantization.py _get_backendrsg   >> # #I . 6 7D ! !:SUZ)[)[*\     s!A)A A"!A"c8V^8dQhR\P/#)module)nnModule)formats"r __annotate__r-s??RYY?c0^RIHpHpV!W!44R#)z.Apply int8 dynamic quantization using torchao.)%Int8DynamicActivationInt8WeightConfig quantize_N)torchao.quantizationrr)rrrs& r_quantize_module_torchaor-sU f;=>rc\P!4R9d"R\PPnR#\PPP R8Xd"R\PPnR#R#)zKSet the quantization engine for torch.ao (QNNPACK for ARM, FBGEMM for x86).qnnpacknonefbgemmN)arm64aarch64)platformmachinetorchbackends quantizedenginerr_ensure_quantization_enginer+4sU11*3  '  ! ! ( (F 2*2  ' 3rczV^8dQhR\PR\\,R\P/#)rflow_lmquantize_groupsreturnrrsetstr)rs"rrr<s- CHrcV'g\PR4V#\4p\PRV4VR8Xd\W4V#\ W4V#)a Apply dynamic int8 quantization to the specified layer groups of a FlowLM model. Automatically selects the best available backend: - torchao (torch 2.10+): optimized C++ kernels, faster on both ARM and x86 - torch.ao (torch 2.5-2.9): deprecated but functional fallback Args: flow_lm: The FlowLM model (model.flow_lm) quantize_groups: Set of group keys to quantize. Valid keys: "attention", "ffn", "flow_net" Returns: The quantized model (modified in-place). zr?r@rArBrC)r-r.rIrDs&& rr7r7ms6!_$))BII;ekkSWX$$++ / ) U__ryykVZ [ O #, emm,ryykEM- emm,ryykEM,r)__doc__loggingr$r&torch.nnr getLogger__name__r4RECOMMENDED_CONFIGrrr+r9r6r7r*rrrUsS     8 $"5)*?3>($r