support torch>=2.9.0 (#3032)

eschmidbauer · web-flow · commit d48d36cb0a1f · 2026-02-11T14:26:13.000-05:00
diff --git a/docs/readthedocs-requirements.txt b/docs/readthedocs-requirements.txt
@@ -4,4 +4,4 @@
 
 -r ../requirements.txt
 -r docs-requirements.txt
-torch==2.8.0
+torch==2.9.0
diff --git a/recipes/LibriSpeech/ASR/transformer/extract_ssl_feats.py b/recipes/LibriSpeech/ASR/transformer/extract_ssl_feats.py
@@ -58,7 +58,9 @@ def audio_pipeline(wav):
     def compute_feats(uid, sig):
         sig = sig.to(hparams["device"]).unsqueeze(0)
         length = torch.ones(1, device=hparams["device"])
-        with torch.no_grad(), torch.cuda.amp.autocast(dtype=hparams["dtype"]):
+        with torch.no_grad(), torch.amp.autocast(
+            hparams["device"].type, dtype=hparams["dtype"]
+        ):
             feats = normalizer(sig, length)
             feats = ssl_encoder(feats, length)
         return feats.squeeze(0).cpu()
diff --git a/requirements.txt b/requirements.txt
@@ -10,7 +10,7 @@ requests>=2.20.0
 scipy>=1.4.1
 sentencepiece>=0.1.91
 soundfile>=0.12.1
-torch>=2.1.0,<2.9
-torchaudio>=2.1.0,<2.9
+torch>=2.1.0
+torchaudio>=2.1.0
 tqdm>=4.42.0
 transformers>=4.30.0
diff --git a/speechbrain/processing/multi_mic.py b/speechbrain/processing/multi_mic.py
@@ -80,7 +80,6 @@
 """
 
 import torch
-from packaging import version
 
 import speechbrain.processing.decomposition as eig
 
@@ -765,11 +764,8 @@ def _gcc_phat(XXs, eps=1e-20):
         # Returning in the temporal domain
         XXs_phat = XXs_phat.transpose(2, 3)
 
-        if version.parse(torch.__version__) >= version.parse("1.8.0"):
-            XXs_phat = torch.complex(XXs_phat[..., 0], XXs_phat[..., 1])
-            xxs = torch.fft.irfft(XXs_phat, n=n_samples)
-        else:
-            xxs = torch.irfft(XXs_phat, signal_ndim=1, signal_sizes=[n_samples])
+        XXs_phat = torch.complex(XXs_phat[..., 0], XXs_phat[..., 1])
+        xxs = torch.fft.irfft(XXs_phat, n=n_samples)
 
         xxs = xxs[..., XXs_idx, :]
 
diff --git a/speechbrain/processing/signal_processing.py b/speechbrain/processing/signal_processing.py
@@ -12,7 +12,6 @@
 import math
 
 import torch
-from packaging import version
 
 
 def compute_amplitude(waveforms, lengths=None, amp_type="avg", scale="linear"):
@@ -280,26 +279,10 @@ def convolve1d(
         kernel = torch.cat((after_index, zeros, before_index), dim=-1)
 
         # Multiply in frequency domain to convolve in time domain
-        if version.parse(torch.__version__) > version.parse("1.6.0"):
-            import torch.fft as fft
+        import torch.fft as fft
 
-            result = fft.rfft(waveform) * fft.rfft(kernel)
-            convolved = fft.irfft(result, n=waveform.size(-1))
-        else:
-            f_signal = torch.rfft(waveform, 1)
-            f_kernel = torch.rfft(kernel, 1)
-            sig_real, sig_imag = f_signal.unbind(-1)
-            ker_real, ker_imag = f_kernel.unbind(-1)
-            f_result = torch.stack(
-                [
-                    sig_real * ker_real - sig_imag * ker_imag,
-                    sig_real * ker_imag + sig_imag * ker_real,
-                ],
-                dim=-1,
-            )
-            convolved = torch.irfft(
-                f_result, 1, signal_sizes=[waveform.size(-1)]
-            )
+        result = fft.rfft(waveform) * fft.rfft(kernel)
+        convolved = fft.irfft(result, n=waveform.size(-1))
 
     # Use the implementation given by torch, which should be efficient on GPU
     else:
diff --git a/speechbrain/utils/checkpoints.py b/speechbrain/utils/checkpoints.py
@@ -276,12 +276,8 @@ def torch_parameter_transfer(obj, path):
     torch.optim.Optimizer: torch_save,
     torch.optim.lr_scheduler.ReduceLROnPlateau: torch_save,
 }
-if version.parse(torch.__version__) < version.parse("2.0.0"):
-    DEFAULT_LOAD_HOOKS[torch.optim.lr_scheduler._LRScheduler] = torch_recovery
-    DEFAULT_SAVE_HOOKS[torch.optim.lr_scheduler._LRScheduler] = torch_save
-else:
-    DEFAULT_LOAD_HOOKS[torch.optim.lr_scheduler.LRScheduler] = torch_recovery
-    DEFAULT_SAVE_HOOKS[torch.optim.lr_scheduler.LRScheduler] = torch_save
+DEFAULT_LOAD_HOOKS[torch.optim.lr_scheduler.LRScheduler] = torch_recovery
+DEFAULT_SAVE_HOOKS[torch.optim.lr_scheduler.LRScheduler] = torch_save
 
 if version.parse(torch.__version__) < version.parse("2.4.0"):
     DEFAULT_LOAD_HOOKS[torch.cuda.amp.grad_scaler.GradScaler] = torch_recovery
