sglang 0.1.14__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

@@ -7,34 +7,35 @@ from typing import Optional, Tuple
 import torch
 from torch import nn
 from transformers import PretrainedConfig
-
-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 vllm.model_executor.layers.activation import SiluAndMul
 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 (
-    VocabParallelEmbedding,
     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
+
 
 class StablelmMLP(nn.Module):
     def __init__(
-        self, config: PretrainedConfig, linear_method: Optional[LinearMethodBase] = None
+        self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -44,10 +45,10 @@ class StablelmMLP(nn.Module):
             config.hidden_size,
             [config.intermediate_size] * 2,
             bias=False,
-            linear_method=linear_method,
+            quant_config=quant_config,
         )
         self.down_proj = RowParallelLinear(
-            config.intermediate_size, config.hidden_size, bias=False
+            config.intermediate_size, config.hidden_size, bias=False, quant_config=quant_config,
         )
         self.act_fn = SiluAndMul()
 
@@ -63,7 +64,7 @@ class StablelmAttention(nn.Module):
         self,
         config: PretrainedConfig,
         layer_id: int = 0,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
@@ -105,13 +106,11 @@ class StablelmAttention(nn.Module):
             self.total_num_heads,
             self.total_num_key_value_heads,
             self.qkv_bias,
-            linear_method=linear_method,
         )
         self.o_proj = RowParallelLinear(
             self.total_num_heads * self.head_dim,
             self.hidden_size,
             bias=False,
-            linear_method=linear_method,
         )
         self.rotary_emb = get_rope(
             self.head_dim,
@@ -146,11 +145,11 @@ class StablelmDecoderLayer(nn.Module):
         self,
         config: PretrainedConfig,
         layer_id: int = 0,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.self_attn = StablelmAttention(config, layer_id=layer_id)
-        self.mlp = StablelmMLP(config, linear_method)
+        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)
@@ -182,7 +181,7 @@ class StablelmDecoderLayer(nn.Module):
 
 class StableLMEpochModel(nn.Module):
     def __init__(
-        self, config: PretrainedConfig, linear_method: Optional[LinearMethodBase] = None
+        self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.embed_tokens = VocabParallelEmbedding(
@@ -191,7 +190,7 @@ class StableLMEpochModel(nn.Module):
         )
         self.layers = nn.ModuleList(
             [
-                StablelmDecoderLayer(config, i, linear_method)
+                StablelmDecoderLayer(config, i, quant_config=quant_config)
                 for i in range(config.num_hidden_layers)
             ]
         )
@@ -224,12 +223,12 @@ class StableLmForCausalLM(nn.Module):
     def __init__(
         self,
         config: PretrainedConfig,
-        linear_method: Optional[LinearMethodBase] = None,
+        quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
         self.config = config
-        self.linear_method = linear_method
-        self.model = StableLMEpochModel(config, linear_method)
+        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)
 

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):

sglang/srt/openai_api_adapter.py

@@ -0,0 +1,356 @@
+"""Conversion between OpenAI APIs and native SRT APIs"""
+import json
+import os
+
+from fastapi import HTTPException, Request
+from fastapi.responses import StreamingResponse
+
+from sglang.srt.conversation import (
+    Conversation,
+    SeparatorStyle,
+    chat_template_exists,
+    generate_chat_conv,
+    register_conv_template,
+)
+from sglang.srt.managers.io_struct import GenerateReqInput
+from sglang.srt.openai_protocol import (
+    ChatCompletionRequest,
+    ChatCompletionResponse,
+    ChatCompletionResponseChoice,
+    ChatCompletionResponseStreamChoice,
+    ChatCompletionStreamResponse,
+    ChatMessage,
+    CompletionRequest,
+    CompletionResponse,
+    CompletionResponseChoice,
+    CompletionResponseStreamChoice,
+    CompletionStreamResponse,
+    DeltaMessage,
+    LogProbs,
+    UsageInfo,
+)
+from sglang.srt.utils import jsonify_pydantic_model
+
+
+chat_template_name = None
+
+def load_chat_template_for_openai_api(chat_template_arg):
+    global chat_template_name
+
+    print(f"Use chat template: {chat_template_arg}")
+    if not chat_template_exists(chat_template_arg):
+        if not os.path.exists(chat_template_arg):
+            raise RuntimeError(
+                f"Chat template {chat_template_arg} is not a built-in template name "
+                "or a valid chat template file path."
+            )
+        with open(chat_template_arg, "r") as filep:
+            template = json.load(filep)
+            try:
+                sep_style = SeparatorStyle[template["sep_style"]]
+            except KeyError:
+                raise ValueError(
+                    f"Unknown separator style: {template['sep_style']}"
+                ) from None
+            register_conv_template(
+                Conversation(
+                    name=template["name"],
+                    system_template=template["system"] + "\n{system_message}",
+                    system_message=template.get("system_message", ""),
+                    roles=(template["user"], template["assistant"]),
+                    sep_style=sep_style,
+                    sep=template.get("sep", "\n"),
+                    stop_str=template["stop_str"],
+                ),
+                override=True,
+            )
+        chat_template_name = template["name"]
+    else:
+        chat_template_name = chat_template_arg
+
+
+async def v1_completions(tokenizer_manager, raw_request: Request):
+    request_json = await raw_request.json()
+    request = CompletionRequest(**request_json)
+
+    # TODO: Validate the request and return HTTPStatus.BAD_REQUEST if invalid.
+    assert request.n == 1
+
+    adapted_request = GenerateReqInput(
+        text=request.prompt,
+        sampling_params={
+            "temperature": request.temperature,
+            "max_new_tokens": request.max_tokens,
+            "stop": request.stop,
+            "top_p": request.top_p,
+            "presence_penalty": request.presence_penalty,
+            "frequency_penalty": request.frequency_penalty,
+            "regex": request.regex,
+        },
+        return_logprob=request.logprobs is not None and request.logprobs > 0,
+        top_logprobs_num=request.logprobs if request.logprobs is not None else 0,
+        return_text_in_logprobs=True,
+        stream=request.stream,
+    )
+    adapted_request.post_init()
+
+    if adapted_request.stream:
+
+        async def generate_stream_resp():
+            stream_buffer = ""
+            n_prev_token = 0
+            async for content in tokenizer_manager.generate_request(adapted_request):
+                text = content["text"]
+                prompt_tokens = content["meta_info"]["prompt_tokens"]
+                completion_tokens = content["meta_info"]["completion_tokens"]
+
+                if not stream_buffer:  # The first chunk
+                    if request.echo:
+                        # Prepend prompt in response text.
+                        text = request.prompt + text
+
+                if request.logprobs:
+                    # The first chunk and echo is enabled.
+                    if not stream_buffer and request.echo:
+                        prefill_token_logprobs = content["meta_info"][
+                            "prefill_token_logprobs"
+                        ]
+                        prefill_top_logprobs = content["meta_info"][
+                            "prefill_top_logprobs"
+                        ]
+                    else:
+                        prefill_token_logprobs = None
+                        prefill_top_logprobs = None
+
+                    logprobs = to_openai_style_logprobs(
+                        prefill_token_logprobs=prefill_token_logprobs,
+                        prefill_top_logprobs=prefill_top_logprobs,
+                        decode_token_logprobs=content["meta_info"][
+                            "decode_token_logprobs"
+                        ][n_prev_token:],
+                        decode_top_logprobs=content["meta_info"]["decode_top_logprobs"][
+                            n_prev_token:
+                        ],
+                    )
+
+                    n_prev_token = len(content["meta_info"]["decode_token_logprobs"])
+                else:
+                    logprobs = None
+
+                delta = text[len(stream_buffer) :]
+                stream_buffer = content["text"]
+                choice_data = CompletionResponseStreamChoice(
+                    index=0,
+                    text=delta,
+                    logprobs=logprobs,
+                    finish_reason=None,
+                )
+                chunk = CompletionStreamResponse(
+                    id=content["meta_info"]["id"],
+                    object="text_completion",
+                    choices=[choice_data],
+                    model=request.model,
+                    usage=UsageInfo(
+                        prompt_tokens=prompt_tokens,
+                        completion_tokens=completion_tokens,
+                        total_tokens=prompt_tokens + completion_tokens,
+                    ),
+                )
+                yield f"data: {jsonify_pydantic_model(chunk)}\n\n"
+            yield "data: [DONE]\n\n"
+
+        return StreamingResponse(generate_stream_resp(), media_type="text/event-stream")
+
+    # Non-streaming response.
+    ret = await tokenizer_manager.generate_request(adapted_request).__anext__()
+    ret = ret[0] if isinstance(ret, list) else ret
+
+    prompt_tokens = ret["meta_info"]["prompt_tokens"]
+    completion_tokens = ret["meta_info"]["completion_tokens"]
+    text = ret["text"]
+    if request.echo:
+        text = request.prompt + text
+
+    if request.logprobs:
+        if request.echo:
+            prefill_token_logprobs = ret["meta_info"]["prefill_token_logprobs"]
+            prefill_top_logprobs = ret["meta_info"]["prefill_top_logprobs"]
+        else:
+            prefill_token_logprobs = None
+            prefill_top_logprobs = None
+
+        logprobs = to_openai_style_logprobs(
+            prefill_token_logprobs=prefill_token_logprobs,
+            prefill_top_logprobs=prefill_top_logprobs,
+            decode_token_logprobs=ret["meta_info"]["decode_token_logprobs"],
+            decode_top_logprobs=ret["meta_info"]["decode_top_logprobs"],
+        )
+    else:
+        logprobs = None
+
+    choice_data = CompletionResponseChoice(
+        index=0,
+        text=text,
+        logprobs=logprobs,
+        finish_reason=None,  # TODO(comaniac): Add finish reason.
+    )
+    response = CompletionResponse(
+        id=ret["meta_info"]["id"],
+        model=request.model,
+        choices=[choice_data],
+        usage=UsageInfo(
+            prompt_tokens=prompt_tokens,
+            completion_tokens=completion_tokens,
+            total_tokens=prompt_tokens + completion_tokens,
+        ),
+    )
+    return response
+
+
+async def v1_chat_completions(tokenizer_manager, raw_request: Request):
+    request_json = await raw_request.json()
+    request = ChatCompletionRequest(**request_json)
+
+    # TODO: Validate the request and return HTTPStatus.BAD_REQUEST if invalid.
+    assert request.n == 1
+
+    # Prep the data needed for the underlying GenerateReqInput:
+    #  - prompt: The full prompt string.
+    #  - stop: Custom stop tokens.
+    #  - image_data: None or a list of image strings (URLs or base64 strings).
+    #    None skips any image processing in GenerateReqInput.
+    if not isinstance(request.messages, str):
+        # Apply chat template and its stop strings.
+        if chat_template_name is None:
+            prompt = tokenizer_manager.tokenizer.apply_chat_template(
+                request.messages, tokenize=False, add_generation_prompt=True
+            )
+            stop = request.stop
+            image_data = None
+        else:
+            conv = generate_chat_conv(request, chat_template_name)
+            prompt = conv.get_prompt()
+            image_data = conv.image_data
+            stop = conv.stop_str or []
+            if request.stop:
+                if isinstance(request.stop, str):
+                    stop.append(request.stop)
+                else:
+                    stop.extend(request.stop)
+    else:
+        # Use the raw prompt and stop strings if the messages is already a string.
+        prompt = request.messages
+        stop = request.stop
+        image_data = None
+
+    adapted_request = GenerateReqInput(
+        text=prompt,
+        image_data=image_data,
+        sampling_params={
+            "temperature": request.temperature,
+            "max_new_tokens": request.max_tokens,
+            "stop": stop,
+            "top_p": request.top_p,
+            "presence_penalty": request.presence_penalty,
+            "frequency_penalty": request.frequency_penalty,
+            "regex": request.regex,
+        },
+        stream=request.stream,
+    )
+    adapted_request.post_init()
+
+    if adapted_request.stream:
+
+        async def generate_stream_resp():
+            is_first = True
+
+            stream_buffer = ""
+            async for content in tokenizer_manager.generate_request(adapted_request):
+                if is_first:
+                    # First chunk with role
+                    is_first = False
+                    choice_data = ChatCompletionResponseStreamChoice(
+                        index=0,
+                        delta=DeltaMessage(role="assistant"),
+                        finish_reason=None,
+                    )
+                    chunk = ChatCompletionStreamResponse(
+                        id=content["meta_info"]["id"],
+                        choices=[choice_data],
+                        model=request.model,
+                    )
+                    yield f"data: {jsonify_pydantic_model(chunk)}\n\n"
+
+                text = content["text"]
+                delta = text[len(stream_buffer) :]
+                stream_buffer = text
+                choice_data = ChatCompletionResponseStreamChoice(
+                    index=0, delta=DeltaMessage(content=delta), finish_reason=None
+                )
+                chunk = ChatCompletionStreamResponse(
+                    id=content["meta_info"]["id"],
+                    choices=[choice_data],
+                    model=request.model,
+                )
+                yield f"data: {jsonify_pydantic_model(chunk)}\n\n"
+            yield "data: [DONE]\n\n"
+
+        return StreamingResponse(generate_stream_resp(), media_type="text/event-stream")
+
+    # Non-streaming response.
+    ret = await tokenizer_manager.generate_request(adapted_request).__anext__()
+    prompt_tokens = ret["meta_info"]["prompt_tokens"]
+    completion_tokens = ret["meta_info"]["completion_tokens"]
+    choice_data = ChatCompletionResponseChoice(
+        index=0,
+        message=ChatMessage(role="assistant", content=ret["text"]),
+        finish_reason=None,  # TODO(comaniac): Add finish reason.
+    )
+    response = ChatCompletionResponse(
+        id=ret["meta_info"]["id"],
+        model=request.model,
+        choices=[choice_data],
+        usage=UsageInfo(
+            prompt_tokens=prompt_tokens,
+            completion_tokens=completion_tokens,
+            total_tokens=prompt_tokens + completion_tokens,
+        ),
+    )
+    return response
+
+
+def to_openai_style_logprobs(
+    prefill_token_logprobs=None,
+    decode_token_logprobs=None,
+    prefill_top_logprobs=None,
+    decode_top_logprobs=None,
+):
+    ret_logprobs = LogProbs()
+
+    def append_token_logprobs(token_logprobs):
+        for logprob, _, token_text in token_logprobs:
+            ret_logprobs.tokens.append(token_text)
+            ret_logprobs.token_logprobs.append(logprob)
+
+            # Not Supported yet
+            ret_logprobs.text_offset.append(-1)
+
+    def append_top_logprobs(top_logprobs):
+        for tokens in top_logprobs:
+            if tokens is not None:
+                ret_logprobs.top_logprobs.append(
+                    {token[2]: token[0] for token in tokens}
+                )
+            else:
+                ret_logprobs.top_logprobs.append(None)
+
+    if prefill_token_logprobs is not None:
+        append_token_logprobs(prefill_token_logprobs)
+    if decode_token_logprobs is not None:
+        append_token_logprobs(decode_token_logprobs)
+    if prefill_top_logprobs is not None:
+        append_top_logprobs(prefill_top_logprobs)
+    if decode_top_logprobs is not None:
+        append_top_logprobs(decode_top_logprobs)
+
+    return ret_logprobs