sglang 0.4.5__py3-none-any.whl → 0.4.5.post2__py3-none-any.whl

sglang/srt/models/qwen2_moe.py

@@ -47,7 +47,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.managers.expert_distribution import ExpertDistributionRecorder
 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
+from sglang.srt.utils import add_prefix, make_layers
 
 expert_distribution_recorder = ExpertDistributionRecorder()
 
@@ -231,6 +231,7 @@ class Qwen2MoeAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 
@@ -261,8 +262,7 @@ class Qwen2MoeDecoderLayer(nn.Module):
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
-        # note: replace config.num_hidden_layers < 80 with True once its available in transformers 4.50.0
-        qkv_bias = getattr(config, "qkv_bias", config.num_hidden_layers < 80)
+        qkv_bias = getattr(config, "qkv_bias", True)
         self.self_attn = Qwen2MoeAttention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
@@ -333,6 +333,7 @@ class Qwen2MoeModel(nn.Module):
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        decoder_layer_type: type[nn.Module] = Qwen2MoeDecoderLayer,
     ) -> None:
         super().__init__()
         self.padding_idx = config.pad_token_id
@@ -343,16 +344,17 @@ class Qwen2MoeModel(nn.Module):
             config.hidden_size,
             prefix=add_prefix("embed_tokens", prefix),
         )
-        self.layers = nn.ModuleList(
-            [
-                Qwen2MoeDecoderLayer(
-                    config,
-                    layer_id,
-                    quant_config=quant_config,
-                    prefix=add_prefix(f"layers.{layer_id}", prefix),
-                )
-                for layer_id in range(config.num_hidden_layers)
-            ]
+        # Use the provided decoder layer type or default to Qwen2MoeDecoderLayer
+        decoder_layer_type = decoder_layer_type or Qwen2MoeDecoderLayer
+        self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: decoder_layer_type(
+                layer_id=idx,
+                config=config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=add_prefix("layers", prefix),
         )
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 

sglang/srt/models/qwen2_vl.py

@@ -152,7 +152,7 @@ class Qwen2VisionBlock(nn.Module):
             embed_dim=dim,
             num_heads=num_heads,
             projection_size=dim,
-            use_qkv_parallel=False,
+            use_qkv_parallel=True,
             use_context_forward=use_context_forward,
             softmax_in_single_precision=softmax_in_single_precision,
             flatten_batch=True,
@@ -351,7 +351,7 @@ class Qwen2VisionTransformer(nn.Module):
 
     @property
     def dtype(self) -> torch.dtype:
-        return next(self.parameters()).dtype
+        return self.patch_embed.proj.weight.dtype
 
     @property
     def device(self) -> torch.device:
@@ -423,6 +423,25 @@ cached_get_processor = lru_cache(get_processor)
 
 
 class Qwen2VLForConditionalGeneration(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    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),
+    }
+
     def calculate_num_image_tokens(self, image_grid_thw: Tuple[int, int, int]):
         processor = cached_get_processor(self.config._name_or_path)
         grid_t, grid_h, grid_w = image_grid_thw
@@ -447,9 +466,9 @@ class Qwen2VLForConditionalGeneration(nn.Module):
         self.visual = Qwen2VisionTransformer(
             config.vision_config,
             norm_eps=getattr(config, "rms_norm_eps", 1e-6),
-            # NOTE: Qwen2-VL vision encoder does not support any
-            # quantization method now.
-            quant_config=None,
+            # NOTE: Qwen2-VL vision encoder currently supports BitsAndBytes 4-bit quantization.
+            # Other quantization methods (e.g., GPTQ, AWQ) are untested and may not be supported.
+            quant_config=quant_config,
             prefix=add_prefix("visual", prefix),
         )
 
@@ -467,6 +486,7 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 prefix=add_prefix("lm_head", prefix),
             )
 
+        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
@@ -521,14 +541,14 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 otherwise it will be `(seq_len,).
                 (Use input_metadata.mrope_positions to replace it)
         """
-        if getattr(self.config, "rope_scaling", {}).get("type", None) == "mrope":
+        if self.is_mrope_enabled:
             positions = forward_batch.mrope_positions
 
         if not (
             forward_batch.forward_mode.is_decode()
             or not forward_batch.contains_image_inputs()
         ):
-            if getattr(self.config, "rope_scaling", {}).get("type", None) == "mrope":
+            if self.is_mrope_enabled:
                 assert positions.ndim == 2 and positions.size(0) == 3, (
                     "multimodal section rotary embedding requires "
                     f"(3, seq_len) positions, but got {positions.size()}"
@@ -577,24 +597,6 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 weight_loader(param, loaded_weight, shard_id)
                 break
             else:
-
-                if "visual" in name and "qkv.weight" in name:
-                    visual_num_heads = self.config.vision_config.num_heads
-                    visual_embed_dim = self.config.vision_config.embed_dim
-                    head_size = visual_embed_dim // visual_num_heads
-                    loaded_weight = loaded_weight.view(
-                        3, visual_num_heads, head_size, visual_embed_dim
-                    )
-                    loaded_weight = loaded_weight.transpose(0, 1)
-                    loaded_weight = loaded_weight.reshape(-1, visual_embed_dim)
-                elif "visual" in name and "qkv.bias" in name:
-                    visual_num_heads = self.config.vision_config.num_heads
-                    visual_embed_dim = self.config.vision_config.embed_dim
-                    head_size = visual_embed_dim // visual_num_heads
-                    loaded_weight = loaded_weight.view(3, visual_num_heads, head_size)
-                    loaded_weight = loaded_weight.transpose(0, 1)
-                    loaded_weight = loaded_weight.reshape(-1)
-
                 if "visual" in name:
                     # adapt to VisionAttention
                     name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")

sglang/srt/models/qwen3.py

@@ -0,0 +1,335 @@
+# Adapted from qwen2.py
+
+from functools import partial
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+
+from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    split_tensor_along_last_dim,
+    tensor_model_parallel_all_gather,
+)
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.pooler import Pooler, PoolingType
+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 ParallelLMHead
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.models.qwen2 import Qwen2MLP as Qwen3MLP
+from sglang.srt.models.qwen2 import Qwen2Model
+from sglang.srt.utils import add_prefix
+
+Qwen3Config = None
+
+
+class Qwen3Attention(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 1000000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        head_dim: Optional[int] = None,
+        max_position_embeddings: int = 32768,
+        quant_config: Optional[QuantizationConfig] = None,
+        rms_norm_eps: float = None,
+        attention_bias: bool = False,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % self.tp_size == 0
+        self.num_heads = self.total_num_heads // self.tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= self.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 % self.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 self.tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size)
+        self.head_dim = head_dim or 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 = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+        self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+        self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("qkv_proj", prefix),
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            prefix=add_prefix("attn", prefix),
+        )
+
+    def _apply_qk_norm(
+        self, q: torch.Tensor, k: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        q_by_head = q.reshape(-1, self.head_dim)
+        q_by_head = self.q_norm(q_by_head)
+        q = q_by_head.view(q.shape)
+        k_by_head = k.reshape(-1, self.head_dim)
+        k_by_head = self.k_norm(k_by_head)
+        k = k_by_head.view(k.shape)
+        return q, k
+
+    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._apply_qk_norm(q, k)
+        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 Qwen3DecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: Qwen3Config,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 1000000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
+        head_dim = getattr(config, "head_dim", None)
+        self.self_attn = Qwen3Attention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            head_dim=head_dim,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            rms_norm_eps=config.rms_norm_eps,
+            attention_bias=config.attention_bias,
+            prefix=add_prefix("self_attn", prefix),
+        )
+        self.mlp = Qwen3MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class Qwen3Model(Qwen2Model):
+    def __init__(
+        self,
+        config: Qwen3Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(
+            config=config,
+            quant_config=quant_config,
+            prefix=prefix,
+            decoder_layer_type=Qwen3DecoderLayer,
+        )
+
+
+class Qwen3ForCausalLM(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    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),
+    }
+
+    def __init__(
+        self,
+        config: Qwen3Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Qwen3Model(
+            config, quant_config=quant_config, prefix=add_prefix("model", prefix)
+        )
+        if config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.lm_head = ParallelLMHead(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("lm_head", prefix),
+            )
+        self.logits_processor = LogitsProcessor(config)
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        get_embedding: bool = False,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
+        if not get_embedding:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head, forward_batch
+            )
+        else:
+            return self.pooler(hidden_states, 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"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name or "projector" 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.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+            if name.startswith("model.vision_tower") and name not in params_dict:
+                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
+                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
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_embed_and_head(self, embed, head):
+        del self.model.embed_tokens.weight
+        del self.lm_head.weight
+        self.model.embed_tokens.weight = embed
+        self.lm_head.weight = head
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        self.model.load_kv_cache_scales(quantization_param_path)
+
+
+EntryClass = Qwen3ForCausalLM