~test conditional workarounds

modeling_sdar.py

@@ -23,15 +23,20 @@
 
 from typing import Callable, Optional, Tuple, Union
 
+from nnll.init_gpu import device
 import torch
 from torch import nn
-
+import torch.nn.functional as F
 from transformers.activations import ACT2FN
-from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
+from transformers.cache_utils import (
+    Cache,
+    DynamicCache,
+    SlidingWindowCache,
+    StaticCache,
+)
 from transformers.generation import GenerationMixin
 from transformers.integrations import use_kernel_forward_from_hub
 from transformers.modeling_attn_mask_utils import AttentionMaskConverter
-from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
 from transformers.modeling_layers import GradientCheckpointingLayer
 from transformers.modeling_outputs import (
     BaseModelOutputWithPast,
@@ -43,31 +48,63 @@ from transformers.modeling_outputs import (
 from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from transformers.processing_utils import Unpack
-from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
-from .configuration_sdar import SDARConfig
+from transformers.utils import (
+    LossKwargs,
+    auto_docstring,
+    can_return_tuple,
+    is_torch_flex_attn_available,
+    logging,
+)
 
-from flash_attn.ops.triton.layer_norm import rms_norm_fn as flash_rms_norm
+from divisor.trado.configuration_sdar import SDARConfig
 
-import torch.nn.functional as F
-try:
-    from flash_attn import flash_attn_func, flash_attn_varlen_func
-    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
-except:
-    pass
+logger = logging.get_logger(__name__)
 
+# Make FlashAttentionKwargs available for all devices (used in type hints)
 try:
-    from liger_kernel.ops.swiglu import LigerSiLUMulFunction  # noqa: F401
-    liger_kernel_is_available = True
+    from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
 except ImportError:
-    liger_kernel_is_available = False
-
+    # Fallback if not available
+    from typing import TypedDict
+
+    FlashAttentionKwargs = TypedDict("FlashAttentionKwargs", {})
+
+# Conditionally import flash attention components (CUDA only)
+flash_rms_norm = None
+flash_attn_func = None
+flash_attn_varlen_func = None
+index_first_axis = None
+pad_input = None
+unpad_input = None
+
+if device.type == "cuda":
+    try:
+        from flash_attn.ops.triton.layer_norm import rms_norm_fn as flash_rms_norm
+    except (ImportError, ModuleNotFoundError):
+        logger.warning("Flash attention RMS norm not available. Falling back to standard implementation.")
+        flash_rms_norm = None
+
+    try:
+        from flash_attn import flash_attn_func, flash_attn_varlen_func
+        from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
+    except (ImportError, ModuleNotFoundError):
+        logger.warning("Flash attention not available. Falling back to standard attention.")
+        flash_attn_func = None
+        flash_attn_varlen_func = None
 
 if is_torch_flex_attn_available():
-    from torch.nn.attention.flex_attention import BlockMask, create_block_mask, flex_attention
-    from transformers.integrations.flex_attention import make_flex_block_causal_mask
+    try:
+        from torch.nn.attention.flex_attention import BlockMask, create_block_mask, flex_attention
+        from transformers.integrations.flex_attention import make_flex_block_causal_mask
+    except ImportError:
+        pass
 
+try:
+    from liger_kernel.ops.swiglu import LigerSiLUMulFunction  # noqa: F401
 
-logger = logging.get_logger(__name__)
+    liger_kernel_is_available = True
+except ImportError:
+    liger_kernel_is_available = False
 
 
 @use_kernel_forward_from_hub("RMSNorm")
@@ -81,16 +118,20 @@ class SDARRMSNorm(nn.Module):
         self.variance_epsilon = eps
 
     def forward(self, hidden_states):
-        return flash_rms_norm(
-            hidden_states, weight=self.weight, bias=None, eps=self.variance_epsilon)
-        '''
+        # Use flash RMS norm if available (CUDA only), otherwise fall back to standard implementation
+        if flash_rms_norm is not None and hidden_states.device.type == "cuda":
+            try:
+                return flash_rms_norm(hidden_states, weight=self.weight, bias=None, eps=self.variance_epsilon)
+            except Exception as e:
+                logger.warning(f"Flash RMS norm failed ({e}). Falling back to standard implementation.")
+                # Fall through to standard implementation
+
+        # Standard RMS norm implementation (fallback for MPS, CPU, or when flash_rms_norm fails)
         input_dtype = hidden_states.dtype
         hidden_states = hidden_states.to(torch.float32)
         variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * \
-            torch.rsqrt(variance + self.variance_epsilon)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
         return self.weight * hidden_states.to(input_dtype)
-        '''
 
     def extra_repr(self):
         return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
@@ -102,27 +143,23 @@ class SDARMLP(nn.Module):
         self.config = config
         self.hidden_size = config.hidden_size
         self.intermediate_size = config.intermediate_size
-        self.gate_proj = nn.Linear(
-            self.hidden_size, self.intermediate_size, bias=False)
-        self.up_proj = nn.Linear(
-            self.hidden_size, self.intermediate_size, bias=False)
-        self.down_proj = nn.Linear(
-            self.intermediate_size, self.hidden_size, bias=False)
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
         self.act_fn = ACT2FN[config.hidden_act]
 
     def forward(self, x):
         if liger_kernel_is_available:
             return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))
         else:
-            down_proj = self.down_proj(self.act_fn(
-                self.gate_proj(x)) * self.up_proj(x))
+            down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
             return down_proj
 
 
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
     x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2:]
+    x2 = x[..., x.shape[-1] // 2 :]
     return torch.cat((-x2, x1), dim=-1)
 
 
@@ -160,8 +197,7 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     batch, num_key_value_heads, slen, head_dim = hidden_states.shape
     if n_rep == 1:
         return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(
-        batch, num_key_value_heads, n_rep, slen, head_dim)
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
     return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
 
 
@@ -183,10 +219,8 @@ def eager_attention_forward(
         causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
         attn_weights = attn_weights + causal_mask
 
-    attn_weights = nn.functional.softmax(
-        attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
-    attn_weights = nn.functional.dropout(
-        attn_weights, p=dropout, training=module.training)
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
     attn_output = torch.matmul(attn_weights, value_states)
     attn_output = attn_output.transpose(1, 2).contiguous()
 
@@ -200,8 +234,7 @@ class SDARAttention(nn.Module):
         super().__init__()
         self.config = config
         self.layer_idx = layer_idx
-        self.head_dim = getattr(
-            config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
         self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
         self.scaling = self.head_dim**-0.5
         self.attention_dropout = config.attention_dropout
@@ -211,28 +244,16 @@ class SDARAttention(nn.Module):
         self.num_attention_heads = config.num_attention_heads
         self.num_key_value_heads = config.num_key_value_heads
 
-        self.q_proj = nn.Linear(
-            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.k_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.v_proj = nn.Linear(
-            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
-        )
-        self.o_proj = nn.Linear(
-            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
-        )
+        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias)
+        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias)
         # unlike olmo, only on the head dim!
         self.q_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
         # thus post q_norm does not need reshape
         self.k_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
         self.sliding_window = config.sliding_window
-        if not (
-            self.config.use_sliding_window
-            and getattr(self.config, "sliding_window", None) is not None
-            and self.layer_idx >= self.config.max_window_layers
-        ):
+        if not (self.config.use_sliding_window and getattr(self.config, "sliding_window", None) is not None and self.layer_idx >= self.config.max_window_layers):
             self.sliding_window = None
 
     def forward(
@@ -248,32 +269,23 @@ class SDARAttention(nn.Module):
         bsz, q_len = input_shape
         hidden_shape = (*input_shape, -1, self.head_dim)
 
-        query_states = self.q_norm(self.q_proj(
-            hidden_states).view(hidden_shape)).transpose(1, 2)
-        key_states = self.k_norm(self.k_proj(
-            hidden_states).view(hidden_shape)).transpose(1, 2)
-        value_states = self.v_proj(hidden_states).view(
-            hidden_shape).transpose(1, 2)
-        
-        
+        query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
 
         cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(
-            query_states, key_states, cos, sin)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 
         if past_key_value is not None and kwargs.get("store_kv", False):
             # sin and cos are specific to RoPE models; cache_position needed for the static cache
-            key_states, value_states = past_key_value.update(
-                key_states, value_states, self.layer_idx)
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
         elif past_key_value is not None and not kwargs.get("store_kv", False) and len(past_key_value) > self.layer_idx:
             # only retrive, do not store kv
             past_key_states, past_value_states = past_key_value[self.layer_idx]
-            key_states = torch.cat(
-                [past_key_states, key_states], dim=-2)
-            value_states = torch.cat(
-                [past_value_states, value_states], dim=-2)
+            key_states = torch.cat([past_key_states, key_states], dim=-2)
+            value_states = torch.cat([past_value_states, value_states], dim=-2)
 
-        '''
+        """
         attention_mask = attention_mask.bool() if attention_mask is not None else None
         if torch.all(attention_mask):  # decoding
             query_states = query_states.transpose(1, 2)
@@ -298,12 +310,12 @@ class SDARAttention(nn.Module):
                 enable_gqa=True
             )
             attn_output = attn_output.transpose(1, 2).contiguous()
-        '''
+        """
 
-        #print(query_states.shape, key_states.shape, value_states.shape)
+        # print(query_states.shape, key_states.shape, value_states.shape)
 
         # --- After RoPE and KV-cache handling, expand KV to all heads ---
-        key_states   = repeat_kv(key_states,   self.num_key_value_groups)  # [B, H, K, D]
+        key_states = repeat_kv(key_states, self.num_key_value_groups)  # [B, H, K, D]
         value_states = repeat_kv(value_states, self.num_key_value_groups)  # [B, H, K, D]
 
         # --- Convert a 0/1 or bool 4D mask into an *additive* mask, and align to [B, H, Q, K] ---
@@ -313,14 +325,14 @@ class SDARAttention(nn.Module):
             am = attention_mask
             # Support either 2D [B, K] or 4D [B, 1/H, Q, K]
             if am.dim() == 2:
-                am = am[:, None, None, :k_len]             # -> [B,1,1,K]
+                am = am[:, None, None, :k_len]  # -> [B,1,1,K]
             else:
-                am = am[:, :, :, :k_len]                   # -> [B,1/H,Q,K]
+                am = am[:, :, :, :k_len]  # -> [B,1/H,Q,K]
 
             finfo_min = torch.finfo(query_states.dtype).min
             # 0/1 or bool -> float additive mask: 1->0, 0->-inf
             if am.dtype == torch.bool:
-                zero   = torch.zeros((), dtype=query_states.dtype, device=am.device)
+                zero = torch.zeros((), dtype=query_states.dtype, device=am.device)
                 neginf = torch.full((), finfo_min, dtype=query_states.dtype, device=am.device)
                 am = torch.where(am, zero, neginf)
             else:
@@ -329,33 +341,59 @@ class SDARAttention(nn.Module):
                 am = torch.where(am > 0, torch.zeros_like(am), torch.full_like(am, finfo_min))
 
             # Expand to all heads
-            #if am.shape[1] == 1 and self.num_attention_heads > 1:
+            # if am.shape[1] == 1 and self.num_attention_heads > 1:
             #    am = am.expand(am.shape[0], self.num_attention_heads, am.shape[2], am.shape[3])
 
-            #attn_mask = am.contiguous()
+            # attn_mask = am.contiguous()
             attn_mask = am
-        
 
         bsz, q_len = input_shape
 
         if q_len == 1 and past_key_value is not None:
-            # --- Decoding: flash-attn ---
-            q = query_states.transpose(1, 2)  # [B,Q,H,D]
-            k = key_states.transpose(1, 2)
-            v = value_states.transpose(1, 2)
-            attn_output = flash_attn_func(
-                q, k, v,
-                causal=True,                # For decoding, explicitly set causal=True
-                softmax_scale=self.scaling
-            )
-            attn_output = attn_output.transpose(1, 2).contiguous()
+            # --- Decoding: try flash-attn if available (CUDA only), otherwise fall back to SDPA ---
+            if flash_attn_func is not None and query_states.device.type == "cuda":
+                try:
+                    q = query_states.transpose(1, 2)  # [B,Q,H,D]
+                    k = key_states.transpose(1, 2)
+                    v = value_states.transpose(1, 2)
+                    attn_output = flash_attn_func(
+                        q,
+                        k,
+                        v,
+                        causal=True,  # For decoding, explicitly set causal=True
+                        softmax_scale=self.scaling,
+                    )
+                    attn_output = attn_output.transpose(1, 2).contiguous()
+                except Exception as e:
+                    logger.warning(f"Flash attention failed during decoding ({e}). Falling back to SDPA.")
+                    # Fall through to SDPA implementation below
+                    attn_output = F.scaled_dot_product_attention(
+                        query=query_states,  # [B,H,Q,D]
+                        key=key_states,  # [B,H,K,D]
+                        value=value_states,  # [B,H,K,D]
+                        attn_mask=attn_mask,  # float additive mask
+                        is_causal=False,  # All constraints are already encoded in the mask
+                        scale=self.scaling,
+                    )
+                    attn_output = attn_output.transpose(1, 2).contiguous()  # -> [B,Q,H,D]
+            else:
+                # Fallback to SDPA for MPS, CPU, or when flash_attn_func is not available
+                attn_output = F.scaled_dot_product_attention(
+                    query=query_states,  # [B,H,Q,D]
+                    key=key_states,  # [B,H,K,D]
+                    value=value_states,  # [B,H,K,D]
+                    attn_mask=attn_mask,  # float additive mask
+                    is_causal=False,  # All constraints are already encoded in the mask
+                    scale=self.scaling,
+                )
+                attn_output = attn_output.transpose(1, 2).contiguous()  # -> [B,Q,H,D]
         else:
             attn_output = F.scaled_dot_product_attention(
-                query=query_states,         # [B,H,Q,D]
-                key=key_states,             # [B,H,K,D]
-                value=value_states,         # [B,H,K,D]
-                attn_mask=attn_mask,        # float additive mask
-                is_causal=False,            # All constraints are already encoded in the mask
+                query=query_states,  # [B,H,Q,D]
+                key=key_states,  # [B,H,K,D]
+                value=value_states,  # [B,H,K,D]
+                attn_mask=attn_mask,  # float additive mask
+                is_causal=False,  # All constraints are already encoded in the mask
                 scale=self.scaling,
             )
             attn_output = attn_output.transpose(1, 2).contiguous()  # -> [B,Q,H,D]
@@ -371,17 +409,10 @@ class SDARDecoderLayer(GradientCheckpointingLayer):
         self.hidden_size = config.hidden_size
         self.self_attn = SDARAttention(config=config, layer_idx=layer_idx)
         self.mlp = SDARMLP(config)
-        self.input_layernorm = SDARRMSNorm(
-            config.hidden_size, eps=config.rms_norm_eps)
-        self.post_attention_layernorm = SDARRMSNorm(
-            config.hidden_size, eps=config.rms_norm_eps)
-        if (
-            config.sliding_window and config._attn_implementation != "flash_attention_2"
-        ):  # diff with Llama is this warning
-            logger.warning_once(
-                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
-                "unexpected results may be encountered."
-            )
+        self.input_layernorm = SDARRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = SDARRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        if config.sliding_window and config._attn_implementation != "flash_attention_2":  # diff with Llama is this warning
+            logger.warning_once(f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; unexpected results may be encountered.")
 
     def forward(
         self,
@@ -394,8 +425,7 @@ class SDARDecoderLayer(GradientCheckpointingLayer):
         store_kv: Optional[bool] = False,
         cache_position: Optional[torch.LongTensor] = None,
         # necessary, but kept here for BC
-        position_embeddings: Optional[Tuple[torch.Tensor,
-                                            torch.Tensor]] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
         **kwargs: Unpack[FlashAttentionKwargs],
     ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
         residual = hidden_states
@@ -463,8 +493,7 @@ class SDARRotaryEmbedding(nn.Module):
         super().__init__()
         # BC: "rope_type" was originally "type"
         if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
-            self.rope_type = config.rope_scaling.get(
-                "rope_type", config.rope_scaling.get("type"))
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
         else:
             self.rope_type = "default"
         self.max_seq_len_cached = config.max_position_embeddings
@@ -473,8 +502,7 @@ class SDARRotaryEmbedding(nn.Module):
         self.config = config
         self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
 
-        inv_freq, self.attention_scaling = self.rope_init_fn(
-            self.config, device)
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)
         self.original_inv_freq = self.inv_freq
 
@@ -482,15 +510,12 @@ class SDARRotaryEmbedding(nn.Module):
     # power user: used with advanced RoPE types (e.g. dynamic rope)
     @dynamic_rope_update
     def forward(self, x, position_ids):
-        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(
-            position_ids.shape[0], -1, 1).to(x.device)
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
         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"
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
         with torch.autocast(device_type=device_type, enabled=False):  # Force float32
-            freqs = (inv_freq_expanded.float() @
-                     position_ids_expanded.float()).transpose(1, 2)
+            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
             sin = emb.sin() * self.attention_scaling
@@ -505,12 +530,8 @@ class SDARModel(SDARPreTrainedModel):
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size
 
-        self.embed_tokens = nn.Embedding(
-            config.vocab_size, config.hidden_size, self.padding_idx)
-        self.layers = nn.ModuleList(
-            [SDARDecoderLayer(config, layer_idx)
-             for layer_idx in range(config.num_hidden_layers)]
-        )
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([SDARDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
         self.norm = SDARRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.rotary_emb = SDARRotaryEmbedding(config=config)
         self.gradient_checkpointing = False
@@ -541,25 +562,19 @@ class SDARModel(SDARPreTrainedModel):
         **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
     ) -> BaseModelOutputWithPast:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         use_cache = use_cache if use_cache is not None else self.config.use_cache
 
         if (input_ids is None) ^ (inputs_embeds is not None):
-            raise ValueError(
-                "You must specify exactly one of input_ids or inputs_embeds")
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
 
         if self.gradient_checkpointing and self.training and use_cache:
-            logger.warning_once(
-                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
-            )
+            logger.warning_once("`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.")
             use_cache = False
 
         # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
         if not isinstance(past_key_values, (type(None), Cache)):
-            raise ValueError(
-                "The `past_key_values` should be either a `Cache` object or `None`.")
+            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")
 
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
@@ -568,11 +583,8 @@ class SDARModel(SDARPreTrainedModel):
             past_key_values = DynamicCache()
 
         if cache_position is None:
-            past_seen_tokens = past_key_values.get_seq_length(
-            ) if past_key_values is not None else 0
-            cache_position = torch.arange(
-                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
-            )
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device)
 
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
@@ -635,8 +647,7 @@ class SDARModel(SDARPreTrainedModel):
     ):
         if self.config._attn_implementation == "flash_attention_2":
             if attention_mask is not None and past_key_values is not None:
-                is_padding_right = attention_mask[:, -
-                                                  1].sum().item() != input_tensor.size()[0]
+                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
                 if is_padding_right:
                     raise ValueError(
                         "You are attempting to perform batched generation with padding_side='right'"
@@ -653,7 +664,10 @@ class SDARModel(SDARPreTrainedModel):
                 attention_mask = create_block_mask(
                     # 2d bool tensor, shape: [2*seqlen, 2*seqlen]
                     lambda b, h, q_idx, kv_idx: attention_mask[q_idx, kv_idx],
-                    B=None, H=None, Q_LEN=seq_len_q, KV_LEN=seq_len_kv,
+                    B=None,
+                    H=None,
+                    Q_LEN=seq_len_q,
+                    KV_LEN=seq_len_kv,
                 )
             else:
                 # Here we pass in flex mask computed externally
@@ -663,18 +677,12 @@ class SDARModel(SDARPreTrainedModel):
         # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
         # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
         # to infer the attention mask.
-        past_seen_tokens = past_key_values.get_seq_length(
-        ) if past_key_values is not None else 0
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
-        using_sliding_window_cache = isinstance(
-            past_key_values, SlidingWindowCache)
+        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
 
         # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
-        if (
-            self.config._attn_implementation == "sdpa"
-            and not (using_static_cache or using_sliding_window_cache)
-            and not output_attentions
-        ):
+        if self.config._attn_implementation == "sdpa" and not (using_static_cache or using_sliding_window_cache) and not output_attentions:
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
                 attention_mask,
                 inputs_embeds=input_tensor,
@@ -692,11 +700,7 @@ class SDARModel(SDARPreTrainedModel):
             target_length = past_key_values.get_max_cache_shape()
         # DynamicCache or no cache
         else:
-            target_length = (
-                attention_mask.shape[-1]
-                if isinstance(attention_mask, torch.Tensor)
-                else past_seen_tokens + sequence_length + 1
-            )
+            target_length = attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1
 
         # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
         causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
@@ -710,17 +714,11 @@ class SDARModel(SDARPreTrainedModel):
             past_key_values=past_key_values,
         )
 
-        if (
-            self.config._attn_implementation == "sdpa"
-            and attention_mask is not None
-            and attention_mask.device.type in ["cuda", "xpu", "npu"]
-            and not output_attentions
-        ):
+        if self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type in ["cuda", "xpu", "npu", "mps"] and not output_attentions:
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
             # Details: https://github.com/pytorch/pytorch/issues/110213
-            causal_mask = AttentionMaskConverter._unmask_unattended(
-                causal_mask, min_dtype)
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
 
         return causal_mask
 
@@ -761,42 +759,29 @@ class SDARModel(SDARPreTrainedModel):
             causal_mask = attention_mask
         else:
             min_dtype = torch.finfo(dtype).min
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
-            )
-            diagonal_attend_mask = torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
-                -1, 1
-            )
+            causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device)
+            diagonal_attend_mask = torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
             text_config = config.get_text_config()
             if getattr(text_config, "use_sliding_window", True) and text_config.sliding_window is not None:
                 # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
                 # the check is needed to verify is current checkpoint was trained with sliding window or not
                 if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
-                    sliding_attend_mask = torch.arange(target_length, device=cache_position.device) <= (
-                        cache_position.reshape(-1, 1) -
-                        text_config.sliding_window
-                    )
+                    sliding_attend_mask = torch.arange(target_length, device=cache_position.device) <= (cache_position.reshape(-1, 1) - text_config.sliding_window)
                     diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
             causal_mask *= diagonal_attend_mask
-            causal_mask = causal_mask[None, None,
-                                      :, :].expand(batch_size, 1, -1, -1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
             if attention_mask is not None:
                 causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                 if attention_mask.shape[-1] > target_length:
                     attention_mask = attention_mask[:, :target_length]
                 mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
-                    causal_mask.device
-                )
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(causal_mask.device)
                 padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(padding_mask, min_dtype)
         return causal_mask
 
 
-class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs):
-    ...
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
 
 
 @auto_docstring
@@ -809,8 +794,7 @@ class SDARForCausalLM(SDARPreTrainedModel, GenerationMixin):
         super().__init__(config)
         self.model = SDARModel(config)
         self.vocab_size = config.vocab_size
-        self.lm_head = nn.Linear(
-            config.hidden_size, config.vocab_size, bias=False)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
 
         # Initialize weights and apply final processing
         self.post_init()
@@ -868,9 +852,7 @@ class SDARForCausalLM(SDARPreTrainedModel, GenerationMixin):
         "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
         ```"""
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
 
         # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
         outputs: BaseModelOutputWithPast = self.model(
@@ -888,8 +870,7 @@ class SDARForCausalLM(SDARPreTrainedModel, GenerationMixin):
 
         hidden_states = outputs.last_hidden_state
         # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
-        slice_indices = slice(-logits_to_keep,
-                              None) if isinstance(logits_to_keep, int) else logits_to_keep
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
         hidden_states = hidden_states[:, slice_indices, :].contiguous()
         fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
         if fuse_linear_and_cross_entropy:
@@ -903,8 +884,7 @@ class SDARForCausalLM(SDARPreTrainedModel, GenerationMixin):
             # FusedLinearCrossEntropyLoss will be implemented by monkey patch when training
             # We don't use it when inferencing
             loss_fct = nn.CrossEntropyLoss()  # nn.CE
-            loss = loss_fct(
-                logits.view(-1, self.config.vocab_size), labels.view(-1))
+            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
 
         return CausalLMOutputWithPast(
             loss=loss,
@@ -919,4 +899,4 @@ __all__ = [
     "SDARForCausalLM",
     "SDARModel",
     "SDARPreTrainedModel",
-]
+]