Replace deprecated pytorch methods

egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/train.py

@@ -67,7 +67,7 @@
 from model import Transducer
 from optim import Eden, Eve
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 
@@ -638,7 +638,7 @@ def train_one_epoch(
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 
-        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+        with torch.amp.autocast("cuda", enabled=params.use_fp16):
             loss, loss_info = compute_loss(
                 params=params,
                 model=model,
@@ -843,7 +843,7 @@ def remove_short_and_long_utt(c: Cut):
             params=params,
         )
 
-    scaler = GradScaler(enabled=params.use_fp16)
+    scaler = GradScaler("cuda", enabled=params.use_fp16)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -912,7 +912,7 @@ def scan_pessimistic_batches_for_oom(
             # warmup = 0.0 is so that the derivs for the pruned loss stay zero
             # (i.e. are not remembered by the decaying-average in adam), because
             # we want to avoid these params being subject to shrinkage in adam.
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/pruned_transducer_stateless2/train.py

@@ -60,7 +60,7 @@
 from model import Transducer
 from optim import Eden, Eve
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 
@@ -688,7 +688,7 @@ def train_one_epoch(
         batch_size = len(batch["supervisions"]["text"])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -888,7 +888,7 @@ def run(rank, world_size, args):
         params=params,
     )
 
-    scaler = GradScaler(enabled=params.use_fp16)
+    scaler = GradScaler("cuda", enabled=params.use_fp16)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -989,7 +989,7 @@ def scan_pessimistic_batches_for_oom(
             # warmup = 0.0 is so that the derivs for the pruned loss stay zero
             # (i.e. are not remembered by the decaying-average in adam), because
             # we want to avoid these params being subject to shrinkage in adam.
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/pruned_transducer_stateless3/model.py

@@ -184,7 +184,7 @@ def forward(
         lm = simple_lm_proj(decoder_out)
         am = simple_am_proj(encoder_out)
 
-        with torch.cuda.amp.autocast(enabled=False):
+        with torch.amp.autocast("cuda", enabled=False):
             simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
                 lm=lm.float(),
                 am=am.float(),
@@ -219,7 +219,7 @@ def forward(
         # prior to do_rnnt_pruning (this is an optimization for speed).
         logits = joiner(am_pruned, lm_pruned, project_input=False)
 
-        with torch.cuda.amp.autocast(enabled=False):
+        with torch.amp.autocast("cuda", enabled=False):
             pruned_loss = k2.rnnt_loss_pruned(
                 logits=logits.float(),
                 symbols=y_padded,

egs/aishell/ASR/pruned_transducer_stateless3/train.py

@@ -79,7 +79,7 @@
 from model import Transducer
 from optim import Eden, Eve
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 
@@ -797,7 +797,7 @@ def train_one_epoch(
         aishell = is_aishell(batch["supervisions"]["cut"][0])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -1096,7 +1096,7 @@ def run(rank, world_size, args):
                 params=params,
             )
 
-    scaler = GradScaler(enabled=params.use_fp16)
+    scaler = GradScaler("cuda", enabled=params.use_fp16)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -1202,7 +1202,7 @@ def scan_pessimistic_batches_for_oom(
             # warmup = 0.0 is so that the derivs for the pruned loss stay zero
             # (i.e. are not remembered by the decaying-average in adam), because
             # we want to avoid these params being subject to shrinkage in adam.
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py

@@ -74,7 +74,7 @@
 from model import Transducer
 from optim import Eden, ScaledAdam
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 from zipformer import Zipformer
@@ -812,7 +812,7 @@ def train_one_epoch(
         batch_size = len(batch["supervisions"]["text"])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -1107,7 +1107,7 @@ def remove_short_and_long_utt(c: Cut):
     #         params=params,
     #     )
 
-    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    scaler = GradScaler("cuda", enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -1206,7 +1206,7 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/pruned_transducer_stateless7/train.py

@@ -70,7 +70,7 @@
 from model import Transducer
 from optim import Eden, ScaledAdam
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 from zipformer import Zipformer
@@ -809,7 +809,7 @@ def train_one_epoch(
         batch_size = len(batch["supervisions"]["text"])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -1107,7 +1107,7 @@ def remove_short_and_long_utt(c: Cut):
     #         params=params,
     #     )
 
-    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    scaler = GradScaler("cuda", enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -1206,7 +1206,7 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py

@@ -64,7 +64,7 @@
 from model import Transducer
 from optim import Eden, ScaledAdam
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 from zipformer import Zipformer
@@ -802,7 +802,7 @@ def train_one_epoch(
         batch_size = len(batch["supervisions"]["text"])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -1102,7 +1102,7 @@ def tokenize_and_encode_text(c: Cut):
             params=params,
         )
 
-    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    scaler = GradScaler("cuda", enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -1202,7 +1202,7 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py

@@ -63,7 +63,7 @@
 from model import Transducer
 from optim import Eden, ScaledAdam
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 from zipformer_for_ncnn_export_only import Zipformer
@@ -813,7 +813,7 @@ def train_one_epoch(
         batch_size = len(batch["supervisions"]["text"])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -1105,7 +1105,7 @@ def remove_short_and_long_utt(c: Cut):
             params=params,
         )
 
-    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    scaler = GradScaler("cuda", enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -1205,7 +1205,7 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py

@@ -63,7 +63,7 @@
 from model import Transducer
 from optim import Eden, ScaledAdam
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 from zipformer import Zipformer
@@ -812,7 +812,7 @@ def train_one_epoch(
         batch_size = len(batch["supervisions"]["text"])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -1104,7 +1104,7 @@ def remove_short_and_long_utt(c: Cut):
     #         params=params,
     #     )
 
-    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    scaler = GradScaler("cuda", enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -1202,7 +1202,7 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,

egs/aishell/ASR/whisper/train.py

@@ -62,7 +62,7 @@
 from lhotse.utils import fix_random_seed
 from optim import Eden, ScaledAdam
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.functional import pad as pad_tensor
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
@@ -514,7 +514,7 @@ def compute_validation_loss(
     tot_loss = MetricsTracker()
 
     for batch_idx, batch in enumerate(valid_dl):
-        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+        with torch.amp.autocast("cuda", enabled=params.use_fp16):
             loss, loss_info = compute_loss(
                 params=params,
                 tokenizer=tokenizer,
@@ -608,7 +608,7 @@ def train_one_epoch(
                 )
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     tokenizer=tokenizer,
@@ -812,7 +812,7 @@ def run(rank, world_size, args):
     train_dl = aishell.train_dataloaders(aishell.train_cuts())
     valid_dl = aishell.valid_dataloaders(aishell.valid_cuts())
 
-    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    scaler = GradScaler("cuda", enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])

egs/aishell/ASR/zipformer/train.py

@@ -71,7 +71,7 @@
 from scaling import ScheduledFloat
 from subsampling import Conv2dSubsampling
 from torch import Tensor
-from torch.cuda.amp import GradScaler
+from torch.amp import GradScaler
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.tensorboard import SummaryWriter
 from zipformer import Zipformer2
@@ -910,7 +910,7 @@ def save_bad_model(suffix: str = ""):
         batch_size = len(batch["supervisions"]["text"])
 
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, loss_info = compute_loss(
                     params=params,
                     model=model,
@@ -1201,7 +1201,7 @@ def remove_short_and_long_utt(c: Cut):
             params=params,
         )
 
-    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
+    scaler = GradScaler("cuda", enabled=params.use_fp16, init_scale=1.0)
     if checkpoints and "grad_scaler" in checkpoints:
         logging.info("Loading grad scaler state dict")
         scaler.load_state_dict(checkpoints["grad_scaler"])
@@ -1302,7 +1302,7 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            with torch.amp.autocast("cuda", enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
                     model=model,