transformers 5.0.0__py3-none-any.whl → 5.0.0rc0__py3-none-any.whl

transformers/models/bamba/configuration_bamba.py

@@ -1,3 +1,4 @@
+# coding=utf-8
 # Copyright 2024 IBM and the HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -13,6 +14,8 @@
 # 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
@@ -113,35 +116,35 @@ class BambaConfig(PreTrainedConfig):
 
     def __init__(
         self,
-        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,
-        z_loss_coefficient: float | None = 0.0,
-        rope_parameters: RopeParameters | None = None,
+        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,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -193,11 +196,13 @@ class BambaConfig(PreTrainedConfig):
         self.rope_parameters = rope_parameters
         kwargs["partial_rotary_factor"] = 0.5  # hardcode for BC
 
-        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)
+        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,
+        )
 
     @property
     def layers_block_type(self):

transformers/models/bamba/modeling_bamba.py

@@ -4,6 +4,7 @@
 #             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
@@ -24,7 +25,7 @@
 # limitations under the License.
 
 from collections.abc import Callable
-from typing import Any, Optional, TypedDict
+from typing import Any, Optional, TypedDict, Union
 
 import torch
 from torch import nn
@@ -34,7 +35,7 @@ from transformers.activations import ACT2FN
 from ... import initialization as init
 from ...cache_utils import Cache
 from ...generation import GenerationMixin
-from ...integrations import lazy_load_kernel, use_kernel_forward_from_hub, use_kernelized_func
+from ...integrations import use_kernel_forward_from_hub
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
@@ -42,10 +43,22 @@ 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 is_flash_attention_requested, maybe_autocast
+from ...utils.import_utils import is_causal_conv1d_available, is_mamba_2_ssm_available
 from .configuration_bamba import BambaConfig
 
 
+if is_mamba_2_ssm_available():
+    from mamba_ssm.ops.triton.selective_state_update import selective_state_update
+    from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
+else:
+    selective_state_update = None
+
+if is_causal_conv1d_available():
+    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+else:
+    causal_conv1d_update, causal_conv1d_fn = None, None
+
+
 logger = logging.get_logger(__name__)
 
 
@@ -54,16 +67,17 @@ 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.
 
-    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.
+    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
@@ -138,7 +152,7 @@ class HybridMambaAttentionDynamicCache:
         key_states: torch.Tensor,
         value_states: torch.Tensor,
         layer_idx: int,
-        cache_kwargs: dict[str, Any] | None = None,
+        cache_kwargs: Optional[dict[str, Any]] = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         # Update the cache
         if self.key_cache[layer_idx].shape[-1] == 0:
@@ -171,7 +185,7 @@ class HybridMambaAttentionDynamicCache:
         kv_length = self.get_seq_length(layer_idx) + query_length
         return kv_length, kv_offset
 
-    def get_seq_length(self, layer_idx: int | None = 0) -> int:
+    def get_seq_length(self, layer_idx: Optional[int] = 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
@@ -197,13 +211,13 @@ class BambaRotaryEmbedding(nn.Module):
         inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
 
         self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+        self.original_inv_freq = inv_freq
 
     @staticmethod
     def compute_default_rope_parameters(
-        config: BambaConfig | None = None,
+        config: Optional[BambaConfig] = None,
         device: Optional["torch.device"] = None,
-        seq_len: int | None = None,
+        seq_len: Optional[int] = None,
     ) -> tuple["torch.Tensor", float]:
         """
         Computes the inverse frequencies according to the original RoPE implementation
@@ -236,7 +250,7 @@ class BambaRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -269,7 +283,7 @@ def eager_attention_forward(
     query: torch.Tensor,
     key: torch.Tensor,
     value: torch.Tensor,
-    attention_mask: torch.Tensor | None,
+    attention_mask: Optional[torch.Tensor],
     scaling: float,
     dropout: float = 0.0,
     **kwargs: Unpack[TransformersKwargs],
@@ -291,7 +305,7 @@ def eager_attention_forward(
 
 
 # Adapted from transformers.models.glm.modular_glm.apply_rotary_pos_emb
-def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     """Applies Rotary Position Embedding to the query and key tensors.
 
     Removes the interleaving of cos and sin from GLM
@@ -301,6 +315,8 @@ def apply_rotary_pos_emb(q, k, cos, sin, 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
@@ -329,7 +345,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
     return q_embed, k_embed
 
 
-@use_kernelized_func(apply_rotary_pos_emb)
 class BambaAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -355,14 +370,15 @@ class BambaAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
+        self.rotary_fn = apply_rotary_pos_emb
 
     def forward(
         self,
         hidden_states: torch.Tensor,
-        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,
+        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,
         **kwargs: Unpack[TransformersKwargs],
     ) -> tuple[torch.Tensor, torch.Tensor]:
         input_shape = hidden_states.shape[:-1]
@@ -484,6 +500,9 @@ def apply_mask_to_padding_states(hidden_states, attention_mask):
     return hidden_states
 
 
+is_fast_path_available = all((selective_state_update, causal_conv1d_fn, causal_conv1d_update))
+
+
 # Adapted from transformers.models.mamba2.modeling_mamba2.Mamba2Mixer
 class BambaMixer(nn.Module):
     """
@@ -555,20 +574,6 @@ class BambaMixer(nn.Module):
 
         self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.use_bias)
 
-        global causal_conv1d_update, causal_conv1d_fn
-        causal_conv1d = lazy_load_kernel("causal-conv1d")
-        causal_conv1d_update = getattr(causal_conv1d, "causal_conv1d_update", None)
-        causal_conv1d_fn = getattr(causal_conv1d, "causal_conv1d_fn", None)
-
-        global selective_state_update, mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
-        mamba_ssm = lazy_load_kernel("mamba-ssm")
-        selective_state_update = getattr(mamba_ssm, "selective_state_update", None)
-        mamba_chunk_scan_combined = getattr(mamba_ssm, "mamba_chunk_scan_combined", None)
-        mamba_split_conv1d_scan_combined = getattr(mamba_ssm, "mamba_split_conv1d_scan_combined", None)
-
-        global is_fast_path_available
-        is_fast_path_available = all((selective_state_update, causal_conv1d_fn, causal_conv1d_update))
-
         if not is_fast_path_available:
             logger.warning_once(
                 "The fast path is not available because one of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
@@ -581,10 +586,10 @@ class BambaMixer(nn.Module):
     def cuda_kernels_forward(
         self,
         hidden_states: torch.Tensor,
-        cache_params: HybridMambaAttentionDynamicCache | None = None,
-        cache_position: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        seq_idx: torch.IntTensor | None = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        seq_idx: Optional[torch.IntTensor] = None,
     ):
         # 1. Gated MLP's linear projection
         hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
@@ -751,9 +756,9 @@ class BambaMixer(nn.Module):
     def torch_forward(
         self,
         input_states,
-        cache_params: HybridMambaAttentionDynamicCache | None = None,
-        cache_position: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
     ):
         batch_size, seq_len, _ = input_states.shape
         dtype = input_states.dtype
@@ -958,10 +963,10 @@ class BambaMixer(nn.Module):
     def forward(
         self,
         hidden_states,
-        cache_params: HybridMambaAttentionDynamicCache | None = None,
-        cache_position: torch.LongTensor | None = None,
-        attention_mask: torch.Tensor | None = None,
-        seq_idx: torch.IntTensor | None = None,
+        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        seq_idx: Optional[torch.IntTensor] = None,
         **kwargs,
     ):
         if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
@@ -1036,15 +1041,15 @@ class BambaDecoderLayer(GradientCheckpointingLayer):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        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,
+        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,
         **kwargs: Unpack[BambaFlashAttentionKwargs],
-    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
             hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
@@ -1157,15 +1162,15 @@ class BambaModel(BambaPreTrainedModel):
     @auto_docstring
     def forward(
         self,
-        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,
+        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,
         **kwargs: Unpack[BambaFlashAttentionKwargs],
     ) -> BaseModelOutputWithPast:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
@@ -1259,7 +1264,7 @@ class BambaModel(BambaPreTrainedModel):
         past_key_values: HybridMambaAttentionDynamicCache,
         output_attentions: bool,
     ):
-        if is_flash_attention_requested(self.config):
+        if self.config._attn_implementation == "flash_attention_2":
             if attention_mask is not None and 0.0 in attention_mask:
                 return attention_mask
             return None
@@ -1399,17 +1404,17 @@ class BambaForCausalLM(BambaPreTrainedModel, GenerationMixin):
     @auto_docstring
     def forward(
         self,
-        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,
+        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,
         **kwargs,
     ) -> CausalLMOutputWithPast:
         r"""
@@ -1483,7 +1488,6 @@ class BambaForCausalLM(BambaPreTrainedModel, GenerationMixin):
         cache_position=None,
         position_ids=None,
         use_cache=True,
-        is_first_iteration=False,
         **kwargs,
     ):
         # Overwritten -- has a unique cache type, `HybridMambaAttentionDynamicCache`
@@ -1516,7 +1520,7 @@ class BambaForCausalLM(BambaPreTrainedModel, GenerationMixin):
                 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:
+        if inputs_embeds is not None and empty_past_kv:
             model_inputs = {"inputs_embeds": inputs_embeds}
         else:
             model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases