[Quantization] Enable BNB support for more MoE models (#21100)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

vllm/model_executor/models/ernie45_moe.py

@@ -51,8 +51,8 @@ from vllm.model_executor.model_loader.weight_utils import (
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
-from .interfaces import SupportsPP
-from .utils import (PPMissingLayer, extract_layer_index,
+from .interfaces import SupportsLoRA, SupportsPP
+from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
                     is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
@@ -427,66 +427,15 @@ class Ernie4_5_MoeModel(nn.Module):
 
         return hidden_states
 
+    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
 
-class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP):
-    packed_modules_mapping = {
-        "qkv_proj": [
-            "q_proj",
-            "k_proj",
-            "v_proj",
-        ],
-        "gate_up_proj": [
-            "gate_proj",
-            "up_proj",
-        ],
-    }
-
-    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 = Ernie4_5_MoeModel(vllm_config=vllm_config,
-                                       prefix=maybe_prefix(prefix, "model"))
-
-        if get_pp_group().is_last_rank:
-            self.lm_head = ParallelLMHead(config.vocab_size,
-                                          config.hidden_size,
-                                          quant_config=quant_config)
-        else:
-            self.lm_head = PPMissingLayer()
-
-        if self.config.tie_word_embeddings:
-            self.lm_head.weight = self.model.embed_tokens.weight
-        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
+        # Params for weights, fp8 weight scales, fp8 activation scales
+        # (param_name, weight_name, expert_id, shard_id)
+        return FusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=self.config.moe_num_experts)
 
     def load_weights(self, weights: Iterable[tuple[str,
                                                    torch.Tensor]]) -> set[str]:
@@ -499,16 +448,9 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP):
             ("gate_up_proj", "up_proj", 1),
         ]
 
-        # Params for weights, fp8 weight scales, fp8 activation scales
-        # (param_name, weight_name, expert_id, shard_id)
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
-            ckpt_gate_proj_name="gate_proj",
-            ckpt_down_proj_name="down_proj",
-            ckpt_up_proj_name="up_proj",
-            num_experts=self.config.moe_num_experts)
-
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
+        expert_params_mapping = self.get_expert_mapping()
         for name, loaded_weight in weights:
             if self.config.tie_word_embeddings and name.endswith(
                     "lm_head.weight"):
@@ -581,3 +523,76 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP):
                     weight_loader(param, loaded_weight)
             loaded_params.add(name)
         return loaded_params
+
+
+class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    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 = Ernie4_5_MoeModel(vllm_config=vllm_config,
+                                       prefix=maybe_prefix(prefix, "model"))
+
+        if get_pp_group().is_last_rank:
+            self.lm_head = ParallelLMHead(config.vocab_size,
+                                          config.hidden_size,
+                                          quant_config=quant_config)
+        else:
+            self.lm_head = PPMissingLayer()
+
+        if self.config.tie_word_embeddings:
+            self.lm_head.weight = self.model.embed_tokens.weight
+        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)
+
+    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
+        return self.model.get_expert_mapping()