+ j]a0t$Rt^RIHt^RIt^RIHt^RIHtH t H t ]'d^RI t .ROt RRR R R R /t R ]R&RRRRRR/tR ]R&]P RRl4t]P RRl4tRRltRtRRltRRltR#) z&Stream utilities for Inductor codegen.) annotationsN) TYPE_CHECKING)DEFAULT_STREAMDEFAULT_STREAM_IDXSTREAM_NAME_TEMPLATEAOTI_SUPPORTED_STREAM_OP_NAMES"AOTI_UNSUPPORTED_STREAM_OP_REASONSz&torch.ops.streams.record_event.default record_eventz$torch.ops.streams.wait_event.default wait_eventz+torch.ops.streams.synchronize_event.defaultsynchronize_eventzdict[str, str]z,torch.ops.streams.synchronize_stream.defaultz~Host-blocking sync ops are not supported inside an AOTI Run(). Use record_event + wait_event for device-side ordering instead.z,torch.ops.streams.synchronize_device.defaultzKHost-blocking device synchronization is not supported inside an AOTI Run().z%torch.ops.streams.wait_stream.defaultzwait_stream is not supported in AOTI cpp_wrapper. Use explicit record_event on the waited-on stream + wait_event on the waiting stream.c V^8dQhRRRR/#) stream_idxintreturnstr)formats"]/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/_inductor/stream_utils.py __annotate__r8s B B B BcJV^8Xd\#\P!VR7#)zGenerate CUDA Stream name from stream index number. Args: stream_idx: Non-negative index number. 0 refers to the default stream, others refer to side streams. r)rrrrs&rget_stream_namer7s"Q#**jAArc V^8dQhRRRR/#r device_idxrrrr)rs"rrrFs%%C%C%rc RV 2#) raw_streamr)rs&r_raw_stream_name_for_devicerEs  | $$rc V^8dQhRRRR/#rr)rs"rrrJs33C3C3rcB^RIHpV!4'dR#\V4#)zCGenerate variable name for a raw stream handle on the given device. _coor_enabledr)"torch.fx.experimental.proxy_tensorr#r)rr#s& rget_raw_stream_namer%JsA &z 22r_coor_device_idxc V^8dQhRRRR/#)r deviceztorch.device | Nonerrr)rs"rrrZs  / C rc^^RIHpV!4'dVe VP#\V4#)zDevice string to emit into generated code (benchmark and autotune harnesses). Under compile-on-one-rank drop the index (``cuda:0`` -> ``cuda``) so the emitted text is rank-identical and its tensors land on the running rank's current device. r")r$r#typer)r(r#s& rcoor_device_strr+Zs( A6-{{ v;rc V^8dQhRRRR/#)r rrrztuple[str | None, int | str]r)rs"rrrgsBB#B2NBrc^RIHpV!4'gRV3#^RIHpVPP P 4p\ RV 2\3#)a(preamble line, index token) for a Triton kernel's benchmark harness. The harness is emitted into the kernel's own source, which is what the kernel hash covers, so under compile-on-one-rank it must resolve the device at run time rather than bake a rank-specific index -- otherwise kernels are not rank-identical whenever ``config.benchmark_kernel`` is on. The preamble line (when not None) must be emitted inside the harness function before the returned token is used. r"N)Vz = )r$r#torch._inductor.virtualizedr.graph device_opscurrent_device_idx_exprCOOR_DEVICE_IDX_VAR)rr#r.exprs& rcoor_benchmark_device_idxr5gsMA ??Z- 77   5 5 7D! "#dV ,.A AAr)rrrrrr%r)__conditional_annotations____doc__ __future__r functoolstypingr torch._inductor.stream_constantsrrrtorch__all__r__annotations__r lru_cacherrr%r3r+r5)r6s@rr@s,"  -n*L23F 23 J3U+ S 6"N   B B %%3) Br