sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc2__py3-none-any.whl

sglang/srt/models/starcoder2.py

@@ -0,0 +1,357 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright 2024 BigCode and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/starcoder2.py
+""" PyTorch Starcoder2 model."""
+from collections.abc import Iterable
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import Starcoder2Config
+
+from sglang.srt.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from sglang.srt.layers.activation import get_act_fn
+from sglang.srt.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.vocab_parallel_embedding import (
+    DEFAULT_VOCAB_PADDING_SIZE,
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.utils import add_prefix, make_layers
+
+
+class Starcoder2Attention(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        layer_id: int = 0,
+    ):
+        super().__init__()
+        self.config = config
+
+        self.hidden_size = config.hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = config.num_attention_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = config.num_key_value_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # 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
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = self.hidden_size // self.total_num_heads
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = config.rope_theta
+        self.max_position_embeddings = config.max_position_embeddings
+        self.use_bias = config.use_bias
+
+        self.qkv_proj = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=self.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            self.hidden_size,
+            bias=self.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.o_proj",
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=self.max_position_embeddings,
+            base=int(self.rope_theta),
+            is_neox_style=True,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Starcoder2MLP(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.c_fc = ColumnParallelLinear(
+            config.hidden_size,
+            config.intermediate_size,
+            bias=config.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_fc",
+        )
+        self.c_proj = RowParallelLinear(
+            config.intermediate_size,
+            config.hidden_size,
+            bias=config.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
+        )
+        self.act = get_act_fn(config.hidden_act)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        hidden_states, _ = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.c_proj(hidden_states)
+        return hidden_states
+
+
+class Starcoder2DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        layer_id: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Starcoder2Attention(
+            config=config,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.mlp = Starcoder2MLP(
+            config, quant_config=quant_config, prefix=f"{prefix}.mlp"
+        )
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
+        self.post_attention_layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.norm_epsilon
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        # Self Attention
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class Starcoder2Model(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+
+        self.config = config
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+            prefix=f"{prefix}.embed_tokens",
+        )
+
+        pp_group = get_pp_group()
+        pp_size = pp_group.world_size
+        pp_rank = pp_group.rank
+        self.start_layer = pp_rank * config.num_hidden_layers // pp_size
+        self.end_layer = (pp_rank + 1) * config.num_hidden_layers // pp_size
+
+        self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: Starcoder2DecoderLayer(
+                config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
+            ),
+            prefix=f"{prefix}.layers",
+        )
+        self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if inputs_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = inputs_embeds
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states = layer(
+                positions,
+                hidden_states,
+                forward_batch,
+            )
+        hidden_states = self.norm(hidden_states)
+        return hidden_states
+
+
+class Starcoder2ForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        self.model = Starcoder2Model(
+            config, quant_config, prefix=add_prefix("model", prefix)
+        )
+        self.vocab_size = config.vocab_size
+        self.unpadded_vocab_size = config.vocab_size
+        if config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.unpadded_vocab_size = config.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,
+                quant_config=quant_config,
+                prefix=f"{prefix}.lm_head",
+            )
+        self.logits_processor = LogitsProcessor(config=config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(
+            input_ids=input_ids,
+            positions=positions,
+            forward_batch=forward_batch,
+            inputs_embeds=inputs_embeds,
+        )
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head, forward_batch
+        )
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+        ]
+        params_dict = dict(self.named_parameters())
+
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freqs" in name:
+                continue
+
+            is_stacked = False
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name in name:
+                    name = name.replace(weight_name, param_name)
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight, shard_id)
+                    is_stacked = True
+                    break
+            if is_stacked:
+                continue
+
+            param = params_dict.get(name)
+            if param is None:
+                continue
+
+            weight_loader = getattr(param, "weight_loader", default_weight_loader)
+            weight_loader(param, loaded_weight)
+
+
+EntryClass = Starcoder2ForCausalLM

sglang/srt/models/step3_vl.py

@@ -133,7 +133,7 @@ class Step3TextMoEMLP(nn.Module):
             use_grouped_topk=False,
         )
 
-        self.experts = get_moe_impl_class()(
+        self.experts = get_moe_impl_class(quant_config)(
             num_experts=config.moe_num_experts,
             top_k=config.moe_top_k,
             hidden_size=config.hidden_size,

sglang/srt/models/torch_native_llama.py

@@ -66,8 +66,8 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix
 
-tp_size = get_tensor_model_parallel_world_size()
-tp_rank = get_tensor_model_parallel_rank()
+tp_size: Optional[int] = None
+tp_rank: Optional[int] = None
 
 
 def gate_up_proj_weight_loader(
@@ -341,6 +341,13 @@ class LlamaModel(nn.Module):
         quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
+
+        global tp_size, tp_rank
+        if tp_size is None:
+            tp_size = get_tensor_model_parallel_world_size()
+        if tp_rank is None:
+            tp_rank = get_tensor_model_parallel_rank()
+
         self.config = config
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size

sglang/srt/models/utils.py

@@ -0,0 +1,51 @@
+# Copyright 2023-2025 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+# ==============================================================================
+
+import torch
+
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.utils import is_cuda
+
+_is_cuda = is_cuda()
+
+
+if _is_cuda:
+    from sgl_kernel import FusedSetKVBufferArg
+
+
+def enable_fused_set_kv_buffer(forward_batch: ForwardBatch):
+    """Enable fused set_kv_buffer only on CUDA with bfloat16 KV cache."""
+    return _is_cuda and forward_batch.token_to_kv_pool.dtype == torch.bfloat16
+
+
+def create_fused_set_kv_buffer_arg(
+    value: torch.Tensor,
+    layer: RadixAttention,
+    forward_batch: ForwardBatch,
+):
+    layer_id = layer.layer_id
+    token_to_kv_pool = forward_batch.token_to_kv_pool
+
+    k_buffer = token_to_kv_pool.get_key_buffer(layer_id)
+    v_buffer = token_to_kv_pool.get_value_buffer(layer_id)
+
+    return FusedSetKVBufferArg(
+        value=value,
+        k_buffer=k_buffer.view(k_buffer.shape[0], -1),
+        v_buffer=v_buffer.view(v_buffer.shape[0], -1),
+        k_scale=layer.k_scale,
+        v_scale=layer.v_scale,
+        cache_loc=forward_batch.out_cache_loc,
+    )

sglang/srt/multimodal/processors/base_processor.py

@@ -234,19 +234,27 @@ class BaseMultimodalProcessor(ABC):
             and isinstance(processor.image_processor, BaseImageProcessorFast)
             and not self.server_args.disable_fast_image_processor
         ):
-            kwargs["device"] = "cuda" if not _is_npu else "npu"
+            if not _is_npu:
+                kwargs["device"] = "cuda"
+            elif processor.__class__.__name__ not in {
+                "Qwen2_5_VLProcessor",
+                "Qwen3VLProcessor",
+            }:
+                # Note: for qwen-vl, processor has some reshape issue because of dims restriction on Ascend.
+                kwargs["device"] = "npu"
         result = processor.__call__(
             text=[input_text],
             padding=True,
             return_tensors="pt",
             **kwargs,
         )
-        # move feature tensors to cpu
-        for feature_name in self.FEATURE_NAMES:
-            if feature_name in result and isinstance(
-                result[feature_name], torch.Tensor
-            ):
-                result[feature_name] = result[feature_name].to("cpu")
+        if not self.server_args.keep_mm_feature_on_device:
+            # move feature tensors to cpu
+            for feature_name in self.FEATURE_NAMES:
+                if feature_name in result and isinstance(
+                    result[feature_name], torch.Tensor
+                ):
+                    result[feature_name] = result[feature_name].to("cpu")
 
         return result
 

sglang/srt/multimodal/processors/dots_vlm.py

@@ -0,0 +1,98 @@
+import asyncio
+import math
+import re
+from typing import Dict, List, Union
+
+from PIL import Image
+
+from sglang.srt.models.dots_ocr import DotsOCRForCausalLM
+from sglang.srt.models.dots_vlm import DotsVLMForCausalLM
+from sglang.srt.multimodal.processors.base_processor import (
+    BaseMultimodalProcessor,
+    MultimodalSpecialTokens,
+)
+from sglang.srt.multimodal.processors.qwen_vl import resize_image_async
+
+
+class DotsVLMImageProcessor(BaseMultimodalProcessor):
+    models = [DotsVLMForCausalLM, DotsOCRForCausalLM]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+        # The single, pre-expanded image token.
+        self.IMAGE_TOKEN = "<|img|><|imgpad|><|endofimg|>"
+        # The regex that matches expanded image tokens.
+        self.IMAGE_TOKEN_REGEX = re.compile(r"<\|img\|>(?:<\|imgpad\|>)+<\|endofimg\|>")
+
+        assert len(_processor.tokenizer.encode("<|img|>")) == 1
+        self.im_start_id = _processor.tokenizer.encode("<|img|>")[0]
+        self.im_end_id = _processor.tokenizer.encode("<|endofimg|>")[0]
+        self.image_token_id = _processor.tokenizer.encode("<|imgpad|>")[0]
+        self.IM_TOKEN_ID = self.image_token_id
+        self.IM_START_ID = self.im_start_id
+        self.IM_END_ID = self.im_end_id
+
+        vision_config = hf_config.vision_config
+        patch_size = vision_config.patch_size
+        merge_size = vision_config.spatial_merge_size
+
+        self.IMAGE_FACTOR = patch_size * merge_size
+        self.MIN_PIXELS = _processor.image_processor.min_pixels
+        self.MAX_PIXELS = _processor.image_processor.max_pixels
+        self.MAX_RATIO = 200
+        self.mm_tokens = MultimodalSpecialTokens(
+            image_token=self.IMAGE_TOKEN,
+            image_token_id=self.image_token_id,
+            image_token_regex=self.IMAGE_TOKEN_REGEX,
+        ).build(_processor)
+
+    async def process_mm_data_async(
+        self,
+        image_data: List[Union[str, bytes, Dict]],
+        input_text,
+        request_obj,
+        max_req_input_len,
+        *args,
+        **kwargs,
+    ):
+        if isinstance(image_data, str):
+            image_data = [image_data]
+
+        if (
+            isinstance(image_data, list)
+            and image_data
+            and isinstance(image_data[0], list)
+        ):
+            image_data = sum(image_data, [])
+
+        base_output = self.load_mm_data(
+            prompt=input_text,
+            image_data=image_data,
+            multimodal_tokens=self.mm_tokens,
+        )
+
+        # Qwen-specific: resize images if they are raw Image objects
+        if base_output.images and isinstance(base_output.images[0], Image.Image):
+            resize_tasks = [
+                resize_image_async(
+                    image,
+                    min_pixels=self.MIN_PIXELS,
+                    max_pixels=self.MAX_PIXELS,
+                    size_factor=self.IMAGE_FACTOR,
+                )
+                for image in base_output.images
+            ]
+            base_output.images = await asyncio.gather(*resize_tasks)
+        combined_mm_item, input_ids, _ = self.process_and_combine_mm_data(
+            base_output, self.mm_tokens
+        )
+        if combined_mm_item is None:
+            return None
+
+        return {
+            "input_ids": input_ids.tolist(),
+            "mm_items": combined_mm_item,
+            "im_start_id": self.im_start_id,
+            "im_end_id": self.im_end_id,
+            "im_token_id": self.image_token_id,
+        }

sglang/srt/multimodal/processors/glm4v.py

@@ -2,7 +2,6 @@ import re
 from typing import List, Union
 
 from decord import VideoReader
-from transformers.video_utils import VideoMetadata
 
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
 from sglang.srt.models.glm4v import Glm4vForConditionalGeneration
@@ -66,17 +65,18 @@ class Glm4vImageProcessor(SGLangBaseProcessor):
         total_num_frames = len(vr)
         duration = total_num_frames / video_fps if video_fps else 0
 
-        metadata = VideoMetadata(
-            total_num_frames=int(total_num_frames),
-            fps=float(video_fps),
-            duration=float(duration),
-            video_backend="decord",
-        )
-
         # Extract all frames
         indices = list(range(total_num_frames))
         frames = vr.get_batch(indices).asnumpy()
-        metadata.frames_indices = indices
+
+        # Return metadata as dict so transformers can properly create VideoMetadata objects
+        metadata = {
+            "total_num_frames": int(total_num_frames),
+            "fps": float(video_fps),
+            "duration": float(duration),
+            "video_backend": "decord",
+            "frames_indices": indices,
+        }
 
         return frames, metadata