test eager and custom allreduce toggles

This commit is contained in:
2026-07-14 19:18:46 +08:00
parent 82c1463b0c
commit a5abd66e21
7 changed files with 777 additions and 2 deletions

View File

@@ -29,6 +29,10 @@ flash attention kernelixformer / cudnnFlashAttnForward
关闭选项),原意是防止 profiling OOM。但 _run_sdpa_fallback 已通过 Q-tiling
解决了该问题chunked prefill 反而会把推理路径从 _run_sdpa_fallback 切换到
_forward_prefix_pytorch属于不必要的行为变更因此一并禁用该自动逻辑。
同时解除 EngineArgs.create_*_config 中强制 enforce_eager=True 和
disable_custom_all_reduce=True 的硬编码,让 Iluvatar 环境可以实际验证
CUDA Graph / custom all-reduce需要回退时仍可通过命令行显式传
--enforce-eager --disable-custom-all-reduce。
Deploy:
python3 modified_scripts/patch_xformers_sdpa_seq.py
@@ -91,6 +95,13 @@ _ARG_NEW_BLOCK = """\
# handles long-context memory without chunked prefill\
"""
_ARG_FORCE_EAGER_OLD = " enforce_eager=True,"
_ARG_FORCE_EAGER_NEW = " enforce_eager=self.enforce_eager,"
_ARG_FORCE_ALLREDUCE_OLD = " disable_custom_all_reduce=True,"
_ARG_FORCE_ALLREDUCE_NEW = (
" disable_custom_all_reduce=self.disable_custom_all_reduce,"
)
FALLBACK_METHOD = '''
def _run_sdpa_fallback(
self,
@@ -275,6 +286,26 @@ def patch_arg_utils(path):
else:
print(" [warn] target block not found — check arg_utils.py version")
if _ARG_FORCE_EAGER_NEW in content:
print(" [skip] enforce_eager already respects CLI")
elif _ARG_FORCE_EAGER_OLD in content:
content = content.replace(_ARG_FORCE_EAGER_OLD,
_ARG_FORCE_EAGER_NEW, 1)
print(" [ok] enforce_eager now respects CLI")
changed = True
else:
print(" [warn] enforce_eager assignment not found")
if _ARG_FORCE_ALLREDUCE_NEW in content:
print(" [skip] custom all-reduce already respects CLI")
elif _ARG_FORCE_ALLREDUCE_OLD in content:
content = content.replace(_ARG_FORCE_ALLREDUCE_OLD,
_ARG_FORCE_ALLREDUCE_NEW, 1)
print(" [ok] custom all-reduce now respects CLI")
changed = True
else:
print(" [warn] custom all-reduce assignment not found")
if changed:
with open(path, "w") as f:
f.write(content)