Add support for Numba FP16 RNNT Loss

nemo/collections/asr/losses/rnnt.py

@@ -38,9 +38,10 @@
 from nemo.collections.asr.losses.rnnt_pytorch import MultiblankRNNTLossPytorch, RNNTLossPytorch, TDTLossPytorch
 from nemo.core.classes import Loss, typecheck
 from nemo.core.neural_types import LabelsType, LengthsType, LogprobsType, LossType, NeuralType
+from nemo.core.utils import numba_utils
 from nemo.core.utils.k2_utils import K2_INSTALLATION_MESSAGE
 from nemo.core.utils.numba_utils import NUMBA_INSTALLATION_MESSAGE
-from nemo.utils import logging, model_utils
+from nemo.utils import logging, logging_mode, model_utils
 
 try:
     import warprnnt_pytorch as warprnnt
@@ -98,7 +99,7 @@
         min_version='0.53.0',
         is_available=NUMBA_RNNT_AVAILABLE,
         installation_msg=NUMBA_INSTALLATION_MESSAGE,
-        force_float32=True,
+        force_float32=not numba_utils.NUMBA_FP16_SUPPORTED,
     ),
     "pytorch": RNNTLossConfig(
         loss_name="pytorch",
@@ -387,7 +388,7 @@
                 for the standard "blank" symbol. In particular, say V is the number of non-blank tokens in
                 the vocabulary, then in the case of,
                 standard RNNT: num_classes = V
-                multiblank RNNT: num_classes = V + number-big-blanks (since we store big-blanks before 
+                multiblank RNNT: num_classes = V + number-big-blanks (since we store big-blanks before
                                  standard blank, and the standard blank is the last symbol in the vocab)
                 TDT: num_classes = V. Note, V here does not include any of the "duration outputs".
 
@@ -413,6 +414,7 @@
         self.reduction = reduction
         self._loss = resolve_rnnt_loss(loss_name, blank_idx=self._blank, loss_kwargs=loss_kwargs)
         self._force_float32 = RNNT_LOSS_RESOLVER[loss_name].force_float32
+        self._fp16_compat_checked = False
 
     def reduce(self, losses, target_lengths):
 
@@ -442,8 +444,21 @@
         max_targets_len = target_lengths.max()
 
         # Force cast joint to float32
-        # TODO: Remove once Numba supports FP16
-        if self._force_float32 and log_probs.dtype != torch.float32:
+        if not self._force_float32 and numba_utils.NUMBA_FP16_SUPPORTED:
+            # Execute the kernel in fp16
+            pass
+        elif self._force_float32 and log_probs.dtype != torch.float32:
+            # Log just once if fp16 tensor was passed and fp16 Numba CUDA loss could not be used.
+            if log_probs.dtype == torch.float16 and not self._fp16_compat_checked:
+                _, reason = numba_utils.is_numba_cuda_fp16_supported(return_reason=True)
+                logging.warning(
+                    f"Provided RNNT Joint tensor is of dtype {log_probs.dtype}, but RNNT loss could not be calculated "
+                    f"in fp16 due to following reason stated below. Loss will be calculated in fp32. \n\n"
+                    f"{reason}"
+                )
+                self._fp16_compat_checked = True
+
+            # Upcast the activation tensor and compute loss and grads in fp32
             logits_orig = log_probs
             log_probs = log_probs.float()
             del logits_orig  # save memory *before* computing the loss

nemo/collections/asr/losses/rnnt_pytorch.py

@@ -47,7 +47,12 @@ def __init__(self, blank, reduction):
         self.reduction = reduction
 
     def forward(self, acts, labels, act_lens, label_lens):
+        # CPU patch for FP16
+        if not acts.is_cuda and acts.dtype == torch.float16:
+            acts = acts.float()
+
         acts = torch.log_softmax(acts, -1)
+
         forward_logprob = self.compute_forward_prob(acts, labels, act_lens, label_lens)
         losses = -forward_logprob
         if self.reduction == 'mean_batch':

nemo/collections/asr/parts/numba/rnnt_loss/rnnt.py

@@ -186,7 +186,7 @@ def rnnt_loss_gpu(
 
     # Select GPU index
     cuda.select_device(acts.device.index)
-    gpu_workspace = torch.zeros(gpu_size, device=acts.device, dtype=acts.dtype, requires_grad=False)
+    gpu_workspace = torch.zeros(gpu_size, device=acts.device, dtype=torch.float32, requires_grad=False)
 
     ### VIEW TENSORS AS VECTORS FOR POINTER INDEXING ###
     acts, acts_shape = rnnt_helper.flatten_tensor(acts)

nemo/collections/asr/parts/numba/rnnt_loss/rnnt_numpy.py

@@ -344,10 +344,15 @@ def forward(self, acts, labels, act_lens, label_lens):
         _assert_no_grad(label_lens)
         certify_inputs(acts, labels, act_lens, label_lens)
 
+        # CPU Patch for fp16 - force cast to fp32
+        if not acts.is_cuda and acts.dtype == torch.float16:
+            acts = acts.float()
+
         if self.clamp > 0.0:
             acts = LogSoftmaxGradModification.apply(acts, self.clamp)
 
         acts = torch.nn.functional.log_softmax(acts, -1)
+
         return self.rnnt(acts, labels, act_lens, label_lens, self.blank, self.fastemit_lambda)
 
 

nemo/collections/asr/parts/numba/rnnt_loss/rnnt_pytorch.py

@@ -57,7 +57,7 @@ def forward(ctx, acts, labels, act_lens, label_lens, blank, reduction, fastemit_
         loss_func = rnnt.rnnt_loss_gpu if is_cuda else rnnt.rnnt_loss_cpu
         grads = torch.zeros_like(acts) if acts.requires_grad else None
         minibatch_size = acts.size(0)
-        costs = torch.zeros(minibatch_size, device=acts.device, dtype=acts.dtype)
+        costs = torch.zeros(minibatch_size, device=acts.device, dtype=torch.float32)
 
         loss_func(
             acts,
@@ -119,7 +119,6 @@ def forward(
         label_lens: Tensor of (batch) containing label length of each example
         fastemit_lambda: Float scaling factor for FastEmit regularization. Refer to
             FastEmit: Low-latency Streaming ASR with Sequence-level Emission Regularization.
-
         durations: list of durations for TDT model, must include 0 and 1, e.g.
             [0, 1, 2, 3, 4].
         sigma: hyper-parameter for logit under-normalization method for training
@@ -417,6 +416,10 @@ def forward(self, acts, labels, act_lens, label_lens):
         label_lens: Tensor of (batch) containing label length of each example
         """
         if not acts.is_cuda:
+            # Force FP32 until log_softmax() is implemented for fp16 on CPU
+            if acts.dtype == torch.float16:
+                acts = acts.float()
+
             # Since CPU requires log_softmax to be computed explicitly, we need to perform grad clipping
             # *after* we have obtained the gradients of loss(logsoftmax()).
             # This is highly wasteful since it requires a copy of the entire joint tensor which is expensive.

nemo/collections/asr/parts/numba/rnnt_loss/utils/cpu_utils/cpu_rnnt.py

@@ -231,8 +231,8 @@ def cost_and_grad_kernel(
         )
 
         # Scale llForward by FastEmit lambda
-        llForward *= 1.0 + self.fastemit_lambda_
-        llBackward *= 1.0 + self.fastemit_lambda_
+        llForward += llForward * self.fastemit_lambda_
+        llBackward += llBackward * self.fastemit_lambda_
 
         diff = (llForward - llBackward).abs()
         if diff > 0.1:
@@ -300,6 +300,10 @@ def compute_betas_and_grads(
         Returns:
             Loglikelihood of the forward variable and inplace updates the grad tensor.
         """
+        # Patch for CPU + fp16
+        if log_probs.dtype == torch.float16 and not log_probs.is_cuda:
+            log_probs = log_probs.float()
+
         idx = CpuRNNT_index(U, self.maxU_, self.minibatch_, self.alphabet_size_, self.batch_first)
         betas[idx(T - 1, U - 1)] = log_probs[idx(T - 1, U - 1) * 2]
 

nemo/collections/asr/parts/numba/rnnt_loss/utils/rnnt_helper.py

@@ -30,6 +30,7 @@
 import math
 from typing import Optional, Tuple
 
+import numba
 import torch
 from numba import cuda
 
@@ -112,7 +113,7 @@
     if idx < length:
         copy_data_1d(source, dest, idx)
         dest[idx] *= -1.0
-        dest[idx] *= 1.0 + fastemit_lambda
+        dest[idx] *= numba.float32(1.0 + fastemit_lambda)
 
 
 def get_workspace_size(

nemo/core/utils/numba_utils.py

@@ -26,6 +26,11 @@
 __NUMBA_DEFAULT_MINIMUM_VERSION__ = "0.53.0"
 __NUMBA_MINIMUM_VERSION__ = os.environ.get("NEMO_NUMBA_MINVER", __NUMBA_DEFAULT_MINIMUM_VERSION__)
 
+__NUMBA_MINIMUM_VERSION_FP16_SUPPORTED__ = "0.57.0"
+NUMBA_FP16_SUPPORTED = model_utils.check_lib_version(
+    'numba', __NUMBA_MINIMUM_VERSION_FP16_SUPPORTED__, operator=operator.ge
+)[0]
+
 
 NUMBA_INSTALLATION_MESSAGE = (
     "Could not import `numba`.\n"
@@ -148,6 +153,35 @@ def numba_cuda_is_supported(min_version: str) -> bool:
         return False
 
 
+def is_numba_cuda_fp16_supported(return_reason: bool = False) -> (bool, str):
+    """
+    Utility method that returns a bool, stating if FP16 is supported for numba cuda kernels or not.
+
+    Returns:
+        bool, whether Numba CUDA will support fp16 or not.
+    """
+    reason = ""
+    use_nvidia_binding = os.environ.get('NUMBA_CUDA_USE_NVIDIA_BINDING', None)
+    if use_nvidia_binding is not None:
+        use_nvidia_binding = use_nvidia_binding.lower() == "1"
+        reason += "Env variable `NUMBA_CUDA_USE_NVIDIA_BINDING` is available and set to `1`. "
+    else:
+        use_nvidia_binding = False
+        reason += "Env variable `NUMBA_CUDA_USE_NVIDIA_BINDING` is not available or has not set to `1`."
+
+    if NUMBA_FP16_SUPPORTED:
+        reason += f"Numba CUDA FP16 is supported in installed numba version."
+    else:
+        reason += f"Numba CUDA FP16 is not supported in installed numba version."
+
+    result = use_nvidia_binding and NUMBA_FP16_SUPPORTED
+
+    if return_reason:
+        return result, reason
+    else:
+        return result
+
+
 def skip_numba_cuda_test_if_unsupported(min_version: str):
     """
     Helper method to skip pytest test case if numba cuda is not supported.