sglang 0.1.21__py3-none-any.whl → 0.1.22__py3-none-any.whl

sglang/srt/models/internlm2.py

@@ -0,0 +1,317 @@
+# -*- coding: utf-8 -*-
+# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/7f62077af5159c625fe3ad1c812e6c1a2b93ba3b/vllm/model_executor/models/internlm2.py
+
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+from vllm.config import CacheConfig
+from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+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.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.controller.model_runner import InputMetadata
+
+
+class InternLM2MLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config
+        )
+        self.w2 = RowParallelLinear(
+            intermediate_size, hidden_size, bias=False, quant_config=quant_config
+        )
+        if hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Only silu is supported for now."
+            )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.w2(x)
+        return x
+
+
+class InternLM2Attention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_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 = 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.wqkv = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.wo = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        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, self.num_kv_heads, layer_id
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.wqkv(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.wo(attn_output)
+        return output
+
+
+class InternLMDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        self.attention = InternLM2Attention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            layer_id=layer_id,
+            quant_config=quant_config,
+        )
+        self.feed_forward = InternLM2MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+        )
+        self.attention_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.ffn_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.attention_norm(hidden_states)
+        else:
+            hidden_states, residual = self.attention_norm(hidden_states, residual)
+        hidden_states = self.attention(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.ffn_norm(hidden_states, residual)
+        hidden_states = self.feed_forward(hidden_states)
+        return hidden_states, residual
+
+
+class InternLM2Model(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.tok_embeddings = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                InternLMDecoderLayer(config, i, quant_config)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_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.tok_embeddings(input_ids)
+        else:
+            hidden_states = input_embeds
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class InternLM2ForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = InternLM2Model(config, quant_config)
+        self.output = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    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.output.weight, input_metadata
+        )
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("gate_up_proj", "w1", 0),
+            ("gate_up_proj", "w3", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                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]
+                if "wqkv" in name:
+                    config = self.config
+                    kv_groups = config.num_attention_heads // config.num_key_value_heads
+                    head_dim = config.hidden_size // config.num_attention_heads
+                    loaded_weight = loaded_weight.view(
+                        -1, 2 + kv_groups, head_dim, loaded_weight.shape[-1]
+                    )
+                    wq, wk, wv = torch.split(loaded_weight, [kv_groups, 1, 1], dim=1)
+                    wq = wq.reshape(-1, wq.shape[-1])
+                    wk = wk.reshape(-1, wk.shape[-1])
+                    wv = wv.reshape(-1, wv.shape[-1])
+                    weight_loader = param.weight_loader
+                    weight_loader(param, wq, "q")
+                    weight_loader(param, wk, "k")
+                    weight_loader(param, wv, "v")
+                else:
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+
+
+EntryClass = InternLM2ForCausalLM

sglang/srt/models/llama2.py

@@ -15,11 +15,6 @@ from vllm.distributed import (
 )
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.layernorm import RMSNorm
-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 (
@@ -32,6 +27,10 @@ from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.managers.controller.model_runner import InputMetadata
 
+MergedColumnParallelLinear = None
+QKVParallelLinear = None
+RowParallelLinear = None
+
 
 class LlamaMLP(nn.Module):
     def __init__(
@@ -163,9 +162,9 @@ class LlamaDecoderLayer(nn.Module):
         if rope_scaling is not None and getattr(
             config, "original_max_position_embeddings", None
         ):
-            rope_scaling[
-                "original_max_position_embeddings"
-            ] = config.original_max_position_embeddings
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings
+            )
         rope_is_neox_style = getattr(config, "rope_is_neox_style", True)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = LlamaAttention(
@@ -267,7 +266,25 @@ class LlamaForCausalLM(nn.Module):
         config: LlamaConfig,
         quant_config: Optional[QuantizationConfig] = None,
         cache_config: Optional[CacheConfig] = None,
+        efficient_weight_load=False,
     ) -> None:
+        global MergedColumnParallelLinear
+        global QKVParallelLinear
+        global RowParallelLinear
+
+        if efficient_weight_load:
+            from sglang.srt.layers.linear import (
+                MergedColumnParallelLinear,
+                QKVParallelLinear,
+                RowParallelLinear,
+            )
+        else:
+            from vllm.model_executor.layers.linear import (
+                MergedColumnParallelLinear,
+                QKVParallelLinear,
+                RowParallelLinear,
+            )
+
         super().__init__()
         self.config = config
         self.quant_config = quant_config
@@ -275,6 +292,7 @@ class LlamaForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -287,7 +305,30 @@ class LlamaForCausalLM(nn.Module):
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )
 
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def get_module_name(self, name):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id, num_shard)
+            ("qkv_proj", "q_proj", "q", 3),
+            ("qkv_proj", "k_proj", "k", 3),
+            ("qkv_proj", "v_proj", "v", 3),
+            ("gate_up_proj", "gate_proj", 0, 2),
+            ("gate_up_proj", "up_proj", 1, 2),
+        ]
+        for param_name, weight_name, shard_id, num_shard in stacked_params_mapping:
+            if weight_name in name:
+                return (
+                    name.replace(weight_name, param_name)[: -len(".weight")],
+                    num_shard,
+                )
+        return name[: -len(".weight")], 1
+
+    def get_num_params(self):
+        params_dict = dict(self.named_parameters())
+        return len(params_dict)
+
+    def load_weights(
+        self, weights: Iterable[Tuple[str, torch.Tensor]], name=None, loaded_weight=None
+    ):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
@@ -297,15 +338,14 @@ class LlamaForCausalLM(nn.Module):
             ("gate_up_proj", "up_proj", 1),
         ]
         params_dict = dict(self.named_parameters())
-        if get_tensor_model_parallel_rank() == 0:
-            weights = tqdm.tqdm(weights, total=int(len(params_dict) * 1.5))
-        for name, loaded_weight in weights:
+
+        def load_weights_per_param(name, loaded_weight):
             if "rotary_emb.inv_freq" in name or "projector" in name:
-                continue
+                return
             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
+                return
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
                     continue
@@ -322,12 +362,21 @@ class LlamaForCausalLM(nn.Module):
             else:
                 # Skip loading extra bias for GPTQ models.
                 if name.endswith(".bias") and name not in params_dict:
-                    continue
+                    return
                 if name.startswith("model.vision_tower") and name not in params_dict:
-                    continue
+                    return
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
+        if name is None or loaded_weight is None:
+            if get_tensor_model_parallel_rank() == 0:
+                weights = tqdm.tqdm(weights, total=int(len(params_dict) * 1.5))
+
+            for name, loaded_weight in weights:
+                load_weights_per_param(name, loaded_weight)
+        else:
+            load_weights_per_param(name, loaded_weight)
+
 
 EntryClass = LlamaForCausalLM

sglang/srt/models/llama_classification.py

@@ -31,6 +31,7 @@ class LlamaForClassification(nn.Module):
         )
         self.eos_token_id = config.eos_token_id
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,

sglang/srt/models/llava.py

@@ -95,6 +95,7 @@ class LlavaLlamaForCausalLM(nn.Module):
 
         return image_features
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.LongTensor,

sglang/srt/models/llavavid.py

@@ -106,6 +106,7 @@ class LlavaVidForCausalLM(nn.Module):
 
         return image_features
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.LongTensor,

sglang/srt/models/minicpm.py

@@ -283,6 +283,7 @@ class MiniCPMForCausalLM(nn.Module):
 
         self.logits_processor = LogitsProcessor(config)
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,

sglang/srt/models/mixtral.py

@@ -460,6 +460,7 @@ class MixtralForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,

sglang/srt/models/mixtral_quant.py

@@ -322,6 +322,7 @@ class QuantMixtralForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,

sglang/srt/models/qwen.py

@@ -237,6 +237,7 @@ class QWenLMHeadModel(nn.Module):
         self.lm_head = ParallelLMHead(vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,

sglang/srt/models/qwen2.py

@@ -261,6 +261,7 @@ class Qwen2ForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -312,6 +313,11 @@ class Qwen2ForCausalLM(nn.Module):
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
+                if (
+                    self.config.tie_word_embeddings
+                    and name == "model.embed_tokens.weight"
+                ):
+                    weight_loader(params_dict["lm_head.weight"], loaded_weight)
 
 
 EntryClass = Qwen2ForCausalLM

sglang/srt/models/qwen2_moe.py

@@ -355,6 +355,7 @@ class Qwen2MoeForCausalLM(nn.Module):
         self.logits_processor = LogitsProcessor(config)
         self.sampler = Sampler()
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -400,9 +401,11 @@ class Qwen2MoeForCausalLM(nn.Module):
             # These are the weights for the experts
             # (param_name, weight_name, expert_id, shard_id)
             (
-                "experts.w13_weight"
-                if weight_name in ["gate_proj", "up_proj"]
-                else "experts.w2_weight",
+                (
+                    "experts.w13_weight"
+                    if weight_name in ["gate_proj", "up_proj"]
+                    else "experts.w2_weight"
+                ),
                 f"experts.{expert_id}.{weight_name}.weight",
                 expert_id,
                 shard_id,
@@ -417,7 +420,7 @@ class Qwen2MoeForCausalLM(nn.Module):
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
-            for (param_name, weight_name, shard_id) in stacked_params_mapping:
+            for param_name, weight_name, shard_id in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name:
                     continue

sglang/srt/models/stablelm.py

@@ -235,6 +235,7 @@ class StableLmForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,