[Misc] Add uninitialized params tracking for AutoWeightsLoader (#10327)

Signed-off-by: Isotr0py <[email protected]>

Author: Isotr0py

vllm/model_executor/model_loader/loader.py

@@ -334,7 +334,17 @@ def load_model(self, vllm_config: VllmConfig) -> nn.Module:
             with target_device:
                 model = _initialize_model(vllm_config=vllm_config)
 
-            model.load_weights(self._get_all_weights(model_config, model))
+            weights_to_load = {name for name, _ in model.named_parameters()}
+            loaded_weights = model.load_weights(
+                self._get_all_weights(model_config, model))
+            # We only enable strict check for non-quantiized models
+            # that have loaded weights tracking currently.
+            if model_config.quantization is None and loaded_weights is not None:
+                weights_not_loaded = weights_to_load - loaded_weights
+                if weights_not_loaded:
+                    raise ValueError(
+                        "Following weights were not initialized from "
+                        f"checkpoint: {weights_not_loaded}")
 
             for _, module in model.named_modules():
                 quant_method = getattr(module, "quant_method", None)

vllm/model_executor/models/arctic.py

@@ -1,5 +1,5 @@
 """Inference-only Snowflake Arctic model."""
-from typing import Iterable, List, Optional, Tuple, Union
+from typing import Iterable, List, Optional, Set, Tuple, Union
 
 import torch
 from torch import nn
@@ -480,7 +480,8 @@ def sample(
         next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
@@ -518,6 +519,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                         ("ws", f"experts.{expert_id}.w3.weight", expert_id))
 
         params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
 
         logger.info(
             "It will take ~10 minutes loading from the 16-bit weights. "
@@ -573,3 +575,5 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                         weight_loader = getattr(param, "weight_loader",
                                                 default_weight_loader)
                         weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params

vllm/model_executor/models/baichuan.py

@@ -18,7 +18,7 @@
 # limitations under the License.
 """Inference-only BaiChuan model compatible with HuggingFace weights."""
 import math
-from typing import Iterable, List, Optional, Tuple, Union
+from typing import Iterable, List, Optional, Set, Tuple, Union
 
 import torch
 from torch import nn
@@ -404,13 +404,15 @@ def sample(
         next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("gate_up_proj", "gate_proj", 0),
             ("gate_up_proj", "up_proj", 1),
         ]
         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
@@ -449,6 +451,8 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
 
 
 class BaichuanForCausalLM(BaiChuanBaseForCausalLM):

vllm/model_executor/models/bert.py

@@ -1,4 +1,4 @@
-from typing import Iterable, List, Optional, Tuple
+from typing import Iterable, List, Optional, Set, Tuple
 
 import torch
 from torch import nn
@@ -337,7 +337,8 @@ def forward(
 
         return self.encoder(hidden_states, kv_caches, attn_metadata)
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "query", "q"),
@@ -346,6 +347,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         ]
 
         params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
             if "pooler" in name:
                 continue
@@ -368,6 +370,8 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
 
 
 class BertEmbeddingModel(nn.Module):

vllm/model_executor/models/blip.py

@@ -1,6 +1,6 @@
 """Minimal implementation of BlipVisionModel intended to be only used 
 within a vision language model."""
-from typing import Iterable, Optional, Tuple, Union
+from typing import Iterable, Optional, Set, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -415,14 +415,16 @@ def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
 
         return self.post_layernorm(hidden_states)
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
             ("qkv_proj", "k_proj", "k"),
             ("qkv_proj", "v_proj", "v"),
         ] if self.shard_weight else []
         params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
         layer_count = len(self.encoder.layers)
 
         for name, loaded_weight in weights:
@@ -440,8 +442,8 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 if weight_name not in name:
                     continue
-
-                param = params_dict[name.replace(weight_name, param_name)]
+                name = name.replace(weight_name, param_name)
+                param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
                 break
@@ -450,3 +452,5 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params

vllm/model_executor/models/blip2.py

@@ -1,5 +1,5 @@
 from functools import cached_property
-from typing import (Iterable, List, Literal, Mapping, Optional, Tuple,
+from typing import (Iterable, List, Literal, Mapping, Optional, Set, Tuple,
                     TypedDict, Union)
 
 import torch
@@ -692,6 +692,7 @@ def sample(
     ) -> Optional[SamplerOutput]:
         return self.language_model.sample(logits, sampling_metadata)
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         loader = AutoWeightsLoader(self)
-        loader.load_weights(weights)
+        return loader.load_weights(weights)

vllm/model_executor/models/bloom.py

@@ -16,7 +16,7 @@
 # limitations under the License.
 """Inference-only BLOOM model compatible with HuggingFace weights."""
 import math
-from typing import Iterable, List, Optional, Tuple, Union
+from typing import Iterable, List, Optional, Set, Tuple, Union
 
 import torch
 from torch import nn
@@ -341,8 +341,10 @@ def sample(
         next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
             if name == "lm_head.weight":
                 continue
@@ -371,3 +373,5 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             weight_loader = getattr(param, "weight_loader",
                                     default_weight_loader)
             weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params

vllm/model_executor/models/chameleon.py

@@ -1,5 +1,5 @@
 from functools import cached_property
-from typing import (Any, Dict, Iterable, List, Literal, Mapping, Optional,
+from typing import (Any, Dict, Iterable, List, Literal, Mapping, Optional, Set,
                     Tuple, TypedDict, Union)
 
 import torch
@@ -1034,7 +1034,8 @@ def sample(
         next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             (".qkv_proj", ".q_proj", "q"),
@@ -1044,6 +1045,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             (".gate_up_proj", ".up_proj", 1),
         ]
         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
@@ -1111,3 +1113,5 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params

vllm/model_executor/models/chatglm.py

@@ -3,7 +3,8 @@
 """Inference-only ChatGLM model compatible with THUDM weights."""
 from argparse import Namespace
 from array import array
-from typing import Dict, Iterable, List, Mapping, Optional, Tuple, TypedDict
+from typing import (Dict, Iterable, List, Mapping, Optional, Set, Tuple,
+                    TypedDict)
 
 import torch
 from PIL import Image
@@ -645,7 +646,8 @@ def sample(
         next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         # Merge two ColumnParallelLinear into one MergedColumnParallelLinear
         merged_weights_dict: Dict[str, Dict[str, Optional[torch.Tensor]]] = {
             "transformer.vision.linear_proj.merged_proj.weight": {
@@ -655,6 +657,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         }
 
         params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
             is_weight_to_be_merge = False
             for _, merged_weight_dict in merged_weights_dict.items():
@@ -677,6 +680,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             weight_loader = getattr(param, "weight_loader",
                                     default_weight_loader)
             weight_loader(param, loaded_weight)
+            loaded_params.add(name)
 
         for combined_name, merged_weight_dict in merged_weights_dict.items():
             if combined_name in params_dict:
@@ -686,3 +690,5 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, combined_weight)
+                loaded_params.add(combined_name)
+        return loaded_params

vllm/model_executor/models/clip.py

@@ -1,6 +1,6 @@
 """Minimal implementation of CLIPVisionModel intended to be only used
 within a vision language model."""
-from typing import Iterable, List, Optional, Tuple, Union
+from typing import Iterable, List, Optional, Set, Tuple, Union
 
 import numpy as np
 import torch
@@ -483,14 +483,16 @@ def device(self):
 
     # (TODO) Add prefix argument for filtering out weights to be loaded
     #        ref: https://github.com/vllm-project/vllm/pull/7186#discussion_r1734163986
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
             ("qkv_proj", "k_proj", "k"),
             ("qkv_proj", "v_proj", "v"),
         ] if self.shard_weight else []
         params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
         layer_count = len(self.vision_model.encoder.layers)
 
         for name, loaded_weight in weights:
@@ -508,8 +510,9 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 if weight_name not in name:
                     continue
+                name = name.replace(weight_name, param_name)
 
-                param = params_dict[name.replace(weight_name, param_name)]
+                param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
                 break
@@ -518,3 +521,5 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)
                 weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params