From a809befdf3210d09d031afb9c6f24fd51d5602f9 Mon Sep 17 00:00:00 2001 From: Ziyue Yang Date: Fri, 21 Jun 2024 09:38:05 +0000 Subject: [PATCH] adapt to public msccl --- msccl/language/ir.py | 3 +++ msccl/ncclize.py | 3 +++ 2 files changed, 6 insertions(+) diff --git a/msccl/language/ir.py b/msccl/language/ir.py index aaee409..1cf52eb 100755 --- a/msccl/language/ir.py +++ b/msccl/language/ir.py @@ -143,6 +143,9 @@ def ir_to_xml(program: Program, old_format=True, use_scratch=True, pretty_print= algo_elem.set("ngpus", str(len(program.gpus))) algo_elem.set("coll", program.collective) algo_elem.set("inplace", str(1 if program.inplace else 0)) + algo_elem.set("outofplace", str(0 if program.inplace else 1)) + algo_elem.set("minBytes", str(0)) + algo_elem.set("maxBytes", str(0)) for gpu in program.gpus: gpu_elem = ET.SubElement(algo_elem, "gpu") gpu_elem.set("id", str(gpu.rank)) diff --git a/msccl/ncclize.py b/msccl/ncclize.py index f2ac2b8..8e158e5 100644 --- a/msccl/ncclize.py +++ b/msccl/ncclize.py @@ -664,6 +664,9 @@ def expand_mappings(mappings): algo_elem.set('nchannels', str(nchannels)) algo_elem.set('ngpus', str(len(gpus))) algo_elem.set('inplace', '0') + algo_elem.set('outofplace', '1') + algo_elem.set('minBytes', '0') + algo_elem.set('maxBytes', '0') algo_elem.set('coll', algorithm.collective.runtime_name) algo_elem.set('nchunksperloop', str(max(max(gpu.input_chunks, gpu.output_chunks) for gpu in gpus.values()))) for rank, gpu in gpus.items():