[Model] use AutoWeightsLoader for phimoe,qwen2_moe,qwen3_moe (#16203)

Signed-off-by: rongfu.leng <[email protected]>

Author: lengrongfu

vllm/model_executor/models/phimoe.py

@@ -49,7 +49,7 @@
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsLoRA, SupportsPP
-from .utils import (is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -448,6 +448,8 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                        (lora_config.max_loras or 1)) if lora_config else 0)
         self.vocab_size = config.vocab_size + lora_vocab
         self.org_vocab_size = config.vocab_size
+        self.config = config
+        self.quant_config = quant_config
 
         self.embed_tokens = VocabParallelEmbedding(
             self.vocab_size,
@@ -504,85 +506,6 @@ def forward(
         hidden_states = self.norm(hidden_states)
         return hidden_states
 
-
-class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
-    fall_back_to_pt_during_load = False
-
-    packed_modules_mapping = {
-        "qkv_proj": [
-            "q_proj",
-            "k_proj",
-            "v_proj",
-        ],
-    }
-
-    # LoRA specific attributes
-    embedding_modules = {
-        "embed_tokens": "input_embeddings",
-        "lm_head": "output_embeddings",
-    }
-    embedding_padding_modules = ["lm_head"]
-
-    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
-        super().__init__()
-        config = vllm_config.model_config.hf_config
-        lora_config = vllm_config.lora_config
-        self.config = config
-        self.lora_config = lora_config
-        self.quant_config = vllm_config.quant_config
-
-        self.model = PhiMoEModel(vllm_config=vllm_config,
-                                 prefix=maybe_prefix(prefix, "model"))
-        self.unpadded_vocab_size = config.vocab_size
-        if lora_config:
-            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
-        self.lm_head = ParallelLMHead(
-            self.unpadded_vocab_size,
-            config.hidden_size,
-            org_num_embeddings=config.vocab_size,
-            padding_size=(
-                DEFAULT_VOCAB_PADDING_SIZE
-                # We need bigger padding if using lora for kernel
-                # compatibility
-                if not lora_config else lora_config.lora_vocab_padding_size),
-            quant_config=None,
-            bias=True,
-        )
-        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
-                                                config.vocab_size)
-        self.sampler = get_sampler()
-
-        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) -> torch.Tensor:
-        logits = self.logits_processor(self.lm_head, hidden_states,
-                                       sampling_metadata)
-        return logits
-
-    def sample(
-        self,
-        logits: Optional[torch.Tensor],
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
@@ -601,9 +524,6 @@ def load_weights(self, weights: Iterable[Tuple[str,
         params_dict = dict(self.named_parameters())
         loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
-
             if (self.quant_config is not None and
                 (scale_name := self.quant_config.get_cache_scale(name))):
                 # Loading kv cache quantization scales
@@ -667,3 +587,90 @@ def load_weights(self, weights: Iterable[Tuple[str,
                     weight_loader(param, loaded_weight)
             loaded_params.add(name)
         return loaded_params
+
+
+class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    fall_back_to_pt_during_load = False
+
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+    }
+
+    # LoRA specific attributes
+    embedding_modules = {
+        "embed_tokens": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+    embedding_padding_modules = ["lm_head"]
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        lora_config = vllm_config.lora_config
+        self.config = config
+        self.lora_config = lora_config
+        self.quant_config = vllm_config.quant_config
+
+        self.model = PhiMoEModel(vllm_config=vllm_config,
+                                 prefix=maybe_prefix(prefix, "model"))
+        self.unpadded_vocab_size = config.vocab_size
+        if lora_config:
+            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+        self.lm_head = ParallelLMHead(
+            self.unpadded_vocab_size,
+            config.hidden_size,
+            org_num_embeddings=config.vocab_size,
+            padding_size=(
+                DEFAULT_VOCAB_PADDING_SIZE
+                # We need bigger padding if using lora for kernel
+                # compatibility
+                if not lora_config else lora_config.lora_vocab_padding_size),
+            quant_config=None,
+            bias=True,
+        )
+        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                config.vocab_size)
+        self.sampler = get_sampler()
+
+        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) -> torch.Tensor:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits
+
+    def sample(
+        self,
+        logits: Optional[torch.Tensor],
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["rotary_emb.inv_freq"]),
+        )
+        return loader.load_weights(weights)

vllm/model_executor/models/qwen2_moe.py

@@ -55,7 +55,8 @@
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP
-from .utils import (extract_layer_index, is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, extract_layer_index,
+                    is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -329,6 +330,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         quant_config = vllm_config.quant_config
 
         self.vocab_size = config.vocab_size
+        self.config = config
 
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
@@ -377,60 +379,6 @@ def forward(
         hidden_states, _ = self.norm(hidden_states, residual)
         return hidden_states
 
-
-class Qwen2MoeForCausalLM(nn.Module, SupportsPP):
-
-    fall_back_to_pt_during_load = False
-
-    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
-        super().__init__()
-        config = vllm_config.model_config.hf_config
-        quant_config = vllm_config.quant_config
-        self.config = config
-        self.quant_config = quant_config
-        self.model = Qwen2MoeModel(vllm_config=vllm_config,
-                                   prefix=maybe_prefix(prefix, "model"))
-        self.lm_head = ParallelLMHead(config.vocab_size,
-                                      config.hidden_size,
-                                      quant_config=quant_config)
-        if self.config.tie_word_embeddings:
-            self.lm_head.weight = self.model.embed_tokens.weight
-        self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
-        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 sample(
-        self,
-        logits: Optional[torch.Tensor],
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
@@ -453,8 +401,6 @@ def load_weights(self, weights: Iterable[Tuple[str,
         params_dict = dict(self.named_parameters())
         loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name:
@@ -531,3 +477,65 @@ def load_weights(self, weights: Iterable[Tuple[str,
                     weight_loader(param, loaded_weight)
             loaded_params.add(name)
         return loaded_params
+
+
+class Qwen2MoeForCausalLM(nn.Module, SupportsPP):
+
+    fall_back_to_pt_during_load = False
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Qwen2MoeModel(vllm_config=vllm_config,
+                                   prefix=maybe_prefix(prefix, "model"))
+        self.lm_head = ParallelLMHead(config.vocab_size,
+                                      config.hidden_size,
+                                      quant_config=quant_config)
+        if self.config.tie_word_embeddings:
+            self.lm_head.weight = self.model.embed_tokens.weight
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.sampler = get_sampler()
+        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 sample(
+        self,
+        logits: Optional[torch.Tensor],
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["rotary_emb.inv_freq"]),
+        )
+        return loader.load_weights(weights)