[NVIDIA] chore: B200 single node DeepSeek v4 SGLang#1131

Merged

cquil11 merged 21 commits intomainfrom

chore/dsv4-sgl-b200

Apr 24, 2026

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
change runner
cquil11
committed
update recipe
cquil11
committed
Fix launch_b200-cw.sh and add b200-cw to runners pool

cquil11
and
claude
committed
update recipe
cquil11
committed
update model storage to nvme
cquil11
committed
fix(launch_b200-cw): skip realpath on worker-local squash; drop stale rmdir

cquil11
and
claude
committed
feat(dsv4_fp4_b200): pick recipe (low-latency/balanced/max-throughput) by CONC

cquil11
and
claude
committed
update b200
cquil11
committed
feat(dsv4-fp4-b200-sglang): split search-space per sglang recipe

cquil11
and
claude
committed
Merge branch 'main' into chore/dsv4-sgl-b200
cquil11
authored

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[NVIDIA] chore: B200 single node DeepSeek v4 SGLang#1131

[NVIDIA] chore: B200 single node DeepSeek v4 SGLang#1131
cquil11 merged 21 commits intomainfrom
chore/dsv4-sgl-b200

Commits on Apr 24, 2026

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

Switch dsv4-fp4-b200-sglang to Pro model, match vllm parallelism

Match DSV4 Pro SGLang recipe literally; port HF cache path

fix: use 'sglang serve' CLI, not python -m sglang.launch_server

fix: mount repo at /ix for deepseek-v4-blackwell image

fix: reinstall sglang from PyPI to work around masked editable install

fix: uninstall editable sglang before reinstalling from PyPI

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

fix: unset baked-in CUDA_VISIBLE_DEVICES for deepseek-v4-blackwell image

fix: apply same /ix mount fix to launch_b200-nb.sh

Drop --container-name arg from launch_b200-nb.sh

change runner

update recipe

Fix launch_b200-cw.sh and add b200-cw to runners pool

update recipe

update model storage to nvme

fix(launch_b200-cw): skip realpath on worker-local squash; drop stale rmdir

feat(dsv4_fp4_b200): pick recipe (low-latency/balanced/max-throughput) by CONC

update b200

feat(dsv4-fp4-b200-sglang): split search-space per sglang recipe

Merge branch 'main' into chore/dsv4-sgl-b200