run-fix-Parameter-init

.github/workflows/doctests.yml

@@ -3,7 +3,7 @@ name: Doctests
 on:
   push:
     branches:
-      - run_doctest*
+      - run-fix-Parameter-init
   repository_dispatch:
   schedule:
     - cron: "17 2 * * *"

.github/workflows/self-scheduled-caller.yml

@@ -7,17 +7,9 @@ on:
     - cron: "17 2 * * *"
   push:
     branches:
-      - run_scheduled_ci*
+      - run-fix-Parameter-init
 
 jobs:
-  model-ci:
-    name: Model CI
-    uses: ./.github/workflows/self-scheduled.yml
-    with:
-      job: run_tests_gpu
-      slack_report_channel: "#transformers-ci-daily-models"
-    secrets: inherit
-
   torch-pipeline:
     name: Torch pipeline CI
     uses: ./.github/workflows/self-scheduled.yml
@@ -25,35 +17,3 @@ jobs:
       job: run_pipelines_torch_gpu
       slack_report_channel: "#transformers-ci-daily-pipeline-torch"
     secrets: inherit
-
-  tf-pipeline:
-    name: TF pipeline CI
-    uses: ./.github/workflows/self-scheduled.yml
-    with:
-      job: run_pipelines_tf_gpu
-      slack_report_channel: "#transformers-ci-daily-pipeline-tf"
-    secrets: inherit
-
-  example-ci:
-    name: Example CI
-    uses: ./.github/workflows/self-scheduled.yml
-    with:
-      job: run_examples_gpu
-      slack_report_channel: "#transformers-ci-daily-examples"
-    secrets: inherit
-
-  deepspeed-ci:
-    name: DeepSpeed CI
-    uses: ./.github/workflows/self-scheduled.yml
-    with:
-      job: run_all_tests_torch_cuda_extensions_gpu
-      slack_report_channel: "#transformers-ci-daily-deepspeed"
-    secrets: inherit
-
-  quantization-ci:
-    name: Quantization CI
-    uses: ./.github/workflows/self-scheduled.yml
-    with:
-      job: run_tests_quantization_torch_gpu
-      slack_report_channel: "#transformers-ci-daily-quantization"
-    secrets: inherit

src/transformers/models/data2vec/modeling_data2vec_audio.py

@@ -822,7 +822,7 @@ class Data2VecAudioModel(Data2VecAudioPreTrainedModel):
 
         # model only needs masking vector if mask prob is > 0.0
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         self.encoder = Data2VecAudioEncoder(config)
 
@@ -858,7 +858,7 @@ class Data2VecAudioModel(Data2VecAudioPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -868,7 +868,7 @@ class Data2VecAudioModel(Data2VecAudioPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/hubert/modeling_hubert.py

@@ -974,7 +974,7 @@ class HubertModel(HubertPreTrainedModel):
         self.feature_projection = HubertFeatureProjection(config)
 
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         if config.do_stable_layer_norm:
             self.encoder = HubertEncoderStableLayerNorm(config)
@@ -1005,7 +1005,7 @@ class HubertModel(HubertPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1015,7 +1015,7 @@ class HubertModel(HubertPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/sew/modeling_sew.py

@@ -834,7 +834,7 @@ class SEWModel(SEWPreTrainedModel):
         self.feature_dropout = nn.Dropout(config.feat_proj_dropout)
 
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         self.encoder = SEWEncoder(config)
 
@@ -862,7 +862,7 @@ class SEWModel(SEWPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -872,7 +872,7 @@ class SEWModel(SEWPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/sew_d/modeling_sew_d.py

@@ -1360,7 +1360,7 @@ class SEWDModel(SEWDPreTrainedModel):
         self.feature_dropout = nn.Dropout(config.feat_proj_dropout)
 
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         self.encoder = SEWDEncoder(config)
 
@@ -1388,7 +1388,7 @@ class SEWDModel(SEWDPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1398,7 +1398,7 @@ class SEWDModel(SEWDPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/speecht5/modeling_speecht5.py

@@ -517,7 +517,7 @@ class SpeechT5SpeechEncoderPrenet(nn.Module):
 
         # model only needs masking vector if mask prob is > 0.0
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         self.pos_conv_embed = SpeechT5PositionalConvEmbedding(config)
         self.pos_sinusoidal_embed = SpeechT5SinusoidalPositionalEmbedding(
@@ -616,7 +616,7 @@ class SpeechT5SpeechEncoderPrenet(nn.Module):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -626,7 +626,7 @@ class SpeechT5SpeechEncoderPrenet(nn.Module):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/unispeech/modeling_unispeech.py

@@ -1090,7 +1090,7 @@ class UniSpeechModel(UniSpeechPreTrainedModel):
         self.feature_projection = UniSpeechFeatureProjection(config)
 
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         if config.do_stable_layer_norm:
             self.encoder = UniSpeechEncoderStableLayerNorm(config)
@@ -1121,7 +1121,7 @@ class UniSpeechModel(UniSpeechPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1131,7 +1131,7 @@ class UniSpeechModel(UniSpeechPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/unispeech_sat/modeling_unispeech_sat.py

@@ -1108,7 +1108,7 @@ class UniSpeechSatModel(UniSpeechSatPreTrainedModel):
         self.feature_extractor = UniSpeechSatFeatureEncoder(config)
         self.feature_projection = UniSpeechSatFeatureProjection(config)
 
-        self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+        self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         if config.do_stable_layer_norm:
             self.encoder = UniSpeechSatEncoderStableLayerNorm(config)
@@ -1139,7 +1139,7 @@ class UniSpeechSatModel(UniSpeechSatPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1149,7 +1149,7 @@ class UniSpeechSatModel(UniSpeechSatPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -1445,7 +1445,7 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
 
         # model only needs masking vector if mask prob is > 0.0
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         if config.do_stable_layer_norm:
             self.encoder = Wav2Vec2EncoderStableLayerNorm(config)
@@ -1496,7 +1496,7 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1506,7 +1506,7 @@ class Wav2Vec2Model(Wav2Vec2PreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/wav2vec2_bert/modeling_wav2vec2_bert.py

@@ -1053,7 +1053,7 @@ class Wav2Vec2BertModel(Wav2Vec2BertPreTrainedModel):
 
         # model only needs masking vector if mask prob is > 0.0
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         self.encoder = Wav2Vec2BertEncoder(config)
 
@@ -1087,7 +1087,7 @@ class Wav2Vec2BertModel(Wav2Vec2BertPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1097,7 +1097,7 @@ class Wav2Vec2BertModel(Wav2Vec2BertPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/wav2vec2_conformer/modeling_wav2vec2_conformer.py

@@ -1235,7 +1235,7 @@ class Wav2Vec2ConformerModel(Wav2Vec2ConformerPreTrainedModel):
 
         # model only needs masking vector if mask prob is > 0.0
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         self.encoder = Wav2Vec2ConformerEncoder(config)
 
@@ -1273,7 +1273,7 @@ class Wav2Vec2ConformerModel(Wav2Vec2ConformerPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1283,7 +1283,7 @@ class Wav2Vec2ConformerModel(Wav2Vec2ConformerPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis

src/transformers/models/wavlm/modeling_wavlm.py

@@ -1107,7 +1107,7 @@ class WavLMModel(WavLMPreTrainedModel):
 
         # model only needs masking vector if mask prob is > 0.0
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
-            self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
+            self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
 
         if config.do_stable_layer_norm:
             self.encoder = WavLMEncoderStableLayerNorm(config)
@@ -1158,7 +1158,7 @@ class WavLMModel(WavLMPreTrainedModel):
 
         if mask_time_indices is not None:
             # apply SpecAugment along time axis with given mask_time_indices
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
         elif self.config.mask_time_prob > 0 and self.training:
             mask_time_indices = _compute_mask_indices(
                 (batch_size, sequence_length),
@@ -1168,7 +1168,7 @@ class WavLMModel(WavLMPreTrainedModel):
                 min_masks=self.config.mask_time_min_masks,
             )
             mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
-            hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
+            hidden_states[mask_time_indices] = self.masked_spec_embed
 
         if self.config.mask_feature_prob > 0 and self.training:
             # generate indices & apply SpecAugment along feature axis