Megatron Export Update (NVIDIA#5343) (NVIDIA#5423)

* export update for Megatron + change ORT optimization Signed-off-by: David Mosallanezhad <[email protected]> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * updated export_utils to use autocast instead of manually casting >:/ Signed-off-by: David Mosallanezhad <[email protected]> * removed dtype from LayerNorm Signed-off-by: David Mosallanezhad <[email protected]> * added comment Signed-off-by: David Mosallanezhad <[email protected]> * reverting changes on FloatCast Signed-off-by: David Mosallanezhad <[email protected]> * Cherry-picked changes from megatron-norm Signed-off-by: Boris Fomitchev <[email protected]> * updated asr_model import to cast_utils Signed-off-by: David Mosallanezhad <[email protected]> * updated del onnx_model place Signed-off-by: David Mosallanezhad <[email protected]> * changed ort optimization to basic -> temp fix Signed-off-by: David Mosallanezhad <[email protected]> Signed-off-by: David Mosallanezhad <[email protected]> Signed-off-by: Boris Fomitchev <[email protected]> Co-authored-by: David Mosallanezhad <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Boris Fomitchev <[email protected]> Signed-off-by: David Mosallanezhad <[email protected]> Signed-off-by: Boris Fomitchev <[email protected]> Signed-off-by: Boris Fomitchev <[email protected]> Co-authored-by: David <[email protected]> Co-authored-by: David Mosallanezhad <[email protected]> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Boris Fomitchev <[email protected]> Co-authored-by: Oleksii Kuchaiev <[email protected]> Co-authored-by: Boris Fomitchev <[email protected]> Signed-off-by: Hainan Xu <[email protected]>
hainan-xv · Nov 29, 2022 · 801a87f · 801a87f
1 parent ed778d8
commit 801a87f
Show file tree

Hide file tree

Showing 5 changed files with 108 additions and 49 deletions.
diff --git a/nemo/collections/asr/models/asr_model.py b/nemo/collections/asr/models/asr_model.py
@@ -21,7 +21,8 @@
 from nemo.core.classes.common import PretrainedModelInfo
 from nemo.core.classes.exportable import Exportable
 from nemo.core.classes.mixins import AccessMixin
-from nemo.utils import cast_all, logging, model_utils
+from nemo.utils import logging, model_utils
+from nemo.utils.cast_utils import cast_all
 
 __all__ = ['ASRModel']
 

diff --git a/nemo/collections/nlp/models/machine_translation/megatron_nmt_model.py b/nemo/collections/nlp/models/machine_translation/megatron_nmt_model.py
@@ -42,6 +42,7 @@
 from nemo.collections.nlp.modules.common.megatron.megatron_export import DecEmb, EncEmb, TokensHeadEmb
 from nemo.collections.nlp.parts.nlp_overrides import GlobalBatchDataFetcher
 from nemo.collections.nlp.parts.utils_funcs import get_last_rank
+from nemo.core.classes import Exportable
 from nemo.utils import AppState, logging, timers
 
 try:
@@ -56,7 +57,7 @@
 __all__ = ["MegatronNMTModel"]
 
 
-class MegatronNMTModel(MegatronLMEncoderDecoderModel):
+class MegatronNMTModel(MegatronLMEncoderDecoderModel, Exportable):
     """
     Megatron NMT training
     """
@@ -750,5 +751,12 @@ def decoder(self):
         return DecEmb(self.enc_dec_model.decoder_embedding, self.enc_dec_model.enc_dec_model.decoder, self.device)
 
     @property
-    def classifier(self):
+    def log_softmax(self):
         return TokensHeadEmb(self.enc_dec_model.decoder_embedding, self.enc_dec_model.tokens_head, self.device)
+
+    @property
+    def input_module(self):
+        return self.encoder
+
+    def list_export_subnets(self):
+        return ['encoder', 'log_softmax', 'decoder']
diff --git a/nemo/collections/nlp/modules/common/megatron/megatron_export.py b/nemo/collections/nlp/modules/common/megatron/megatron_export.py
@@ -49,21 +49,23 @@ def forward(self, dec_output):
         if isinstance(dec_output, list):
             dec_output = dec_output[0]
 
-        dec_output = torch.permute(dec_output, (1, 0, 2))
-
         if self.tokens_head_bias is not None:
             return F.linear(dec_output, self.decoder_embedding.word_embeddings.weight, self.tokens_head_bias)
         return F.linear(dec_output, self.decoder_embedding.word_embeddings.weight)
 
-    def input_example(self, max_batch=1, max_dim=1024, seq_len=6):
+    def input_example(self, max_batch=1, max_dim=768, seq_len=6):
         return [
-            torch.randint(low=-3, high=3, size=(seq_len, max_batch, max_dim), device=self.device, dtype=torch.float32)
+            torch.randint(low=-3, high=3, size=(max_batch, seq_len, max_dim), device=self.device, dtype=torch.float32)
         ]
 
+    def freeze(self):
+        for param in self.parameters():
+            param.requires_grad = False
+
     @property
     def input_types(self) -> Optional[Dict[str, NeuralType]]:
         return {
-            "hidden_states": NeuralType(('T', 'B', 'D'), ChannelType()),
+            "hidden_states": NeuralType(('B', 'T', 'D'), ChannelType()),
         }
 
     @property
@@ -107,18 +109,28 @@ def forward(self, input_ids, decoder_mask, encoder_mask, encoder_embeddings, dec
         # dec_input, dec_attn_mask, enc_output, enc_attn_mask | dec_input, dec_attn_mask, enc_output, enc_attn_mask
         _ = dec_mems
 
-        return self.decoder(dec_input, decoder_mask, encoder_embeddings, encoder_mask).float()
+        return (
+            self.decoder(dec_input, decoder_mask, encoder_embeddings.permute(1, 0, 2), encoder_mask)
+            .float()
+            .permute(1, 0, 2)
+        )
 
-    def input_example(self, max_batch=1, max_dim=1024, seq_len=6):
+    def freeze(self):
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def input_example(self, max_batch=1, max_dim=768, seq_len=6):
         enc_output = torch.randint(
-            low=-3, high=3, size=(seq_len, max_batch, max_dim), device=self.device, dtype=torch.float32
+            low=-3, high=3, size=(max_batch, seq_len, max_dim), device=self.device, dtype=torch.float32
         )
         enc_attn_mask = torch.tensor([[1 for _ in range(seq_len)]]).to(self.device)
 
         dec_len = random.randint(10, 128)
         dec_input = torch.randint(low=0, high=1000, size=(max_batch, dec_len), device=self.device)
         dec_attn_mask = torch.tensor([[1 for _ in range(dec_len)]]).to(self.device)
-        decoder_mems = torch.zeros([8, 6, 1024], dtype=torch.float32).to(self.device)
+
+        # constant decoder_mems as placeholder for now
+        decoder_mems = torch.zeros([8, 6, max_dim], dtype=torch.float32).to(self.device)
 
         # input_ids, decoder_mask, encoder_mask, encoder_embeddings
         return (dec_input, dec_attn_mask, enc_attn_mask, enc_output, decoder_mems)
@@ -128,14 +140,14 @@ def input_types(self) -> Optional[Dict[str, NeuralType]]:
         return {
             "input_ids": NeuralType(('B', 'T', 'D'), ChannelType()),
             "decoder_mask": NeuralType(('B', 'T'), MaskType()),
-            "encoder_mask": NeuralType(('T', 'B', 'D'), ChannelType()),
+            "encoder_mask": NeuralType(('B', 'T', 'D'), ChannelType()),
             "encoder_embeddings": NeuralType(('B', 'T'), MaskType()),
-            "decoder_mems": NeuralType(('T', 'B', 'D'), ChannelType()),
+            "decoder_mems": NeuralType(('B', 'T', 'D'), ChannelType()),
         }
 
     @property
     def output_types(self) -> Optional[Dict[str, NeuralType]]:
-        return {"last_hidden_states": NeuralType(('T', 'B', 'D'), ChannelType())}
+        return {"last_hidden_states": NeuralType(('B', 'T', 'D'), ChannelType())}
 
     @property
     def input_names(self) -> List[str]:
@@ -172,15 +184,19 @@ def forward(self, input_ids, encoder_mask):
         enc_input = self.encoder_embedding(input_ids, position_ids, token_type_ids=None)
 
         # pass input through the encoder
-        return self.encoder(enc_input=enc_input, enc_attn_mask=encoder_mask,).type(torch.float32)
+        return self.encoder(enc_input=enc_input, enc_attn_mask=encoder_mask,).permute(1, 0, 2)
 
-    def input_example(self):
+    def input_example(self, max_batch=1, max_dim=30000, seq_len=6):
         seq_len = random.randint(0, 128)
         return (
-            torch.randint(0, 30000, (1, seq_len)).to(self.device),
-            torch.ones((1, seq_len), dtype=int).to(self.device),
+            torch.randint(0, max_dim, (max_batch, seq_len)).to(self.device),
+            torch.ones((max_batch, seq_len), dtype=int).to(self.device),
         )
 
+    def freeze(self):
+        for param in self.parameters():
+            param.requires_grad = False
+
     @property
     def input_types(self) -> Optional[Dict[str, NeuralType]]:
         return {
@@ -190,7 +206,7 @@ def input_types(self) -> Optional[Dict[str, NeuralType]]:
 
     @property
     def output_types(self) -> Optional[Dict[str, NeuralType]]:
-        return {"last_hidden_states": NeuralType(('T', 'B', 'D'), ChannelType())}
+        return {"last_hidden_states": NeuralType(('B', 'T', 'D'), ChannelType())}
 
     @property
     def input_names(self) -> List[str]:

diff --git a/nemo/utils/export_utils.py b/nemo/utils/export_utils.py
@@ -59,21 +59,40 @@ def forward(self, x):
         return F.linear(x, self.weight, self.bias), None
 
 
-# ScaledMaskedSoftmax replacement
-def mask_func(attention_scores, attention_mask):
-    attention_scores.masked_fill_(attention_mask, -10000.0)
-    return attention_scores
-
-
-def exportable_ScaledMaskedSoftmax(input, mask, scale):
-    if scale is not None:
-        input = input * scale
-
-    mask_output = mask_func(input, mask) if mask is not None else input
-    probs = torch.nn.Softmax(dim=-1)(mask_output)
-
-    probs = probs.half()
-    return probs
+class ExportableMatchedScaleMaskSoftmax(nn.Module):
+    def __init__(self, mod):
+        super(ExportableMatchedScaleMaskSoftmax, self).__init__()
+        self.init_module(mod.input_in_fp16, mod.input_in_bf16, mod.mask_func, mod.softmax_in_fp32, mod.scale)
+
+    def init_module(
+        self, input_in_fp16, input_in_bf16, mask_func, softmax_in_fp32, scale,
+    ):
+        self.input_in_fp16 = input_in_fp16
+        self.input_in_bf16 = input_in_bf16
+        self.softmax_in_fp32 = softmax_in_fp32
+        self.mask_func = mask_func
+        self.scale = scale
+
+        self.input_in_float16 = self.input_in_fp16 or self.input_in_bf16
+
+    def forward(self, input, mask):
+        if self.input_in_float16 and self.softmax_in_fp32:
+            input = input.float()
+
+        if self.scale is not None:
+            input = input * self.scale
+        mask_output = self.mask_func(input, mask) if mask is not None else input
+        probs = torch.nn.Softmax(dim=-1)(mask_output)
+        all_k_masked = mask.all(axis=-1)
+        zero_attention_mask = (1.0 - all_k_masked.float())[:, :, :, None]
+        probs = probs * zero_attention_mask
+
+        if self.input_in_float16 and self.softmax_in_fp32:
+            if self.input_in_fp16:
+                probs = probs.half()
+            else:
+                probs = probs.bfloat16()
+        return probs
 
 
 def get_export_format(filename: str):
@@ -159,10 +178,12 @@ def verify_runtime(model, output, input_examples, input_names, check_tolerance=0
         logging.warning(f"ONNX generated at {output}, not verified - please install onnxruntime_gpu package.\n")
         onnx.checker.check_model(onnx_model, full_check=True)
         return
-    del onnx_model
     onnx_session_opt = onnxruntime.SessionOptions()
-    onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
-    sess = onnxruntime.InferenceSession(output, sess_options=onnx_session_opt, providers=['CUDAExecutionProvider'])
+    onnx_session_opt.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_BASIC
+    sess = onnxruntime.InferenceSession(
+        onnx_model.SerializeToString(), sess_options=onnx_session_opt, providers=['CUDAExecutionProvider']
+    )
+    del onnx_model
     all_good = True
     for input_example in input_examples:
         input_list, input_dict = parse_input_example(input_example)
@@ -227,27 +248,24 @@ def run_ort_and_compare(sess, ort_input, output_example, check_tolerance=0.01):
     from apex.transformer.tensor_parallel.layers import RowParallelLinear, ColumnParallelLinear
     from apex.transformer.functional.fused_softmax import FusedScaleMaskSoftmax
 
-    def replace_FusedLayerNorm(n: nn.Module) -> Optional[nn.BatchNorm2d]:
+    def replace_FusedLayerNorm(n: nn.Module) -> Optional[nn.LayerNorm]:
         """
         Replaces Apex's FusedLayerNorm with nn.LayerNorm. This is required for ONNX export.
         Args:
            n: the FusedLayerNorm pytorch module to replace
         Returns:
            Equivalent LayerNorm module
         """
-        if (
-            not isinstance(n, FusedLayerNorm)
-            and not isinstance(n, FastLayerNorm)
-            and not isinstance(n, MixedFusedLayerNorm)
-        ):
-            return None
 
-        dev = next(n.parameters()).device
+        p = next(n.parameters())
         if isinstance(n, FusedLayerNorm) or isinstance(n, MixedFusedLayerNorm):
-            mod = nn.LayerNorm(n.normalized_shape, eps=n.eps, elementwise_affine=n.elementwise_affine,).to(dev)
+            shape, eps, affine = n.normalized_shape, n.eps, n.elementwise_affine
         elif isinstance(n, FastLayerNorm):
-            mod = nn.LayerNorm(n.weight.shape, eps=n.epsilon, elementwise_affine=True, dtype=torch.float16,).to(dev)
+            shape, eps, affine = n.weight.shape, n.epsilon, True
+        else:
+            return None
 
+        mod = nn.LayerNorm(shape, eps=eps, elementwise_affine=affine, device=p.device, dtype=p.dtype)
         n_state = n.state_dict()
         mod.load_state_dict(n_state)
         return mod
@@ -264,7 +282,7 @@ def replace_RowParallelLinear(n: nn.Module) -> Optional[nn.Linear]:
             raise ValueError("This function can only change the RowParallelLinear module.")
 
         dev = next(n.parameters()).device
-        mod = LinearWithBiasSkip(n.weight, n.bias, n.skip_bias_add).to(dev)
+        mod = LinearWithBiasSkip(n.weight, n.bias, n.skip_bias_add).to(device=dev)
 
         n_state = n.state_dict()
         mod.load_state_dict(n_state)
@@ -340,6 +358,20 @@ def expansion_fn(mod: nn.Module) -> Optional[nn.Module]:
     return expansion_fn
 
 
+def replace_MatchedScaleMaskSoftmax(n: nn.Module) -> Optional[nn.Linear]:
+    """
+    Replaces MatchedScaleMaskSoftmax with exportable softmax layer
+    Args:
+        n: module to replace
+    Returns:
+        exportable module
+    """
+
+    mod = ExportableMatchedScaleMaskSoftmax(n.input_in_fp16, n.input_in_bf16, n.mask_func, n.softmax_in_fp32, n.scale)
+
+    return mod
+
+
 def wrap_module(BaseT: Type[nn.Module], DestT: Type[nn.Module]) -> Callable[[nn.Module], Optional[nn.Module]]:
     """
     Generic function generator to replace BaseT module with DestT wrapper. 
@@ -408,6 +440,7 @@ def script_module(m: nn.Module):
     "BatchNorm1d": wrap_module(nn.BatchNorm1d, CastToFloat),
     "BatchNorm2d": wrap_module(nn.BatchNorm2d, CastToFloat),
     "LayerNorm": wrap_module(nn.LayerNorm, CastToFloat),
+    "MatchedScaleMaskSoftmax": wrap_module(nn.Softmax, ExportableMatchedScaleMaskSoftmax),
 }
 
 script_replacements = {

diff --git a/scripts/export.py b/scripts/export.py
@@ -139,6 +139,7 @@ def nemo_export(argv):
         with autocast(), torch.no_grad(), torch.inference_mode():
             model.to(device=args.device).freeze()
             model.eval()
+            input_example = None
             if check_trace and len(in_args) > 0:
                 input_example = model.input_module.input_example(**in_args)
                 check_trace = [input_example]