sglang 0.5.2rc1__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/models/llama4.py

@@ -423,6 +423,12 @@ class Llama4DecoderLayer(nn.Module):
             return self.config.num_local_experts > 0
         return (layer_id + 1) % self.config.interleave_moe_layer_step == 0
 
+    def get_intermediate_size(self) -> int:
+        if isinstance(self.feed_forward, Llama4MoE):
+            return self.config.intermediate_size
+        else:
+            return self.config.intermediate_size_mlp
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -540,6 +546,9 @@ class Llama4ForCausalLM(LlamaForCausalLM):
     def get_input_embeddings(self):
         return self.model.embed_tokens
 
+    def get_layers(self):
+        return self.model.layers
+
     def _init_model(
         self,
         config: Llama4TextConfig,

sglang/srt/models/llama_eagle3.py

@@ -109,6 +109,16 @@ class LlamaModel(nn.Module):
     ) -> None:
         super().__init__()
         self.config = config
+
+        self.is_mrope_enabled = (
+            hasattr(config, "rope_scaling")
+            and config.rope_scaling is not None
+            and "mrope_section" in config.rope_scaling
+        )
+        # fix rope_scaling for qwen2.5-vl
+        if self.is_mrope_enabled:
+            config.rope_scaling["rope_type"] = "default"
+
         self.vocab_size = config.vocab_size
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
@@ -144,6 +154,9 @@ class LlamaModel(nn.Module):
         else:
             embeds = input_embeds
 
+        if self.is_mrope_enabled:
+            positions = forward_batch.mrope_positions
+
         hidden_states = forward_batch.spec_info.hidden_states
         if hidden_states.shape[-1] != embeds.shape[-1]:
             hidden_states = self.fc(hidden_states)

sglang/srt/models/longcat_flash.py

@@ -260,7 +260,7 @@ class LongcatFlashMoE(nn.Module):
         )
         self.topk.forward = self.topk.forward_native
 
-        self.experts = get_moe_impl_class()(
+        self.experts = get_moe_impl_class(quant_config)(
             num_experts=self.num_experts,
             top_k=self.top_k,
             layer_id=self.layer_id,
@@ -853,7 +853,7 @@ class LongcatFlashForCausalLM(nn.Module):
 
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
-        expert_params_mapping = get_moe_impl_class().make_expert_params_mapping(
+        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",

sglang/srt/models/minicpmv.py

@@ -54,6 +54,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.utils import set_default_torch_dtype
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.idefics2 import Idefics2VisionTransformer
+from sglang.srt.models.llama import LlamaConfig, LlamaForCausalLM
 from sglang.srt.models.qwen2 import Qwen2Config, Qwen2ForCausalLM
 from sglang.srt.utils import add_prefix, flatten_nested_list
 
@@ -581,7 +582,7 @@ class MiniCPMBaseModel(nn.Module):
 
     def init_llm(
         self,
-        config: Qwen2Config,
+        config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> nn.Module:
@@ -774,7 +775,168 @@ class MiniCPMV2_6(MiniCPMBaseModel):
         return pattern.pad_input_tokens(input_ids, image_inputs)
 
 
-_SUPPORT_VERSION = {(2, 6): MiniCPMV2_6}
+class MiniCPMV4_0(MiniCPMBaseModel):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+    # LoRA specific attributes
+    supported_lora_modules = [
+        # vision encoder
+        "fc1",
+        "fc2",
+        "out_proj",
+        # language model
+        "qkv_proj",  # same name with vision encoder
+        "o_proj",
+        "gate_up_proj",
+        "down_proj",
+        # resampler
+        "kv_proj",
+    ]
+
+    # BitandBytes specific attributes
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
+    embedding_modules = {}
+    embedding_padding_modules = []
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__(config=config, quant_config=quant_config, prefix=prefix)
+        assert self.version == (4, 0)
+
+    def init_llm(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> nn.Module:
+        return LlamaForCausalLM(config=config, quant_config=quant_config, prefix=prefix)
+
+    def init_vision_module(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig],
+        prefix: str = "",
+    ) -> nn.Module:
+        model = Idefics2VisionTransformer(
+            config=config.vision_config, quant_config=quant_config, prefix=prefix
+        )
+        if self.config.drop_vision_last_layer:
+            model.encoder.layers = model.encoder.layers[:-1]
+
+        setattr(model, "embed_dim", model.embeddings.embed_dim)
+        setattr(model, "patch_size", model.embeddings.patch_size)
+        return model
+
+    def init_resampler(
+        self,
+        embed_dim: int,
+        vision_dim: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> nn.Module:
+        with set_default_torch_dtype(torch.float16):
+            # The resampler in 2.6 remains consistent with the one in 2.5.
+            resampler = Resampler2_5(
+                num_queries=self.config.query_num,
+                embed_dim=embed_dim,
+                num_heads=embed_dim // 128,
+                kv_dim=vision_dim,
+                quant_config=quant_config,
+                prefix=prefix,
+            )
+
+        return resampler.to(device="cuda", dtype=torch.get_default_dtype())
+
+    def get_vision_embedding(
+        self,
+        pixel_values: List[torch.Tensor],
+        patch_attn_mask: Optional[torch.Tensor] = None,
+        tgt_sizes: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        vision_embedding = self.vpm(
+            pixel_values,
+            patch_attention_mask=patch_attn_mask,
+            tgt_sizes=tgt_sizes,
+        )
+        return vision_embedding
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        # list of tensors
+        pixel_values = flatten_nested_list([item.feature for item in items])
+        tgt_sizes = torch.stack(
+            flatten_nested_list([item.tgt_size for item in items]), dim=0
+        )
+        assert len(pixel_values) == tgt_sizes.shape[0]
+
+        device = self.vpm.embeddings.position_embedding.weight.device
+        dtype = self.vpm.embeddings.position_embedding.weight.dtype
+        all_pixel_values_lst = [
+            i.flatten(end_dim=1).permute(1, 0) for i in pixel_values
+        ]
+
+        max_patches = (tgt_sizes[:, 0] * tgt_sizes[:, 1]).max().item()
+        assert isinstance(max_patches, int)
+        all_pixel_values = torch.nn.utils.rnn.pad_sequence(
+            all_pixel_values_lst, batch_first=True, padding_value=0.0
+        )
+
+        B, L, _ = all_pixel_values.shape
+        all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
+        patch_attn_mask = torch.zeros(
+            (B, 1, max_patches), dtype=torch.bool, device=device
+        )
+
+        tgt_sizes_tensor = tgt_sizes.clone().to(device=patch_attn_mask.device)
+        mask_shapes = tgt_sizes_tensor[:, 0] * tgt_sizes_tensor[:, 1]
+        patch_attn_mask[:, 0, :] = torch.arange(
+            patch_attn_mask.size(2), device=patch_attn_mask.device
+        ).unsqueeze(0) < mask_shapes.unsqueeze(1)
+
+        vision_embedding = self.vpm(
+            all_pixel_values.type(dtype),
+            patch_attention_mask=patch_attn_mask,
+            tgt_sizes=tgt_sizes,
+        )
+        return self.resampler(vision_embedding, tgt_sizes)
+
+    def pad_input_ids(self, input_ids: List[int], image_inputs: MultimodalInputs):
+        # Get all special token IDs
+        im_start_id: int = image_inputs.im_start_id
+        im_end_id: int = image_inputs.im_end_id
+        slice_start_id: int = image_inputs.slice_start_id
+        slice_end_id: int = image_inputs.slice_end_id
+
+        media_token_pairs = [(im_start_id, im_end_id), (slice_start_id, slice_end_id)]
+        pattern = MultiModalityDataPaddingPatternTokenPairs(media_token_pairs)
+
+        return pattern.pad_input_tokens(input_ids, image_inputs)
+
+
+_SUPPORT_VERSION = {
+    (2, 6): MiniCPMV2_6,
+    (4, 0): MiniCPMV4_0,
+}
 
 
 class MiniCPMV:
@@ -809,7 +971,7 @@ class MiniCPMV:
         # Dispatch class based on version
         instance_class = _SUPPORT_VERSION.get(version)
         if instance_class is None:
-            raise ValueError("Currently, MiniCPMV only supports versions 2.6")
+            raise ValueError("Currently, MiniCPMV only supports versions 2.6 and 4.0")
 
         try:
             minicpmv = instance_class(

sglang/srt/models/mllama4.py

@@ -961,5 +961,30 @@ class Llama4ForConditionalGeneration(nn.Module):
     def set_embed(self, embed):
         return self.language_model.set_embed(embed)
 
+    def get_hidden_dim(self, module_name, layer_idx):
+        # return input_dim, output_dim
+        if module_name == "qkv_proj":
+            return (
+                self.config.hidden_size,
+                self.config.head_dim
+                * (
+                    self.config.num_attention_heads
+                    + self.config.num_key_value_heads * 2
+                ),
+            )
+        elif module_name == "o_proj":
+            return (
+                self.config.head_dim * self.config.num_attention_heads,
+                self.config.hidden_size,
+            )
+        elif module_name == "gate_up_proj":
+            return self.config.hidden_size, self.config.intermediate_size * 2
+        elif module_name == "down_proj":
+            decoder_layer = self.language_model.get_layers()[layer_idx]
+            intermediate_size = decoder_layer.get_intermediate_size()
+            return intermediate_size, self.config.hidden_size
+        else:
+            raise NotImplementedError()
+
 
 EntryClass = Llama4ForConditionalGeneration