transformers 5.0.0rc2__py3-none-any.whl → 5.1.0__py3-none-any.whl

transformers/models/bamba/configuration_bamba.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 IBM and the HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,8 +13,6 @@
 # limitations under the License.
 """Bamba model configuration"""
 
-from typing import Optional
-
 from ...configuration_utils import PreTrainedConfig
 from ...modeling_rope_utils import RopeParameters
 from ...utils import logging
@@ -103,6 +100,12 @@ class BambaConfig(PreTrainedConfig):
             Flag indicating whether or not to use bias in the convolution layer of the mamba mixer block.
         mamba_proj_bias (`bool`, *optional*, defaults to `False`):
             Flag indicating whether or not to use bias in the input and output projections (["in_proj", "out_proj"]) of the mamba mixer block
+        time_step_min (`float`, *optional*, defaults to 0.001):
+            Minimum `time_step` used to bound `dt_proj.bias`.
+        time_step_max (`float`, *optional*, defaults to 0.1):
+            Maximum `time_step` used to bound `dt_proj.bias`.
+        time_step_limit (`tuple`, *optional*, defaults to `(0.0, inf)`):
+            Accepted range of time step values for clamping.
         z_loss_coefficient (`float`, *optional*, defaults to 0.0):
             Coefficient for auxiliary z-loss used to control logit growth during training
         rope_parameters (`RopeParameters`, *optional*):
@@ -116,35 +119,38 @@ class BambaConfig(PreTrainedConfig):
 
     def __init__(
         self,
-        vocab_size: Optional[int] = 128000,
-        tie_word_embeddings: Optional[bool] = False,
-        hidden_size: Optional[int] = 4096,
-        intermediate_size: Optional[int] = 14336,
-        num_hidden_layers: Optional[int] = 32,
-        num_attention_heads: Optional[int] = 32,
-        num_key_value_heads: Optional[int] = 8,
-        hidden_act: Optional[str] = "silu",
-        initializer_range: Optional[float] = 0.02,
-        rms_norm_eps: Optional[float] = 1e-5,
-        use_cache: Optional[bool] = True,
-        num_logits_to_keep: Optional[int] = 1,
-        pad_token_id: Optional[int] = 0,
-        bos_token_id: Optional[int] = 1,
-        eos_token_id: Optional[int] = 2,
-        max_position_embeddings: Optional[int] = 262144,
-        attention_dropout: Optional[float] = 0.0,
-        attn_layer_indices: Optional[list[int]] = None,
-        mamba_n_heads: Optional[int] = 128,
-        mamba_d_head: Optional[str] = "auto",
-        mamba_n_groups: Optional[int] = 1,
-        mamba_d_state: Optional[int] = 256,
-        mamba_d_conv: Optional[int] = 4,
-        mamba_expand: Optional[int] = 2,
-        mamba_chunk_size: Optional[int] = 256,
-        mamba_conv_bias: Optional[bool] = True,
-        mamba_proj_bias: Optional[bool] = False,
-        z_loss_coefficient: Optional[float] = 0.0,
-        rope_parameters: Optional[RopeParameters] = None,
+        vocab_size: int | None = 128000,
+        tie_word_embeddings: bool | None = False,
+        hidden_size: int | None = 4096,
+        intermediate_size: int | None = 14336,
+        num_hidden_layers: int | None = 32,
+        num_attention_heads: int | None = 32,
+        num_key_value_heads: int | None = 8,
+        hidden_act: str | None = "silu",
+        initializer_range: float | None = 0.02,
+        rms_norm_eps: float | None = 1e-5,
+        use_cache: bool | None = True,
+        num_logits_to_keep: int | None = 1,
+        pad_token_id: int | None = 0,
+        bos_token_id: int | None = 1,
+        eos_token_id: int | None = 2,
+        max_position_embeddings: int | None = 262144,
+        attention_dropout: float | None = 0.0,
+        attn_layer_indices: list[int] | None = None,
+        mamba_n_heads: int | None = 128,
+        mamba_d_head: str | None = "auto",
+        mamba_n_groups: int | None = 1,
+        mamba_d_state: int | None = 256,
+        mamba_d_conv: int | None = 4,
+        mamba_expand: int | None = 2,
+        mamba_chunk_size: int | None = 256,
+        mamba_conv_bias: bool | None = True,
+        mamba_proj_bias: bool | None = False,
+        time_step_min: float | None = 0.001,
+        time_step_max: float | None = 0.1,
+        time_step_limit: tuple[float, float] | None = (0.0, float("inf")),
+        z_loss_coefficient: float | None = 0.0,
+        rope_parameters: RopeParameters | None = None,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -192,17 +198,18 @@ class BambaConfig(PreTrainedConfig):
         self.mamba_chunk_size = mamba_chunk_size
         self.mamba_conv_bias = mamba_conv_bias
         self.mamba_proj_bias = mamba_proj_bias
+        self.time_step_min = time_step_min
+        self.time_step_max = time_step_max
+        self.time_step_limit = tuple(time_step_limit) if time_step_limit is not None else None
         self.z_loss_coefficient = z_loss_coefficient
         self.rope_parameters = rope_parameters
         kwargs["partial_rotary_factor"] = 0.5  # hardcode for BC
 
-        super().__init__(
-            pad_token_id=pad_token_id,
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
+        self.tie_word_embeddings = tie_word_embeddings
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        super().__init__(**kwargs)
 
     @property
     def layers_block_type(self):

transformers/models/bamba/modeling_bamba.py

@@ -4,7 +4,6 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_bamba.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
-# coding=utf-8
 # Copyright 2024 IBM and the HuggingFace Inc. team. All rights reserved.
 #
 # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
@@ -25,7 +24,7 @@
 # limitations under the License.
 
 from collections.abc import Callable
-from typing import Any, Optional, TypedDict, Union
+from typing import Any, Optional, TypedDict
 
 import torch
 from torch import nn
@@ -42,8 +41,8 @@ from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 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 TransformersKwargs, auto_docstring, can_return_tuple, logging
-from ...utils.generic import maybe_autocast
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, is_torchdynamo_compiling, logging
+from ...utils.generic import is_flash_attention_requested, maybe_autocast
 from .configuration_bamba import BambaConfig
 
 
@@ -55,17 +54,16 @@ class BambaFlashAttentionKwargs(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`):
+        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
@@ -140,7 +138,7 @@ class HybridMambaAttentionDynamicCache:
         key_states: torch.Tensor,
         value_states: torch.Tensor,
         layer_idx: int,
-        cache_kwargs: Optional[dict[str, Any]] = None,
+        cache_kwargs: dict[str, Any] | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         # Update the cache
         if self.key_cache[layer_idx].shape[-1] == 0:
@@ -173,7 +171,7 @@ class HybridMambaAttentionDynamicCache:
         kv_length = self.get_seq_length(layer_idx) + query_length
         return kv_length, kv_offset
 
-    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+    def get_seq_length(self, layer_idx: int | None = 0) -> int:
         """Returns the sequence length of the cached states. A layer index can be optionally passed."""
         # take any layer that contains cache and not empty tensor
         layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
@@ -203,9 +201,9 @@ class BambaRotaryEmbedding(nn.Module):
 
     @staticmethod
     def compute_default_rope_parameters(
-        config: Optional[BambaConfig] = None,
+        config: BambaConfig | None = None,
         device: Optional["torch.device"] = None,
-        seq_len: Optional[int] = None,
+        seq_len: int | None = None,
     ) -> tuple["torch.Tensor", float]:
         """
         Computes the inverse frequencies according to the original RoPE implementation
@@ -271,7 +269,7 @@ def eager_attention_forward(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attention_mask: Optional[torch.Tensor],
+    attention_mask: torch.Tensor | None,
     scaling: float,
     dropout: float = 0.0,
     **kwargs: Unpack[TransformersKwargs],
@@ -293,7 +291,7 @@ def eager_attention_forward(
 
 
 # Adapted from transformers.models.glm.modular_glm.apply_rotary_pos_emb
-def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
     Removes the interleaving of cos and sin from GLM
@@ -303,8 +301,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
         k (`torch.Tensor`): The key tensor.
         cos (`torch.Tensor`): The cosine part of the rotary embedding.
         sin (`torch.Tensor`): The sine part of the rotary embedding.
-        position_ids (`torch.Tensor`, *optional*):
-            Deprecated and unused.
         unsqueeze_dim (`int`, *optional*, defaults to 1):
             The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
             sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
@@ -363,10 +359,10 @@ class BambaAttention(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_values: Optional[Cache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        attention_mask: torch.Tensor | None = None,
+        past_key_values: Cache | None = None,
+        cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[TransformersKwargs],
     ) -> tuple[torch.Tensor, torch.Tensor]:
         input_shape = hidden_states.shape[:-1]
@@ -384,9 +380,9 @@ class BambaAttention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -522,10 +518,9 @@ class BambaMixer(nn.Module):
         self.head_dim = config.mamba_d_head
         self.chunk_size = config.mamba_chunk_size
 
-        # FIXME:
-        self.time_step_limit = (0.0, float("inf"))
-        self.time_step_min = 0.001
-        self.time_step_max = 0.1
+        self.time_step_limit = config.time_step_limit
+        self.time_step_min = config.time_step_min
+        self.time_step_max = config.time_step_max
 
         self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
         self.conv1d = nn.Conv1d(
@@ -585,10 +580,10 @@ class BambaMixer(nn.Module):
     def cuda_kernels_forward(
         self,
         hidden_states: torch.Tensor,
-        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        seq_idx: Optional[torch.IntTensor] = None,
+        cache_params: HybridMambaAttentionDynamicCache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        seq_idx: torch.IntTensor | None = None,
     ):
         # 1. Gated MLP's linear projection
         hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
@@ -755,9 +750,9 @@ class BambaMixer(nn.Module):
     def torch_forward(
         self,
         input_states,
-        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
+        cache_params: HybridMambaAttentionDynamicCache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
     ):
         batch_size, seq_len, _ = input_states.shape
         dtype = input_states.dtype
@@ -962,13 +957,13 @@ class BambaMixer(nn.Module):
     def forward(
         self,
         hidden_states,
-        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        seq_idx: Optional[torch.IntTensor] = None,
+        cache_params: HybridMambaAttentionDynamicCache | None = None,
+        cache_position: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        seq_idx: torch.IntTensor | None = None,
         **kwargs,
     ):
-        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
+        if is_fast_path_available and "cuda" in self.in_proj.weight.device.type and not is_torchdynamo_compiling():
             return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask, seq_idx)
         if seq_idx is not None:
             raise NotImplementedError(
@@ -1040,15 +1035,15 @@ class BambaDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[HybridMambaAttentionDynamicCache] = None,
-        output_attentions: Optional[bool] = False,
-        use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: HybridMambaAttentionDynamicCache | None = None,
+        output_attentions: bool | None = False,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
         **kwargs: Unpack[BambaFlashAttentionKwargs],
-    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
+    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
         """
         Args:
             hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
@@ -1161,15 +1156,15 @@ class BambaModel(BambaPreTrainedModel):
     @auto_docstring
     def forward(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[HybridMambaAttentionDynamicCache] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: HybridMambaAttentionDynamicCache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
         **kwargs: Unpack[BambaFlashAttentionKwargs],
     ) -> BaseModelOutputWithPast:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
@@ -1263,7 +1258,7 @@ class BambaModel(BambaPreTrainedModel):
         past_key_values: HybridMambaAttentionDynamicCache,
         output_attentions: bool,
     ):
-        if self.config._attn_implementation == "flash_attention_2":
+        if is_flash_attention_requested(self.config):
             if attention_mask is not None and 0.0 in attention_mask:
                 return attention_mask
             return None
@@ -1386,7 +1381,7 @@ class BambaModel(BambaPreTrainedModel):
 @auto_docstring
 class BambaForCausalLM(BambaPreTrainedModel, GenerationMixin):
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
-    _tp_plan = {"lm_head": "colwise_rep"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
 
     def __init__(self, config):
@@ -1403,17 +1398,17 @@ class BambaForCausalLM(BambaPreTrainedModel, GenerationMixin):
     @auto_docstring
     def forward(
         self,
-        input_ids: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[HybridMambaAttentionDynamicCache] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        labels: Optional[torch.LongTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        logits_to_keep: Union[int, torch.Tensor] = 0,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: HybridMambaAttentionDynamicCache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
         **kwargs,
     ) -> CausalLMOutputWithPast:
         r"""
@@ -1492,55 +1487,24 @@ class BambaForCausalLM(BambaPreTrainedModel, GenerationMixin):
     ):
         # Overwritten -- has a unique cache type, `HybridMambaAttentionDynamicCache`
 
-        empty_past_kv = past_key_values is None
-
-        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
-        # Exception 1: when passing input_embeds, input_ids may be missing entries
-        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
-        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
-        #              (we can't check exception 3 while compiling)
-        if not empty_past_kv:
-            if (
-                inputs_embeds is not None  # Exception 1
-                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
-            ):
-                input_ids = input_ids[:, -cache_position.shape[0] :]
-            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
-                input_ids = input_ids[:, cache_position]
-        else:
+        if past_key_values is None:
             past_key_values = HybridMambaAttentionDynamicCache(
                 self.config, input_ids.shape[0], self.dtype, device=self.device
             )
 
-        if attention_mask is not None and position_ids is None:
-            # create position_ids on the fly for batch generation
-            position_ids = attention_mask.long().cumsum(-1) - 1
-            position_ids.masked_fill_(attention_mask == 0, 1)
-            if not empty_past_kv:
-                position_ids = position_ids[:, -input_ids.shape[1] :]
-
-        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
-        if inputs_embeds is not None and is_first_iteration:
-            model_inputs = {"inputs_embeds": inputs_embeds}
-        else:
-            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
-
-        model_inputs.update(
-            {
-                "position_ids": position_ids,
-                "past_key_values": past_key_values,
-                "use_cache": use_cache,
-                "attention_mask": attention_mask,
-                "logits_to_keep": self.config.num_logits_to_keep,
-                "cache_position": cache_position,
-            }
+        kwargs["logits_to_keep"] = self.config.num_logits_to_keep
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            position_ids=position_ids,
+            use_cache=use_cache,
+            is_first_iteration=is_first_iteration,
+            **kwargs,
         )
 
-        # Forward ALL kwargs that are uninitialized (e.g. `use_cache`).
-        for key, value in kwargs.items():
-            if key not in model_inputs:
-                model_inputs[key] = value
-
         return model_inputs