dsv4 B200 MTP SGLang launch#1139

Closed

yhyang201 wants to merge 14 commits intoSemiAnalysisAI:mainfrom

yhyang201:chore/dsv4-sgl-b200

Commits on Apr 24, 2026

Add dsv4-fp4-b200-sglang single-node config

cquil11
and
claude
committed
Switch dsv4-fp4-b200-sglang to Pro model, match vllm parallelism

cquil11
and
claude
committed
Match DSV4 Pro SGLang recipe literally; port HF cache path

cquil11
and
claude
committed
fix: use 'sglang serve' CLI, not python -m sglang.launch_server

cquil11
and
claude
committed
fix: mount repo at /ix for deepseek-v4-blackwell image

cquil11
and
claude
committed
fix: reinstall sglang from PyPI to work around masked editable install

cquil11
and
claude
committed
fix: uninstall editable sglang before reinstalling from PyPI

cquil11
and
claude
committed
fix: mount repo at /ix for deepseek-v4-blackwell; drop pip workaround

cquil11
and
claude
committed
fix: unset baked-in CUDA_VISIBLE_DEVICES for deepseek-v4-blackwell image

cquil11
and
claude
committed
fix: apply same /ix mount fix to launch_b200-nb.sh

cquil11
and
claude
committed
Drop --container-name arg from launch_b200-nb.sh

cquil11
and
claude
committed
Update dsv4 B200 SGLang launch: sglang serve + EAGLE speculative decoding

yhyang201
and
claude
committed
Add spec-decoding: mtp to dsv4-fp4-b200-sglang config

yhyang201
and
claude
committed
Add dsv4_fp4_b200_mtp.sh for spec-decoding benchmarks

yhyang201
and
claude
committed