Support device vs gpu http mode #626

Open

YqGe585 wants to merge 39 commits intoPFCCLab:mainfrom

YqGe585:accuracy

Commits on Apr 21, 2026

support accuracy by API
YqGe585
committed
add http

root
authored and
YqGe585
committed
support http mode
YqGe585
committed
fix: device_vs_gpu HTTP 模式精度对比的三个问题

YqGe585
and
claude
committed
fix: api_config_skip.txt 覆盖导致 write_to_log("skip") 条目丢失

YqGe585
and
claude
committed
fix: worker进程在设置CUDA_VISIBLE_DEVICES之前import paddle导致CUDA context固定在GPU 0

YqGe585
and
claude
committed
feat: device_vs_gpu http模式支持float8 dtype

YqGe585
and
claude
committed
fix: 修复float8 backward grad生成和skip/error误判两个问题

YqGe585
and
claude
committed
fix: skip fp8 on XPU before HTTP request, fix backward accuracy_error logging

YqGe585
and
claude
committed
fix: convert sparse grads to dense before paddle.save serialization

YqGe585
and
claude
committed
fix: correctly convert SelectedRows grad to dense for embedding(sparse=True)

YqGe585
and
claude
committed
fix: normalize named tuple output before paddle.save in http_server

YqGe585
and
claude
committed
fix: handle zero-size tensor in pad and take index generation

YqGe585
and
claude
committed
fix: use iinfo for integer dtypes in pow/rpow dtype_max calculation

YqGe585
and
claude
committed
update dtype_atol_rtol config: add float32/float64/int/bool tolerances

YqGe585
and
claude
committed
feat: introduce remote_error log type for HTTP mode remote execution failures

YqGe585
and
claude
committed
add rrelu to nondeterministic.py
YqGe585
committed
update http_config.yaml: set remote_host to 10.78.119.13

YqGe585
and
claude
committed

Commits on Apr 22, 2026

Commits on Apr 23, 2026

Commits on Apr 24, 2026

Commits on Apr 27, 2026

fix: only skip complex scalar cases when combined with float64 tensor

committed