sglang 0.3.0__py3-none-any.whl → 0.3.1__py3-none-any.whl

sglang/srt/models/xverse_moe.py

@@ -0,0 +1,445 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+"""Inference-only XVERSE MoE model."""
+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_rank,
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_reduce,
+)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import fused_moe
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    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.model_executor.forward_batch_info import InputMetadata
+
+
+class XverseMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        reduce_results: bool = True,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            reduce_results=reduce_results,
+        )
+        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.down_proj(x)
+        return x
+
+
+class XverseMoE(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.rank = get_tensor_model_parallel_rank()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.n_routed_experts = config.num_experts
+        self.top_k = config.moe_top_k
+        if self.tp_size > self.n_routed_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {self.n_routed_experts}."
+            )
+
+        self.experts = nn.ModuleList(
+            [
+                XverseMLP(
+                    hidden_size=config.hidden_size,
+                    intermediate_size=config.intermediate_size,
+                    hidden_act=config.hidden_act,
+                    quant_config=quant_config,
+                    reduce_results=False,
+                )
+                for _ in range(self.n_routed_experts)
+            ]
+        )
+        self.pack_params()
+
+        self.router = ReplicatedLinear(
+            config.hidden_size, self.n_routed_experts, bias=False, quant_config=None
+        )
+
+        if config.num_shared_experts is not None:
+            intermediate_size = config.intermediate_size * config.num_shared_experts
+            self.shared_experts = XverseMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+            )
+
+    def pack_params(self):
+        w1 = []
+        w2 = []
+        for expert in self.experts:
+            w1.append(expert.gate_up_proj.weight)
+            w2.append(expert.down_proj.weight)
+        self.w1 = torch._utils._flatten_dense_tensors(w1)
+        w1s = torch._utils._unflatten_dense_tensors(self.w1, w1)
+        for data, param in zip(w1s, w1):
+            param.data = data
+        self.w1 = self.w1.view(len(w1), *w1s[0].shape)
+
+        self.w2 = torch._utils._flatten_dense_tensors(w2)
+        w2s = torch._utils._unflatten_dense_tensors(self.w2, w2)
+        for data, param in zip(w2s, w2):
+            param.data = data
+
+        self.w2 = self.w2.view(len(w2), *w2s[0].shape)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        if self.config.num_shared_experts is not None:
+            shared_output = self.shared_experts(hidden_states)
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.router(hidden_states)
+        final_hidden_states = fused_moe(
+            hidden_states,
+            self.w1,
+            self.w2,
+            router_logits,
+            self.top_k,
+            renormalize=getattr(self.config, "norm_topk_prob", False),
+            inplace=True,
+        )
+
+        if self.config.num_shared_experts is not None:
+            final_hidden_states = final_hidden_states + shared_output
+        final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
+
+        return final_hidden_states.view(num_tokens, hidden_dim)
+
+
+class XverseAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        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.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        self.o_proj = 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,
+            num_kv_heads=self.num_kv_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 XverseDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int,
+        cache_config: Optional[CacheConfig] = 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)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        num_key_value_heads = getattr(
+            config, "num_key_value_heads", config.num_attention_heads
+        )
+        self.self_attn = XverseAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+        )
+        if config.num_experts is not None:
+            self.mlp = XverseMoE(config=config, quant_config=quant_config)
+        else:
+            self.mlp = XverseMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+            )
+        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,
+        input_metadata: InputMetadata,
+        residual: Optional[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,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class XverseModel(nn.Module):
+
+    fall_back_to_pt_during_load = False
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                XverseDecoderLayer(
+                    config, layer_id, cache_config, quant_config=quant_config
+                )
+                for layer_id 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,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        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 XverseMoeForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = XverseModel(config, cache_config, quant_config)
+        self.lm_head = ParallelLMHead(
+            config.vocab_size, config.hidden_size, quant_config=quant_config
+        )
+        self.logits_processor = LogitsProcessor(config)
+
+        self.param_dict = dict(self.named_parameters())
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    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 = self.param_dict
+
+        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
+                # Skip experts that are not assigned to this worker.
+                if (
+                    "mlp.experts." in name or "mlp.shared_experts." in name
+                ) 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
+                # Skip experts that are not assigned to this worker.
+                if (
+                    "mlp.experts." in name or "mlp.shared_experts." in name
+                ) 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 = XverseMoeForCausalLM

sglang/srt/openai_api/adapter.py

@@ -22,12 +22,19 @@ import os
 import time
 import uuid
 from http import HTTPStatus
-from typing import Dict, List, Optional
+from typing import Dict, List
 
 from fastapi import HTTPException, Request, UploadFile
 from fastapi.responses import JSONResponse, StreamingResponse
 from pydantic import ValidationError
 
+try:
+    from outlines.fsm.json_schema import convert_json_schema_to_str
+except ImportError:
+    # Before outlines 0.0.47, convert_json_schema_to_str is under
+    # outlines.integrations.utils
+    from outlines.integrations.utils import convert_json_schema_to_str
+
 from sglang.srt.conversation import (
     Conversation,
     SeparatorStyle,
@@ -88,19 +95,6 @@ file_id_storage: Dict[str, str] = {}
 storage_dir = None
 
 
-def format_finish_reason(finish_reason) -> Optional[str]:
-    if finish_reason.startswith("None"):
-        return None
-    elif finish_reason.startswith("FINISH_MATCHED"):
-        return "stop"
-    elif finish_reason.startswith("FINISH_LENGTH"):
-        return "length"
-    elif finish_reason.startswith("FINISH_ABORT"):
-        return "abort"
-    else:
-        return "unknown"
-
-
 def create_error_response(
     message: str,
     err_type: str = "BadRequestError",
@@ -478,7 +472,7 @@ def v1_generate_request(
     first_prompt_type = type(all_requests[0].prompt)
     for request in all_requests:
         assert (
-            type(request.prompt) == first_prompt_type
+            type(request.prompt) is first_prompt_type
         ), "All prompts must be of the same type in file input settings"
         if len(all_requests) > 1 and request.n > 1:
             raise ValueError(
@@ -611,8 +605,10 @@ def v1_generate_response(request, ret, tokenizer_manager, to_file=False):
                 "index": 0,
                 "text": text,
                 "logprobs": logprobs,
-                "finish_reason": format_finish_reason(
-                    ret_item["meta_info"]["finish_reason"]
+                "finish_reason": (
+                    ret_item["meta_info"]["finish_reason"]["type"]
+                    if ret_item["meta_info"]["finish_reason"]
+                    else ""
                 ),
             }
         else:
@@ -620,8 +616,10 @@ def v1_generate_response(request, ret, tokenizer_manager, to_file=False):
                 index=idx,
                 text=text,
                 logprobs=logprobs,
-                finish_reason=format_finish_reason(
-                    ret_item["meta_info"]["finish_reason"]
+                finish_reason=(
+                    ret_item["meta_info"]["finish_reason"]["type"]
+                    if ret_item["meta_info"]["finish_reason"]
+                    else ""
                 ),
             )
 
@@ -755,8 +753,10 @@ async def v1_completions(tokenizer_manager, raw_request: Request):
                         index=index,
                         text=delta,
                         logprobs=logprobs,
-                        finish_reason=format_finish_reason(
-                            content["meta_info"]["finish_reason"]
+                        finish_reason=(
+                            content["meta_info"]["finish_reason"]["type"]
+                            if content["meta_info"]["finish_reason"]
+                            else ""
                         ),
                     )
                     chunk = CompletionStreamResponse(
@@ -832,6 +832,7 @@ def v1_chat_generate_request(
     return_logprobs = []
     logprob_start_lens = []
     top_logprobs_nums = []
+    modalities_list = []
 
     # NOTE: with openai API, the prompt's logprobs are always not computed
 
@@ -864,10 +865,12 @@ def v1_chat_generate_request(
                 )
                 stop = request.stop
                 image_data = None
+                modalities = []
             else:
                 conv = generate_chat_conv(request, chat_template_name)
                 prompt = conv.get_prompt()
                 image_data = conv.image_data
+                modalities = conv.modalities
                 stop = conv.stop_str or []
                 if request.stop:
                     if isinstance(request.stop, str):
@@ -880,27 +883,33 @@ def v1_chat_generate_request(
             prompt_ids = request.messages
             stop = request.stop
             image_data = None
+            modalities = []
         input_ids.append(prompt_ids)
         return_logprobs.append(request.logprobs)
         logprob_start_lens.append(-1)
-        top_logprobs_nums.append(request.top_logprobs)
-        sampling_params_list.append(
-            {
-                "temperature": request.temperature,
-                "max_new_tokens": request.max_tokens,
-                "min_new_tokens": request.min_tokens,
-                "stop": stop,
-                "stop_token_ids": request.stop_token_ids,
-                "top_p": request.top_p,
-                "presence_penalty": request.presence_penalty,
-                "frequency_penalty": request.frequency_penalty,
-                "repetition_penalty": request.repetition_penalty,
-                "regex": request.regex,
-                "json_schema": request.json_schema,
-                "n": request.n,
-            }
-        )
+        top_logprobs_nums.append(request.top_logprobs or 0)
+
+        sampling_params = {
+            "temperature": request.temperature,
+            "max_new_tokens": request.max_tokens,
+            "min_new_tokens": request.min_tokens,
+            "stop": stop,
+            "stop_token_ids": request.stop_token_ids,
+            "top_p": request.top_p,
+            "presence_penalty": request.presence_penalty,
+            "frequency_penalty": request.frequency_penalty,
+            "repetition_penalty": request.repetition_penalty,
+            "regex": request.regex,
+            "n": request.n,
+        }
+        if request.response_format and request.response_format.type == "json_schema":
+            sampling_params["json_schema"] = convert_json_schema_to_str(
+                request.response_format.json_schema.schema_
+            )
+        sampling_params_list.append(sampling_params)
+
         image_data_list.append(image_data)
+        modalities_list.extend(modalities)
     if len(all_requests) == 1:
         input_ids = input_ids[0]
         if isinstance(input_ids, str):
@@ -912,6 +921,7 @@ def v1_chat_generate_request(
         return_logprobs = return_logprobs[0]
         logprob_start_lens = logprob_start_lens[0]
         top_logprobs_nums = top_logprobs_nums[0]
+        modalities_list = modalities_list[:1]
     else:
         if isinstance(input_ids[0], str):
             prompt_kwargs = {"text": input_ids}
@@ -928,6 +938,7 @@ def v1_chat_generate_request(
         stream=all_requests[0].stream,
         return_text_in_logprobs=True,
         rid=request_ids,
+        modalities=modalities_list,
     )
     if len(all_requests) == 1:
         return adapted_request, all_requests[0]
@@ -981,8 +992,10 @@ def v1_chat_generate_response(request, ret, to_file=False):
                 "index": 0,
                 "message": {"role": "assistant", "content": ret_item["text"]},
                 "logprobs": choice_logprobs,
-                "finish_reason": format_finish_reason(
-                    ret_item["meta_info"]["finish_reason"]
+                "finish_reason": (
+                    ret_item["meta_info"]["finish_reason"]["type"]
+                    if ret_item["meta_info"]["finish_reason"]
+                    else ""
                 ),
             }
         else:
@@ -990,8 +1003,10 @@ def v1_chat_generate_response(request, ret, to_file=False):
                 index=idx,
                 message=ChatMessage(role="assistant", content=ret_item["text"]),
                 logprobs=choice_logprobs,
-                finish_reason=format_finish_reason(
-                    ret_item["meta_info"]["finish_reason"]
+                finish_reason=(
+                    ret_item["meta_info"]["finish_reason"]["type"]
+                    if ret_item["meta_info"]["finish_reason"]
+                    else ""
                 ),
             )
 
@@ -1116,8 +1131,10 @@ async def v1_chat_completions(tokenizer_manager, raw_request: Request):
                         choice_data = ChatCompletionResponseStreamChoice(
                             index=index,
                             delta=DeltaMessage(role="assistant"),
-                            finish_reason=format_finish_reason(
-                                content["meta_info"]["finish_reason"]
+                            finish_reason=(
+                                content["meta_info"]["finish_reason"]["type"]
+                                if content["meta_info"]["finish_reason"]
+                                else ""
                             ),
                             logprobs=choice_logprobs,
                         )
@@ -1134,8 +1151,10 @@ async def v1_chat_completions(tokenizer_manager, raw_request: Request):
                     choice_data = ChatCompletionResponseStreamChoice(
                         index=index,
                         delta=DeltaMessage(content=delta),
-                        finish_reason=format_finish_reason(
-                            content["meta_info"]["finish_reason"]
+                        finish_reason=(
+                            content["meta_info"]["finish_reason"]["type"]
+                            if content["meta_info"]["finish_reason"]
+                            else ""
                         ),
                         logprobs=choice_logprobs,
                     )