sglang 0.3.5.post1__tar.gz → 0.3.5.post2__tar.gz

{sglang-0.3.5.post1 → sglang-0.3.5.post2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.3.5.post1
+Version: 0.3.5.post2
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -233,7 +233,7 @@ Requires-Dist: torchao; extra == "runtime-common"
 Requires-Dist: uvicorn; extra == "runtime-common"
 Requires-Dist: uvloop; extra == "runtime-common"
 Requires-Dist: pyzmq>=25.1.2; extra == "runtime-common"
-Requires-Dist: outlines>=0.0.44; extra == "runtime-common"
+Requires-Dist: outlines<0.1.0,>=0.0.44; extra == "runtime-common"
 Requires-Dist: modelscope; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"

{sglang-0.3.5.post1 → sglang-0.3.5.post2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.3.5.post1"
+version = "0.3.5.post2"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -19,7 +19,7 @@ dependencies = ["requests", "tqdm", "numpy", "IPython"]
 runtime_common = ["aiohttp", "decord", "fastapi", "hf_transfer", "huggingface_hub", "interegular",
     "orjson", "packaging", "pillow", "prometheus-client>=0.20.0", "psutil", "pydantic", "python-multipart",
     "torchao", "uvicorn", "uvloop", "pyzmq>=25.1.2",
-    "outlines>=0.0.44", "modelscope"]
+    "outlines>=0.0.44,<0.1.0", "modelscope"]
 srt = ["sglang[runtime_common]", "torch", "vllm==0.6.3.post1"]
 
 # HIP (Heterogeneous-computing Interface for Portability) for AMD

sglang-0.3.5.post2/sglang/bench_offline_throughput.py

@@ -0,0 +1,309 @@
+"""
+Benchmark the throughput of using the offline LLM engine.
+This script does not launch a server.
+It accepts server arguments (the same as launch_server.py) and benchmark arguments (the same as bench_serving.py).
+
+# Usage
+## Sharegpt dataset with default args
+python -m sglang.bench_offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct
+
+## Random dataset with default args
+python -m sglang.bench_offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --dataset-name random
+
+## Shared prefix dataset with default args
+python -m sglang.bench_offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --dataset-name generated-shared-prefix
+
+## Sharegpt dataset on runtime backend
+python -m sglang.bench_offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --backend runtime
+"""
+
+import argparse
+import dataclasses
+import json
+import logging
+import random
+import time
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from sglang.api import Engine
+from sglang.bench_serving import (
+    get_dataset,
+    get_tokenizer,
+    sample_random_requests,
+    set_ulimit,
+)
+from sglang.srt.server import Runtime
+from sglang.srt.server_args import ServerArgs
+
+
+@dataclasses.dataclass
+class BenchArgs:
+    backend: str = "engine"
+    result_filename: str = ""
+    dataset_name: str = "sharegpt"
+    dataset_path: str = ""
+    num_prompts: int = 1000
+    sharegpt_output_len: Optional[int] = None
+    random_input_len: int = 1024
+    random_output_len: int = 1024
+    random_range_ratio: float = 0.0
+    gen_num_groups: int = 64
+    gen_prompts_per_group: int = 16
+    gen_system_prompt_len: int = 2048
+    gen_question_len: int = 128
+    gen_output_len: int = 256
+    disable_ignore_eos: bool = False
+    seed: int = 1
+
+    @staticmethod
+    def add_cli_args(parser: argparse.ArgumentParser):
+        parser.add_argument("--backend", type=str, default=BenchArgs.backend)
+        parser.add_argument(
+            "--result-filename", type=str, default=BenchArgs.result_filename
+        )
+        parser.add_argument(
+            "--dataset-name",
+            type=str,
+            default="sharegpt",
+            choices=["sharegpt", "random", "generated-shared-prefix"],
+            help="Name of the dataset to benchmark on.",
+        )
+        parser.add_argument(
+            "--dataset-path", type=str, default="", help="Path to the dataset."
+        )
+        parser.add_argument(
+            "--num-prompts",
+            type=int,
+            default=BenchArgs.num_prompts,
+            help="Number of prompts to process. Default is 1000.",
+        )
+        parser.add_argument(
+            "--sharegpt-output-len",
+            type=int,
+            default=BenchArgs.sharegpt_output_len,
+            help="Output length for each request. Overrides the output length from the ShareGPT dataset.",
+        )
+        parser.add_argument(
+            "--random-input-len",
+            type=int,
+            default=BenchArgs.random_input_len,
+            help="Number of input tokens per request, used only for random dataset.",
+        )
+        parser.add_argument(
+            "--random-output-len",
+            type=int,
+            default=BenchArgs.random_output_len,
+            help="Number of output tokens per request, used only for random dataset.",
+        )
+        parser.add_argument(
+            "--random-range-ratio",
+            type=float,
+            default=BenchArgs.random_range_ratio,
+            help="Range of sampled ratio of input/output length, "
+            "used only for random dataset.",
+        )
+        parser.add_argument(
+            "--gen-num-groups",
+            type=int,
+            default=BenchArgs.gen_num_groups,
+            help="Number of groups with shared prefix, used"
+            "only for generate-shared-prefix",
+        )
+        parser.add_argument(
+            "--gen-prompts-per-group",
+            type=int,
+            default=BenchArgs.gen_prompts_per_group,
+            help="Number of prompts per group of shared prefix, used"
+            "only for generate-shared-prefix",
+        )
+        parser.add_argument(
+            "--gen-system-prompt-len",
+            type=int,
+            default=BenchArgs.gen_system_prompt_len,
+            help="System prompt length, used" "only for generate-shared-prefix",
+        )
+        parser.add_argument(
+            "--gen-question-len",
+            type=int,
+            default=BenchArgs.gen_question_len,
+            help="Question length, used" "only for generate-shared-prefix",
+        )
+        parser.add_argument(
+            "--gen-output-len",
+            type=int,
+            default=BenchArgs.gen_output_len,
+            help="Target length in tokens for outputs in generated-shared-prefix dataset",
+        )
+        parser.add_argument(
+            "--disable-ignore-eos",
+            type=bool,
+            default=BenchArgs.disable_ignore_eos,
+            help="Disable ignore EOS token",
+        )
+        parser.add_argument("--seed", type=int, default=1, help="The random seed.")
+
+    @classmethod
+    def from_cli_args(cls, args: argparse.Namespace):
+        attrs = [attr.name for attr in dataclasses.fields(cls)]
+        return cls(**{attr: getattr(args, attr) for attr in attrs})
+
+
+def throughput_test_once(
+    backend_name: str,
+    backend,
+    reqs: List[Tuple[str, int, int]],
+    ignore_eos: bool,
+):
+    measurement_results = {
+        "backend": backend_name,
+        "successful_requests": len(reqs),
+        "total_latency": -1,
+        "total_input_tokens": sum(r[1] for r in reqs),
+        "total_output_tokens": -1,
+        "request_throughput": -1,
+        "input_throughput": -1,
+        "output_throughput": -1,
+        "total_throughput": -1,
+    }
+
+    prompt = [r[0] for r in reqs]
+    sampling_params = [
+        {
+            "temperature": 0,
+            "max_new_tokens": r[2],
+            "ignore_eos": ignore_eos,
+        }
+        for r in reqs
+    ]
+
+    st = time.perf_counter()
+    gen_out = backend.generate(prompt=prompt, sampling_params=sampling_params)
+    latency = time.perf_counter() - st
+
+    if backend_name == "runtime":
+        gen_out = json.loads(gen_out)
+
+    measurement_results["total_latency"] = latency
+    measurement_results["total_output_tokens"] = sum(
+        o["meta_info"]["completion_tokens"] for o in gen_out
+    )
+    measurement_results["request_throughput"] = (
+        measurement_results["successful_requests"] / latency
+    )
+    measurement_results["input_throughput"] = (
+        measurement_results["total_input_tokens"] / latency
+    )
+    measurement_results["output_throughput"] = (
+        measurement_results["total_output_tokens"] / latency
+    )
+    measurement_results["total_throughput"] = (
+        measurement_results["total_input_tokens"]
+        + measurement_results["total_output_tokens"]
+    ) / latency
+
+    return measurement_results
+
+
+def throughput_test(
+    server_args: ServerArgs,
+    bench_args: BenchArgs,
+):
+    if bench_args.backend == "engine":
+        backend = Engine(**dataclasses.asdict(server_args))
+        if not backend:
+            raise ValueError("Please provide valid engine arguments")
+    elif bench_args.backend == "runtime":
+        backend = Runtime(**dataclasses.asdict(server_args))
+    else:
+        raise ValueError('Please set backend to either "engine" or "runtime"')
+
+    tokenizer_id = server_args.model_path
+    tokenizer = get_tokenizer(tokenizer_id)
+
+    # Set global environmnets
+    set_ulimit()
+    random.seed(bench_args.seed)
+    np.random.seed(bench_args.seed)
+
+    # Read dataset
+    input_requests = get_dataset(bench_args, tokenizer)
+
+    warmup_requests = sample_random_requests(
+        input_len=20,
+        output_len=4,
+        num_prompts=2,
+        range_ratio=0.8,
+        tokenizer=tokenizer,
+        dataset_path=bench_args.dataset_path,
+    )
+
+    # Warm up
+    throughput_test_once(
+        backend_name=bench_args.backend,
+        backend=backend,
+        reqs=warmup_requests,
+        ignore_eos=not bench_args.disable_ignore_eos,
+    )
+
+    result = throughput_test_once(
+        backend_name=bench_args.backend,
+        backend=backend,
+        reqs=input_requests,
+        ignore_eos=not bench_args.disable_ignore_eos,
+    )
+
+    if bench_args.result_filename:
+        with open(bench_args.result_filename, "a") as fout:
+            fout.write(json.dumps(result) + "\n")
+
+    print(
+        "\n{s:{c}^{n}}".format(s=" Offline Throughput Benchmark Result ", n=50, c="=")
+    )
+    print("{:<40} {:<10}".format("Backend:", result["backend"]))
+    print("{:<40} {:<10}".format("Successful requests:", result["successful_requests"]))
+    print("{:<40} {:<10.2f}".format("Benchmark duration (s):", result["total_latency"]))
+    print("{:<40} {:<10}".format("Total input tokens:", result["total_input_tokens"]))
+    print(
+        "{:<40} {:<10}".format("Total generated tokens:", result["total_output_tokens"])
+    )
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Request throughput (req/s):", result["request_throughput"]
+        )
+    )
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Input token throughput (tok/s):", result["input_throughput"]
+        )
+    )
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Output token throughput (tok/s):", result["output_throughput"]
+        )
+    )
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Total token throughput (tok/s):", result["total_throughput"]
+        )
+    )
+    print("=" * 50)
+
+    return result
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ServerArgs.add_cli_args(parser)
+    BenchArgs.add_cli_args(parser)
+    args = parser.parse_args()
+    server_args = ServerArgs.from_cli_args(args)
+    bench_args = BenchArgs.from_cli_args(args)
+
+    logging.basicConfig(
+        level=getattr(logging, server_args.log_level.upper()),
+        format="%(message)s",
+    )
+
+    throughput_test(server_args, bench_args)

{sglang-0.3.5.post1 → sglang-0.3.5.post2}/sglang/bench_serving.py

@@ -421,6 +421,37 @@ def get_tokenizer(
     )
 
 
+def get_dataset(args, tokenizer):
+    if args.dataset_name == "sharegpt":
+        input_requests = sample_sharegpt_requests(
+            dataset_path=args.dataset_path,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            fixed_output_len=args.sharegpt_output_len,
+        )
+    elif args.dataset_name == "random":
+        input_requests = sample_random_requests(
+            input_len=args.random_input_len,
+            output_len=args.random_output_len,
+            num_prompts=args.num_prompts,
+            range_ratio=args.random_range_ratio,
+            tokenizer=tokenizer,
+            dataset_path=args.dataset_path,
+        )
+    elif args.dataset_name == "generated-shared-prefix":
+        input_requests = sample_generated_shared_prefix_requests(
+            num_groups=args.gen_num_groups,
+            prompts_per_group=args.gen_prompts_per_group,
+            system_prompt_len=args.gen_system_prompt_len,
+            question_len=args.gen_question_len,
+            output_len=args.gen_output_len,
+            tokenizer=tokenizer,
+        )
+    else:
+        raise ValueError(f"Unknown dataset: {args.dataset_name}")
+    return input_requests
+
+
 ASYNC_REQUEST_FUNCS = {
     "sglang": async_request_sglang_generate,
     "sglang-native": async_request_sglang_generate,
@@ -443,6 +474,8 @@ class BenchmarkMetrics:
     input_throughput: float
     output_throughput: float
     output_throughput_retokenized: float
+    total_throughput: float
+    total_throughput_retokenized: float
     mean_ttft_ms: float
     median_ttft_ms: float
     std_ttft_ms: float
@@ -590,7 +623,6 @@ def sample_random_requests(
             (data["conversations"][0]["value"], data["conversations"][1]["value"])
             for data in dataset
         ]
-
         # Shuffle the dataset.
         random.shuffle(dataset)
 
@@ -764,6 +796,9 @@ def calculate_metrics(
         input_throughput=total_input / dur_s,
         output_throughput=sum(output_lens) / dur_s,
         output_throughput_retokenized=sum(retokenized_output_lens) / dur_s,
+        total_throughput=(total_input + sum(output_lens)) / dur_s,
+        total_throughput_retokenized=(total_input + sum(retokenized_output_lens))
+        / dur_s,
         mean_ttft_ms=np.mean(ttfts or 0)
         * 1000,  # ttfts is empty if streaming is not supported by backend
         median_ttft_ms=np.median(ttfts or 0) * 1000,
@@ -881,6 +916,11 @@ async def benchmark(
             "Output token throughput (tok/s):", metrics.output_throughput
         )
     )
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Total token throughput (tok/s):", metrics.total_throughput
+        )
+    )
     print("{s:{c}^{n}}".format(s="End-to-End Latency", n=50, c="-"))
     print(
         "{:<40} {:<10.2f}".format("Mean E2E Latency (ms):", metrics.mean_e2e_latency_ms)
@@ -1098,35 +1138,7 @@ def run_benchmark(args_: argparse.Namespace):
 
     tokenizer = get_tokenizer(tokenizer_id)
 
-    if args.dataset_name == "sharegpt":
-        assert args.random_input_len is None and args.random_output_len is None
-        input_requests = sample_sharegpt_requests(
-            dataset_path=args.dataset_path,
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            fixed_output_len=args.sharegpt_output_len,
-        )
-    elif args.dataset_name == "random":
-        assert args.random_input_len is not None and args.random_output_len is not None
-        input_requests = sample_random_requests(
-            input_len=args.random_input_len,
-            output_len=args.random_output_len,
-            num_prompts=args.num_prompts,
-            range_ratio=args.random_range_ratio,
-            tokenizer=tokenizer,
-            dataset_path=args.dataset_path,
-        )
-    elif args.dataset_name == "generated-shared-prefix":
-        input_requests = sample_generated_shared_prefix_requests(
-            num_groups=args.gen_num_groups,
-            prompts_per_group=args.gen_prompts_per_group,
-            system_prompt_len=args.gen_system_prompt_len,
-            question_len=args.gen_question_len,
-            output_len=args.gen_output_len,
-            tokenizer=tokenizer,
-        )
-    else:
-        raise ValueError(f"Unknown dataset: {args.dataset_name}")
+    input_requests = get_dataset(args, tokenizer)
 
     if not args.multi:
         return asyncio.run(
@@ -1229,10 +1241,12 @@ if __name__ == "__main__":
     parser.add_argument(
         "--random-input-len",
         type=int,
+        default=1024,
         help="Number of input tokens per request, used only for random dataset.",
     )
     parser.add_argument(
         "--random-output-len",
+        default=1024,
         type=int,
         help="Number of output tokens per request, used only for random dataset.",
     )

{sglang-0.3.5.post1 → sglang-0.3.5.post2}/sglang/srt/constrained/base_grammar_backend.py

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 
-"""The baseclass of backends for grammar-guided constrained decoding."""
+"""The baseclass of a backend for grammar-guided constrained decoding."""
 
 from concurrent.futures import Future, ThreadPoolExecutor
 from dataclasses import dataclass
@@ -52,7 +52,7 @@ class BaseGrammarBackend:
         else:
             entry.value = self.init_value_impl(key)
             entry.event.set()
-        return entry.value.copy()
+        return entry.value.copy() if entry.value else None
 
     def init_value_impl(self, key: Tuple[str, str]) -> BaseGrammarObject:
         raise NotImplementedError()
@@ -62,7 +62,8 @@ class BaseGrammarBackend:
             entry = self.cache.get(key)
             if not entry or not entry.event.is_set():
                 return None
-            return self.cache[key].value.copy()
+            val = self.cache[key].value
+            return val.copy() if val else None
 
     def get_future_value(self, key: Tuple[str, str]) -> Future:
         return self.executor.submit(self.init_value, key)

{sglang-0.3.5.post1 → sglang-0.3.5.post2}/sglang/srt/constrained/outlines_backend.py

@@ -19,9 +19,12 @@ import json
 import logging
 from typing import Dict, List, Optional, Tuple, Union
 
+import interegular
 import torch
 from outlines.fsm.guide import RegexGuide
+from outlines.fsm.json_schema import build_regex_from_schema
 from outlines.models.transformers import TransformerTokenizer
+from pydantic import BaseModel
 
 from sglang.srt.constrained.base_grammar_backend import (
     BaseGrammarBackend,
@@ -32,26 +35,6 @@ from sglang.srt.constrained.outlines_jump_forward import OutlinesJumpForwardMap
 logger = logging.getLogger(__name__)
 
 
-try:
-    from outlines.fsm.json_schema import build_regex_from_object
-except ImportError:
-    # Since outlines 0.0.32, build_regex_from_object is replaced by build_regex_from_schema,
-    # which only accepts string schema as input.
-    from outlines.fsm.json_schema import build_regex_from_schema
-    from pydantic import BaseModel
-
-    def build_regex_from_object(
-        object: Union[str, BaseModel, Dict], whitespace_pattern: Optional[str] = None
-    ):
-        if isinstance(object, type(BaseModel)):
-            schema = json.dumps(object.model_json_schema())
-        elif isinstance(object, Dict):
-            schema = json.dumps(object)
-        else:
-            schema = object
-        return build_regex_from_schema(schema, whitespace_pattern)
-
-
 class OutlinesGrammar(BaseGrammarObject):
     def __init__(
         self,
@@ -147,19 +130,36 @@ class OutlinesGrammarBackend(BaseGrammarBackend):
                     key_string,
                     whitespace_pattern=self.whitespace_pattern,
                 )
-            except NotImplementedError as e:
+            except (NotImplementedError, json.decoder.JSONDecodeError) as e:
                 logger.warning(
-                    f"skip invalid json schema: json_schema={key_string}, {e=}"
+                    f"Skip invalid json_schema: json_schema={key_string}, {e=}"
                 )
-                return None, key_string
+                return None
         elif key_type == "regex":
             regex = key_string
         else:
             raise ValueError(f"Invalid key_type: {key_type}")
 
-        guide = RegexGuide(regex, self.outlines_tokenizer)
+        try:
+            guide = RegexGuide(regex, self.outlines_tokenizer)
+        except interegular.patterns.InvalidSyntax as e:
+            logger.warning(f"skip invalid regex schema: {regex=}, {e=}")
+            return None
+
         if self.allow_jump_forward:
             jump_forward_map = OutlinesJumpForwardMap(regex)
         else:
             jump_forward_map = None
         return OutlinesGrammar(guide, jump_forward_map)
+
+
+def build_regex_from_object(
+    object: Union[str, BaseModel, Dict], whitespace_pattern: Optional[str] = None
+):
+    if isinstance(object, type(BaseModel)):
+        schema = json.dumps(object.model_json_schema())
+    elif isinstance(object, Dict):
+        schema = json.dumps(object)
+    else:
+        schema = object
+    return build_regex_from_schema(schema, whitespace_pattern)

{sglang-0.3.5.post1 → sglang-0.3.5.post2}/sglang/srt/constrained/xgrammar_backend.py

@@ -15,16 +15,29 @@ limitations under the License.
 
 """Constrained decoding with xgrammar backend."""
 
+import logging
 from typing import List, Tuple
 
 import torch
-from xgrammar import CachedGrammarCompiler, CompiledGrammar, GrammarMatcher
+
+try:
+    from xgrammar import CachedGrammarCompiler, CompiledGrammar, GrammarMatcher
+
+    import_error = None
+except ImportError as e:
+    CachedGrammarCompiler = CompiledGrammar = GrammarMatcher = TokenizerInfo = (
+        ImportError
+    )
+    import_error = e
 
 from sglang.srt.constrained.base_grammar_backend import (
     BaseGrammarBackend,
     BaseGrammarObject,
 )
 
+logger = logging.getLogger(__name__)
+
+
 MAX_ROLLBACK_TOKENS = 10
 
 
@@ -91,15 +104,37 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
         vocab_size: int,
     ):
         super().__init__()
+
+        if import_error:
+            logger.warning(
+                f"Ignore import error for the grammar backend: {import_error}"
+            )
+            self.grammar_cache = None
+            return
+
         self.grammar_cache = CachedGrammarCompiler(tokenizer_or_vocab=tokenizer)
         self.vocab_size = vocab_size
 
     def init_value_impl(self, key: Tuple[str, str]) -> XGrammarGrammar:
+        if import_error:
+            raise import_error
+
         key_type, key_string = key
         if key_type == "json":
-            ctx = self.grammar_cache.get_compiled_grammar_for_json_schema(key_string)
+            try:
+                ctx = self.grammar_cache.get_compiled_grammar_for_json_schema(
+                    key_string
+                )
+            except RuntimeError as e:
+                logging.warning(
+                    f"Skip invalid json_schema: json_schema={key_string}, {e=}"
+                )
+                return None
         elif key_type == "regex":
-            raise ValueError("regex hasn't been supported by xgrammar yet")
+            logger.warning(
+                "regex hasn't been supported by xgrammar yet. This is skipped."
+            )
+            return None
         else:
             raise ValueError(f"Invalid key_type: {key_type}")
 
@@ -111,4 +146,5 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
         return XGrammarGrammar(matcher, self.vocab_size, ctx)
 
     def reset(self):
-        self.grammar_cache.clear()
+        if self.grammar_cache:
+            self.grammar_cache.clear()

{sglang-0.3.5.post1 → sglang-0.3.5.post2}/sglang/srt/layers/fused_moe/patch.py

@@ -1,4 +1,4 @@
-from typing import Optional
+from typing import Callable, Optional
 
 import torch
 from torch.nn import functional as F
@@ -98,7 +98,9 @@ def fused_moe_forward_native(
     renormalize: bool,
     topk_group: Optional[int] = None,
     num_expert_group: Optional[int] = None,
+    custom_routing_function: Optional[Callable] = None,
 ) -> torch.Tensor:
+    assert custom_routing_function is None
     topk_weights, topk_ids = select_experts_native(
         hidden_states=x,
         router_logits=router_logits,
@@ -114,4 +116,4 @@ def fused_moe_forward_native(
     x1 = F.silu(torch.einsum("ti,taoi -> tao", x, w1_weights))
     x3 = torch.einsum("ti, taoi -> tao", x, w3_weights)
     expert_outs = torch.einsum("tao, taio -> tai", (x1 * x3), w2_weights)
-    return torch.einsum("tai,ta -> ti", expert_outs, topk_weights)
+    return torch.einsum("tai,ta -> ti", expert_outs, topk_weights.to(expert_outs.dtype))