sglang 0.1.14__py3-none-any.whl → 0.1.21__py3-none-any.whl

sglang/srt/models/llama2.py

@@ -1,34 +1,36 @@
 # Adapted from
-# https://github.com/vllm-project/vllm/blob/671af2b1c0b3ed6d856d37c21a561cc429a10701/vllm/model_executor/models/llama.py#L1
+# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/llama.py#L1
 """Inference-only LLaMA model compatible with HuggingFace weights."""
-from typing import Any, Dict, List, Optional, Tuple
+
+from typing import Any, Dict, Iterable, 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
+import tqdm
 from torch import nn
 from transformers import LlamaConfig
+from vllm.config import CacheConfig
+from vllm.distributed import (
+    get_tensor_model_parallel_rank,
+    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 (
-    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 (
-    get_tensor_model_parallel_world_size,
-)
-from vllm.model_executor.weight_utils import (
-    default_weight_loader,
-    hf_model_weights_iterator,
-)
+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 LlamaMLP(nn.Module):
@@ -37,17 +39,20 @@ class LlamaMLP(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(
@@ -72,8 +77,9 @@ class LlamaAttention(nn.Module):
         layer_id: int = 0,
         rope_theta: float = 10000,
         rope_scaling: Optional[Dict[str, Any]] = None,
+        rope_is_neox_style: bool = True,
         max_position_embeddings: int = 8192,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -104,13 +110,13 @@ class LlamaAttention(nn.Module):
             self.total_num_heads,
             self.total_num_kv_heads,
             bias=False,
-            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(
@@ -119,6 +125,7 @@ class LlamaAttention(nn.Module):
             max_position=max_position_embeddings,
             base=rope_theta,
             rope_scaling=rope_scaling,
+            is_neox_style=rope_is_neox_style,
         )
         self.attn = RadixAttention(
             self.num_heads,
@@ -147,12 +154,19 @@ class LlamaDecoderLayer(nn.Module):
         self,
         config: LlamaConfig,
         layer_id: int = 0,
-        linear_method: Optional[LinearMethodBase] = None,
+        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)
+        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_is_neox_style = getattr(config, "rope_is_neox_style", True)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = LlamaAttention(
             hidden_size=self.hidden_size,
@@ -161,14 +175,15 @@ class LlamaDecoderLayer(nn.Module):
             layer_id=layer_id,
             rope_theta=rope_theta,
             rope_scaling=rope_scaling,
+            rope_is_neox_style=rope_is_neox_style,
             max_position_embeddings=max_position_embeddings,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.mlp = LlamaMLP(
             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(
@@ -204,7 +219,7 @@ class LlamaModel(nn.Module):
     def __init__(
         self,
         config: LlamaConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -216,7 +231,7 @@ class LlamaModel(nn.Module):
         )
         self.layers = nn.ModuleList(
             [
-                LlamaDecoderLayer(config, i, linear_method)
+                LlamaDecoderLayer(config, i, quant_config=quant_config)
                 for i in range(config.num_hidden_layers)
             ]
         )
@@ -250,12 +265,13 @@ class LlamaForCausalLM(nn.Module):
     def __init__(
         self,
         config: LlamaConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
-        self.linear_method = linear_method
-        self.model = LlamaModel(config, linear_method)
+        self.quant_config = quant_config
+        self.model = LlamaModel(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
@@ -271,13 +287,7 @@ class LlamaForCausalLM(nn.Module):
             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,
-    ):
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
@@ -287,9 +297,9 @@ class LlamaForCausalLM(nn.Module):
             ("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 get_tensor_model_parallel_rank() == 0:
+            weights = tqdm.tqdm(weights, total=int(len(params_dict) * 1.5))
+        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:

sglang/srt/models/llama_classification.py

@@ -0,0 +1,107 @@
+from typing import Iterable, Optional, Tuple
+
+import torch
+import tqdm
+from torch import nn
+from transformers import LlamaConfig
+from vllm.config import CacheConfig
+from vllm.distributed import get_tensor_model_parallel_rank
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.logits_processor import LogitProcessorOutput
+from sglang.srt.managers.controller.model_runner import InputMetadata
+from sglang.srt.models.llama2 import LlamaModel
+
+
+class LlamaForClassification(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = LlamaModel(config, quant_config=quant_config)
+
+        self.classification_head = nn.Linear(
+            config.hidden_size, config.classification_out_size
+        )
+        self.eos_token_id = config.eos_token_id
+
+    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)
+        is_eos_token = input_ids == self.eos_token_id
+        hidden_states = hidden_states[is_eos_token]
+        scores = self.classification_head(hidden_states)
+
+        if scores.shape[0] != input_metadata.batch_size:
+            print("Warning: the EOS tokens are missing in some sentences.")
+            scores = torch.ones(
+                (input_metadata.batch_size, self.config.classification_out_size)
+            ).to(input_ids.device)
+
+        return LogitProcessorOutput(
+            next_token_logits=scores,
+            next_token_logprobs=scores,
+            normalized_prompt_logprobs=scores,
+            prefill_token_logprobs=torch.ones_like(input_ids),
+            prefill_top_logprobs=None,
+            decode_top_logprobs=None,
+        )
+
+    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())
+        if get_tensor_model_parallel_rank() == 0:
+            weights = tqdm.tqdm(weights, total=int(len(params_dict) * 1.5))
+        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 "lm_head" 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
+                if name.startswith("model.vision_tower") 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
+                if name.startswith("model.vision_tower") 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 = LlamaForClassification

sglang/srt/models/llava.py

@@ -1,32 +1,40 @@
 """Inference-only LLaVa model compatible with HuggingFace weights."""
 
-from typing import List, Optional
+from typing import Iterable, List, Optional, Tuple
 
 import numpy as np
 import torch
-from sglang.srt.managers.router.infer_batch import ForwardMode
-from sglang.srt.managers.router.model_runner import InputMetadata
+from torch import nn
+from transformers import (
+    CLIPVisionConfig,
+    CLIPVisionModel,
+    LlavaConfig,
+    MistralConfig,
+    Qwen2Config,
+)
+from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
+from vllm.config import CacheConfig
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.managers.controller.infer_batch import ForwardMode
+from sglang.srt.managers.controller.model_runner import InputMetadata
 from sglang.srt.mm_utils import (
     get_anyres_image_grid_shape,
     unpad_image,
     unpad_image_shape,
 )
 from sglang.srt.models.llama2 import LlamaForCausalLM
-from torch import nn
-from transformers import CLIPVisionModel, LlamaConfig, LlavaConfig
-from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
-from vllm.model_executor.layers.linear import LinearMethodBase
-from vllm.model_executor.weight_utils import (
-    default_weight_loader,
-    hf_model_weights_iterator,
-)
+from sglang.srt.models.mistral import MistralForCausalLM
+from sglang.srt.models.qwen2 import Qwen2ForCausalLM
 
 
 class LlavaLlamaForCausalLM(nn.Module):
     def __init__(
         self,
         config: LlavaConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -34,7 +42,7 @@ class LlavaLlamaForCausalLM(nn.Module):
         self.config.vision_config.hidden_size = config.mm_hidden_size
         self.config.text_config.hidden_size = config.hidden_size
         self.multi_modal_projector = LlavaMultiModalProjector(config)
-        self.language_model = LlamaForCausalLM(config, linear_method)
+        self.language_model = LlamaForCausalLM(config, quant_config=quant_config)
         if "unpad" in getattr(config, "mm_patch_merge_type", ""):
             self.language_model.model.image_newline = nn.Parameter(
                 torch.empty(config.text_config.hidden_size, dtype=torch.float16)
@@ -235,13 +243,7 @@ class LlavaLlamaForCausalLM(nn.Module):
         elif input_metadata.forward_mode == ForwardMode.DECODE:
             return self.language_model(input_ids, positions, input_metadata)
 
-    def load_weights(
-        self,
-        model_name_or_path: str,
-        cache_dir: Optional[str] = None,
-        load_format: str = "auto",
-        revision: Optional[str] = None,
-    ):
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         # load clip vision model by cfg['mm_vision_tower']:
         #   huggingface_name or path_of_clip_relative_to_llava_model_dir
         vision_path = self.config.mm_vision_tower
@@ -274,9 +276,8 @@ class LlavaLlamaForCausalLM(nn.Module):
             "model.vision_tower.vision_tower": "vision_tower",  # Update the vision tower weights if we find them in the checkpoint (it may be finetuned).
         }
         params_dict = dict(self.named_parameters())
-        for name, loaded_weight in hf_model_weights_iterator(
-            model_name_or_path, cache_dir, load_format, revision
-        ):
+        weights = list(weights)
+        for name, loaded_weight in weights:
             # FIXME: why projector weights read two times?
             if "projector" in name or "vision_tower" in name:
                 for weight_name, param_name in projector_weights.items():
@@ -287,9 +288,7 @@ class LlavaLlamaForCausalLM(nn.Module):
                 weight_loader(param, loaded_weight)
 
         # load language model
-        self.language_model.load_weights(
-            model_name_or_path, cache_dir, load_format, revision
-        )
+        self.language_model.load_weights(weights)
 
         monkey_path_clip_vision_embed_forward()
 
@@ -298,6 +297,72 @@ class LlavaLlamaForCausalLM(nn.Module):
         return self.image_size // self.patch_size
 
 
+class LlavaQwenForCausalLM(LlavaLlamaForCausalLM):
+    def __init__(
+        self,
+        config: LlavaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
+    ) -> None:
+        super().__init__(config, quant_config=quant_config, cache_config=cache_config)
+        self.config = config
+        self.vision_tower = None
+        if getattr(self.config, "vision_config", None) is None:
+            self.config.vision_config = CLIPVisionConfig(self.config.mm_vision_tower)
+
+        if getattr(self.config, "text_config", None) is None:
+            self.config.text_config = Qwen2Config(self.config._name_or_path)
+
+        self.config.vision_config.hidden_size = config.mm_hidden_size
+        self.config.text_config.hidden_size = config.hidden_size
+
+        if getattr(self.config, "projector_hidden_act", None) is None:
+            self.config.projector_hidden_act = "gelu"
+
+        if getattr(self.config, "image_token_index", None) is None:
+            self.config.image_token_index = 151646
+
+        self.multi_modal_projector = LlavaMultiModalProjector(config)
+        self.language_model = Qwen2ForCausalLM(config, quant_config=quant_config)
+        if "unpad" in getattr(config, "mm_patch_merge_type", ""):
+            self.language_model.model.image_newline = nn.Parameter(
+                torch.empty(config.text_config.hidden_size, dtype=torch.float16)
+            )
+
+
+class LlavaMistralForCausalLM(LlavaLlamaForCausalLM):
+    def __init__(
+        self,
+        config: LlavaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
+    ) -> None:
+        super().__init__(config, quant_config=quant_config, cache_config=cache_config)
+        self.config = config
+        self.vision_tower = None
+        if getattr(self.config, "vision_config", None) is None:
+            self.config.vision_config = CLIPVisionConfig(self.config.mm_vision_tower)
+
+        if getattr(self.config, "text_config", None) is None:
+            self.config.text_config = MistralConfig(self.config._name_or_path)
+
+        self.config.vision_config.hidden_size = config.mm_hidden_size
+        self.config.text_config.hidden_size = config.hidden_size
+
+        if getattr(self.config, "projector_hidden_act", None) is None:
+            self.config.projector_hidden_act = "gelu"
+
+        if getattr(self.config, "image_token_index", None) is None:
+            self.config.image_token_index = 32000
+
+        self.multi_modal_projector = LlavaMultiModalProjector(config)
+        self.language_model = MistralForCausalLM(config, quant_config=quant_config)
+        if "unpad" in getattr(config, "mm_patch_merge_type", ""):
+            self.language_model.model.image_newline = nn.Parameter(
+                torch.empty(config.text_config.hidden_size, dtype=torch.float16)
+            )
+
+
 first_call = True
 
 
@@ -330,4 +395,4 @@ def monkey_path_clip_vision_embed_forward():
     )
 
 
-EntryClass = LlavaLlamaForCausalLM
+EntryClass = [LlavaLlamaForCausalLM, LlavaQwenForCausalLM, LlavaMistralForCausalLM]