Harden H3 detail-pass cleanup

dumas_h3_longvideos.py

@@ -6461,36 +6461,41 @@ class H3LongVideos:
             raise
         refined_out = out
         detail_pass = _coerce_bool_flag(detail_pass)
-        if detail_pass:
-            detail_latent = _latent_with_replaced_samples(latent, out)
-            try:
-                detail_start = time.perf_counter()
-                (refined_out,) = nodes.common_ksampler(
+        if detail_pass:
+            detail_latent = _latent_with_replaced_samples(latent, out)
+            try:
+                detail_start = time.perf_counter()
+                (refined_out,) = nodes.common_ksampler(
                     model, seed, int(detail_steps), cfg, detail_sampler_name, detail_scheduler,
                     positive, negative, detail_latent, denoise=float(detail_denoise))
                 timing["detail_sample"] += time.perf_counter() - detail_start
-            except Exception as e:
-                if _is_oom(e):
-                    e._h3_stage = "sampling"
-                raise
-            refined_out = _video_only_refined_latent(out, refined_out)
-        # Keep a CPU copy of the sampled latent BEFORE decoding, for the `latent`
-        # output. Latents are ~1000x smaller than the frames they decode to (a
-        # 1344x768 124f shot is ~1.5MB against ~1.5GB), so carrying one per shot for
-        # the whole chain is free. Detached and moved off the card immediately, for
-        # the same reason the decoded frames are.
-        decode_video_start = time.perf_counter()
-        shot_latent = _copy_sample_latent(refined_out)
-        video = _decode_video(vae, refined_out, tiled, free_first=model,
-                              tile_t=decode_tile_frames, tile_xy=decode_tile_size)
-        timing["decode_video"] += time.perf_counter() - decode_video_start
-        decode_audio_start = time.perf_counter()
-        audio = _decode_audio(audio_vae, out)
-        timing["decode_audio"] += time.perf_counter() - decode_audio_start
-        cleanup_start = time.perf_counter()
-        del out, refined_out, positive, latent
-        _deep_cleanup()
-        timing["cleanup"] += time.perf_counter() - cleanup_start
+            except Exception as e:
+                if _is_oom(e):
+                    e._h3_stage = "sampling"
+                raise
+            refined_out = _video_only_refined_latent(out, refined_out)
+            del detail_latent
+        # Keep a CPU copy of the sampled latent BEFORE decoding, for the `latent`
+        # output. Latents are ~1000x smaller than the frames they decode to (a
+        # 1344x768 124f shot is ~1.5MB against ~1.5GB), so carrying one per shot for
+        # the whole chain is free. Detached and moved off the card immediately, for
+        # the same reason the decoded frames are.
+        decode_video_start = time.perf_counter()
+        shot_latent = _copy_sample_latent(refined_out)
+        decode_audio_start = time.perf_counter()
+        audio = _decode_audio(audio_vae, out)
+        timing["decode_audio"] += time.perf_counter() - decode_audio_start
+        # Audio is much smaller than the video decode. Drop the first-pass
+        # conditioning before the VAE work so the optional detail pass does not
+        # keep both sampled latents resident across the heaviest allocation.
+        del out, positive, latent
+        video = _decode_video(vae, refined_out, tiled, free_first=model,
+                              tile_t=decode_tile_frames, tile_xy=decode_tile_size)
+        timing["decode_video"] += time.perf_counter() - decode_video_start
+        cleanup_start = time.perf_counter()
+        del refined_out
+        _deep_cleanup()
+        timing["cleanup"] += time.perf_counter() - cleanup_start
         if timing_sink is not None:
             timing["total"] = sum(timing.values())
             timing_sink.append(timing)