Update mimi export test (#9755)

Summary:

Feed embeddings to mimi decoder directly.

Reviewed By: iseeyuan

Differential Revision: D72091091

Author: limintang

examples/models/moshi/mimi/test_mimi.py

@@ -28,7 +28,6 @@
     "https": "http://fwdproxy:8080",
 }
 
-
 def compute_sqnr(x: torch.Tensor, y: torch.Tensor) -> float:
     assert x.shape == y.shape, "Tensor shapes do not match"
     x = x.float()
@@ -39,7 +38,6 @@ def compute_sqnr(x: torch.Tensor, y: torch.Tensor) -> float:
     sqnr = 10 * torch.log10(original_power / error_power)
     return sqnr.item()
 
-
 def read_mp3_from_url(url):
     try:
         response = requests.get(url)
@@ -173,15 +171,23 @@ def __init__(self, mimi: nn.Module):
                 self.mimi_model = mimi
 
             def forward(self, x):
-                return self.mimi_model.decode(x)
+                x = x.transpose(1, 2)
+                x = self.mimi_model.upsample(x)
+                (emb,) = self.mimi_model.decoder_transformer(x)
+                emb.transpose(1, 2)
+                with self.mimi_model._context_for_encoder_decoder:
+                    out = self.mimi_model.decoder(emb)
+                return out
 
-        sample_pcm = torch.tensor(self.sample_pcm, device=self.device)[None]
-        pcm_chunk_size = int(self.mimi.sample_rate / self.mimi.frame_rate)
-        chunk = sample_pcm[..., 0:pcm_chunk_size]
-        input = self.mimi.encode(chunk)
+        emb_input = torch.rand(1, 1, 512, device="cpu")
 
         mimi_decode = MimiDecode(self.mimi)
-        exported_decode: ExportedProgram = export(mimi_decode, (input,), strict=False)
+        mimi_decode.eval()
+        mimi_decode(emb_input)
+
+        exported_decode: ExportedProgram = export(
+            mimi_decode, (emb_input,), strict=False
+        )
         quantization_config = get_symmetric_quantization_config(
             is_per_channel=True,
             is_dynamic=True,
@@ -190,12 +196,12 @@ def forward(self, x):
         quantizer.set_global(quantization_config)
         m = exported_decode.module()
         m = prepare_pt2e(m, quantizer)
-        m(input)
+        m(emb_input)
         m = convert_pt2e(m)
         print("quantized graph:")
         print(m.graph)
         # Export quantized module
-        exported_decode: ExportedProgram = export(m, (input,), strict=False)
+        exported_decode: ExportedProgram = export(m, (emb_input,), strict=False)
         # Lower
         edge_manager = to_edge_transform_and_lower(
             exported_decode,
@@ -208,16 +214,16 @@ def forward(self, x):
         with open(output_file, "wb") as file:
             exec_prog.write_to_file(file)
 
-        eager_res = mimi_decode(input)
+        eager_res = mimi_decode(emb_input)
         runtime = Runtime.get()
         program = runtime.load_program(output_file)
         method = program.load_method("forward")
-        flattened_x = tree_flatten(input)[0]
+        flattened_x = tree_flatten(emb_input)[0]
         res = method.execute(flattened_x)
         # Compare results
         sqnr = compute_sqnr(eager_res, res[0])
         print(f"SQNR: {sqnr}")
-        torch.testing.assert_close(eager_res, res[0], atol=1e-3, rtol=1e-3)
+        torch.testing.assert_close(eager_res, res[0], atol=4e-3, rtol=1e-3)
 
 
 if __name__ == "__main__":