Fix documentation for Numba (NVIDIA#7065)

* Fix documentation for Numba Signed-off-by: smajumdar <[email protected]> * Update force float32 flag dynamically Signed-off-by: smajumdar <[email protected]> * Update force float32 flag dynamically Signed-off-by: smajumdar <[email protected]> * Fix nemo version Signed-off-by: smajumdar <[email protected]> --------- Signed-off-by: smajumdar <[email protected]> Co-authored-by: Eric Harper <[email protected]> Signed-off-by: zhehuaichen <[email protected]>
zhehuaichen · Oct 4, 2023 · 6fdf815 · 6fdf815
1 parent 65a6c43
commit 6fdf815
Show file tree

Hide file tree

Showing 6 changed files with 18 additions and 13 deletions.
diff --git a/README.rst b/README.rst
@@ -132,8 +132,8 @@ Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training t
 Requirements
 ------------
 
-1) Python 3.8 or above
-2) Pytorch 1.10.0 or above
+1) Python 3.9 or above
+2) Pytorch 1.13.1 or above
 3) NVIDIA GPU for training
 
 Documentation

diff --git a/docs/source/nlp/api.rst b/docs/source/nlp/api.rst
@@ -124,7 +124,7 @@ Datasets
 .. autoclass:: nemo.collections.nlp.data.language_modeling.megatron.gpt_sft_dataset.GPTSFTDataset
     :show-inheritance: 
 
-.. autoclass:: nemo.collections.nlp.data.language_modeling.megatron.gpt_sft_dataset.GPTSFTChatDataset
+.. autoclass:: nemo.collections.nlp.data.language_modeling.megatron.gpt_sft_chat_dataset.GPTSFTChatDataset
     :show-inheritance: 
 
 .. autoclass:: nemo.collections.nlp.data.language_modeling.megatron.retro_dataset.RETRODataset

diff --git a/docs/source/starthere/intro.rst b/docs/source/starthere/intro.rst
@@ -34,9 +34,9 @@ Prerequisites
 
 Before you begin using NeMo, it's assumed you meet the following prerequisites.
 
-#. You have Python version 3.6, 3.7 or 3.8.
+#. You have Python version 3.9, 3.10.
 
-#. You have Pytorch version 1.8.1.
+#. You have Pytorch version 1.13.1 or 2.0+.
 
 #. You have access to an NVIDIA GPU for training.
 

diff --git a/nemo/collections/asr/losses/rnnt.py b/nemo/collections/asr/losses/rnnt.py
@@ -99,7 +99,7 @@ class RNNTLossConfig:
         min_version='0.53.0',
         is_available=NUMBA_RNNT_AVAILABLE,
         installation_msg=NUMBA_INSTALLATION_MESSAGE,
-        force_float32=not numba_utils.NUMBA_FP16_SUPPORTED,
+        force_float32=False,  # This is only temporarily false, will be dynamically updated during resolution
     ),
     "pytorch": RNNTLossConfig(
         loss_name="pytorch",
@@ -258,6 +258,9 @@ def resolve_rnnt_loss(loss_name: str, blank_idx: int, loss_kwargs: dict = None)
         _warn_unused_additional_kwargs(loss_name, loss_kwargs)
 
     elif loss_name == 'warprnnt_numba':
+        # Update loss config's forced float32 flag if set to None
+        loss_config.force_float32 = not numba_utils.is_numba_cuda_fp16_supported()
+
         fastemit_lambda = loss_kwargs.pop('fastemit_lambda', 0.0)
         clamp = loss_kwargs.pop('clamp', -1.0)
         loss_func = RNNTLossNumba(blank=blank_idx, reduction='none', fastemit_lambda=fastemit_lambda, clamp=clamp)
@@ -444,7 +447,7 @@ def forward(self, log_probs, targets, input_lengths, target_lengths):
         max_targets_len = target_lengths.max()
 
         # Force cast joint to float32
-        if not self._force_float32 and numba_utils.NUMBA_FP16_SUPPORTED:
+        if not self._force_float32 and numba_utils.is_numba_cuda_fp16_supported():
             # Execute the kernel in fp16
             pass
         elif self._force_float32 and log_probs.dtype != torch.float32:

diff --git a/nemo/core/utils/numba_utils.py b/nemo/core/utils/numba_utils.py
@@ -29,9 +29,6 @@
 __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 = (
@@ -171,12 +168,16 @@ def is_numba_cuda_fp16_supported(return_reason: bool = False) -> Union[bool, Tup
         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:
+    numba_fp16_version_correct = model_utils.check_lib_version(
+        'numba', __NUMBA_MINIMUM_VERSION_FP16_SUPPORTED__, operator=operator.ge
+    )[0]
+
+    if numba_fp16_version_correct:
         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
+    result = use_nvidia_binding and numba_fp16_version_correct
 
     if return_reason:
         return result, reason

diff --git a/nemo/utils/model_utils.py b/nemo/utils/model_utils.py
@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import copy
+import importlib
 import os
 from dataclasses import dataclass, is_dataclass
 from enum import Enum
@@ -554,7 +555,7 @@ def check_lib_version(lib_name: str, checked_version: str, operator) -> Tuple[Op
         if '.' in lib_name:
             mod = import_class_by_path(lib_name)
         else:
-            mod = __import__(lib_name)
+            mod = importlib.import_module(lib_name)
 
         if hasattr(mod, '__version__'):
             lib_ver = version.Version(mod.__version__)