Optimize mem indices mangement (#619)

benchmark/latency_throughput/bench_one.py

@@ -17,7 +17,8 @@ def run_one_batch_size(bs):
 
     if args.input_len:
         input_ids = [
-            [int(x) for x in np.random.randint(0, high=16384, size=(args.input_len,))] for _ in range(bs)
+            [int(x) for x in np.random.randint(0, high=16384, size=(args.input_len,))]
+            for _ in range(bs)
         ]
     else:
         text = [f"{i, }" for i in range(bs)]
@@ -116,9 +117,11 @@ if __name__ == "__main__":
     parser.add_argument("--port", type=int, default=None)
     parser.add_argument("--backend", type=str, default="srt")
     parser.add_argument("--input-len", type=int, default=None)
-    parser.add_argument("--batch-size", type=int, nargs='*', default=[1])
+    parser.add_argument("--batch-size", type=int, nargs="*", default=[1])
     parser.add_argument("--max-tokens", type=int, default=256)
-    parser.add_argument("--vllm-model-name", type=str, default="meta-llama/Meta-Llama-3-70B")
+    parser.add_argument(
+        "--vllm-model-name", type=str, default="meta-llama/Meta-Llama-3-70B"
+    )
     args = parser.parse_args()
 
     if args.port is None:

python/sglang/backend/runtime_endpoint.py

@@ -12,7 +12,6 @@ from sglang.utils import http_request
 
 
 class RuntimeEndpoint(BaseBackend):
-
     def __init__(
         self,
         base_url: str,
@@ -38,7 +37,8 @@ class RuntimeEndpoint(BaseBackend):
         self.model_info = res.json()
 
         self.chat_template = get_chat_template_by_model_path(
-            self.model_info["model_path"])
+            self.model_info["model_path"]
+        )
 
     def get_model_name(self):
         return self.model_info["model_path"]
@@ -124,7 +124,12 @@ class RuntimeEndpoint(BaseBackend):
         else:
             raise RuntimeError(f"Invalid dtype: {sampling_params.dtype}")
 
-        for item in ["return_logprob", "logprob_start_len", "top_logprobs_num", "return_text_in_logprobs"]:
+        for item in [
+            "return_logprob",
+            "logprob_start_len",
+            "top_logprobs_num",
+            "return_text_in_logprobs",
+        ]:
             value = getattr(sampling_params, item, None)
             if value is not None:
                 data[item] = value
@@ -171,7 +176,12 @@ class RuntimeEndpoint(BaseBackend):
         else:
             raise RuntimeError(f"Invalid dtype: {sampling_params.dtype}")
 
-        for item in ["return_logprob", "logprob_start_len", "top_logprobs_num", "return_text_in_logprobs"]:
+        for item in [
+            "return_logprob",
+            "logprob_start_len",
+            "top_logprobs_num",
+            "return_text_in_logprobs",
+        ]:
             value = getattr(sampling_params, item, None)
             if value is not None:
                 data[item] = value

python/sglang/bench_latency.py

@@ -32,7 +32,6 @@ import logging
 import multiprocessing
 import time
 
-
 import numpy as np
 import torch
 import torch.distributed as dist

python/sglang/global_config.py

@@ -44,4 +44,5 @@ class GlobalConfig:
         # adjust_cache: Adjust the position embedding of KV cache.
         self.concate_and_append_mode = "no_adjust"
 
+
 global_config = GlobalConfig()

python/sglang/lang/chat_template.py

@@ -84,7 +84,7 @@ register_chat_template(
             "system": ("SYSTEM:", "\n"),
             "user": ("USER:", "\n"),
             "assistant": ("ASSISTANT:", "\n"),
-        }
+        },
     )
 )
 
@@ -177,7 +177,7 @@ register_chat_template(
             "assistant": ("", "<|im_end|>\n"),
         },
         style=ChatTemplateStyle.PLAIN,
-        stop_str=("<|im_end|>",)
+        stop_str=("<|im_end|>",),
     )
 )
 

python/sglang/lang/ir.py

@@ -24,9 +24,9 @@ class SglSamplingParams:
     presence_penalty: float = 0.0
     ignore_eos: bool = False
     return_logprob: Optional[bool] = None
-    logprob_start_len: Optional[int] = None,
-    top_logprobs_num: Optional[int] = None,
-    return_text_in_logprobs: Optional[bool] = None,
+    logprob_start_len: Optional[int] = (None,)
+    top_logprobs_num: Optional[int] = (None,)
+    return_text_in_logprobs: Optional[bool] = (None,)
 
     # for constrained generation, not included in to_xxx_kwargs
     dtype: Optional[str] = None

python/sglang/srt/managers/controller/cuda_graph_runner.py

@@ -8,7 +8,10 @@ from vllm.distributed.parallel_state import graph_capture
 from sglang.global_config import global_config
 from sglang.srt.layers.logits_processor import LogitProcessorOutput
 from sglang.srt.managers.controller.infer_batch import (
-    Batch, ForwardMode, InputMetadata, init_flashinfer_args
+    Batch,
+    ForwardMode,
+    InputMetadata,
+    init_flashinfer_args,
 )
 
 
@@ -24,18 +27,28 @@ class CudaGraphRunner:
         # Common inputs
         self.max_bs = max_batch_size_to_capture
         self.input_ids = torch.zeros((self.max_bs,), dtype=torch.int32, device="cuda")
-        self.req_pool_indices = torch.zeros((self.max_bs,), dtype=torch.int32, device="cuda")
+        self.req_pool_indices = torch.zeros(
+            (self.max_bs,), dtype=torch.int32, device="cuda"
+        )
         self.seq_lens = torch.ones((self.max_bs,), dtype=torch.int32, device="cuda")
-        self.position_ids_offsets = torch.zeros((self.max_bs,), dtype=torch.int32, device="cuda")
-        self.out_cache_loc = torch.zeros((self.max_bs,), dtype=torch.int32, device="cuda")
+        self.position_ids_offsets = torch.zeros(
+            (self.max_bs,), dtype=torch.int32, device="cuda"
+        )
+        self.out_cache_loc = torch.zeros(
+            (self.max_bs,), dtype=torch.int32, device="cuda"
+        )
 
         # FlashInfer inputs
-        self.flashinfer_workspace_buffer = self.model_runner.flashinfer_workspace_buffers[0]
+        self.flashinfer_workspace_buffer = (
+            self.model_runner.flashinfer_workspace_buffers[0]
+        )
         self.flashinfer_kv_indptr = torch.zeros(
             (self.max_bs + 1,), dtype=torch.int32, device="cuda"
         )
         self.flashinfer_kv_indices = torch.zeros(
-            (self.max_bs * model_runner.model_config.context_len,), dtype=torch.int32, device="cuda"
+            (self.max_bs * model_runner.model_config.context_len,),
+            dtype=torch.int32,
+            device="cuda",
         )
         self.flashinfer_kv_last_page_len = torch.ones(
             (self.max_bs,), dtype=torch.int32, device="cuda"
@@ -49,7 +62,12 @@ class CudaGraphRunner:
         with graph_capture() as graph_capture_context:
             self.stream = graph_capture_context.stream
             for bs in batch_size_list:
-                graph, input_buffers, output_buffers, flashinfer_handler = self.capture_one_batch_size(bs)
+                (
+                    graph,
+                    input_buffers,
+                    output_buffers,
+                    flashinfer_handler,
+                ) = self.capture_one_batch_size(bs)
                 self.graphs[bs] = graph
                 self.input_buffers[bs] = input_buffers
                 self.output_buffers[bs] = output_buffers
@@ -71,17 +89,19 @@ class CudaGraphRunner:
 
         # FlashInfer inputs
         if not _grouped_size_compiled_for_decode_kernels(
-            self.model_runner.model_config.num_attention_heads // self.model_runner.tp_size,
+            self.model_runner.model_config.num_attention_heads
+            // self.model_runner.tp_size,
             self.model_runner.model_config.get_num_kv_heads(self.model_runner.tp_size),
         ):
             use_tensor_cores = True
         else:
             use_tensor_cores = False
         flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
-            self.flashinfer_workspace_buffer, "NHD",
+            self.flashinfer_workspace_buffer,
+            "NHD",
             use_cuda_graph=True,
             use_tensor_cores=use_tensor_cores,
-            paged_kv_indptr_buffer=self.flashinfer_kv_indptr[:bs+1],
+            paged_kv_indptr_buffer=self.flashinfer_kv_indptr[: bs + 1],
             paged_kv_indices_buffer=self.flashinfer_kv_indices,
             paged_kv_last_page_len_buffer=self.flashinfer_kv_last_page_len[:bs],
         )
@@ -163,10 +183,14 @@ class CudaGraphRunner:
         else:
             output = LogitProcessorOutput(
                 next_token_logits=output.next_token_logits[:raw_bs],
-                next_token_logprobs=output.next_token_logprobs[:raw_bs] if output.next_token_logprobs is not None else None,
+                next_token_logprobs=output.next_token_logprobs[:raw_bs]
+                if output.next_token_logprobs is not None
+                else None,
                 normalized_prompt_logprobs=None,
                 prefill_token_logprobs=None,
                 prefill_top_logprobs=None,
-                decode_top_logprobs=output.decode_top_logprobs[:raw_bs] if output.decode_top_logprobs is not None else None,
+                decode_top_logprobs=output.decode_top_logprobs[:raw_bs]
+                if output.decode_top_logprobs is not None
+                else None,
             )
         return output

python/sglang/srt/managers/controller/infer_batch.py

@@ -668,7 +668,9 @@ class Batch:
             sampled_index = torch.multinomial(probs_sort, num_samples=1)
         except RuntimeError as e:
             warnings.warn(f"Ignore errors in sampling: {e}")
-            sampled_index = torch.ones(probs_sort.shape[:-1] + (1,), dtype=torch.int64, device=probs.device)
+            sampled_index = torch.ones(
+                probs_sort.shape[:-1] + (1,), dtype=torch.int64, device=probs.device
+            )
         batch_next_token_ids = torch.gather(probs_idx, dim=1, index=sampled_index).view(
             -1
         )
@@ -749,8 +751,14 @@ class InputMetadata:
         skip_flashinfer_init=False,
     ):
         if not skip_flashinfer_init and not model_runner.server_args.disable_flashinfer:
-            init_flashinfer_args(forward_mode, model_runner, req_pool_indices, seq_lens, prefix_lens,
-                                 model_runner.flashinfer_decode_wrapper)
+            init_flashinfer_args(
+                forward_mode,
+                model_runner,
+                req_pool_indices,
+                seq_lens,
+                prefix_lens,
+                model_runner.flashinfer_decode_wrapper,
+            )
 
         batch_size = len(req_pool_indices)
 
@@ -807,16 +815,24 @@ class InputMetadata:
         )
 
         if model_runner.server_args.disable_flashinfer:
-            (ret.triton_max_seq_len,
-             ret.triton_max_extend_len,
-             ret.triton_start_loc,
-             ret.triton_prefix_lens) = init_triton_args(forward_mode, seq_lens, prefix_lens)
+            (
+                ret.triton_max_seq_len,
+                ret.triton_max_extend_len,
+                ret.triton_start_loc,
+                ret.triton_prefix_lens,
+            ) = init_triton_args(forward_mode, seq_lens, prefix_lens)
 
         return ret
 
 
-def init_flashinfer_args(forward_mode, model_runner, req_pool_indices, seq_lens, prefix_lens,
-                         flashinfer_decode_wrapper):
+def init_flashinfer_args(
+    forward_mode,
+    model_runner,
+    req_pool_indices,
+    seq_lens,
+    prefix_lens,
+    flashinfer_decode_wrapper,
+):
     num_qo_heads = model_runner.model_config.num_attention_heads // model_runner.tp_size
     num_kv_heads = model_runner.model_config.get_num_kv_heads(model_runner.tp_size)
     head_dim = model_runner.model_config.head_dim
@@ -827,9 +843,7 @@ def init_flashinfer_args(forward_mode, model_runner, req_pool_indices, seq_lens,
     else:
         paged_kernel_lens = prefix_lens
 
-    kv_indptr = torch.zeros(
-        (batch_size + 1,), dtype=torch.int32, device="cuda"
-    )
+    kv_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
     kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0)
     req_pool_indices_cpu = req_pool_indices.cpu().numpy()
     paged_kernel_lens_cpu = paged_kernel_lens.cpu().numpy()
@@ -842,9 +856,7 @@ def init_flashinfer_args(forward_mode, model_runner, req_pool_indices, seq_lens,
         ],
         dim=0,
     ).contiguous()
-    kv_last_page_len = torch.ones(
-        (batch_size,), dtype=torch.int32, device="cuda"
-    )
+    kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda")
 
     if forward_mode == ForwardMode.DECODE:
         flashinfer_decode_wrapper.end_forward()
@@ -859,9 +871,7 @@ def init_flashinfer_args(forward_mode, model_runner, req_pool_indices, seq_lens,
         )
     else:
         # extend part
-        qo_indptr = torch.zeros(
-            (batch_size + 1,), dtype=torch.int32, device="cuda"
-        )
+        qo_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
         qo_indptr[1:] = torch.cumsum(seq_lens - prefix_lens, dim=0)
 
         model_runner.flashinfer_prefill_wrapper_ragged.end_forward()

python/sglang/srt/managers/controller/model_runner.py

@@ -16,7 +16,12 @@ from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.models import ModelRegistry
 
 from sglang.global_config import global_config
-from sglang.srt.managers.controller.infer_batch import Batch, ForwardMode, InputMetadata, global_server_args_dict
+from sglang.srt.managers.controller.infer_batch import (
+    Batch,
+    ForwardMode,
+    InputMetadata,
+    global_server_args_dict,
+)
 from sglang.srt.memory_pool import ReqToTokenPool, TokenToKVPool
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import (
@@ -83,7 +88,9 @@ class ModelRunner:
 
         # Set some global args
         global_server_args_dict["disable_flashinfer"] = server_args.disable_flashinfer
-        global_server_args_dict["attention_reduce_in_fp32"] = server_args.attention_reduce_in_fp32
+        global_server_args_dict[
+            "attention_reduce_in_fp32"
+        ] = server_args.attention_reduce_in_fp32
 
         # Load the model and create memory pool
         self.load_model()
@@ -217,7 +224,9 @@ class ModelRunner:
             self.flashinfer_workspace_buffers[1], "NHD"
         )
         self.flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
-            self.flashinfer_workspace_buffers[0], "NHD", use_tensor_cores=use_tensor_cores
+            self.flashinfer_workspace_buffers[0],
+            "NHD",
+            use_tensor_cores=use_tensor_cores,
         )
 
     def init_cuda_graphs(self):
@@ -229,7 +238,9 @@ class ModelRunner:
 
         logger.info(f"[gpu_id={self.gpu_id}] Capture cuda graph begin.")
         batch_size_list = [1, 2, 4] + [i * 8 for i in range(1, 16)]
-        self.cuda_graph_runner = CudaGraphRunner(self, max_batch_size_to_capture=max(batch_size_list))
+        self.cuda_graph_runner = CudaGraphRunner(
+            self, max_batch_size_to_capture=max(batch_size_list)
+        )
         self.cuda_graph_runner.capture(batch_size_list)
 
     @torch.inference_mode()

python/sglang/srt/managers/controller/radix_cache.py

@@ -125,7 +125,8 @@ class RadixCache:
             if x.lock_ref > 0:
                 continue
 
-            num_evicted += evict_callback(x.value)
+            evict_callback(x.value)
+            num_evicted += len(x.value)
             self._delete_leaf(x)
 
             if len(x.parent.children) == 0:

python/sglang/srt/managers/controller/tp_worker.py

@@ -314,7 +314,9 @@ class ModelTpServer:
         self.forward_queue.append(req)
 
     def get_new_fill_batch(self) -> Optional[Batch]:
-        running_bs = len(self.running_batch.reqs) if self.running_batch is not None else 0
+        running_bs = (
+            len(self.running_batch.reqs) if self.running_batch is not None else 0
+        )
         if running_bs >= self.max_running_requests:
             return
 

python/sglang/srt/memory_pool.py

@@ -39,10 +39,12 @@ class ReqToTokenPool:
 class TokenToKVPool:
     def __init__(self, size, dtype, head_num, head_dim, layer_num):
         self.size = size
-        # mem_state is the reference counter.
+        # This can be promised:
+        # assert torch.all(mem_state <= 1) and torch.all(mem_state >= 0)
         # We also add one slot. This slot is used for writing dummy output from padded tokens.
-        self.mem_state = torch.zeros((self.size + 1,), dtype=torch.int16, device="cuda")
-        self.total_ref_ct = 0
+        self.mem_state = torch.zeros((self.size + 1,), dtype=torch.bool, device="cuda")
+        self.total_size = self.size
+        self.total_alloc = 0
 
         # [size, key/value, head_num, head_dim] for each layer
         self.kv_data = [
@@ -71,7 +73,9 @@ class TokenToKVPool:
 
         addition_size = need_size - buffer_len
         alloc_size = max(addition_size, self.prefetch_chunk_size)
-        select_index = torch.nonzero(self.mem_state == 0).squeeze(1)[:alloc_size].to(torch.int32)
+        select_index = (
+            torch.nonzero(self.mem_state == 0).squeeze(1)[:alloc_size].to(torch.int32)
+        )
 
         if select_index.shape[0] < addition_size:
             return None
@@ -105,26 +109,22 @@ class TokenToKVPool:
         return select_index.to(torch.int32), start_loc, start_loc + need_size
 
     def used_size(self):
-        return len(torch.nonzero(self.mem_state).squeeze(1))
+        return self.total_alloc
 
     def available_size(self):
-        return torch.sum(self.mem_state == 0).item() + len(self.prefetch_buffer)
+        return self.total_size - self.total_alloc + len(self.prefetch_buffer)
 
     def add_refs(self, token_index: torch.Tensor):
-        self.total_ref_ct += len(token_index)
-        self.mem_state[token_index] += 1
+        self.total_alloc += len(token_index)
+        self.mem_state[token_index] ^= True
 
     def dec_refs(self, token_index: torch.Tensor):
-        self.total_ref_ct -= len(token_index)
-        self.mem_state[token_index] -= 1
-
-        num_freed = torch.sum(self.mem_state[token_index] == 0)
-
-        return num_freed
+        self.total_alloc -= len(token_index)
+        self.mem_state[token_index] ^= True
 
     def clear(self):
         self.mem_state.fill_(0)
-        self.total_ref_ct = 0
+        self.total_alloc = 0
 
         # We also add one slot. This slot is used for writing dummy output from padded tokens.
-        self.add_refs(torch.tensor([0], dtype=torch.int32))
+        self.mem_state[0] = True

python/sglang/srt/models/minicpm.py

@@ -5,12 +5,9 @@ from typing import Any, Dict, Iterable, Optional, Tuple
 
 import torch
 from torch import nn
-
 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,
@@ -31,7 +28,6 @@ from sglang.srt.managers.controller.model_runner import InputMetadata
 
 
 class MiniCPMMLP(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
@@ -67,7 +63,6 @@ class MiniCPMMLP(nn.Module):
 
 
 class MiniCPMAttention(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
@@ -152,7 +147,6 @@ class MiniCPMAttention(nn.Module):
 
 
 class MiniCPMDecoderLayer(nn.Module):
-
     def __init__(
         self,
         config,
@@ -217,7 +211,6 @@ class MiniCPMDecoderLayer(nn.Module):
 
 
 class MiniCPMModel(nn.Module):
-
     def __init__(
         self,
         config,
@@ -274,7 +267,7 @@ class MiniCPMForCausalLM(nn.Module):
     ) -> None:
         super().__init__()
         self.config = config
-        
+
         self.num_experts = getattr(self.config, "num_experts", 0)
         self.quant_config = quant_config
         self.model = MiniCPMModel(config, quant_config=quant_config)

python/sglang/srt/models/qwen2_moe.py

@@ -8,24 +8,28 @@ import torch
 import torch.nn.functional as F
 from torch import nn
 from transformers import PretrainedConfig
-
 from vllm.config import CacheConfig
-from vllm.distributed import (get_tensor_model_parallel_world_size,
-                              tensor_model_parallel_all_reduce)
+from vllm.distributed import (
+    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 FusedMoE
 from vllm.model_executor.layers.layernorm import RMSNorm
-from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
-                                               QKVParallelLinear,
-                                               ReplicatedLinear,
-                                               RowParallelLinear)
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
-from vllm.model_executor.layers.quantization.base_config import (
-    QuantizationConfig)
+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.sampler import Sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
-    ParallelLMHead, VocabParallelEmbedding)
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors, SamplerOutput
@@ -34,8 +38,8 @@ 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 Qwen2MoeMLP(nn.Module):
 
+class Qwen2MoeMLP(nn.Module):
     def __init__(
         self,
         hidden_size: int,
@@ -46,17 +50,20 @@ class Qwen2MoeMLP(nn.Module):
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
-            hidden_size, [intermediate_size] * 2,
+            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)
-        self.down_proj = RowParallelLinear(intermediate_size,
-                                           hidden_size,
-                                           bias=False,
-                                           quant_config=quant_config,
-                                           reduce_results=reduce_results)
+            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.")
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Only silu is supported for now."
+            )
         self.act_fn = SiluAndMul()
 
     def forward(self, x):
@@ -67,7 +74,6 @@ class Qwen2MoeMLP(nn.Module):
 
 
 class Qwen2MoeSparseMoeBlock(nn.Module):
-
     def __init__(
         self,
         config: PretrainedConfig,
@@ -79,20 +85,22 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         if self.tp_size > config.num_experts:
             raise ValueError(
                 f"Tensor parallel size {self.tp_size} is greater than "
-                f"the number of experts {config.num_experts}.")
-
-        self.experts = FusedMoE(num_experts=config.num_experts,
-                                top_k=config.num_experts_per_tok,
-                                hidden_size=config.hidden_size,
-                                intermediate_size=config.moe_intermediate_size,
-                                reduce_results=False,
-                                renormalize=config.norm_topk_prob,
-                                quant_config=quant_config)
-
-        self.gate = ReplicatedLinear(config.hidden_size,
-                                     config.num_experts,
-                                     bias=False,
-                                     quant_config=None)
+                f"the number of experts {config.num_experts}."
+            )
+
+        self.experts = FusedMoE(
+            num_experts=config.num_experts,
+            top_k=config.num_experts_per_tok,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.moe_intermediate_size,
+            reduce_results=False,
+            renormalize=config.norm_topk_prob,
+            quant_config=quant_config,
+        )
+
+        self.gate = ReplicatedLinear(
+            config.hidden_size, config.num_experts, bias=False, quant_config=None
+        )
         if config.shared_expert_intermediate_size > 0:
             self.shared_expert = Qwen2MoeMLP(
                 hidden_size=config.hidden_size,
@@ -103,9 +111,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
             )
         else:
             self.shared_expert = None
-        self.shared_expert_gate = torch.nn.Linear(config.hidden_size,
-                                                  1,
-                                                  bias=False)
+        self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
@@ -114,24 +120,24 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         if self.shared_expert is not None:
             shared_output = self.shared_expert(hidden_states)
             if self.shared_expert_gate is not None:
-                shared_output = F.sigmoid(
-                    self.shared_expert_gate(hidden_states)) * shared_output
+                shared_output = (
+                    F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_output
+                )
 
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
-        final_hidden_states = self.experts(hidden_states=hidden_states,
-                                           router_logits=router_logits)
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states, router_logits=router_logits
+        )
         if shared_output is not None:
             final_hidden_states = final_hidden_states + shared_output
         if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(
-                final_hidden_states)
+            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(num_tokens, hidden_dim)
 
 
 class Qwen2MoeAttention(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
@@ -190,17 +196,19 @@ class Qwen2MoeAttention(nn.Module):
             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)
+        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
+        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)
@@ -211,7 +219,6 @@ class Qwen2MoeAttention(nn.Module):
 
 
 class Qwen2MoeDecoderLayer(nn.Module):
-
     def __init__(
         self,
         config: PretrainedConfig,
@@ -223,8 +230,7 @@ class Qwen2MoeDecoderLayer(nn.Module):
         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)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = Qwen2MoeAttention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
@@ -239,13 +245,13 @@ class Qwen2MoeDecoderLayer(nn.Module):
 
         # Note: Qwen/Qwen2-57B-A14B-Instruct does not have
         # `mlp_only_layers` in the config.
-        mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
-                           config.mlp_only_layers)
+        mlp_only_layers = (
+            [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
+        )
         if (layer_id not in mlp_only_layers) and (
-                config.num_experts > 0 and
-            (layer_id + 1) % config.decoder_sparse_step == 0):
-            self.mlp = Qwen2MoeSparseMoeBlock(config=config,
-                                              quant_config=quant_config)
+            config.num_experts > 0 and (layer_id + 1) % config.decoder_sparse_step == 0
+        ):
+            self.mlp = Qwen2MoeSparseMoeBlock(config=config, quant_config=quant_config)
         else:
             self.mlp = Qwen2MoeMLP(
                 hidden_size=config.hidden_size,
@@ -253,10 +259,10 @@ class Qwen2MoeDecoderLayer(nn.Module):
                 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)
+        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,
@@ -270,23 +276,20 @@ class Qwen2MoeDecoderLayer(nn.Module):
             residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
         else:
-            hidden_states, residual = self.input_layernorm(
-                hidden_states, residual)
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
         hidden_states = self.self_attn(
             positions=positions,
             hidden_states=hidden_states,
-            input_metadata=input_metadata
+            input_metadata=input_metadata,
         )
 
         # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(
-            hidden_states, residual)
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
         hidden_states = self.mlp(hidden_states)
         return hidden_states, residual
 
 
 class Qwen2MoeModel(nn.Module):
-
     def __init__(
         self,
         config: PretrainedConfig,
@@ -301,13 +304,14 @@ class Qwen2MoeModel(nn.Module):
             config.vocab_size,
             config.hidden_size,
         )
-        self.layers = nn.ModuleList([
-            Qwen2MoeDecoderLayer(config,
-                                 layer_id,
-                                 cache_config,
-                                 quant_config=quant_config)
-            for layer_id in range(config.num_hidden_layers)
-        ])
+        self.layers = nn.ModuleList(
+            [
+                Qwen2MoeDecoderLayer(
+                    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(
@@ -315,7 +319,7 @@ class Qwen2MoeModel(nn.Module):
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        input_embeds: torch.Tensor = None
+        input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
@@ -324,10 +328,9 @@ class Qwen2MoeModel(nn.Module):
         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, residual = layer(
+                positions, hidden_states, input_metadata, residual
+            )
         hidden_states, _ = self.norm(hidden_states, residual)
         return hidden_states
 
@@ -346,9 +349,9 @@ class Qwen2MoeForCausalLM(nn.Module):
         self.config = config
         self.quant_config = quant_config
         self.model = Qwen2MoeModel(config, cache_config, quant_config)
-        self.lm_head = ParallelLMHead(config.vocab_size,
-                                      config.hidden_size,
-                                      quant_config=quant_config)
+        self.lm_head = ParallelLMHead(
+            config.vocab_size, config.hidden_size, quant_config=quant_config
+        )
         self.logits_processor = LogitsProcessor(config)
         self.sampler = Sampler()
 
@@ -357,17 +360,22 @@ class Qwen2MoeForCausalLM(nn.Module):
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        input_embeds: torch.Tensor = None
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    def compute_logits(
+        self,
+        input_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, input_metadata,
-                                   input_embeds)
-        return self.logits_processor(input_ids, hidden_states, self.lm_head.weight,
-                                     input_metadata)
-
-    def compute_logits(self, input_ids: torch.Tensor, hidden_states: torch.Tensor,
-                       input_metadata: InputMetadata) -> torch.Tensor:
-        logits = self.logits_processor(input_ids, hidden_states, self.lm_head.weight,
-                                       input_metadata)
+        logits = self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
         return logits
 
     def sample(
@@ -391,11 +399,18 @@ class Qwen2MoeForCausalLM(nn.Module):
         expert_params_mapping = [
             # 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",
-             f"experts.{expert_id}.{weight_name}.weight", expert_id, shard_id)
-            for expert_id in range(self.config.num_experts) for shard_id,
-            weight_name in enumerate(["gate_proj", "down_proj", "up_proj"])
+            (
+                "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,
+            )
+            for expert_id in range(self.config.num_experts)
+            for shard_id, weight_name in enumerate(
+                ["gate_proj", "down_proj", "up_proj"]
+            )
         ]
 
         params_dict = dict(self.named_parameters())
@@ -433,11 +448,13 @@ class Qwen2MoeForCausalLM(nn.Module):
                     name = name.replace(weight_name, param_name)
                     param = params_dict[name]
                     weight_loader = param.weight_loader
-                    weight_loader(param,
-                                  loaded_weight,
-                                  weight_name,
-                                  shard_id=shard_id,
-                                  expert_id=expert_id)
+                    weight_loader(
+                        param,
+                        loaded_weight,
+                        weight_name,
+                        shard_id=shard_id,
+                        expert_id=expert_id,
+                    )
                     break
                 else:
                     # Skip loading extra bias for GPTQ models.
@@ -447,8 +464,10 @@ class Qwen2MoeForCausalLM(nn.Module):
                         continue
 
                     param = params_dict[name]
-                    weight_loader = getattr(param, "weight_loader",
-                                            default_weight_loader)
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
                     weight_loader(param, loaded_weight)
 
+
 EntryClass = Qwen2MoeForCausalLM

python/sglang/srt/utils.py

@@ -474,9 +474,9 @@ def monkey_patch_vllm_dummy_weight_loader():
         DummyModelLoader,
         LoRAConfig,
         ModelConfig,
+        MultiModalConfig,
         ParallelConfig,
         SchedulerConfig,
-        MultiModalConfig,
         _initialize_model,
         initialize_dummy_weights,
         nn,