Re-enable cuda graphs in training modes. (NVIDIA#9338) (NVIDIA#9343)

* Re-enable cuda graphs in training modes.

"global" capture mode was sporadically crashing because of pinning
host memory in other threads spawned by the data loader when
num_workers > 0.

Add relevant changs to TDT cuda graphs decoding as well.

I didn't test the TDT change because I'm not sure how. But it seems low risk.



* Apply isort and black reformatting



---------

Signed-off-by: Daniel Galvez <dgalvez@nvidia.com>
Signed-off-by: galv <galv@users.noreply.github.com>
Co-authored-by: Daniel Galvez <galv@users.noreply.github.com>
Co-authored-by: Somshubra Majumdar <titu1994@gmail.com>

examples/asr/transcribe_speech.py

@@ -164,9 +164,7 @@ class TranscriptionConfig:
 
     # Decoding strategy for RNNT models
     # enable CUDA graphs for transcription
-    rnnt_decoding: RNNTDecodingConfig = RNNTDecodingConfig(
-        fused_batch_size=-1, greedy=GreedyBatchedRNNTInferConfig(use_cuda_graph_decoder=True)
-    )
+    rnnt_decoding: RNNTDecodingConfig = RNNTDecodingConfig(fused_batch_size=-1)
 
     # Decoding strategy for AED models
     multitask_decoding: MultiTaskDecodingConfig = MultiTaskDecodingConfig()

examples/asr/transcribe_speech_parallel.py

@@ -101,10 +101,8 @@ class ParallelTranscriptionConfig:
     use_cer: bool = False
 
     # decoding strategy for RNNT models
-    # enable CUDA graphs for transcription
-    rnnt_decoding: RNNTDecodingConfig = RNNTDecodingConfig(
-        fused_batch_size=-1, greedy=GreedyBatchedRNNTInferConfig(use_cuda_graph_decoder=True)
-    )
+    # Double check whether fused_batch_size=-1 is right
+    rnnt_decoding: RNNTDecodingConfig = RNNTDecodingConfig(fused_batch_size=-1)
 
     # decoder type: ctc or rnnt, can be used to switch between CTC and RNNT decoder for Hybrid RNNT/CTC models
     decoder_type: Optional[str] = None

nemo/collections/asr/parts/submodules/cuda_graph_rnnt_greedy_decoding.py

@@ -37,7 +37,7 @@
 
 def create_outer_for_loop_kernel():
     """
-    Creates a kernel that evaluates whether or not to enter the for loop body. 
+    Creates a kernel that evaluates whether or not to enter the for loop body.
     Effectively substitutes for `for time_idx in range(trip_count)`
     such that that for loop can run on a GPU.
     """
@@ -171,8 +171,10 @@ def _reinitialize(self, max_time, batch_size, encoder_output, encoder_output_len
 
         # Always create a new stream, because the per-thread default stream disallows stream capture to a graph.
         stream_for_graph = torch.cuda.Stream(self.device)
-        with torch.cuda.stream(stream_for_graph), torch.inference_mode(), torch.cuda.graph(
-            self.graph, stream=stream_for_graph
+        with (
+            torch.cuda.stream(stream_for_graph),
+            torch.inference_mode(),
+            torch.cuda.graph(self.graph, stream=stream_for_graph, capture_error_mode="thread_local"),
         ):
             # This is failing...
             self.f = torch.zeros(

nemo/collections/asr/parts/submodules/rnnt_decoding.py

@@ -331,7 +331,7 @@ def __init__(self, decoding_cfg, decoder, joint, blank_id: int):
                         preserve_frame_confidence=self.preserve_frame_confidence,
                         confidence_method_cfg=self.confidence_method_cfg,
                         loop_labels=self.cfg.greedy.get('loop_labels', True),
-                        use_cuda_graph_decoder=self.cfg.greedy.get('use_cuda_graph_decoder', False),
+                        use_cuda_graph_decoder=self.cfg.greedy.get('use_cuda_graph_decoder', True),
                     )
                 else:
                     self.decoding = rnnt_greedy_decoding.GreedyBatchedTDTInfer(
@@ -347,7 +347,7 @@ def __init__(self, decoding_cfg, decoder, joint, blank_id: int):
                         preserve_frame_confidence=self.preserve_frame_confidence,
                         include_duration_confidence=self.tdt_include_duration_confidence,
                         confidence_method_cfg=self.confidence_method_cfg,
-                        use_cuda_graph_decoder=self.cfg.greedy.get('use_cuda_graph_decoder', False),
+                        use_cuda_graph_decoder=self.cfg.greedy.get('use_cuda_graph_decoder', True),
                     )
 
             else:

nemo/collections/asr/parts/submodules/rnnt_greedy_decoding.py

@@ -592,7 +592,7 @@ def __init__(
         preserve_frame_confidence: bool = False,
         confidence_method_cfg: Optional[DictConfig] = None,
         loop_labels: bool = True,
-        use_cuda_graph_decoder: bool = False,
+        use_cuda_graph_decoder: bool = True,
     ):
         super().__init__(
             decoder_model=decoder_model,
@@ -2360,7 +2360,7 @@ class GreedyBatchedRNNTInferConfig:
     tdt_include_duration_confidence: bool = False
     confidence_method_cfg: Optional[ConfidenceMethodConfig] = field(default_factory=lambda: ConfidenceMethodConfig())
     loop_labels: bool = True
-    use_cuda_graph_decoder: bool = False
+    use_cuda_graph_decoder: bool = True
 
     def __post_init__(self):
         # OmegaConf.structured ensures that post_init check is always executed
@@ -2712,7 +2712,7 @@ def __init__(
         preserve_frame_confidence: bool = False,
         include_duration_confidence: bool = False,
         confidence_method_cfg: Optional[DictConfig] = None,
-        use_cuda_graph_decoder: bool = False,
+        use_cuda_graph_decoder: bool = True,
     ):
         super().__init__(
             decoder_model=decoder_model,

nemo/collections/asr/parts/submodules/rnnt_loop_labels_computer.py

@@ -630,42 +630,49 @@ def _partial_graphs_compile(self):
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.before_outer_loop, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.before_outer_loop, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._before_outer_loop()
 
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.before_inner_loop, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.before_inner_loop, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._before_inner_loop_get_decoder_output()
             self._before_inner_loop_get_joint_output()
 
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.inner_loop_code, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.inner_loop_code, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._inner_loop_code()
 
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.after_inner_loop, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.after_inner_loop, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._after_inner_loop()
 
     def _full_graph_compile(self):
         """Compile full graph for decoding"""
         # Always create a new stream, because the per-thread default stream disallows stream capture to a graph.
         stream_for_graph = torch.cuda.Stream(self.state.device)
-        stream_for_graph.wait_stream(torch.cuda.default_stream(self.state.device))
         self.full_graph = torch.cuda.CUDAGraph()
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.full_graph, stream=stream_for_graph),
+            torch.cuda.graph(self.full_graph, stream=stream_for_graph, capture_error_mode="thread_local"),
         ):
             self._before_outer_loop()
 

nemo/collections/asr/parts/submodules/tdt_loop_labels_computer.py

@@ -691,29 +691,37 @@ def _partial_graphs_compile(self):
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.before_outer_loop, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.before_outer_loop, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._before_outer_loop()
 
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.before_inner_loop, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.before_inner_loop, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._before_inner_loop_get_decoder_output()
             self._before_inner_loop_get_joint_output()
 
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.inner_loop_code, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.inner_loop_code, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._inner_loop_code()
 
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.separate_graphs.after_inner_loop, stream=stream_for_graph),
+            torch.cuda.graph(
+                self.separate_graphs.after_inner_loop, stream=stream_for_graph, capture_error_mode="thread_local"
+            ),
         ):
             self._after_inner_loop()
 
@@ -726,7 +734,7 @@ def _full_graph_compile(self):
         with (
             torch.cuda.stream(stream_for_graph),
             torch.inference_mode(),
-            torch.cuda.graph(self.full_graph, stream=stream_for_graph),
+            torch.cuda.graph(self.full_graph, stream=stream_for_graph, capture_error_mode="thread_local"),
         ):
             self._before_outer_loop()