support deepseekv2 for maca backend.

lmdeploy/pytorch/backends/dlinfer/attention.py

@@ -132,6 +132,7 @@ def forward(
             is_decoding=is_decoding,
             block_size=block_size,
             attn_mask=attn_mask,
+            softmax_scale=self.scale,
             is_unpaged_prefill=is_unpaged_prefill,
             kv_scales=kv_scales,
             kv_zeros=kv_zeros,

lmdeploy/pytorch/backends/dlinfer/rotary_embedding.py

@@ -5,7 +5,7 @@
 import torch
 from torch import nn
 
-from ..default.rotary_embedding import LlamaDynamicNTKScalingRotaryEmbedding
+from ..default.rotary_embedding import LlamaDynamicNTKScalingRotaryEmbedding, YarnRotaryEmbeddingImpl
 from ..rotary_embedding import (Llama3Parameters, LongRoPEScalingParameters, RopeType, RotaryEmbeddingBuilder,
                                 RotaryEmbeddingImpl, YarnParameters)
 
@@ -128,6 +128,25 @@ def __init__(
         self.register_buffer('inv_freq', inv_freq_llama)
 
 
+class DlinferYarnRotaryEmbeddingImpl(YarnRotaryEmbeddingImpl):
+    """yarn rotary embedding implementation."""
+
+    def __init__(self,
+                 dim: int,
+                 base: int = 10000,
+                 scaling_factor: float = 1.0,
+                 original_max_position_embeddings: int = 4096,
+                 yarn_params: YarnParameters = None):
+        super().__init__(dim, base, scaling_factor, original_max_position_embeddings, yarn_params)
+
+    def forward(self, x: torch.Tensor, position_ids: torch.Tensor):
+        """forward."""
+        dtype = x.dtype
+        if self.inv_freq.device != x.device:
+            self.inv_freq = self.inv_freq.to(x.device)
+        return _rotary_embedding_fwd(position_ids, self.inv_freq, scaling_factor=1.0, mscale=self.mscale, dtype=dtype)
+
+
 class DlinferRotaryEmbeddingBuilder(RotaryEmbeddingBuilder):
     """rotary embedding dlinfer builder."""
 
@@ -150,5 +169,11 @@ def build(
         elif emb_type == RopeType.Llama3:
             return DlinferLlama3RotaryEmbeddingImpl(dim, base, scaling_factor, llama3_params.low_freq_factor,
                                                     llama3_params.high_freq_factor, max_position_embeddings)
+        elif emb_type == RopeType.Yarn:
+            return DlinferYarnRotaryEmbeddingImpl(dim,
+                                                  base,
+                                                  scaling_factor,
+                                                  max_position_embeddings,
+                                                  yarn_params=yarn_params)
         else:
             raise NotImplementedError(f'Unsupported embedding type: {emb_type}')

lmdeploy/pytorch/kernels/dlinfer/pagedattention.py

@@ -20,6 +20,7 @@ def prefill_attention(
     max_kv_seq_len: int,
     block_size: int,
     attn_mask: Sequence[Optional[Tensor]],
+    softmax_scale: Optional[float],
     is_unpaged_prefill: Optional[bool],
     kv_scales: Optional[Tensor],
     kv_zeros: Optional[Tensor],
@@ -39,6 +40,7 @@ def prefill_attention(
             num_q_heads,
             num_kv_heads,
             attn_mask,
+            softmax_scale=softmax_scale,
             attn_output=attn_output,
         )
     else:
@@ -59,6 +61,7 @@ def prefill_attention(
             num_q_heads,
             num_kv_heads,
             attn_mask,
+            softmax_scale=softmax_scale,
             attn_output=attn_output,
             kv_scales=kv_scales,
             kv_zeros=kv_zeros,
@@ -75,6 +78,7 @@ def paged_token_attention(
     max_kv_seq_len,
     block_offsets,
     block_size,
+    softmax_scale: Optional[float],
     kv_scales: Optional[Tensor],
     kv_zeros: Optional[Tensor],
     quant_bits: Optional[int],
@@ -91,6 +95,7 @@ def paged_token_attention(
         max_kv_seq_len,
         num_q_heads,
         num_kv_heads,
+        softmax_scale=softmax_scale,
         attn_output=attn_output,
         kv_scales=kv_scales,
         kv_zeros=kv_zeros,
@@ -115,6 +120,7 @@ def paged_attention_fwd(
     is_decoding: bool,
     block_size: int,
     attn_mask: Sequence[Optional[Tensor]] = (),
+    softmax_scale: Optional[float] = None,
     is_unpaged_prefill: Optional[bool] = None,
     kv_scales: Optional[Tensor] = None,
     kv_zeros: Optional[Tensor] = None,
@@ -137,6 +143,7 @@ def paged_attention_fwd(
             max_kv_seq_len,
             block_size,
             attn_mask,
+            softmax_scale,
             is_unpaged_prefill,
             kv_scales=kv_scales,
             kv_zeros=kv_zeros,
@@ -152,6 +159,7 @@ def paged_attention_fwd(
             max_kv_seq_len,
             block_offsets,
             block_size,
+            softmax_scale=softmax_scale,
             kv_scales=kv_scales,
             kv_zeros=kv_zeros,
             quant_bits=quant_bits,