From d5681bbdc067eb502798231d2aa5b4c3be82169f Mon Sep 17 00:00:00 2001
From: Sunny-bot1 <592045536@qq.com>
Date: Fri, 11 Jul 2025 16:17:50 +0800
Subject: [PATCH 1/3] fix topp default value

---
 fastdeploy/engine/sampling_params.py       | 4 ++--
 fastdeploy/worker/gcu_model_runner.py      | 4 ++--
 fastdeploy/worker/gpu_model_runner.py      | 2 +-
 fastdeploy/worker/iluvatar_model_runner.py | 4 ++--
 fastdeploy/worker/xpu_model_runner.py      | 4 ++--
 5 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/fastdeploy/engine/sampling_params.py b/fastdeploy/engine/sampling_params.py
index a7912407a8c..d81f9f99902 100644
--- a/fastdeploy/engine/sampling_params.py
+++ b/fastdeploy/engine/sampling_params.py
@@ -82,7 +82,7 @@ class SamplingParams:
     frequency_penalty: float = None
     repetition_penalty: float = None
     temperature: float = None
-    top_p: float = 1.0
+    top_p: float = None
     top_k: int = 0
     seed: Optional[int] = None
     stop: Optional[Union[str, List[str]]] = None
@@ -132,7 +132,7 @@ def from_optional(cls,
                    repetition_penalty=repetition_penalty
                    if repetition_penalty is not None else 1.0,
                    temperature=temperature if temperature is not None else 1.0,
-                   top_p=top_p if top_p is not None else 1.0,
+                   top_p=top_p,
                    top_k=top_k if top_k is not None else 0,
                    seed=seed,
                    stop=stop,
diff --git a/fastdeploy/worker/gcu_model_runner.py b/fastdeploy/worker/gcu_model_runner.py
index 5ad0cec76ed..eee5dbf8e1e 100644
--- a/fastdeploy/worker/gcu_model_runner.py
+++ b/fastdeploy/worker/gcu_model_runner.py
@@ -24,6 +24,7 @@
 
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
+from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.model_executor.guided_decoding import get_guided_backend
 from fastdeploy.model_executor.guided_decoding.base_guided_decoding import \
     LogitsProcessorBase
@@ -39,7 +40,6 @@
 from fastdeploy.model_executor.pre_and_post_process import (post_process,
                                                             pre_process,
                                                             rebuild_padding)
-from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.worker.model_runner_base import ModelRunnerBase
 from fastdeploy.worker.output import ModelOutputData, ModelRunnerOutput
 
@@ -235,7 +235,7 @@ def insert_prefill_inputs(self, req_dicts: List[Request]):
                 request.eos_token_ids.append(request.eos_token_ids[0])
             self.share_inputs["eos_token_id"][:] = np.array(
                 request.eos_token_ids, dtype="int64").reshape(-1, 1)
-            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 1.0)
+            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 0.7)
             self.share_inputs["top_k"][idx:idx + 1] = request.get("top_k", 0)
             self.share_inputs["temperature"][idx:idx + 1] = request.get(
                 "temperature", 0.95)
diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
index 3a61cbde1bd..f22d7cc6eff 100644
--- a/fastdeploy/worker/gpu_model_runner.py
+++ b/fastdeploy/worker/gpu_model_runner.py
@@ -245,7 +245,7 @@ def insert_prefill_inputs(self, req_dicts: List[Request]):
                 request.eos_token_ids.append(request.eos_token_ids[0])
             self.share_inputs["eos_token_id"][:] = np.array(
                 request.eos_token_ids, dtype="int64").reshape(-1, 1)
-            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 1.0)
+            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 0.7)
             self.share_inputs["top_k"][idx:idx + 1] = request.get("top_k", 0)
             self.share_inputs["temperature"][idx:idx + 1] = request.get(
                 "temperature", 0.95)
diff --git a/fastdeploy/worker/iluvatar_model_runner.py b/fastdeploy/worker/iluvatar_model_runner.py
index 069f7b37c24..8d9477b78fd 100644
--- a/fastdeploy/worker/iluvatar_model_runner.py
+++ b/fastdeploy/worker/iluvatar_model_runner.py
@@ -24,6 +24,7 @@
 
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
+from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.model_executor.layers.attention import get_attention_backend
 from fastdeploy.model_executor.layers.attention.base_attention_backend import \
     AttentionBackend
@@ -37,7 +38,6 @@
                                                             pre_process,
                                                             rebuild_padding,
                                                             step_cuda)
-from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.worker.model_runner_base import ModelRunnerBase
 from fastdeploy.worker.output import ModelOutputData, ModelRunnerOutput
 
@@ -224,7 +224,7 @@ def insert_prefill_inputs(self, req_dicts: List[Request]):
                 request.eos_token_ids.append(request.eos_token_ids[0])
             self.share_inputs["eos_token_id"][:] = np.array(
                 request.eos_token_ids, dtype="int64").reshape(-1, 1)
-            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 1.0)
+            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 0.7)
             self.share_inputs["top_k"][idx:idx + 1] = request.get("top_k", 0)
             self.share_inputs["temperature"][idx:idx + 1] = request.get(
                 "temperature", 0.95)
diff --git a/fastdeploy/worker/xpu_model_runner.py b/fastdeploy/worker/xpu_model_runner.py
index 7fb585f8a53..9099339766d 100644
--- a/fastdeploy/worker/xpu_model_runner.py
+++ b/fastdeploy/worker/xpu_model_runner.py
@@ -23,6 +23,7 @@
 
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
+from fastdeploy.model_executor.forward_meta import ForwardMeta, XPUForwardMeta
 from fastdeploy.model_executor.layers.attention import get_attention_backend
 from fastdeploy.model_executor.layers.attention.base_attention_backend import \
     AttentionBackend
@@ -31,7 +32,6 @@
 from fastdeploy.model_executor.layers.sample.sampler import Sampler
 from fastdeploy.model_executor.model_loader import get_model_from_loader
 from fastdeploy.utils import get_logger
-from fastdeploy.model_executor.forward_meta import ForwardMeta, XPUForwardMeta
 from fastdeploy.worker.model_runner_base import ModelRunnerBase
 from fastdeploy.worker.output import ModelOutputData, ModelRunnerOutput
 
@@ -295,7 +295,7 @@ def process_prefill_inputs(self, req_dicts: List[Request]):
             self.share_inputs["eos_token_id"][:] = np.array(
                 request.eos_token_ids, dtype="int64").reshape(-1, 1)
             self.share_inputs["pre_ids"][idx:idx + 1] = -1
-            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 1.0)
+            self.share_inputs["top_p"][idx:idx + 1] = request.get("top_p", 0.7)
             self.share_inputs["top_k"][idx:idx + 1] = request.get("top_k", 0)
             self.share_inputs["temperature"][idx:idx + 1] = request.get(
                 "temperature", 0.95)

From cb7ad2f8e62abd8a5f26e2e9f09d73940fb1aef5 Mon Sep 17 00:00:00 2001
From: lizexu123 <2694294196@qq.com>
Date: Tue, 29 Jul 2025 09:21:09 +0000
Subject: [PATCH 2/3] merge develop

---
 fastdeploy/engine/args_utils.py       |  28 +++++--
 fastdeploy/engine/sampling_params.py  | 110 +++++++++++++++-----------
 fastdeploy/worker/gpu_model_runner.py |  12 +--
 3 files changed, 90 insertions(+), 60 deletions(-)

diff --git a/fastdeploy/engine/args_utils.py b/fastdeploy/engine/args_utils.py
index b1e464a6ac0..4bb39e9427c 100644
--- a/fastdeploy/engine/args_utils.py
+++ b/fastdeploy/engine/args_utils.py
@@ -320,6 +320,11 @@ class EngineArgs:
     Must be explicitly enabled via the `--enable-logprob` startup parameter to output logprob values.
     """
 
+    seed : Optional[int] = None
+    """
+    Random seed to use for initialization. If not set, a random seed is used.
+    """
+
     def __post_init__(self):
         """
         Post-initialization processing to set default tokenizer if not provided.
@@ -465,7 +470,15 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             default=EngineArgs.enable_logprob,
             help="Enable output of token-level log probabilities.",
         )
-
+        model_group.add_argument("--enable-logprob",
+                                 action="store_true",
+                                 default=EngineArgs.enable_logprob,
+                                 help="Enable output of token-level log probabilities."
+                                 )
+        model_group.add_argument("--seed",
+                                type=int,
+                                defualt=None,
+                                help="Random seed for initialization. If not specified, a random seed will be used.")
         # Parallel processing parameters group
         parallel_group = parser.add_argument_group("Parallel Configuration")
         parallel_group.add_argument(
@@ -775,13 +788,12 @@ def create_model_config(self) -> ModelConfig:
         """
         Create and return a ModelConfig object based on the current settings.
         """
-        return ModelConfig(
-            model_name_or_path=self.model,
-            config_json_file=self.model_config_name,
-            quantization=self.quantization,
-            dynamic_load_weight=self.dynamic_load_weight,
-            load_strategy=self.load_strategy,
-        )
+        return ModelConfig(model_name_or_path=self.model,
+                           config_json_file=self.model_config_name,
+                           quantization=self.quantization,
+                           dynamic_load_weight=self.dynamic_load_weight,
+                           load_strategy=self.load_strategy,
+                           seed=self.seed)
 
     def create_cache_config(self, model_cfg) -> CacheConfig:
         """
diff --git a/fastdeploy/engine/sampling_params.py b/fastdeploy/engine/sampling_params.py
index 688351883b2..67343a88796 100644
--- a/fastdeploy/engine/sampling_params.py
+++ b/fastdeploy/engine/sampling_params.py
@@ -15,10 +15,20 @@
 """
 
 from __future__ import annotations
-
 import random
+from functools import cached_property
 from dataclasses import dataclass, fields
 from typing import Any, List, Optional, Union
+from enum import Enum, IntEnum
+
+
+_SAMPLING_EPS = 1e-5
+_MAX_TEMP = 1e-2
+
+class SamplingType(IntEnum):
+    GREEDY = 0
+    RANDOM = 1
+    RANDOM_SEED = 2
 
 
 @dataclass
@@ -108,50 +118,49 @@ def from_dict(cls, req_dict: dict[str, Any]) -> SamplingParams:
         )
 
     @classmethod
-    def from_optional(
-        cls,
-        n,
-        best_of,
-        presence_penalty,
-        frequency_penalty,
-        repetition_penalty,
-        temperature,
-        top_p,
-        top_k,
-        min_p,
-        seed=None,
-        stop=None,
-        stop_token_ids=None,
-        max_tokens=None,
-        reasoning_max_tokens=None,
-        min_tokens=1,
-        logprobs=None,
-        bad_words=None,
-    ) -> SamplingParams:
+    def from_optional(cls,
+                      n,
+                      best_of,
+                      presence_penalty,
+                      frequency_penalty,
+                      repetition_penalty,
+                      temperature,
+                      top_p,
+                      top_k,
+                      min_p,
+                      seed=None,
+                      stop=None,
+                      stop_token_ids=None,
+                      max_tokens=None,
+                      reasoning_max_tokens=None,
+                      min_tokens=1,
+                      logprobs=None,
+                      bad_words=None) -> "SamplingParams":
         """Create instance from command line arguments"""
-        return cls(
-            n=1 if n is None else n,
-            best_of=best_of,
-            presence_penalty=(presence_penalty if presence_penalty is not None else 0.0),
-            frequency_penalty=(frequency_penalty if frequency_penalty is not None else 0.0),
-            repetition_penalty=(repetition_penalty if repetition_penalty is not None else 1.0),
-            temperature=temperature if temperature is not None else 1.0,
-            top_p=top_p,
-            top_k=top_k if top_k is not None else 0,
-            min_p=min_p if min_p is not None else 0.0,
-            seed=seed,
-            stop=stop,
-            stop_token_ids=stop_token_ids,
-            max_tokens=max_tokens if max_tokens is not None else 8192,
-            reasoning_max_tokens=reasoning_max_tokens,
-            min_tokens=min_tokens,
-            logprobs=logprobs,
-            bad_words=bad_words,
-        )
+        return cls(n=1 if n is None else n,
+                   best_of=best_of,
+                   presence_penalty=presence_penalty
+                   if presence_penalty is not None else 0.0,
+                   frequency_penalty=frequency_penalty
+                   if frequency_penalty is not None else 0.0,
+                   repetition_penalty=repetition_penalty
+                   if repetition_penalty is not None else 1.0,
+                   temperature=temperature if temperature is not None else 1.0,
+                   top_p=top_p,
+                   top_k=top_k if top_k is not None else 0,
+                   min_p=min_p if min_p is not None else 0.0,
+                   seed=seed,
+                   stop=stop,
+                   stop_token_ids=stop_token_ids,
+                   max_tokens=max_tokens if max_tokens is not None else 8192,
+                   reasoning_max_tokens=reasoning_max_tokens,
+                   min_tokens=min_tokens,
+                   logprobs=logprobs,
+                   bad_words=bad_words)
 
     def __post_init__(self):
-        if self.seed is None:
-            self.seed = random.randint(0, 922337203685477580)
+        if self.seed == -1:
+            self.seed = None
         if self.max_tokens is not None and self.reasoning_max_tokens is None:
             self.reasoning_max_tokens = max(int(self.max_tokens * 0.8), 1)
         self._verify_args()
@@ -175,7 +184,8 @@ def _verify_args(self) -> None:
         if self.top_k < -1:
             raise ValueError(f"top_k must be 0 (disable), or at least 1, " f"got {self.top_k}.")
         if not isinstance(self.top_k, int):
-            raise TypeError(f"top_k must be an integer, got {type(self.top_k).__name__}")
+            raise TypeError(
+                f"top_k must be an integer, got {type(self.top_k).__name__}")
         if not 0.0 <= self.min_p <= 1.0:
             raise ValueError("min_p must be in [0,1],got f{self.min_p}")
 
@@ -194,10 +204,16 @@ def _verify_args(self) -> None:
         if self.logprobs is not None and self.logprobs < 0:
             raise ValueError(f"logprobs must be non-negative, got {self.logprobs}.")
         if self.logprobs is not None and self.logprobs > 20:
-            raise ValueError("Invalid value for 'top_logprobs': must be less than or equal to 20.")
-
-        if not 0 <= self.seed <= 922337203685477580:
-            raise ValueError("seed must be in [0, 922337203685477580], got " f"{self.seed}.")
+            raise ValueError(
+                "Invalid value for 'top_logprobs': must be less than or equal to 20.")
+        
+    @cached_property
+    def sampling_type(self)->SamplingType:
+        if self.temperature < _SAMPLING_EPS:
+            return SamplingType.GREEDY
+        if self.seed is not None:
+            return SamplingType.RANDOM_SEED
+        return SamplingType.RANDOM
 
     def update_from_tokenizer(self, tokenizer):
         """
diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
index 432a12ddff0..b4486801756 100644
--- a/fastdeploy/worker/gpu_model_runner.py
+++ b/fastdeploy/worker/gpu_model_runner.py
@@ -78,11 +78,14 @@ def __init__(
         self.enable_mm = self.model_config.enable_mm
         self.rank = rank
         self.local_rank = local_rank
+        self.generators={}
         self.device_id = device_id
         self.speculative_method = self.fd_config.speculative_config.method
         self.speculative_decoding = self.speculative_method is not None
         self.enable_logprob = fd_config.model_config.enable_logprob
 
+        self.gene
+
         self.guided_backend = None
         if self.fd_config.parallel_config.guided_decoding_backend != "off":
             self.guided_backend = get_guided_backend(fd_config=self.fd_config)
@@ -128,9 +131,7 @@ def __init__(
         self._init_share_inputs(self.parallel_config.max_num_seqs)
         self.infer_seed_increment = paddle.full(
             shape=[self.parallel_config.max_num_seqs, 1],
-            fill_value=4,
-            dtype="int64",
-        )
+            dtype="int64")
         self.restore_chunked_prefill_request = dict()
 
         # Initialize attention Backend
@@ -500,7 +501,7 @@ def _init_share_inputs(self, max_num_seqs: int):
         """
         Initialize all share buffers for model inputs.
         """
-        self.MAX_INFER_SEED = 9223372036854775806
+
         self.share_inputs = {}
 
         self.share_inputs["pre_ids"] = paddle.full(
@@ -925,9 +926,10 @@ def _dummy_run(
                     ids_remove_padding=self.share_inputs["ids_remove_padding"],
                     forward_meta=self.forward_meta,
                 )
+                logits = self.model.compute_logits(model_output)
 
                 hidden_states = rebuild_padding(
-                    model_output,
+                    logits,
                     self.share_inputs["cum_offsets"],
                     self.share_inputs["seq_lens_this_time"],
                     self.share_inputs["seq_lens_decoder"],

From bb92db2f995008ccf3343fe9415f6b6fe1f704bc Mon Sep 17 00:00:00 2001
From: lizexu123 <2694294196@qq.com>
Date: Wed, 30 Jul 2025 10:23:43 +0000
Subject: [PATCH 3/3] seed

---
 fastdeploy/config.py                            |  1 +
 fastdeploy/engine/args_utils.py                 | 17 +++++++----------
 .../model_executor/layers/sample/meta_data.py   |  2 --
 fastdeploy/utils.py                             |  9 +++++++++
 fastdeploy/worker/gpu_model_runner.py           |  8 +++-----
 fastdeploy/worker/gpu_worker.py                 |  4 +++-
 6 files changed, 23 insertions(+), 18 deletions(-)

diff --git a/fastdeploy/config.py b/fastdeploy/config.py
index c8428d1f974..0afb57c4b7c 100644
--- a/fastdeploy/config.py
+++ b/fastdeploy/config.py
@@ -116,6 +116,7 @@ def __init__(
         self.enable_redundant_experts = False
         self.redundant_experts_num = 0
         self.quantization = None
+        self.seed = 0
         for key, value in args.items():
             if hasattr(self, key):
                 setattr(self, key, value)
diff --git a/fastdeploy/engine/args_utils.py b/fastdeploy/engine/args_utils.py
index ada82c003f4..cff949e2747 100644
--- a/fastdeploy/engine/args_utils.py
+++ b/fastdeploy/engine/args_utils.py
@@ -313,7 +313,7 @@ class EngineArgs:
     Must be explicitly enabled via the `--enable-logprob` startup parameter to output logprob values.
     """
 
-    seed : Optional[int] = None
+    seed: Optional[int] = None
     """
     Random seed to use for initialization. If not set, a random seed is used.
     """
@@ -469,15 +469,12 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             default=EngineArgs.enable_logprob,
             help="Enable output of token-level log probabilities.",
         )
-        model_group.add_argument("--enable-logprob",
-                                 action="store_true",
-                                 default=EngineArgs.enable_logprob,
-                                 help="Enable output of token-level log probabilities."
-                                 )
-        model_group.add_argument("--seed",
-                                type=int,
-                                defualt=None,
-                                help="Random seed for initialization. If not specified, a random seed will be used.")
+        model_group.add_argument(
+            "--seed",
+            type=int,
+            default=None,
+            help="Random seed for initialization. If not specified, a random seed will be used.",
+        )
         # Parallel processing parameters group
         parallel_group = parser.add_argument_group("Parallel Configuration")
         parallel_group.add_argument(
diff --git a/fastdeploy/model_executor/layers/sample/meta_data.py b/fastdeploy/model_executor/layers/sample/meta_data.py
index 69b2e3e1987..f8781de77c4 100644
--- a/fastdeploy/model_executor/layers/sample/meta_data.py
+++ b/fastdeploy/model_executor/layers/sample/meta_data.py
@@ -27,13 +27,11 @@ class SamplingMetadata:
     """
 
     temperature: paddle.Tensor
-
     pre_token_ids: paddle.Tensor
     eos_token_ids: paddle.Tensor
     frequency_penalties: paddle.Tensor
     presence_penalties: paddle.Tensor
     repetition_penalties: paddle.Tensor
-
     min_dec_lens: paddle.Tensor
 
     bad_words_token_ids: paddle.Tensor
diff --git a/fastdeploy/utils.py b/fastdeploy/utils.py
index 0ceba665a20..29328fa4e03 100644
--- a/fastdeploy/utils.py
+++ b/fastdeploy/utils.py
@@ -29,6 +29,8 @@
 from pathlib import Path
 from typing import Literal, TypeVar, Union
 
+import numpy as np
+import paddle
 import requests
 import yaml
 from aistudio_sdk.snapshot_download import snapshot_download as aistudio_download
@@ -291,6 +293,13 @@ def extract_tar(tar_path, output_dir):
         raise RuntimeError(f"Extraction failed: {e!s}")
 
 
+def set_random_seed(seed: int) -> None:
+    if seed is not None:
+        random.seed(seed)
+        np.random.seed(seed)
+        paddle.seed(seed)
+
+
 def download_model(url, output_dir, temp_tar):
     """
     下载模型，并将其解压到指定目录。
diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
index 2c3ddfc25d9..03b5cd69ef7 100644
--- a/fastdeploy/worker/gpu_model_runner.py
+++ b/fastdeploy/worker/gpu_model_runner.py
@@ -78,14 +78,11 @@ def __init__(
         self.enable_mm = self.model_config.enable_mm
         self.rank = rank
         self.local_rank = local_rank
-        self.generators={}
         self.device_id = device_id
         self.speculative_method = self.fd_config.speculative_config.method
         self.speculative_decoding = self.speculative_method is not None
         self.enable_logprob = fd_config.model_config.enable_logprob
 
-        self.gene
-
         self.guided_backend = None
         if self.fd_config.parallel_config.guided_decoding_backend != "off":
             self.guided_backend = get_guided_backend(fd_config=self.fd_config)
@@ -130,8 +127,8 @@ def __init__(
         # Initialize share inputs
         self._init_share_inputs(self.parallel_config.max_num_seqs)
         self.infer_seed_increment = paddle.full(
-            shape=[self.parallel_config.max_num_seqs, 1],
-            dtype="int64")
+            shape=[self.parallel_config.max_num_seqs, 1], fill_value=4, dtype="int64"
+        )
         self.restore_chunked_prefill_request = dict()
 
         # Initialize attention Backend
@@ -270,6 +267,7 @@ def insert_tasks_v1(self, req_dicts: List[Request]):
             self.share_inputs["first_token_ids"][idx : idx + 1] = self.share_inputs["input_ids"][idx : idx + 1, :1]
             self.share_inputs["ori_seq_lens_encoder"][idx : idx + 1] = length
 
+            print("seed", request.get("seed"))
             if request.get("seed") is not None:
                 self.share_inputs["infer_seed"][idx : idx + 1] = request.get("seed")
 
diff --git a/fastdeploy/worker/gpu_worker.py b/fastdeploy/worker/gpu_worker.py
index 084b4f0f2d8..c1c8552444a 100644
--- a/fastdeploy/worker/gpu_worker.py
+++ b/fastdeploy/worker/gpu_worker.py
@@ -26,7 +26,7 @@
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
 from fastdeploy.platforms import current_platform
-from fastdeploy.utils import get_logger
+from fastdeploy.utils import get_logger, set_random_seed
 from fastdeploy.worker.gpu_model_runner import GPUModelRunner
 from fastdeploy.worker.output import ModelRunnerOutput
 from fastdeploy.worker.worker_base import WorkerBase
@@ -69,6 +69,7 @@ def init_device(self):
         else:
             raise RuntimeError(f"Not support device type: {self.device_config.device}")
 
+        set_random_seed(self.fd_config.model_config.seed)
         # Construct model runner
         self.model_runner: GPUModelRunner = GPUModelRunner(
             fd_config=self.fd_config,
@@ -123,6 +124,7 @@ def determine_available_memory(self) -> int:
 
         # 2. Profile run
         self.model_runner.profile_run()
+        set_random_seed(self.fd_config.model_config.seed)
 
         # 3. Statistical memory information
         paddle_reserved_mem_after_run = paddle.device.cuda.max_memory_reserved(local_rank)