diff --git a/src/h3_blackwell_runtime/sampler.py b/src/h3_blackwell_runtime/sampler.py
index 1a3755e..1079be5 100644
--- a/src/h3_blackwell_runtime/sampler.py
+++ b/src/h3_blackwell_runtime/sampler.py
@@ -64,6 +64,8 @@ def sample_video_res_multistep(model, packer: H3PromptPacker, text: torch.Tensor
         native_audio = audio_carried * (sigma_audio / sigma)
         hidden, times, segments, positions, video_segment, audio_segment = packer(text, video, native_audio, float(sigma))
         raw_video, raw_audio = model(hidden, times, positions, segments, video_segment, audio_segment)
+        raw_video = raw_video.to(torch.bfloat16).float()
+        raw_audio = raw_audio.to(torch.bfloat16).float()
         velocity_video = -unpatchify_video(raw_video, video.shape[2], video.shape[-2], video.shape[-1])
         velocity_audio_native = -_unpack_audio(raw_audio)
         carry = sigma_audio / sigma
diff --git a/tools/compare_fl2va_steps.py b/tools/compare_fl2va_steps.py
index cfd461d..ae5914a 100644
--- a/tools/compare_fl2va_steps.py
+++ b/tools/compare_fl2va_steps.py
@@ -39,7 +39,7 @@ for index, reference in enumerate(steps):
 
     state_video = reference["x"].to("cuda").reshape(-1)[:video_count].reshape(video_shape)
     reference_denoised = reference["denoised"].to("cuda").reshape(-1)[:video_count].reshape(video_shape)
-    converted_denoised = state_video + sigmas[index] * h3_output["video"].to("cuda")
+    converted_denoised = state_video + sigmas[index] * h3_output["video"].to("cuda").to(torch.bfloat16).float()
     denoised_delta = (converted_denoised.float() - reference_denoised.float()).abs()
 
     if index + 1 < len(steps):