58 lines
1.5 KiB
YAML
58 lines
1.5 KiB
YAML
# OpenAI-compatible FastH3 server for one DGX Spark (one GB10).
|
|
# Install: docs/getting_started/installation/spark.md
|
|
#
|
|
# GB10 has no FA4 / ptx_blackwell VSA kernel. Prefix the serve command with:
|
|
# FASTVIDEO_VSA_PTX_BLACKWELL=0 FASTVIDEO_FA4=0 FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3
|
|
#
|
|
# default_request.sampling is an example. Legal H3 frame counts are 17n+5,
|
|
# max 345 (15 s). A 345-frame request on one Spark can OOM; prefer 124 or 243,
|
|
# or a two-Spark Ray cluster.
|
|
generator:
|
|
model_path: FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
|
|
engine:
|
|
num_gpus: 1
|
|
use_fsdp_inference: false
|
|
parallelism:
|
|
tp_size: 1
|
|
sp_size: 1
|
|
offload:
|
|
dit: false
|
|
dit_layerwise: false
|
|
text_encoder: true
|
|
image_encoder: true
|
|
vae: true
|
|
pin_cpu_memory: true
|
|
lazy_module_load: true
|
|
compile:
|
|
enabled: false
|
|
vae_enabled: true
|
|
pipeline:
|
|
workload_type: t2v
|
|
experimental:
|
|
attention_backend: VIDEO_SPARSE_ATTN_H3
|
|
VSA_sparsity: 0.9
|
|
VSA_tile_size: 64
|
|
inference_torch_compile: false
|
|
vae_parallel_decode: true
|
|
vae_parallel_decode_strategy: gather
|
|
|
|
server:
|
|
host: 0.0.0.0
|
|
port: 8000
|
|
output_dir: outputs/openai_fasth3_spark
|
|
served_model_name: fasth3
|
|
|
|
default_request:
|
|
negative_prompt: ""
|
|
sampling:
|
|
height: 768
|
|
width: 1344
|
|
num_frames: 124
|
|
fps: 24
|
|
num_inference_steps: 5
|
|
guidance_scale: 1.0
|
|
batch_cfg: false
|
|
seed: 1000
|
|
output:
|
|
return_frames: false
|