Optimize perf of Qwen3

rjg-lyh · rjg-lyh · commit ec4fc9ea7d8d · 2025-06-16T21:43:41.000+08:00
Signed-off-by: rjg-lyh &lt;1318825571@qq.com&gt;
diff --git a/vllm_ascend/models/__init__.py b/vllm_ascend/models/__init__.py
@@ -11,6 +11,7 @@ def register_model():
     from .qwen2_5_vl import \
         AscendQwen2_5_VLForConditionalGeneration  # noqa: F401
     from .qwen2_vl import AscendQwen2VLForConditionalGeneration  # noqa: F401
+    from .qwen3 import CustomQwen3ForCausalLM  # noqa: F401
 
     ModelRegistry.register_model(
         "DeepSeekMTPModel",
@@ -47,3 +48,7 @@ def register_model():
     ModelRegistry.register_model(
         "Qwen3MoeForCausalLM",
         "vllm_ascend.models.qwen3_moe:CustomQwen3MoeForCausalLM")
+
+    ModelRegistry.register_model(
+        "Qwen3ForCausalLM",
+        "vllm_ascend.models.qwen3:CustomQwen3ForCausalLM")
diff --git a/vllm_ascend/models/qwen3.py b/vllm_ascend/models/qwen3.py
@@ -0,0 +1,209 @@
+from collections.abc import Iterable
+from typing import Optional, Union
+
+import torch
+from torch import nn
+from transformers import Qwen3Config
+
+from vllm.attention import Attention, AttentionType
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group
+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.vocab_parallel_embedding import ParallelLMHead
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP
+from vllm.model_executor.models.qwen2 import Qwen2MLP as Qwen3MLP
+from vllm.model_executor.models.qwen2 import Qwen2Model
+from vllm.model_executor.models.qwen3 import Qwen3ForCausalLM, Qwen3Attention
+from vllm.model_executor.models.utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
+
+from vllm_ascend.ops.layernorm import AddRMSNormQuant
+
+class CustomQwen3DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: Qwen3Config,
+        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)
+
+        # By default, Qwen3 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-Qwen3-7B-instruct)
+        if getattr(config, "is_causal", True):
+            attn_type = AttentionType.DECODER
+        else:
+            attn_type = AttentionType.ENCODER_ONLY
+
+        self.self_attn = Qwen3Attention(
+            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,
+            rms_norm_eps=config.rms_norm_eps,
+            qkv_bias=getattr(config, 'attention_bias', False),
+            head_dim=getattr(config, 'head_dim', None),
+            cache_config=cache_config,
+            quant_config=quant_config,
+            rope_scaling=rope_scaling,
+            prefix=f"{prefix}.self_attn",
+            attn_type=attn_type,
+        )
+        self.mlp = Qwen3MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=f"{prefix}.mlp",
+        )
+        if quant_config is None:
+            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)
+        else:
+            from vllm_ascend.quantization.quant_config import AscendQuantConfig
+            assert isinstance(quant_config, AscendQuantConfig)
+            self.input_layernorm = AddRMSNormQuant(config.hidden_size,
+                                                   self.self_attn.qkv_proj.aclnn_input_scale,
+                                                   self.self_attn.qkv_proj.aclnn_input_offset,
+                                                   eps=config.rms_norm_eps)
+            self.post_attention_layernorm = AddRMSNormQuant(config.hidden_size,
+                                                            self.mlp.gate_up_proj.aclnn_input_scale,
+                                                            self.mlp.gate_up_proj.aclnn_input_offset,
+                                                            eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        residual: Optional[torch.Tensor],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+        )
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+ALL_DECODER_LAYER_TYPES = {
+    "attention": CustomQwen3DecoderLayer,
+}
+
+
+@support_torch_compile(
+    dynamic_arg_dims={
+        "input_ids": 0,
+        # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl,
+        # otherwise (seq_len, ).
+        "positions": -1,
+        "intermediate_tensors": 0,
+        "inputs_embeds": 0,
+    })
+class CustomQwen3Model(Qwen2Model):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config,
+                         prefix=prefix,
+                         decoder_layer_type=CustomQwen3DecoderLayer)
+
+
+class CustomQwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    # add `CustomQwen3Model` to init self.model
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.lora_config = lora_config
+
+        self.quant_config = quant_config
+        self.model = CustomQwen3Model(vllm_config=vllm_config,
+                                prefix=maybe_prefix(prefix, "model"))
+
+        if get_pp_group().is_last_rank:
+            if config.tie_word_embeddings:
+                self.lm_head = self.model.embed_tokens
+            else:
+                self.lm_head = ParallelLMHead(config.vocab_size,
+                                              config.hidden_size,
+                                              quant_config=quant_config,
+                                              prefix=maybe_prefix(
+                                                  prefix, "lm_head"))
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, intermediate_tensors,
+                                   inputs_embeds)
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head."]
+                           if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(weights)
diff --git a/vllm_ascend/ops/layernorm.py b/vllm_ascend/ops/layernorm.py
@@ -21,6 +21,46 @@
 from vllm.model_executor.layers.layernorm import RMSNorm
 
 
+class AddRMSNormQuant(RMSNorm):
+    """Root mean square normalization.
+
+    Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight.
+    Refer to https://arxiv.org/abs/1910.07467
+    """
+    def __init__(
+        self,
+        hidden_size: int,
+        scale: torch.Tensor,
+        offset: torch.Tensor,
+        eps: float = 1e-6,
+        var_hidden_size: Optional[int] = None,
+        has_weight: bool = True,
+        dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        super().__init__(hidden_size,
+                         eps,
+                         var_hidden_size,
+                         has_weight,
+                         dtype)
+        self.scale = scale
+        self.offset = offset
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
+        import torch_npu
+
+        if residual is not None:
+            x, _, residual = torch_npu.npu_add_rms_norm_quant(x, residual, self.weight,
+                                                              self.scale, self.offset,
+                                                              epsilon=self.variance_epsilon)
+            return x, residual
+
+        x, residual = torch_npu.npu_rms_norm(x, self.weight, self.variance_epsilon)
+        return x
+
 def forward_oot(
     self,
     x: torch.Tensor,
diff --git a/vllm_ascend/quantization/w8a8.py b/vllm_ascend/quantization/w8a8.py
@@ -33,6 +33,7 @@ class AscendW8A8LinearMethod:
     Args:
         w_sym: whether the linear weight is symmetrically quantized.
     """
+    params_dtype: torch.dtype = torch.bfloat16
 
     def __init__(self) -> None:
         # aclnn quant matmul requires to transpose matrix B, set to true by default.
@@ -54,6 +55,7 @@ def get_pertensor_param(params_dtype: torch.dtype) -> Dict[str, Any]:
         params_dict = {}
         params_dict["input_scale"] = torch.empty(1, dtype=params_dtype)
         params_dict["input_offset"] = torch.empty(1, dtype=torch.int8)
+        AscendW8A8LinearMethod.params_dtype = params_dtype
         return params_dict
 
     @staticmethod
@@ -75,6 +77,7 @@ def get_perchannel_param(
         params_dict["weight_offset"] = torch.empty(output_size,
                                                    1,
                                                    dtype=params_dtype)
+        AscendW8A8LinearMethod.params_dtype = params_dtype
         return params_dict
 
     @staticmethod
@@ -84,8 +87,7 @@ def apply(
         bias: Optional[torch.Tensor] = None,
         tp_rank: Optional[int] = 0,
     ) -> torch.Tensor:
-        original_dtype = x.dtype
-        if original_dtype != torch.int8:
+        if x.dtype != torch.int8:
             x = quant_per_tensor(
                 x,
                 layer.aclnn_input_scale,
@@ -97,7 +99,7 @@ def apply(
             layer.weight,
             layer.deq_scale,
             bias=quant_bias,
-            output_dtype=original_dtype,
+            output_dtype=AscendW8A8LinearMethod.params_dtype,
         )
 
     def process_weights_after_loading(self, layer):
