export: put dummy tensors on the model device

kaggle/export_browser.py

@@ -89,17 +89,18 @@ def stage_lm(mm, torch) -> dict:
             return (out.hidden_states[-1], *flat)
 
     wrapper = Wrapper().float().eval()
+    dev = next(lm.parameters()).device
     B, S, P = 1, 4, 6
     H = cfg.hidden_size
-    embeds = torch.randn(B, S, H)
-    mask = torch.zeros(B, 1, S, P + S)
+    embeds = torch.randn(B, S, H, device=dev)
+    mask = torch.zeros(B, 1, S, P + S, device=dev)
     for i in range(S):
         mask[:, :, i, :P + i + 1] = 1.0
-    pos = torch.arange(P, P + S).unsqueeze(0)
+    pos = torch.arange(P, P + S, device=dev).unsqueeze(0)
     past = []
     for _ in range(n):
-        past.append(torch.randn(B, kv_heads, P, head_dim))
-        past.append(torch.randn(B, kv_heads, P, head_dim))
+        past.append(torch.randn(B, kv_heads, P, head_dim, device=dev))
+        past.append(torch.randn(B, kv_heads, P, head_dim, device=dev))
 
     with torch.no_grad():
         ref = wrapper(embeds, mask, pos, *past)
@@ -141,7 +142,8 @@ def stage_agg(mm, torch) -> dict:
 
     lin = mm.linear_proj_audio
     dim_in = mm.latent_dim
-    dummy = torch.randn(1, 1, dim_in)
+    dev = next(lin.parameters()).device
+    dummy = torch.randn(1, 1, dim_in, device=dev)
     with torch.no_grad():
         ref = lin(dummy)
     path = OUT / "aggregator.onnx"
@@ -174,7 +176,8 @@ def stage_voc(mm, torch) -> dict:
 
     voc = Voc().float().eval()
     T = 8
-    dummy = torch.randn(1, T, mm.latent_dim)
+    dev = next(audio.parameters()).device
+    dummy = torch.randn(1, T, mm.latent_dim, device=dev)
     with torch.no_grad():
         ref = voc(dummy)
     log(f"voc ref shape {tuple(ref.shape)}")