ar_generate: sample under no_grad, detach latents (fixes leak)

kaggle/train_voices.py

@@ -310,8 +310,11 @@ def ar_generate(mm, dit, n_steps, text, task_prompt, prompt_latent, prompt_text,
         past = hidden.past_key_values
         z = hidden.hidden_states[-1][:, -1:, :]
         noise = torch.randn(1, 1, mm.latent_dim, device=device, dtype=dtype)
-        latent = sample_latent(dit, z, hist, noise, n_steps).to(device=device, dtype=dtype)
-        out.append(latent)
+        # no_grad matters: the DiT params require grad, so an unguarded sample
+        # builds a graph per frame that stays alive through `hist`.
+        with torch.no_grad():
+            latent = sample_latent(dit, z, hist, noise, n_steps).to(device=device, dtype=dtype)
+        out.append(latent.detach())
         inputs_embeds = mm.linear_proj_audio(latent)
         position_ids = position_ids[:, -1:] + 1
         attention_mask = torch.ones(inputs_embeds.shape[0], 1, device=device)