Kandinsky6

​

1. Quick start

uv pip install -e "python[diffusion]" --prerelease=allow

​

2. Model capabilities

​

3. Offline generation

sglang generate \
  --model-path kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers \
  --performance-mode speed --attention-backend fa --warmup-mode off \
  --prompt "A woman chops vegetables in a sunlit kitchen, with soft jazz playing." \
  --seed 42 --save-output

​

Python API

from sglang.multimodal_gen import DiffGenerator

generator = DiffGenerator.from_pretrained(
    model_path="kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers",
    num_gpus=1,
    performance_mode="speed",
    attention_backend="fa",
    warmup_mode="off",
)
try:
    result = generator.generate(sampling_params_kwargs={
        "prompt": "A woman chops vegetables in a sunlit kitchen, with soft jazz playing.",
        "seed": 42,
        "save_output": True,
        "output_path": "kandinsky6_samples",
        "return_file_paths_only": False,
        "return_frames": True,
    })
    print(result.output_file_path)
finally:
    generator.shutdown()

​

4. Parallelism and memory

• Ulysses: full-size Pro-distill generation on two B200 or four GB300 GPUs matches single-GPU output on the same hardware, including every decoded RGB frame and the audio waveform. It shards the video stream; audio and text streams remain replicated. Video K/V is gathered for audio-to-video attention.
• CFG parallel: Pro can distribute its conditional and unconditional branches with --num-gpus 2 --ulysses-degree 1 --enable-cfg-parallel. This is a different topology from Ulysses. It does not accelerate Pro-distill, which has no guidance branch.
• TP and Ring: implemented, but full-model outputs differ from the single-GPU baseline. Do not treat them as verified bitwise-equivalent modes. Ring2, Ring4, and TP2 x Ring2 complete full-size Pro-distill generation on GB300 and repeat exactly within each topology. Ring2 serving also covers image conditioning, Cache-DiT, and dynamic/merged LoRA restoration. Use --num-gpus 2 --ulysses-degree 1 --ring-degree 2 to select Ring2; validate both video and audio quality before deployment.
• TP x SP: four-GB300 TP2 x Ulysses2 full generation is tested. It reduces weight memory but was slower than Ulysses4 and produced different video and audio. Prefer Ulysses when memory permits; validate quality before deploying TP.
• FSDP + Ulysses: add --use-fsdp-inference true to shard DiT weights without sharding linear reductions. Four-GB300 Ulysses4 generation matches resident Ulysses4 in every decoded RGB frame and the audio waveform. This is a lower-memory alternative with additional weight-gather latency, not a faster default. Request-level Cache-DiT switching is also tested with FSDP.
• Encoder parallelism: --encoder-parallel fold shards Reason1 and CLIP across the DiT replica, including a Ulysses-only replica. Four-GPU generation is tested, but folding changes reductions and the generated video and audio. Keep auto for the baseline recipes and validate quality before forcing folding. See Encoder parallelism.
• Offload: keep components resident when memory allows. Use --cpu-offload-components transformer text_encoder vae for whole-component offload, or --layerwise-offload-components all for layerwise offload. Layerwise offload includes both DiT text towers and the audio VAE’s encoder, decoder, and vocoder blocks. Use component-specific residency settings to keep selected components resident. Both require host RAM and add transfers; lower device memory does not imply lower total memory or lower latency.

​

Deployment limits

​

Choosing a topology

​

RTX PRO 6000 and SageAttention3

​

VAE decode

​

Cache-DiT

​

LoRA adapters

curl --fail-with-body http://localhost:30000/v1/set_lora \
  -H "Content-Type: application/json" \
  -d '{
    "lora_nickname": "kandinsky6",
    "lora_path": "/path/to/adapter.safetensors",
    "target": "transformer",
    "strength": 1.0,
    "merge_mode": "dynamic"
  }'

curl --fail-with-body http://localhost:30000/v1/unmerge_lora_weights \
  -H "Content-Type: application/json" -d '{"target": "transformer"}'

​

5. Download a completed video

import json
import shutil
import time
import urllib.request

base_url = "http://localhost:30000"
video_id = "REPLACE_WITH_RETURNED_ID"
deadline = time.monotonic() + 1800

while time.monotonic() < deadline:
    with urllib.request.urlopen(f"{base_url}/v1/videos/{video_id}", timeout=30) as response:
        job = json.load(response)
    if job["status"] == "failed":
        raise RuntimeError(job)
    if job["status"] == "completed":
        with urllib.request.urlopen(
            f"{base_url}/v1/videos/{video_id}/content", timeout=120
        ) as response, open("output.mp4", "wb") as output:
            shutil.copyfileobj(response, output)
        break
    time.sleep(2)
else:
    raise TimeoutError("Video generation did not finish within 30 minutes")

curl --fail-with-body -X DELETE http://localhost:30000/v1/videos/REPLACE_WITH_RETURNED_ID

​

6. ComfyUI