sglang 0.1.13__py3-none-any.whl → 0.1.15__py3-none-any.whl

sglang/srt/models/qwen2.py

@@ -1,34 +1,36 @@
 # Adapted from llama2.py
 # Modify details for the adaptation of Qwen2 model.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, Optional, Tuple
 
 import torch
-from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.managers.router.model_runner import InputMetadata
 from torch import nn
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
-    LinearMethodBase,
     MergedColumnParallelLinear,
     QKVParallelLinear,
     RowParallelLinear,
 )
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from vllm.model_executor.parallel_utils.parallel_state import (
+from vllm.distributed import (
     get_tensor_model_parallel_world_size,
 )
-from vllm.model_executor.weight_utils import (
+from sglang.srt.weight_utils import (
     default_weight_loader,
     hf_model_weights_iterator,
 )
 
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+
 Qwen2Config = None
 
 
@@ -38,17 +40,17 @@ class Qwen2MLP(nn.Module):
         hidden_size: int,
         intermediate_size: int,
         hidden_act: str,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
             hidden_size,
             [intermediate_size] * 2,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.down_proj = RowParallelLinear(
-            intermediate_size, hidden_size, bias=False, linear_method=linear_method
+            intermediate_size, hidden_size, bias=False, quant_config=quant_config,
         )
         if hidden_act != "silu":
             raise ValueError(
@@ -74,7 +76,7 @@ class Qwen2Attention(nn.Module):
         rope_theta: float = 1000000,
         rope_scaling: Optional[Dict[str, Any]] = None,
         max_position_embeddings: int = 32768,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -105,13 +107,13 @@ class Qwen2Attention(nn.Module):
             self.total_num_heads,
             self.total_num_kv_heads,
             bias=True,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.o_proj = RowParallelLinear(
             self.total_num_heads * self.head_dim,
             hidden_size,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
 
         self.rotary_emb = get_rope(
@@ -148,7 +150,7 @@ class Qwen2DecoderLayer(nn.Module):
         self,
         config: Qwen2Config,
         layer_id: int = 0,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -163,13 +165,13 @@ class Qwen2DecoderLayer(nn.Module):
             rope_theta=rope_theta,
             rope_scaling=rope_scaling,
             max_position_embeddings=max_position_embeddings,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.mlp = Qwen2MLP(
             hidden_size=self.hidden_size,
             intermediate_size=config.intermediate_size,
             hidden_act=config.hidden_act,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.post_attention_layernorm = RMSNorm(
@@ -205,7 +207,7 @@ class Qwen2Model(nn.Module):
     def __init__(
         self,
         config: Qwen2Config,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -217,7 +219,7 @@ class Qwen2Model(nn.Module):
         )
         self.layers = nn.ModuleList(
             [
-                Qwen2DecoderLayer(config, i, linear_method)
+                Qwen2DecoderLayer(config, i, quant_config=quant_config)
                 for i in range(config.num_hidden_layers)
             ]
         )
@@ -251,12 +253,12 @@ class Qwen2ForCausalLM(nn.Module):
     def __init__(
         self,
         config: Qwen2Config,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
-        self.linear_method = linear_method
-        self.model = Qwen2Model(config, linear_method)
+        self.quant_config = quant_config
+        self.model = Qwen2Model(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 

sglang/srt/models/stablelm.py

@@ -0,0 +1,292 @@
+# This code is based on:
+# https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/stablelm.py
+"""Inference-only StableLM-2 (https://huggingface.co/stabilityai/stablelm-2-1_6b)
+model compatible with HuggingFace weights."""
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.distributed import (
+    get_tensor_model_parallel_world_size,
+)
+from sglang.srt.weight_utils import (
+    default_weight_loader,
+    hf_model_weights_iterator,
+)
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+
+
+class StablelmMLP(nn.Module):
+    def __init__(
+        self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_up_proj = MergedColumnParallelLinear(
+            config.hidden_size,
+            [config.intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.down_proj = RowParallelLinear(
+            config.intermediate_size, config.hidden_size, bias=False, quant_config=quant_config,
+        )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class StablelmAttention(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        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
+        self.num_heads = self.total_num_heads // tp_size
+
+        self.total_num_key_value_heads = config.num_key_value_heads
+        if self.total_num_key_value_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_key_value_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_key_value_heads == 0
+        self.num_key_value_heads = max(1, self.total_num_key_value_heads // tp_size)
+        self.head_dim = self.hidden_size // self.total_num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        rope_pct = getattr(
+            config, "rope_pct", getattr(config, "partial_rotary_factor", 1)
+        )
+        self.rotary_ndims = int(self.head_dim * rope_pct)
+        self.scaling = self.head_dim**-0.5
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_key_value_heads * self.head_dim
+        self.qkv_bias = getattr(config, "use_qkv_bias", False)
+        if (self.head_dim * self.num_heads * tp_size) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads "
+                f"(got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        self.qkv_proj = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_key_value_heads,
+            self.qkv_bias,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            self.hidden_size,
+            bias=False,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.rotary_ndims,
+            max_position=self.config.max_position_embeddings,
+            base=self.config.rope_theta,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_key_value_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> 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, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class StablelmDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.self_attn = StablelmAttention(config, layer_id=layer_id)
+        self.mlp = StablelmMLP(config, quant_config=quant_config)
+        norm_eps = getattr(config, "norm_eps", getattr(config, "layer_norm_eps", 1e-05))
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=norm_eps)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> Tuple[torch.Tensor, 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,
+            input_metadata=input_metadata,
+        )
+        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, residual
+
+
+class StableLMEpochModel(nn.Module):
+    def __init__(
+        self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                StablelmDecoderLayer(config, i, quant_config=quant_config)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        norm_eps = getattr(config, "norm_eps", getattr(config, "layer_norm_eps", 1e-05))
+        self.norm = nn.LayerNorm(config.hidden_size, eps=norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        if input_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_embeds
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                input_metadata,
+            )
+        hidden_states = self.norm(hidden_states)
+        return hidden_states
+
+
+class StableLmForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = StableLMEpochModel(config, quant_config=quant_config)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.logits_processor = LogitsProcessor(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    def load_weights(
+        self,
+        model_name_or_path: str,
+        cache_dir: Optional[str] = None,
+        load_format: str = "auto",
+        revision: Optional[str] = None,
+    ):
+        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 hf_model_weights_iterator(
+            model_name_or_path, cache_dir, load_format, revision
+        ):
+            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
+            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)
+
+
+EntryClass = StableLmForCausalLM

sglang/srt/models/yivl.py

@@ -5,16 +5,17 @@ from typing import List, Optional
 
 import torch
 import torch.nn as nn
+from transformers import CLIPVisionModel, LlavaConfig
+from sglang.srt.weight_utils import (
+    default_weight_loader,
+    hf_model_weights_iterator,
+)
+
 from sglang.srt.models.llava import (
     LlavaLlamaForCausalLM,
     clip_vision_embed_forward,
     monkey_path_clip_vision_embed_forward,
 )
-from transformers import CLIPVisionModel, LlavaConfig
-from vllm.model_executor.weight_utils import (
-    default_weight_loader,
-    hf_model_weights_iterator,
-)
 
 
 class YiVLForCausalLM(LlavaLlamaForCausalLM):