[WIP][V0.9.1] add support for flashcomm2 in qwen2

vllm_ascend/models/qwen2.py

@@ -1,23 +1,28 @@
 from collections.abc import Iterable
-from typing import Optional, Union
+from typing import Any, Optional, Union
 
 import torch
+import torch.distributed as dist
 import torch.nn.functional as F
 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,
-                              get_tensor_model_parallel_world_size,
+                              get_tp_group, get_tensor_model_parallel_world_size,
                               tensor_model_parallel_all_gather,
                               tensor_model_parallel_all_reduce,
                               tensor_model_parallel_reduce_scatter)
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (ReplicatedLinear,
+                                               RowParallelLinear)
 from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
 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 Qwen2Model, Qwen2Attention, Qwen2MLP
 from vllm.model_executor.models.utils import (AutoWeightsLoader,
                                               PPMissingLayer, maybe_prefix)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
@@ -27,6 +32,20 @@
 from vllm_ascend.attention.attention_v1 import AscendAttentionState
 
 
+def pad(tensor, x):
+    length = tensor.size(0)
+    pad_size = (x - (length % x)) % x
+    if pad_size > 0:
+        return F.pad(tensor, (0, 0, 0, pad_size)), pad_size
+    return tensor, pad_size
+
+
+def unpad(tensor, pad_size):
+    if pad_size > 0:
+        return tensor[:-pad_size, :]
+    return tensor
+
+
 def all_gather_and_maybe_unpad(
     hidden_states: torch.Tensor,
     pad_size: int,
@@ -47,30 +66,215 @@ def maybe_pad_and_reduce_scatter(
     return hidden_states
 
 
-class CustomQwen2DecoderLayer(Qwen2DecoderLayer):
+class CustomQwen2MLP(Qwen2MLP):
 
     def __init__(
         self,
-        config: Qwen2Config,
-        cache_config: Optional[CacheConfig] = None,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
-        super().__init__(config=config,
-                         cache_config=cache_config,
+        super().__init__(hidden_size=hidden_size,
+                         intermediate_size=intermediate_size,
+                         hidden_act=hidden_act,
                          quant_config=quant_config,
                          prefix=prefix)
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+        self.enable_fc = ascend_envs.VLLM_ASCEND_ENABLE_FLASHCOMM
+        if self.enable_fc == 2:
+            # if flashcomm2 enabled, replace Linear+AllReduce with All2All+Linear
+            self.down_proj = ReplicatedLinear(
+                intermediate_size,
+                hidden_size,
+                bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.down_proj",
+            )
+        else:
+            self.down_proj = RowParallelLinear(
+                intermediate_size,
+                hidden_size,
+                bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.down_proj",
+            )
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        pad_size = 0
+        if self.enable_fc == 2:
+            # pad input because AllGather requires token_num to be divisible by tp_size
+            x, pad_size = pad(x, self.tp_size)
+            output = torch.empty(x.shape, dtype=x.dtype, device=x.device)
+            dist.all_to_all_single(output,
+                                   x,
+                                   group=get_tp_group().device_group)
+            x = output.reshape(self.tp_size, -1, output.size(-1)) \
+                        .transpose(0, 1) \
+                        .reshape(-1, output.size(-1)*self.tp_size)
+        x, _ = self.down_proj(x)
+        return x, pad_size
+
+
+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)
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+        self.enable_fc = ascend_envs.VLLM_ASCEND_ENABLE_FLASHCOMM
+        if self.enable_fc == 2:
+            self.o_proj = ReplicatedLinear(
+                self.total_num_heads * self.head_dim,
+                hidden_size,
+                bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.o_proj",
+            )
+        else:
+            self.o_proj = RowParallelLinear(
+                self.total_num_heads * self.head_dim,
+                hidden_size,
+                bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.o_proj",
+            )
+
+    def forward(self, positions: torch.Tensor, hidden_states: torch.Tensor,
+        cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions,
+                               q,
+                               k,
+                               cos=cos,
+                               sin=sin,
+                               skip_index_select=True)
+        attn_output = self.attn(q, k, v)
+        pad_size = 0
+        if self.enable_fc == 2:
+            # pad input because AllGather requires token_num to be divisible by tp_size
+            attn_output, pad_size = pad(attn_output, self.tp_size)
+            output = torch.empty(attn_output.shape,
+                                 dtype=attn_output.dtype,
+                                 device=attn_output.device)
+            dist.all_to_all_single(output,
+                                   attn_output,
+                                   group=get_tp_group().device_group)
+            attn_output = output.reshape(self.tp_size, -1, output.size(-1)) \
+                                .transpose(0, 1) \
+                                .reshape(-1, output.size(-1)*self.tp_size)
+        output, _ = self.o_proj(attn_output)
+        return output, pad_size
+
+
+class CustomQwen2DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: Qwen2Config,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        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 = CustomQwen2MLP(
+            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.enable_fc = ascend_envs.VLLM_ASCEND_ENABLE_FLASHCOMM
         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
         self.mlp.down_proj.reduce_results = False
 
+    def pre_attention_process(self, hidden_states, residual, pad_size=0):
+        hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = tensor_model_parallel_all_gather(hidden_states, 0)
+        hidden_states = unpad(hidden_states, pad_size)
+        return hidden_states, residual
+
+    def pre_mlp_process(self, hidden_states, residual, pad_size=0):
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        hidden_states = tensor_model_parallel_all_gather(hidden_states, 0)
+        hidden_states = unpad(hidden_states, pad_size)
+        return hidden_states, residual
+
     def forward(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         residual: Optional[torch.Tensor],
         flashcomm_v1_enabled: bool,
+        cos,
+        sin,
         pad_size: int,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         # Self Attention
@@ -81,34 +285,47 @@ def forward(
                     residual = F.pad(residual, (0, 0, 0, pad_size))
                 residual = torch.chunk(residual, self.tp_size,
                                        dim=0)[self.tp_rank]
+            if self.enable_fc == 2:
+                residual, pad_size = pad(residual, self.tp_size)
+                chunk_size = residual.size(0) // self.tp_size
+                residual = residual[chunk_size * self.tp_rank:chunk_size *
+                                    (self.tp_rank + 1)]
             hidden_states = self.input_layernorm(hidden_states)
         else:
-            hidden_states, residual = self.input_layernorm(
-                hidden_states, residual)
+            if self.enable_fc == 2:
+                hidden_states, residual = self.pre_attention_process(
+                    hidden_states, residual, pad_size)
+            else:
+                hidden_states, residual = self.input_layernorm(
+                    hidden_states, residual)
             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, pad_size = 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)
-        else:
+        elif self.enable_fc != 2:
             hidden_states = tensor_model_parallel_all_reduce(hidden_states)
         # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(
-            hidden_states, residual)
+        if self.enable_fc == 2:
+            hidden_states, residual = self.pre_mlp_process(
+                hidden_states, residual, pad_size)
+        else:
+            hidden_states, residual = self.post_attention_layernorm(
+                hidden_states, residual)
         if flashcomm_v1_enabled:
             hidden_states = all_gather_and_maybe_unpad(hidden_states, pad_size)
-        hidden_states = self.mlp(hidden_states)
+        hidden_states, pad_size = self.mlp(hidden_states)
         if flashcomm_v1_enabled:
             hidden_states = maybe_pad_and_reduce_scatter(
                 hidden_states, pad_size)
-        else:
+        elif self.enable_fc != 2:
             hidden_states = tensor_model_parallel_all_reduce(hidden_states)
-        return hidden_states, residual
+        return hidden_states, residual, pad_size
 
 
 @support_torch_compile(
@@ -132,6 +349,8 @@ 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
+        self.enable_fc = ascend_envs.VLLM_ASCEND_ENABLE_FLASHCOMM
 
     def forward(
         self,
@@ -161,12 +380,24 @@ def forward(
             num_tokens = hidden_states.size(0)
             pad_size = (self.tp_size -
                         (num_tokens % self.tp_size)) % self.tp_size
+
+        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)
+        # BSNH
+        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(
+            hidden_states, residual, pad_size = layer(
                 positions,
                 hidden_states,
                 residual,
                 flashcomm_v1_enabled,
+                cos,
+                sin,
                 pad_size,
             )
         if not get_pp_group().is_last_rank:
@@ -177,6 +408,12 @@ def forward(
         hidden_states, _ = self.norm(hidden_states, residual)
         if flashcomm_v1_enabled:
             hidden_states = all_gather_and_maybe_unpad(hidden_states, pad_size)
+        if self.enable_fc == 2:
+            hidden_states = tensor_model_parallel_all_gather(hidden_states, 0)
+            residual = tensor_model_parallel_all_gather(residual, 0)
+            if pad_size > 0:
+                hidden_states = hidden_states[:-pad_size]
+                residual = residual[:-pad_size]
         return hidden_states