Add padding-free to Granite hybrid moe models (#39677)

* start fixing kwarg handling

* fmt

* updates padding free tests

* docs

* add missing kwargs modeling_granitemoe.py

* run modular util

* rm unrelated changes from modular util

docs/source/en/model_doc/granitemoehybrid.md

@@ -48,6 +48,32 @@ for i in output:
 
 This HF implementation is contributed by [Sukriti Sharma](https://huggingface.co/SukritiSharma) and [Alexander Brooks](https://huggingface.co/abrooks9944).
 
+## Notes
+
+- `GraniteMoeHybridForCausalLM` supports padding-free training which concatenates distinct training examples while still processing inputs as separate batches. It can significantly accelerate inference by [~2x](https://github.com/huggingface/transformers/pull/35861#issue-2807873129) (depending on model and data distribution) and reduce memory-usage if there are examples of varying lengths by avoiding unnecessary compute and memory overhead from padding tokens.
+
+  Padding-free training requires the `flash-attn`, `mamba-ssm`, and `causal-conv1d` packages and the following arguments must be passed to the model in addition to `input_ids` and `labels`.
+
+  - `position_ids: torch.LongTensor`: the position index of each token in each sequence.
+  - `seq_idx: torch.IntTensor`: the index of each sequence in the batch.
+  - Each of the [`FlashAttentionKwargs`]
+    - `cu_seq_lens_q: torch.LongTensor`: the cumulative sequence lengths of all queries.
+    - `cu_seq_lens_k: torch.LongTensor`: the cumulative sequence lengths of all keys.
+    - `max_length_q: int`: the longest query length in the batch.
+    - `max_length_k: int`: the longest key length in the batch.
+
+  The `attention_mask` inputs should not be provided. The [`DataCollatorWithFlattening`] programmatically generates the set of additional arguments above using `return_seq_idx=True` and `return_flash_attn_kwargs=True`. See the [Improving Hugging Face Training Efficiency Through Packing with Flash Attention](https://huggingface.co/blog/packing-with-FA2) blog post for additional information.
+
+  ```python
+  from transformers import DataCollatorWithFlattening
+
+  # Example of using padding-free training
+  data_collator = DataCollatorWithFlattening(
+      tokenizer=tokenizer,
+      return_seq_idx=True,
+      return_flash_attn_kwargs=True
+  )
+  ```
 
 ## GraniteMoeHybridConfig
 
@@ -61,4 +87,4 @@ This HF implementation is contributed by [Sukriti Sharma](https://huggingface.co
 ## GraniteMoeHybridForCausalLM
 
 [[autodoc]] GraniteMoeHybridForCausalLM
-    - forward
+    - forward

src/transformers/models/granitemoe/modeling_granitemoe.py

@@ -641,6 +641,7 @@ class GraniteMoeModel(GraniteMoePreTrainedModel):
         output_router_logits: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
     ) -> Union[tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -947,6 +948,7 @@ class GraniteMoeForCausalLM(GraniteMoePreTrainedModel, GenerationMixin):
             output_router_logits=output_router_logits,
             return_dict=return_dict,
             cache_position=cache_position,
+            **kwargs,
         )
 
         # Only compute necessary logits

src/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py

@@ -19,7 +19,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import Any, Callable, Optional, Union
+from typing import Any, Callable, Optional, TypedDict, Union
 
 import torch
 import torch.nn.functional as F
@@ -34,6 +34,7 @@ from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast, MoeCausalLMOutputWithPast, MoeModelOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
 from ...utils import auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
 from ...utils.import_utils import is_causal_conv1d_available, is_mamba_2_ssm_available
 from .configuration_granitemoehybrid import GraniteMoeHybridConfig
@@ -918,6 +919,31 @@ class GraniteMoeHybridMLP(nn.Module):
         return hidden_states
 
 
+class GraniteFlashAttentionKwargs(TypedDict, total=False):
+    """
+    Keyword arguments for advanced Flash Attention, causal-conv1d, and mamba_ssm kernel usage.
+    Use cases include padding-free training and fewer `torch.compile` graph breaks.
+
+    Attributes:
+        cu_seq_lens_q (`torch.LongTensor`)
+            Gets cumulative sequence length for query state.
+        cu_seq_lens_k (`torch.LongTensor`)
+            Gets cumulative sequence length for key state.
+        max_length_q (`int`):
+            Maximum sequence length for query state.
+        max_length_k (`int`):
+            Maximum sequence length for key state.
+        seq_idx (`torch.IntTensor):
+            Index of each packed sequence.
+    """
+
+    cu_seq_lens_q: torch.LongTensor
+    cu_seq_lens_k: torch.LongTensor
+    max_length_q: int
+    max_length_k: int
+    seq_idx: torch.IntTensor
+
+
 class GraniteMoeHybridRMSNorm(nn.Module):
     def __init__(self, hidden_size, eps=1e-6):
         """
@@ -1125,7 +1151,7 @@ class GraniteMoeHybridDecoderLayer(GradientCheckpointingLayer):
         cache_position: Optional[torch.LongTensor] = None,
         output_router_logits: Optional[bool] = False,
         position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
-        **kwargs,
+        **kwargs: Unpack[GraniteFlashAttentionKwargs],
     ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
@@ -1149,8 +1175,8 @@ class GraniteMoeHybridDecoderLayer(GradientCheckpointingLayer):
                 Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                 with `head_dim` being the embedding dimension of each attention head.
             kwargs (`dict`, *optional*):
-                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
-                into the model
+                Arbitrary kwargs.Can be used to provide `GraniteFlashAttentionKwargs` for
+                padding-free training and/or improve torch.compile performance.
         """
         residual = hidden_states
         hidden_states = self.input_layernorm(hidden_states)
@@ -1161,6 +1187,7 @@ class GraniteMoeHybridDecoderLayer(GradientCheckpointingLayer):
                 cache_position=cache_position,
                 cache_params=past_key_value,
                 attention_mask=attention_mask,
+                **kwargs,
             )
             # No attention weights for state space layers
             self_attn_weights = None
@@ -1303,6 +1330,7 @@ class GraniteMoeHybridModel(GraniteMoeHybridPreTrainedModel):
         output_router_logits: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[GraniteFlashAttentionKwargs],
     ) -> Union[tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -1374,6 +1402,7 @@ class GraniteMoeHybridModel(GraniteMoeHybridPreTrainedModel):
                 cache_position=cache_position,
                 output_router_logits=output_router_logits,
                 position_embeddings=position_embeddings,
+                **kwargs,
             )
 
             hidden_states = layer_outputs[0]
@@ -1706,6 +1735,7 @@ class GraniteMoeHybridForCausalLM(GraniteMoeHybridPreTrainedModel, GenerationMix
             output_router_logits=output_router_logits,
             return_dict=return_dict,
             cache_position=cache_position,
+            **kwargs,
         )
 
         # Only compute necessary logits

src/transformers/models/granitemoehybrid/modular_granitemoehybrid.py

@@ -20,10 +20,12 @@ from torch import nn
 
 from ...cache_utils import Cache
 from ...modeling_outputs import BaseModelOutputWithPast, MoeModelOutputWithPast
+from ...processing_utils import Unpack
 from ...utils import auto_docstring, can_return_tuple, logging
 from ..bamba.configuration_bamba import BambaConfig
 from ..bamba.modeling_bamba import BambaMixer, BambaRMSNormGated, HybridMambaAttentionDynamicCache
 from ..granitemoeshared.modeling_granitemoeshared import (
+    GraniteFlashAttentionKwargs,
     GraniteMoeSharedAttention,
     GraniteMoeSharedDecoderLayer,
     GraniteMoeSharedForCausalLM,
@@ -84,7 +86,7 @@ class GraniteMoeHybridDecoderLayer(GraniteMoeSharedDecoderLayer):
         cache_position: Optional[torch.LongTensor] = None,
         output_router_logits: Optional[bool] = False,
         position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
-        **kwargs,
+        **kwargs: Unpack[GraniteFlashAttentionKwargs],
     ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
@@ -108,8 +110,8 @@ class GraniteMoeHybridDecoderLayer(GraniteMoeSharedDecoderLayer):
                 Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                 with `head_dim` being the embedding dimension of each attention head.
             kwargs (`dict`, *optional*):
-                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
-                into the model
+                Arbitrary kwargs.Can be used to provide `GraniteFlashAttentionKwargs` for
+                padding-free training and/or improve torch.compile performance.
         """
         residual = hidden_states
         hidden_states = self.input_layernorm(hidden_states)
@@ -120,6 +122,7 @@ class GraniteMoeHybridDecoderLayer(GraniteMoeSharedDecoderLayer):
                 cache_position=cache_position,
                 cache_params=past_key_value,
                 attention_mask=attention_mask,
+                **kwargs,
             )
             # No attention weights for state space layers
             self_attn_weights = None
@@ -198,6 +201,7 @@ class GraniteMoeHybridModel(GraniteMoeSharedModel):
         output_router_logits: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[GraniteFlashAttentionKwargs],
     ) -> Union[tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -269,6 +273,7 @@ class GraniteMoeHybridModel(GraniteMoeSharedModel):
                 cache_position=cache_position,
                 output_router_logits=output_router_logits,
                 position_embeddings=position_embeddings,
+                **kwargs,
             )
 
             hidden_states = layer_outputs[0]

src/transformers/models/granitemoeshared/modeling_granitemoeshared.py

@@ -19,7 +19,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import Callable, Optional, Union
+from typing import Callable, Optional, TypedDict, Union
 
 import torch
 import torch.nn.functional as F
@@ -33,6 +33,7 @@ from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast, MoeCausalLMOutputWithPast, MoeModelOutputWithPast
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from ...processing_utils import Unpack
 from ...utils import auto_docstring, is_torch_flex_attn_available, logging
 from .configuration_granitemoeshared import GraniteMoeSharedConfig
 
@@ -46,6 +47,31 @@ if is_torch_flex_attn_available():
 logger = logging.get_logger(__name__)
 
 
+class GraniteFlashAttentionKwargs(TypedDict, total=False):
+    """
+    Keyword arguments for advanced Flash Attention, causal-conv1d, and mamba_ssm kernel usage.
+    Use cases include padding-free training and fewer `torch.compile` graph breaks.
+
+    Attributes:
+        cu_seq_lens_q (`torch.LongTensor`)
+            Gets cumulative sequence length for query state.
+        cu_seq_lens_k (`torch.LongTensor`)
+            Gets cumulative sequence length for key state.
+        max_length_q (`int`):
+            Maximum sequence length for query state.
+        max_length_k (`int`):
+            Maximum sequence length for key state.
+        seq_idx (`torch.IntTensor):
+            Index of each packed sequence.
+    """
+
+    cu_seq_lens_q: torch.LongTensor
+    cu_seq_lens_k: torch.LongTensor
+    max_length_q: int
+    max_length_k: int
+    seq_idx: torch.IntTensor
+
+
 class GraniteMoeSharedMLP(nn.Module):
     """
     MLP layer for shared experts
@@ -431,7 +457,7 @@ class GraniteMoeSharedDecoderLayer(GradientCheckpointingLayer):
         cache_position: Optional[torch.LongTensor] = None,
         output_router_logits: Optional[bool] = False,
         position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
-        **kwargs,
+        **kwargs: Unpack[GraniteFlashAttentionKwargs],
     ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
@@ -455,8 +481,8 @@ class GraniteMoeSharedDecoderLayer(GradientCheckpointingLayer):
                 Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                 with `head_dim` being the embedding dimension of each attention head.
             kwargs (`dict`, *optional*):
-                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
-                into the model
+                Arbitrary kwargs. Can be used to provide `GraniteFlashAttentionKwargs` for
+                padding-free training and/or improve torch.compile performance.
         """
         residual = hidden_states
 
@@ -593,6 +619,7 @@ class GraniteMoeSharedModel(GraniteMoeSharedPreTrainedModel):
         output_router_logits: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
     ) -> Union[tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -979,6 +1006,7 @@ class GraniteMoeSharedForCausalLM(GraniteMoeSharedPreTrainedModel, GenerationMix
             output_router_logits=output_router_logits,
             return_dict=return_dict,
             cache_position=cache_position,
+            **kwargs,
         )
 
         # Only compute necessary logits

src/transformers/models/granitemoeshared/modular_granitemoeshared.py

@@ -13,13 +13,14 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import Optional
+from typing import Optional, TypedDict
 
 import torch
 from torch import nn
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache
+from ...processing_utils import Unpack
 from ...utils import logging
 from ..granitemoe.modeling_granitemoe import (
     GraniteMoeDecoderLayer,
@@ -33,6 +34,31 @@ from .configuration_granitemoeshared import GraniteMoeSharedConfig
 logger = logging.get_logger(__name__)
 
 
+class GraniteFlashAttentionKwargs(TypedDict, total=False):
+    """
+    Keyword arguments for advanced Flash Attention, causal-conv1d, and mamba_ssm kernel usage.
+    Use cases include padding-free training and fewer `torch.compile` graph breaks.
+
+    Attributes:
+        cu_seq_lens_q (`torch.LongTensor`)
+            Gets cumulative sequence length for query state.
+        cu_seq_lens_k (`torch.LongTensor`)
+            Gets cumulative sequence length for key state.
+        max_length_q (`int`):
+            Maximum sequence length for query state.
+        max_length_k (`int`):
+            Maximum sequence length for key state.
+        seq_idx (`torch.IntTensor):
+            Index of each packed sequence.
+    """
+
+    cu_seq_lens_q: torch.LongTensor
+    cu_seq_lens_k: torch.LongTensor
+    max_length_q: int
+    max_length_k: int
+    seq_idx: torch.IntTensor
+
+
 class GraniteMoeSharedMLP(nn.Module):
     """
     MLP layer for shared experts
@@ -75,7 +101,7 @@ class GraniteMoeSharedDecoderLayer(GraniteMoeDecoderLayer):
         cache_position: Optional[torch.LongTensor] = None,
         output_router_logits: Optional[bool] = False,
         position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
-        **kwargs,
+        **kwargs: Unpack[GraniteFlashAttentionKwargs],
     ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
@@ -99,8 +125,8 @@ class GraniteMoeSharedDecoderLayer(GraniteMoeDecoderLayer):
                 Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                 with `head_dim` being the embedding dimension of each attention head.
             kwargs (`dict`, *optional*):
-                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
-                into the model
+                Arbitrary kwargs. Can be used to provide `GraniteFlashAttentionKwargs` for
+                padding-free training and/or improve torch.compile performance.
         """
         residual = hidden_states
 

tests/models/bamba/test_modeling_bamba.py

@@ -551,6 +551,15 @@ class BambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
                     inputs_dict["attention_mask"] = inputs_dict["attention_mask"].flip(1)
                 dummy_attention_mask = inputs_dict["attention_mask"]
                 inputs_dict["input_ids"][~dummy_attention_mask.bool()] = config.get_text_config().pad_token_id
+                # Ensure inputs_dict also has labels in it, as their presence/absence can induce
+                # dtype conversions. This also lets us compare losses.
+                labels = inputs_dict["input_ids"].clone()
+                # Mask padding tokens
+                labels[~dummy_attention_mask.bool()] = -100
+                # Also need to mask the first non-trivial token to match the padding-free batch.
+                first_nonneg_idx = (labels >= 0).int().argmax(dim=1)
+                labels[torch.arange(labels.size(0), device=labels.device), first_nonneg_idx] = -100
+                inputs_dict["labels"] = labels
 
                 model = (
                     model_class.from_pretrained(
@@ -586,6 +595,10 @@ class BambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
                 tol = torch.finfo(torch.float16).eps
                 torch.testing.assert_close(logits_padded, logits_padfree, rtol=tol, atol=tol)
 
+                loss_padded = res_padded.loss
+                loss_padfree = res_padfree.loss
+                torch.testing.assert_close(loss_padded, loss_padfree)
+
 
 @slow
 @require_torch