# Required patches (fish-speech main e5e2926 / sglang-omni main, 2026-07-18)
## fish-speech (training)
1. fish_speech/models/text2semantic/llama.py: both `checkpoint(...)` calls -> use_reentrant=False
2. fish_speech/models/text2semantic/lit_module.py: add `strict_loading = False` to TextToSemantic
3. (optional) lora_slow_patch.py adds slow_* target names that do NOT imply fast_*
## sglang-omni (serving on Blackwell/sm_120)
1. rm -r site-packages/deep_gemm (import-time assert without CUDA toolkit)
2. apt: cuda-nvcc-13-0 libcublas-dev-13-0 libcusparse-dev-13-0 libcusolver-dev-13-0 libcurand-dev-13-0 ninja-build (torch cu130)
3. sglang_omni/models/fishaudio_s2_pro/engine_builder.py: attention_backend "fa3" -> "triton"; disable_cuda_graph True; mem_fraction_static 0.55
4. run patch_audio_decoder.py (adds FISH_FORCE_SDPA env gate replacing sgl_kernel flash_attn_with_kvcache)
Launch env: CUDA_HOME=/usr/local/cuda-13.0 SGLANG_ENABLE_JIT_DEEPGEMM=0 FISH_FORCE_SDPA=1