[0.9.1]optmize rope in qwen2 (#1782)

### What this PR does / why we need it?

Optimize rope by extracting index_select from layers into model, which
can reduce (layer_num -1) * 2 Gather ops in each prefill/decode stage.

### Does this PR introduce _any_ user-facing change?

No

### How was this patch tested?

Pass cos and sin and set skip_index_select=True to enable this
optimization. As shown in the following code :
`q, k = self.rotary_emb(positions, q, k, cos=cos, sin=sin,
skip_index_select=True)`

**Performance results:**
**origin:**
Successful requests:                     400
Benchmark duration (s):                  243.10
Total input tokens:                      1200000
Total generated tokens:                  60000
Request throughput (req/s):              1.65
Output token throughput (tok/s):         246.81
Total Token throughput (tok/s):          5183.02
**optimized:**
Successful requests:                     400
Benchmark duration (s):                  237.42
Total input tokens:                      1200000
Total generated tokens:                  60000
Request throughput (req/s):              1.68
Output token throughput (tok/s):         252.72
Total Token throughput (tok/s):          5307.03

Signed-off-by: David9857 <985700846@qq.com>

Author: David9857

vllm_ascend/models/qwen2.py

@@ -1,11 +1,12 @@
 from collections.abc import Iterable
-from typing import Optional, Union
+from typing import Any, Optional, Union
 
 import torch
 import torch.nn.functional as F
 import vllm.envs as envs
 from torch import nn
 from transformers import Qwen2Config
+from vllm.attention import AttentionType
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
@@ -14,11 +15,14 @@
                               tensor_model_parallel_all_reduce,
                               tensor_model_parallel_reduce_scatter)
 from vllm.forward_context import get_forward_context
+from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
 from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
 from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP
-from vllm.model_executor.models.qwen2 import Qwen2DecoderLayer, Qwen2Model
+from vllm.model_executor.models.qwen2 import (Qwen2Attention, Qwen2MLP,
+                                              Qwen2Model)
 from vllm.model_executor.models.utils import (AutoWeightsLoader,
                                               PPMissingLayer, maybe_prefix)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
@@ -48,7 +52,59 @@ def maybe_pad_and_reduce_scatter(
     return hidden_states
 
 
-class CustomQwen2DecoderLayer(Qwen2DecoderLayer):
+class CustomQwen2Attention(Qwen2Attention):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        max_position: int = 4096 * 32,
+        rope_theta: float = 10000,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        rope_scaling: Optional[tuple] = None,
+        prefix: str = "",
+        attn_type: str = AttentionType.DECODER,
+        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
+    ) -> None:
+        super().__init__(
+            hidden_size=hidden_size,
+            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            max_position=max_position,
+            rope_theta=rope_theta,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            rope_scaling=rope_scaling,
+            prefix=prefix,
+            attn_type=attn_type,
+            dual_chunk_attention_config=dual_chunk_attention_config)
+
+    def forward(self,
+                positions: torch.Tensor,
+                hidden_states: torch.Tensor,
+                cos: Optional[torch.Tensor] = None,
+                sin: Optional[torch.Tensor] = None) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        if type(self.rotary_emb) is RotaryEmbedding:
+            # We optimized RotaryEmbedding by moving index_select of cos & sin outside.
+            # if cos & sin are provided, set is_cos_sin_cached to True to skip index_select.
+            q, k = self.rotary_emb(positions,
+                                   q,
+                                   k,
+                                   cos=cos,
+                                   sin=sin,
+                                   is_cos_sin_cached=True)
+        else:
+            q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class CustomQwen2DecoderLayer(nn.Module):
 
     def __init__(
         self,
@@ -57,10 +113,49 @@ def __init__(
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
-        super().__init__(config=config,
-                         cache_config=cache_config,
-                         quant_config=quant_config,
-                         prefix=prefix)
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        # Requires transformers > 4.32.0
+        rope_theta = getattr(config, "rope_theta", 1000000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        dual_chunk_attention_config = getattr(config,
+                                              "dual_chunk_attention_config",
+                                              None)
+
+        # By default, Qwen2 uses causal attention as it is a decoder-only model.
+        # You can override the HF config with `is_causal=False` to enable
+        # bidirectional attention, which is used in some embedding models
+        # (e.g. Alibaba-NLP/gte-Qwen2-7B-instruct)
+        if getattr(config, "is_causal", True):
+            attn_type = AttentionType.DECODER
+        else:
+            attn_type = AttentionType.ENCODER_ONLY
+
+        self.self_attn = CustomQwen2Attention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            max_position=config.max_position_embeddings,
+            num_kv_heads=config.num_key_value_heads,
+            rope_theta=rope_theta,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            rope_scaling=rope_scaling,
+            prefix=f"{prefix}.self_attn",
+            attn_type=attn_type,
+            dual_chunk_attention_config=dual_chunk_attention_config,
+        )
+        self.mlp = Qwen2MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=f"{prefix}.mlp",
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size,
+                                       eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size,
+                                                eps=config.rms_norm_eps)
+
         self.tp_rank = get_tensor_model_parallel_rank()
         self.tp_size = get_tensor_model_parallel_world_size()
         self.self_attn.o_proj.reduce_results = False
@@ -73,6 +168,8 @@ def forward(
         residual: Optional[torch.Tensor],
         flashcomm_v1_enabled: bool,
         pad_size: int,
+        cos: Optional[torch.Tensor] = None,
+        sin: Optional[torch.Tensor] = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         # Self Attention
         if residual is None:
@@ -89,10 +186,10 @@ def forward(
             if flashcomm_v1_enabled:
                 hidden_states = all_gather_and_maybe_unpad(
                     hidden_states, pad_size)
-        hidden_states = self.self_attn(
-            positions=positions,
-            hidden_states=hidden_states,
-        )
+        hidden_states = self.self_attn(positions=positions,
+                                       hidden_states=hidden_states,
+                                       cos=cos,
+                                       sin=sin)
         if flashcomm_v1_enabled:
             hidden_states = maybe_pad_and_reduce_scatter(
                 hidden_states, pad_size)
@@ -133,6 +230,7 @@ def __init__(
                          prefix=prefix,
                          decoder_layer_type=decoder_layer_type)
         self.tp_size = get_tensor_model_parallel_world_size()
+        self.cos_sin_cache = self.layers[0].self_attn.rotary_emb.cos_sin_cache
 
     def forward(
         self,
@@ -163,13 +261,28 @@ def forward(
             num_tokens = hidden_states.size(0)
             pad_size = (self.tp_size -
                         (num_tokens % self.tp_size)) % self.tp_size
+
+        # Generate cos and sin outside layers to avoid repeated calculation.
+        cos, sin = None, None
+        if type(self.layers[0].self_attn.rotary_emb) is RotaryEmbedding:
+            cos_sin = self.cos_sin_cache.index_select(0, positions)
+            last_dim = cos_sin.size()[-1]
+            cos, sin = cos_sin.reshape(-1, 2,
+                                       last_dim // 2).repeat(1, 1,
+                                                             2).chunk(2,
+                                                                      dim=-2)
+            cos, sin = cos.view(1, -1, 1, last_dim).contiguous(), sin.view(
+                1, -1, 1, last_dim).contiguous()
+
         for layer in self.layers[self.start_layer:self.end_layer]:
             hidden_states, residual = layer(
                 positions,
                 hidden_states,
                 residual,
                 flashcomm_v1_enabled,
                 pad_size,
+                cos=cos,
+                sin=sin,
             )
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm_ascend/models/qwen3.py

@@ -170,7 +170,7 @@ def forward(
                                k,
                                cos=cos,
                                sin=sin,
-                               skip_index_select=True)
+                               is_cos_sin_cached=True)
         attn_output = self.attn(q, k, v)
         pad_size = 0
         if self.enable_fc == 2:

vllm_ascend/ops/rotary_embedding.py

@@ -39,7 +39,7 @@ def rope_forward_oot(
         cos: torch.Tensor = None,
         sin: torch.Tensor = None,
         is_neox_style_override: Optional[bool] = None,
-        skip_index_select: bool = False) -> Tuple[torch.Tensor, torch.Tensor]:
+        is_cos_sin_cached: bool = False) -> Tuple[torch.Tensor, torch.Tensor]:
     import torch_npu
     query_shape, key_shape = query.shape, key.shape
     if self.cos_sin_cache.device != query.device:
@@ -64,16 +64,14 @@ def rope_forward_oot(
         raise NotImplementedError(
             "Batched rotary embedding is currently not supported on NPU.")
     else:
-        if skip_index_select and neox_style and self.head_size == self.rotary_dim:
-            # TODO: Remove the contiguous in the future.
-            # BSNH
+        if is_cos_sin_cached and neox_style and self.head_size == self.rotary_dim and self.head_size == 128:
+            # If cos and sin are generated outside, use npu_apply_rotary_pos_emb to avoid redundant calculation.
+            # This method requires head_size and rotary_dim equal 128 and neox_style is True
             query = query.contiguous().view(1, query.shape[0], -1,
                                             self.head_size)
             key = key.contiguous().view(1, key.shape[0], -1, self.head_size)
-            # requires head_size=128 and neox_style=True
             torch_npu.npu_apply_rotary_pos_emb(query, key, cos, sin)
         else:
-            # TODO: Remove the contiguous in the future.
             query = query.contiguous().view(query.shape[0], -1)
             key = key.contiguous().view(key.shape[0], -1)
             torch_npu._npu_rotary_embedding(