Revert FP8 integration (#8520)

* Revert FP8 integration

Signed-off-by: Mingyuan Ma <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Signed-off-by: Mingyuan Ma <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

examples/multimodal/text_to_image/stable_diffusion/conf/sd_train.yaml

@@ -121,7 +121,6 @@ model:
     use_flash_attention: True
     enable_amp_o2_fp16: False
     resblock_gn_groups: 32
-    use_te_fp8: False
 
   first_stage_config:
     _target_: nemo.collections.multimodal.models.text_to_image.stable_diffusion.ldm.autoencoder.AutoencoderKL
@@ -192,7 +191,7 @@ model:
       synthetic_data_length: 10000
       train:
           dataset_path:
-            - /datasets/coyo/wdinfo/coyo-700m/wdinfo-selene.pkl
+            - /datasets/coyo/test.pkl
           augmentations:
             resize_smallest_side: 512
             center_crop_h_w: 512, 512

nemo/collections/multimodal/modules/stable_diffusion/attention.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import math
-import os
 from inspect import isfunction
 
 import torch
@@ -22,13 +21,6 @@
 from torch import einsum, nn
 from torch._dynamo import disable
 
-if os.environ.get("USE_NATIVE_GROUP_NORM", "0") == "1":
-    from nemo.gn_native import GroupNormNormlization as GroupNorm
-else:
-    from apex.contrib.group_norm import GroupNorm
-
-from transformer_engine.pytorch.module import LayerNormLinear, LayerNormMLP
-
 from nemo.collections.multimodal.modules.stable_diffusion.diffusionmodules.util import checkpoint
 from nemo.collections.nlp.modules.common.megatron.adapters.parallel_adapters import (
     AdapterName,
@@ -103,19 +95,13 @@ def forward(self, x):
 
 
 class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0, use_te=False):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
         super().__init__()
         inner_dim = int(dim * mult)
         dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(LinearWrapper(dim, inner_dim), nn.GELU()) if not glu else GEGLU(dim, inner_dim)
 
-        if use_te:
-            activation = 'gelu' if not glu else 'geglu'
-            # TODO: more parameters to be confirmed, dropout, seq_length
-            self.net = LayerNormMLP(hidden_size=dim, ffn_hidden_size=inner_dim, activation=activation,)
-        else:
-            norm = nn.LayerNorm(dim)
-            project_in = nn.Sequential(LinearWrapper(dim, inner_dim), nn.GELU()) if not glu else GEGLU(dim, inner_dim)
-            self.net = nn.Sequential(norm, project_in, nn.Dropout(dropout), LinearWrapper(inner_dim, dim_out))
+        self.net = nn.Sequential(project_in, nn.Dropout(dropout), LinearWrapper(inner_dim, dim_out))
 
     def forward(self, x):
         return self.net(x)
@@ -238,7 +224,6 @@ def __init__(
         dropout=0.0,
         use_flash_attention=False,
         lora_network_alpha=None,
-        use_te=False,
     ):
         super().__init__()
 
@@ -252,16 +237,10 @@ def __init__(
         self.scale = dim_head ** -0.5
         self.heads = heads
 
+        self.to_q = LinearWrapper(query_dim, self.inner_dim, bias=False, lora_network_alpha=lora_network_alpha)
         self.to_k = LinearWrapper(context_dim, self.inner_dim, bias=False, lora_network_alpha=lora_network_alpha)
         self.to_v = LinearWrapper(context_dim, self.inner_dim, bias=False, lora_network_alpha=lora_network_alpha)
 
-        if use_te:
-            self.norm_to_q = LayerNormLinear(query_dim, self.inner_dim, bias=False)
-        else:
-            norm = nn.LayerNorm(query_dim)
-            to_q = LinearWrapper(query_dim, self.inner_dim, bias=False, lora_network_alpha=lora_network_alpha)
-            self.norm_to_q = nn.Sequential(norm, to_q)
-
         self.to_out = nn.Sequential(
             LinearWrapper(self.inner_dim, query_dim, lora_network_alpha=lora_network_alpha), nn.Dropout(dropout)
         )
@@ -276,7 +255,7 @@ def __init__(
     def forward(self, x, context=None, mask=None):
         h = self.heads
 
-        q = self.norm_to_q(x)
+        q = self.to_q(x)
         context = default(context, x)
         k = self.to_k(context)
         v = self.to_v(context)
@@ -356,7 +335,6 @@ def __init__(
         use_flash_attention=False,
         disable_self_attn=False,
         lora_network_alpha=None,
-        use_te=False,
     ):
         super().__init__()
         self.disable_self_attn = disable_self_attn
@@ -368,9 +346,8 @@ def __init__(
             use_flash_attention=use_flash_attention,
             context_dim=context_dim if self.disable_self_attn else None,
             lora_network_alpha=lora_network_alpha,
-            use_te=use_te,
         )  # is a self-attention
-        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff, use_te=use_te)
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
         self.attn2 = CrossAttention(
             query_dim=dim,
             context_dim=context_dim,
@@ -379,8 +356,10 @@ def __init__(
             dropout=dropout,
             use_flash_attention=use_flash_attention,
             lora_network_alpha=lora_network_alpha,
-            use_te=use_te,
         )  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
         self.use_checkpoint = use_checkpoint
 
     def forward(self, x, context=None):
@@ -390,9 +369,9 @@ def forward(self, x, context=None):
             return self._forward(x, context)
 
     def _forward(self, x, context=None):
-        x = self.attn1(x, context=context if self.disable_self_attn else None) + x
-        x = self.attn2(x, context=context) + x
-        x = self.ff(x) + x
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
         return x
 
 
@@ -418,7 +397,6 @@ def __init__(
         use_checkpoint=False,
         use_flash_attention=False,
         lora_network_alpha=None,
-        use_te=False,
     ):
         super().__init__()
         if exists(context_dim) and not isinstance(context_dim, list):
@@ -444,7 +422,6 @@ def __init__(
                     use_flash_attention=use_flash_attention,
                     disable_self_attn=disable_self_attn,
                     lora_network_alpha=lora_network_alpha,
-                    use_te=use_te,
                 )
                 for d in range(depth)
             ]

nemo/collections/multimodal/modules/stable_diffusion/diffusionmodules/openaimodel.py

@@ -12,19 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import math
-import os
 from abc import abstractmethod
-from contextlib import nullcontext
 
 import numpy as np
 import torch
 import torch as th
 import torch.nn as nn
 import torch.nn.functional as F
 
-# FP8 related import
-import transformer_engine
-
 from nemo.collections.multimodal.modules.stable_diffusion.attention import SpatialTransformer
 from nemo.collections.multimodal.modules.stable_diffusion.diffusionmodules.util import (
     avg_pool_nd,
@@ -50,39 +45,6 @@ def convert_module_to_fp16(module):
     convert_module_to_dtype(module, torch.float16)
 
 
-def convert_module_to_fp32(module):
-    convert_module_to_dtype(module, torch.float32)
-
-
-def convert_module_to_fp8(model):
-    def _set_module(model, submodule_key, module):
-        tokens = submodule_key.split('.')
-        sub_tokens = tokens[:-1]
-        cur_mod = model
-        for s in sub_tokens:
-            cur_mod = getattr(cur_mod, s)
-        setattr(cur_mod, tokens[-1], module)
-
-    import copy
-
-    from transformer_engine.pytorch.module import Linear as te_Linear
-
-    for n, v in model.named_modules():
-        if isinstance(v, torch.nn.Linear):
-            # if n in ['class_embed', 'bbox_embed.layers.0', 'bbox_embed.layers.1', 'bbox_embed.layers.2']: continue
-            logging.info(f'[INFO] Replace Linear: {n}, weight: {v.weight.shape}')
-            if v.bias is None:
-                is_bias = False
-            else:
-                is_bias = True
-            newlinear = te_Linear(v.in_features, v.out_features, bias=is_bias)
-            newlinear.weight = copy.deepcopy(v.weight)
-            if v.bias is not None:
-                newlinear.bias = copy.deepcopy(v.bias)
-            _set_module(model, n, newlinear)
-
-
-## go
 class AttentionPool2d(nn.Module):
     """
     Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
@@ -509,7 +471,6 @@ def __init__(
         use_flash_attention: bool = False,
         enable_amp_o2_fp16: bool = False,
         lora_network_alpha=None,
-        use_te_fp8: bool = False,
     ):
         super().__init__()
         if use_spatial_transformer:
@@ -565,7 +526,6 @@ def __init__(
         input_block_chans = [model_channels]
         ch = model_channels
         ds = 1
-        self.use_te_fp8 = use_te_fp8
         for level, mult in enumerate(channel_mult):
             for _ in range(num_res_blocks):
                 layers = [
@@ -608,7 +568,6 @@ def __init__(
                             use_linear=use_linear_in_transformer,
                             use_checkpoint=use_checkpoint,
                             use_flash_attention=use_flash_attention,
-                            use_te=self.use_te_fp8,
                             lora_network_alpha=lora_network_alpha,
                         )
                     )
@@ -674,7 +633,6 @@ def __init__(
                 use_linear=use_linear_in_transformer,
                 use_checkpoint=use_checkpoint,
                 use_flash_attention=use_flash_attention,
-                use_te=self.use_te_fp8,
                 lora_network_alpha=lora_network_alpha,
             ),
             ResBlock(
@@ -702,7 +660,6 @@ def __init__(
                         dims=dims,
                         use_checkpoint=use_checkpoint,
                         use_scale_shift_norm=use_scale_shift_norm,
-                        resblock_gn_groups=resblock_gn_groups,
                     )
                 ]
                 ch = model_channels * mult
@@ -733,7 +690,6 @@ def __init__(
                             use_linear=use_linear_in_transformer,
                             use_checkpoint=use_checkpoint,
                             use_flash_attention=use_flash_attention,
-                            use_te=self.use_te_fp8,
                             lora_network_alpha=lora_network_alpha,
                         )
                     )
@@ -790,32 +746,6 @@ def __init__(
         if enable_amp_o2_fp16:
             self.convert_to_fp16()
 
-        elif self.use_te_fp8:
-            assert enable_amp_o2_fp16 is False, "fp8 training can't work with fp16 O2 amp recipe"
-            convert_module_to_fp8(self)
-
-            fp8_margin = int(os.getenv("FP8_MARGIN", '0'))
-            fp8_interval = int(os.getenv("FP8_INTERVAL", '1'))
-            fp8_format = os.getenv("FP8_FORMAT", "hybrid")
-            fp8_amax_history_len = int(os.getenv("FP8_HISTORY_LEN", '1024'))
-            fp8_amax_compute_algo = os.getenv("FP8_COMPUTE_ALGO", 'max')
-            fp8_wgrad = os.getenv("FP8_WGRAD", '1') == '1'
-
-            fp8_format_dict = {
-                'hybrid': transformer_engine.common.recipe.Format.HYBRID,
-                'e4m3': transformer_engine.common.recipe.Format.E4M3,
-            }
-            fp8_format = fp8_format_dict[fp8_format]
-
-            self.fp8_recipe = transformer_engine.common.recipe.DelayedScaling(
-                margin=fp8_margin,
-                interval=fp8_interval,
-                fp8_format=fp8_format,
-                amax_history_len=fp8_amax_history_len,
-                amax_compute_algo=fp8_amax_compute_algo,
-                override_linear_precision=(False, False, not fp8_wgrad),
-            )
-
     def _input_blocks_mapping(self, input_dict):
         res_dict = {}
         for key_, value_ in input_dict.items():
@@ -1030,7 +960,7 @@ def convert_to_fp16(self):
         """
         self.apply(convert_module_to_fp16)
 
-    def _forward(self, x, timesteps=None, context=None, y=None, **kwargs):
+    def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
         """
         Apply the model to an input batch.
 
@@ -1069,13 +999,6 @@ def _forward(self, x, timesteps=None, context=None, y=None, **kwargs):
         else:
             return self.out(h)
 
-    def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
-        with transformer_engine.pytorch.fp8_autocast(
-            enabled=self.use_te_fp8, fp8_recipe=self.fp8_recipe,
-        ) if self.use_te_fp8 else nullcontext():
-            out = self._forward(x, timesteps, context, y, **kwargs)
-        return out
-
 
 class EncoderUNetModel(nn.Module):
     """

nemo/collections/nlp/parts/nlp_overrides.py

@@ -1003,30 +1003,6 @@ def should_process(key):
                     new_state_dict[key_] = state_dict[key_]
             state_dict = new_state_dict
 
-        if conf.get('unet_config') and conf.get('unet_config').get('use_te_fp8') == False:
-            # remove _extra_state in fp8 if there is.
-            new_state_dict = {}
-            for key in state_dict.keys():
-                if 'extra_state' in key:
-                    continue
-
-                ### LayerNormLinear
-                # norm_to_q.layer_norm_{weight|bias} -> norm_to_q.0.{weight|bias}
-                # norm_to_q.weight -> norm_to_q.1.weight
-                new_key = key.replace('norm_to_q.layer_norm_', 'norm_to_q.0.')
-                new_key = new_key.replace('norm_to_q.weight', 'norm_to_q.1.weight')
-
-                ### LayerNormMLP
-                # ff.net.layer_norm_{weight|bias} -> ff.net.0.{weight|bias}
-                # ff.net.fc1_{weight|bias} -> ff.net.1.proj.{weight|bias}
-                # ff.net.fc2_{weight|bias} -> ff.net.3.{weight|bias}
-                new_key = new_key.replace('ff.net.layer_norm_', 'ff.net.0.')
-                new_key = new_key.replace('ff.net.fc1_', 'ff.net.1.proj.')
-                new_key = new_key.replace('ff.net.fc2_', 'ff.net.3.')
-
-                new_state_dict[new_key] = state_dict[key]
-            state_dict = new_state_dict
-
         return state_dict
 
     def _load_state_dict_from_disk(self, model_weights, map_location=None):