30 lines
1.3 KiB
YAML
30 lines
1.3 KiB
YAML
|
# Cosmos3 omni deploy overlay: single co-located diffusion stage (both MoT
|
||
|
|
# towers in one process). Topology (COSMOS3_OMNI_DEPLOY_PIPELINE) is declared in
|
||
|
|
# vllm_omni/model_executor/models/cosmos3/pipeline.py.
|
||
|
|
#
|
||
|
|
# T2I / video / policy Cosmos3 checkpoints share HF metadata
|
||
|
|
# (model_type=cosmos3_omni, model_index.json _class_name=
|
||
|
|
# Cosmos3OmniDiffusersPipeline). This pipeline cannot be auto-detected and must
|
||
|
|
# be selected explicitly so --deploy-config is applied (#6874):
|
||
|
|
#
|
||
|
|
# vllm serve nvidia/Cosmos3-Super --omni \
|
||
|
|
# --deploy-config /absolute/path/to/vllm_omni/deploy/cosmos3_omni.yaml
|
||
|
|
#
|
||
|
|
# Same overlay works for Nano (including video). Without --deploy-config,
|
||
|
|
# behaviour is unchanged (CLI single-stage fallback). --no-guardrails remains
|
||
|
|
# the CLI-only equivalent of model_config.guardrails below. Per-checkpoint
|
||
|
|
# knobs can still be layered with --stage-overrides.
|
||
|
|
|
||
|
|
pipeline: cosmos3_omni_deploy
|
||
|
|
# async_chunk defaults to True, which the deploy validator rejects for a
|
||
|
|
# single-stage pipeline (no next-stage processor). Must be false here.
|
||
|
|
async_chunk: false
|
||
|
|
distributed_executor_backend: mp
|
||
|
|
|
||
|
|
stages:
|
||
|
|
- stage_id: 0
|
||
|
|
model_class_name: Cosmos3OmniDiffusersPipeline
|
||
|
|
model_config:
|
||
|
|
# Gated nvidia/Cosmos-1.0-Guardrail is often unavailable; default this
|
||
|
|
# recommended layout off. Equivalent to launching with --no-guardrails.
|
||
|
|
guardrails: false
|