sync v0.18.0

vllm/model_executor/models/dots_ocr.py

@@ -754,7 +754,7 @@ class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -790,4 +790,4 @@ class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
             language_model="language_model",
             connector="vision_tower.merger",
             tower_model="vision_tower.",
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/eagle2_5_vl.py

@@ -429,7 +429,7 @@ class Eagle2_5_VLForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -437,7 +437,6 @@ class Eagle2_5_VLForConditionalGeneration(
     ) -> IntermediateTensors:
         """Forward pass through the model."""
         if intermediate_tensors is not None:
-            input_ids = None
             inputs_embeds = None
 
         forward_kwargs = {
@@ -468,4 +467,4 @@ class Eagle2_5_VLForConditionalGeneration(
             language_model="language_model",
             connector="mlp1",
             tower_model="vision_model",
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/ernie45_moe.py

@@ -465,7 +465,7 @@ class Ernie4_5_MoeModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -727,7 +727,7 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExpe
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -752,4 +752,4 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExpe
         return loader.load_weights(weights)
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
-        return self.model.get_expert_mapping()
+        return self.model.get_expert_mapping()
\ No newline at end of file

vllm/model_executor/models/ernie45_vl.py

@@ -1680,7 +1680,7 @@ class Ernie4_5_VLMoeForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -1716,4 +1716,4 @@ class Ernie4_5_VLMoeForConditionalGeneration(
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
\ No newline at end of file

vllm/model_executor/models/ernie45_vl_moe.py

@@ -563,7 +563,7 @@ class Ernie4_5_VLMoeModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -644,7 +644,7 @@ class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -798,4 +798,4 @@ class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP):
                     )
                     weight_loader(param, loaded_weight)
             loaded_params.add(name)
-        return loaded_params
+        return loaded_params
\ No newline at end of file

vllm/model_executor/models/ernie_mtp.py

@@ -164,7 +164,7 @@ class ErnieMTP(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
@@ -275,4 +275,4 @@ class ErnieMTP(nn.Module):
         name = name.replace(
             "model.mtp_block.0.", f"model.layers.{layer_idx}.mtp_block."
         )
-        return name
+        return name
\ No newline at end of file

vllm/model_executor/models/exaone.py

@@ -496,7 +496,7 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -521,4 +521,4 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             # processed with quantization, LoRA, fine-tuning, etc.
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/exaone4.py

@@ -490,7 +490,7 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -515,4 +515,4 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             # processed with quantization, LoRA, fine-tuning, etc.
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/exaone_moe.py

@@ -549,7 +549,7 @@ class ExaoneMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -576,4 +576,4 @@ class ExaoneMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
                 ["lm_head.", "mtp."] if self.config.tie_word_embeddings else ["mtp."]
             ),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/falcon.py

@@ -402,7 +402,7 @@ class FalconModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -540,4 +540,4 @@ class FalconForCausalLM(nn.Module, SupportsPP):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/falcon_h1.py

@@ -465,7 +465,7 @@ class FalconH1Model(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -608,7 +608,7 @@ class FalconH1ForCausalLM(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/fuyu.py

@@ -340,7 +340,7 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -365,4 +365,4 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/gemma.py

@@ -297,7 +297,7 @@ class GemmaModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -399,7 +399,7 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -421,4 +421,4 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/gemma2.py

@@ -406,7 +406,7 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -428,4 +428,4 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/gemma3.py

@@ -494,7 +494,7 @@ class Gemma3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -517,4 +517,4 @@ class Gemma3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/gemma3_mm.py

@@ -606,7 +606,7 @@ class Gemma3ForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -643,4 +643,42 @@ class Gemma3ForConditionalGeneration(
             language_model="language_model",
             connector="multi_modal_projector",
             tower_model="vision_tower",
-        )
+        )
\ No newline at end of file
+
+    def get_num_mm_encoder_tokens(self, num_image_tokens: int) -> int:
+        """
+        Calculate the number of tokens output by the vision encoder.
+
+        The vision encoder processes images into patch embeddings. For Gemma3,
+        the relationship between prompt placeholder tokens and actual vision
+        encoder output tokens depends on the patch grid size.
+
+        Args:
+            num_image_tokens: Number of image placeholder tokens in the prompt
+                              (typically mm_tokens_per_image per image)
+
+        Returns:
+            Number of tokens output by the vision encoder
+        """
+        # For Gemma3, the vision encoder outputs tokens_per_side x tokens_per_side
+        # tokens per image. Since num_image_tokens represents the number of
+        # connector output tokens (mm_tokens_per_image = 256), and tokens_per_side
+        # is sqrt(256) = 16, we need to account for the token expansion.
+        # Based on empirical testing, the multiplier of 16 works correctly.
+        return num_image_tokens * 16
+
+    def get_num_mm_connector_tokens(self, num_vision_tokens: int) -> int:
+        """
+        Calculate the number of tokens output by the multimodal connector.
+
+        The connector applies projection and normalization but maintains the
+        token count for Gemma3.
+
+        Args:
+            num_vision_tokens: Number of tokens from vision encoder
+
+        Returns:
+            Number of tokens after connector processing
+        """
+        # The Gemma3 connector maintains a 1:1 token mapping
+        return num_vision_tokens
\ No newline at end of file

vllm/model_executor/models/gemma3n.py

@@ -704,7 +704,7 @@ class Gemma3nSelfDecoder(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         inputs_embeds: torch.Tensor | None = None,
         per_layer_inputs: torch.Tensor | None = None,
@@ -887,7 +887,7 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
 
     def fast_prefill_forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         inputs_embeds: torch.Tensor | None = None,
         per_layer_inputs: torch.Tensor | None = None,
@@ -964,7 +964,7 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
 
     def normal_forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         inputs_embeds: torch.Tensor | None = None,
         per_layer_inputs: torch.Tensor | None = None,
@@ -1131,7 +1131,7 @@ class Gemma3nForCausalLM(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         *,
         per_layer_inputs: torch.Tensor | None = None,

vllm/model_executor/models/gemma3n_mm.py

@@ -713,7 +713,7 @@ class Gemma3nForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -822,4 +822,4 @@ class Gemma3nForConditionalGeneration(
             sample_rate=16000,
             # TODO enable chunking after more thorough testing.
             min_energy_split_window_size=None,
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/glm4.py

@@ -39,13 +39,22 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader,
+    maybe_remap_kv_scale_name,
+)
 from vllm.sequence import IntermediateTensors
 from vllm.v1.attention.backend import AttentionType
 
 from .interfaces import SupportsLoRA, SupportsPP
 from .llama import LlamaMLP as Glm4MLP
 from .llama import LlamaModel
-from .utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
+from .utils import (
+    AutoWeightsLoader,
+    PPMissingLayer,
+    is_pp_missing_parameter,
+    maybe_prefix,
+)
 
 
 class Glm4Attention(nn.Module):
@@ -78,7 +87,15 @@ class Glm4Attention(nn.Module):
             # Number of KV heads is less than TP size, so we replicate
             # the KV heads across multiple tensor parallel GPUs.
             assert tp_size % self.total_num_kv_heads == 0
-        config.rope_parameters.setdefault("partial_rotary_factor", 0.5)
+
+        rope_params = getattr(config, "rope_parameters", None)
+        if isinstance(rope_params, dict) and "partial_rotary_factor" in rope_params:
+            config.rope_parameters.setdefault(
+                "partial_rotary_factor", rope_params["partial_rotary_factor"]
+            )
+        else:
+            config.rope_parameters.setdefault("partial_rotary_factor", 0.5)
+
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
         self.head_dim = head_dim or hidden_size // self.total_num_heads
         self.q_size = self.num_heads * self.head_dim
@@ -220,6 +237,73 @@ class Glm4Model(LlamaModel):
             vllm_config=vllm_config, prefix=prefix, layer_type=Glm4DecoderLayer
         )
 
+    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"),
+            (".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:
+            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
+            if spec_layer is not None:
+                continue
+            if "rotary_emb.inv_freq" in name:
+                continue
+            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+            if self.quant_config is not None and (
+                scale_name := self.quant_config.get_cache_scale(name)
+            ):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                loaded_weight = (
+                    loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
+                )
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                continue
+            if "scale" in name or "zero_point" in name:
+                # Remapping the name of FP8 kv-scale or zero point.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+            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)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
 
 class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
     packed_modules_mapping = {
@@ -270,7 +354,7 @@ class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -292,4 +376,17 @@ class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file
+
+
+def get_spec_layer_idx_from_weight_name(
+    config: Glm4Config, weight_name: str
+) -> int | None:
+    if hasattr(config, "num_nextn_predict_layers") and (
+        config.num_nextn_predict_layers > 0
+    ):
+        layer_idx = config.num_hidden_layers
+        for i in range(config.num_nextn_predict_layers):
+            if f"layers.{layer_idx + i}." in weight_name:
+                return layer_idx + i
+    return None
\ No newline at end of file

vllm/model_executor/models/glm4_1v.py

@@ -24,7 +24,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""Inference-only GLM-4V model compatible with HuggingFace weights."""
+"""Inference-only GLM-4.1V & GLM-4.6V-Flash, AutoGLM-Phone-9B model
+compatible with HuggingFace weights."""
 
 import math
 from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence
@@ -1447,7 +1448,7 @@ class Glm4vForConditionalGeneration(
                 prefix=maybe_prefix(prefix, "visual"),
             )
 
-        if config.model_type == "glm4v":
+        if config.model_type in ("glm4v", "glm_ocr"):
             architectures = ["Glm4ForCausalLM"]
         elif config.model_type == "glm4v_moe":
             architectures = ["Glm4MoeForCausalLM"]
@@ -1664,7 +1665,7 @@ class Glm4vForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,