remove flow_cache

cosyvoice/utils/file_utils.py

@@ -56,7 +56,7 @@ def convert_onnx_to_trt(trt_model, trt_kwargs, onnx_model, fp16):
     network = builder.create_network(network_flags)
     parser = trt.OnnxParser(network, logger)
     config = builder.create_builder_config()
-    config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 30)  # 1GB
+    config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 31)  # 1GB
     if fp16:
         config.set_flag(trt.BuilderFlag.FP16)
     profile = builder.create_optimization_profile()

cosyvoice/utils/mask.py

@@ -86,7 +86,7 @@ def subsequent_mask(
     return mask
 
 
-def subsequent_chunk_mask(
+def subsequent_chunk_mask_deprecated(
         size: int,
         chunk_size: int,
         num_left_chunks: int = -1,
@@ -124,6 +124,40 @@ def subsequent_chunk_mask(
     return ret
 
 
+def subsequent_chunk_mask(
+        size: int,
+        chunk_size: int,
+        num_left_chunks: int = -1,
+        device: torch.device = torch.device("cpu"),
+) -> torch.Tensor:
+    """Create mask for subsequent steps (size, size) with chunk size,
+       this is for streaming encoder
+
+    Args:
+        size (int): size of mask
+        chunk_size (int): size of chunk
+        num_left_chunks (int): number of left chunks
+            <0: use full chunk
+            >=0: use num_left_chunks
+        device (torch.device): "cpu" or "cuda" or torch.Tensor.device
+
+    Returns:
+        torch.Tensor: mask
+
+    Examples:
+        >>> subsequent_chunk_mask(4, 2)
+        [[1, 1, 0, 0],
+         [1, 1, 0, 0],
+         [1, 1, 1, 1],
+         [1, 1, 1, 1]]
+    """
+    # NOTE this modified implementation meets onnx export requirements, but it doesn't support num_left_chunks
+    pos_idx = torch.arange(size, device=device)
+    block_value = (torch.div(pos_idx, chunk_size, rounding_mode='trunc') + 1) * chunk_size
+    ret = pos_idx.unsqueeze(0) < block_value.unsqueeze(1)
+    return ret
+
+
 def add_optional_chunk_mask(xs: torch.Tensor,
                             masks: torch.Tensor,
                             use_dynamic_chunk: bool,
@@ -196,9 +230,6 @@ def add_optional_chunk_mask(xs: torch.Tensor,
     else:
         chunk_masks = masks
     assert chunk_masks.dtype == torch.bool
-    if (chunk_masks.sum(dim=-1) == 0).sum().item() != 0:
-        print('get chunk_masks all false at some timestep, force set to true, make sure they are masked in futuer computation!')
-        chunk_masks[chunk_masks.sum(dim=-1) == 0] = True
     return chunk_masks