Skip to content

Updating Cosmos3 Nano DFlash recipe with optional fixes #2205

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
adsridhar wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Updating Cosmos3 Nano DFlash recipe with optional fixes #2205

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 26 additions & 0 deletions examples/speculative_decoding/recipes/train_dflash_cosmos3_nano.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -249,6 +249,28 @@
" --overwrite\n"
]
},
{
"cell_type": "markdown",
"id": "pai-sglang-runtime-notes",
"metadata": {},
"source": [
"#### Native-video SGLang runtime notes\n",
"\n",
"The PAI path sends native video to SGLang. If the selected SGLang container does not include OpenCV, enable the worker's optional installation before running the compute cell:\n",
"\n",
"```bash\n",
"export INSTALL_OPENCV_HEADLESS=1\n",
"```\n",
"\n",
"For the pinned SGLang container on affected B200/Blackwell environments, use Triton multimodal attention:\n",
"\n",
"```bash\n",
"export SGLANG_EXTRA_ARGS=\"${SGLANG_EXTRA_ARGS:-} --mm-attention-backend triton_attn\"\n",
"```\n",
"\n",
"These are opt-in settings: leave them unset when the container already provides OpenCV or the default multimodal attention backend works on the allocated GPU.\n"
]
},
{
"cell_type": "code",
"execution_count": null,
Expand All @@ -259,6 +281,10 @@
"%%bash\n",
"# Compute node: run a representative PAI shard slice in an existing Slurm GPU allocation.\n",
"set -euo pipefail\n",
"# If the selected SGLang image lacks OpenCV for native-video decoding, uncomment:\n",
"# export INSTALL_OPENCV_HEADLESS=1\n",
"# On affected B200/Blackwell environments with the pinned SGLang image, uncomment:\n",
"# export SGLANG_EXTRA_ARGS=\"${SGLANG_EXTRA_ARGS:-} --mm-attention-backend triton_attn\"\n",
"SPEC=\"$(cd .. && pwd)\"\n",
": \"${DATA_ROOT:?Run the Step 2 configuration cell before this cell.}\"\n",
": \"${MODEL_PATH:?Set MODEL_PATH and run Step 2 before this compute-node cell.}\"\n",
Expand Down