[DSv4][bring-up] FP8 H100 Dynamo-vLLM multinode benchmark / DSv4 FP8 H100 Dynamo-vLLM 多节点基准测试（bring-up） - #2196

Closed

Oseltamivir wants to merge 22 commits into

mainfrom

dsv4-h100-dynamo-sglang-bringup

Commits on Jul 14, 2026

feat(dsv4): scaffold FP8 H100 dynamo-sglang multinode bring-up
Oseltamivir
committed
9cada95
refine(dsv4): extrapolate H100 dynamo-sglang topologies + 8k1k
Oseltamivir
committed
8b28eec
Merge branch 'main' into dsv4-h100-dynamo-sglang-bringup
Oseltamivir
authored
e92892c
fix(dsv4): route H100 dynamo-sglang to a runner label that exists
Oseltamivir
committed
18eabcb
refactor(dsv4): follow vLLM H100 recipe
Oseltamivir
committed
190057f
fix(h100): install make before srt-slurm setup
Oseltamivir
committed
b04ebb7
revert(h100): avoid host package installation
Oseltamivir
committed
6a4d351
fix(dsv4): use H100 checkpoint and healthy nodes
Oseltamivir
committed
edbaf8d
fix(dsv4): use standard vLLM container setup
Oseltamivir
committed
94f1d11
fix(dsv4): use TP8 PP2 on H100
Oseltamivir
committed
f0a3df2
fix(dsv4): resolve H100 TP8/PP2 CUDA OOM during model load
Oseltamivir
committed
034d219
fix(dsv4): H100 use DP2+EP16 (multi_node_dep) instead of PP2
Oseltamivir
committed
464f4c6
fix(dsv4): rebase H100 perf-changelog onto origin/main (append-only)
Oseltamivir
committed
b8056c6
fix(dsv4): H100 use pure DP16+EP16 (tp1), not TP8xDP2
Oseltamivir
committed
185c83f
fix(dsv4): H100 aggregated DEP32 over 4 nodes (fit ~1TB FP8 on 80GB)
Oseltamivir
committed
3e2cbfc
fix(dsv4): H100 aggregated TP8xPP2 (2 nodes), not DP32
Oseltamivir
committed
fe5a2fe

Commits on Jul 15, 2026

Commits on Jul 16, 2026