fix

modeling_deepseek.py

@@ -651,12 +651,280 @@ class DeepseekV3Attention(nn.Module):
         return attn_output, attn_weights, past_key_value
 
 
+# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with Llama->DeepseekV3
 class DeepseekV3FlashAttention2(DeepseekV3Attention):
     """
-    Omitted for brevity - see original code above if you want flash attention integration
+    DeepseekV3 flash attention module. This module inherits from `DeepseekV3Attention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
     """
-    # Implementation remains the same as above...
-    pass
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # DeepseekV3FlashAttention2 attention does not support output_attentions
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+
+            # overwrite attention_mask with padding_mask
+            attention_mask = kwargs.pop("padding_mask")
+
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        if self.q_lora_rank is None:
+            q = self.q_proj(hidden_states)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q = q.view(bsz, q_len, self.num_heads, self.q_head_dim).transpose(1, 2)
+        q_nope, q_pe = torch.split(
+            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
+        )
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        compressed_kv, k_pe = torch.split(
+            compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+        )
+        k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)
+        kv = (
+            self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
+            .view(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
+            .transpose(1, 2)
+        )
+
+        k_nope, value_states = torch.split(
+            kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1
+        )
+        kv_seq_len = value_states.shape[-2]
+
+        kv_seq_len = value_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
+
+        query_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
+        query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
+        query_states[:, :, :, self.qk_nope_head_dim :] = q_pe
+
+        key_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
+        key_states[:, :, :, : self.qk_nope_head_dim] = k_nope
+        key_states[:, :, :, self.qk_nope_head_dim :] = k_pe
+
+        if self.q_head_dim != self.v_head_dim:
+            value_states = F.pad(value_states, [0, self.q_head_dim - self.v_head_dim])
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (DeepseekV3RMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            # Handle the case where the model is quantized
+            if hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            elif torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            else:
+                target_dtype = (
+                    self.q_proj.weight.dtype
+                    if self.q_lora_rank is None
+                    else self.q_a_proj.weight.dtype
+                )
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            softmax_scale=self.softmax_scale,
+        )
+        if self.q_head_dim != self.v_head_dim:
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        attn_output = attn_output.reshape(
+            bsz, q_len, self.num_heads * self.v_head_dim
+        ).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in DeepseekV3FlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            (
+                query_states,
+                key_states,
+                value_states,
+                indices_q,
+                cu_seq_lens,
+                max_seq_lens,
+            ) = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(
+                attn_output_unpad, indices_q, batch_size, query_length
+            )
+        else:
+            attn_output = flash_attn_func(
+                query_states,
+                key_states,
+                value_states,
+                dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+        return attn_output
+
+    def _upad_input(
+        self, query_layer, key_layer, value_layer, attention_mask, query_length
+    ):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim),
+                indices_k,
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(
+                query_layer, attention_mask
+            )
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
 
 
 ATTENTION_CLASSES = {
@@ -664,7 +932,6 @@ ATTENTION_CLASSES = {
     "flash_attention_2": DeepseekV3FlashAttention2,
 }
 
-
 class DeepseekV3DecoderLayer(nn.Module):
     def __init__(self, config: DeepseekV3Config, layer_idx: int):
         super().__init__()