daniel/h3-blackwell-runtime
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Match Comfy joint AV noise initialization

Browse source
This commit is contained in:
Daniel Maddern 2026-08-13 16:12:10 +07:00
parent 82e8ddecbc
commit 09f5d2ef65
2 changed files with 11 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
src/h3_blackwell_runtime/t2v.py
View file

@ -28,3 +28,12 @@ def empty_av_latents(width: int, height: int, frames: int, *, device: torch.devi
video = torch.zeros((1, 24, video_steps, height // 16, width // 16), device=device)
audio = torch.zeros((1, 32, 2, audio_steps), device=device)
return video, audio, frames
def random_av_latents(width: int, height: int, frames: int, seed: int, *, device: torch.device | str = "cuda") -> tuple[torch.Tensor, torch.Tensor, int]:
"""Generate Comfy-equivalent CPU-seeded joint AV noise latents."""
video, audio, frames = empty_av_latents(width, height, frames, device="cpu")
generator = torch.manual_seed(seed)
video = torch.randn(video.shape, dtype=torch.float32, generator=generator, device="cpu").to(video.dtype)
audio = torch.randn(audio.shape, dtype=torch.float32, generator=generator, device="cpu").to(audio.dtype)
return video.to(device), audio.to(device), frames

6
tools/direct_t2v_preview.py
View file

@ -14,7 +14,7 @@ from h3_blackwell_runtime.denoiser import H3PackedDenoiser
from h3_blackwell_runtime.packing import H3PromptPacker
from h3_blackwell_runtime.qwen3vl_text import Qwen3VLPromptConditioner
from h3_blackwell_runtime.sampler import sample_video_res_multistep
from h3_blackwell_runtime.t2v import empty_av_latents
from h3_blackwell_runtime.t2v import random_av_latents
from h3_blackwell_runtime.token_refiner import H3TokenRefiner
from h3_blackwell_runtime.vae_decoder import MiniMaxH3VideoVAE
@ -30,14 +30,12 @@ parser.add_argument("--seed", type=int, default=440204)
parser.add_argument("--attention", choices=("sage2", "sdpa", "sage3"), default="sage2")
args = parser.parse_args()
torch.manual_seed(args.seed)
checkpoint = H3Checkpoint("/models/minimax_h3_fl2va_pruned_nvfp4.safetensors")
conditioner = Qwen3VLPromptConditioner(
"/text-encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
"/opt/h3-blackwell-runtime/src/h3_blackwell_runtime/qwen25_tokenizer",
)
video, audio, frames = empty_av_latents(args.width, args.height, args.frames)
video.normal_()
video, audio, frames = random_av_latents(args.width, args.height, args.frames, args.seed)
model = H3PackedDenoiser.from_checkpoint(checkpoint, attention_backend=args.attention).eval()
text = H3TokenRefiner(checkpoint, attention_backend=args.attention)(conditioner(args.prompt))
latent = sample_video_res_multistep(model, H3PromptPacker(checkpoint), text, video, audio, steps=args.steps)