[Feature] support seed parameter (#3161)

lizexu123 · web-flow · commit afff4d37ea1f · 2025-08-06T15:20:47.000+08:00
* support seed

* fix

* add SamplingMetadata seed test

* The next_tokens values are inconsistent!

* add air and rejection seed test

* fix

* add SamplingParams seed test

* fix seed=0

* Default to defualt

* fix

* fix args_utils

* fix review

* fix review

* fix

* fix

* add xpu,gcu,iluvatar support seed

* fix
diff --git a/fastdeploy/config.py b/fastdeploy/config.py
@@ -122,6 +122,7 @@ def __init__(
         self.enable_mm = False
         self.enable_redundant_experts = False
         self.redundant_experts_num = 0
+        self.seed = 0
         self.quantization = None
         for key, value in args.items():
             if hasattr(self, key):
diff --git a/fastdeploy/engine/args_utils.py b/fastdeploy/engine/args_utils.py
@@ -316,6 +316,11 @@ class EngineArgs:
     Must be explicitly enabled via the `--enable-logprob` startup parameter to output logprob values.
     """
 
+    seed: int = 0
+    """
+    Random seed to use for initialization. If not set, defaults to 0.
+    """
+
     enable_early_stop: bool = False
     """
     Flag to enable early stop. Default is False (disabled).
@@ -484,6 +489,12 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             default=EngineArgs.enable_logprob,
             help="Enable output of token-level log probabilities.",
         )
+        model_group.add_argument(
+            "--seed",
+            type=int,
+            default=EngineArgs.seed,
+            help="Random seed for initialization. If not specified, defaults to 0.",
+        )
         model_group.add_argument(
             "--enable-early-stop",
             action="store_true",
diff --git a/fastdeploy/model_executor/layers/sample/meta_data.py b/fastdeploy/model_executor/layers/sample/meta_data.py
@@ -43,6 +43,7 @@ class SamplingMetadata:
     top_p: paddle.Tensor
     top_k: Optional[paddle.Tensor] = None
     min_p: Optional[paddle.Tensor] = None
+    seed: Optional[paddle.Tensor] = None
     max_num_logprobs: Optional[int] = None
     enable_early_stop: Optional[int] = False
     stop_flags: Optional[paddle.Tensor] = None
diff --git a/fastdeploy/model_executor/layers/sample/sampler.py b/fastdeploy/model_executor/layers/sample/sampler.py
@@ -282,7 +282,9 @@ def forward_cuda(
 
         probs = min_p_sampling(probs, sampling_metadata.min_p)
 
-        _, next_tokens = top_k_top_p_sampling(probs, sampling_metadata.top_p, sampling_metadata.top_k)
+        _, next_tokens = top_k_top_p_sampling(
+            probs, sampling_metadata.top_p, sampling_metadata.top_k, seed=sampling_metadata.seed[0, 0]
+        )
 
         logprobs_tensors = (
             None if num_logprobs is None else self.gather_logprobs(raw_logprobs, num_logprobs, token_ids=next_tokens)
diff --git 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
@@ -295,6 +297,13 @@ 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/gcu_model_runner.py b/fastdeploy/worker/gcu_model_runner.py
@@ -540,6 +540,7 @@ def _prepare_inputs(self) -> None:
             top_p=self.share_inputs["top_p"],
             top_k=self.share_inputs["top_k"],
             min_p=self.share_inputs["min_p"],
+            seed=self.share_inputs["infer_seed"],
             step_idx=self.share_inputs["step_idx"],
             pre_token_ids=self.share_inputs["pre_ids"],
             prompt_ids=self.share_inputs["prompt_ids"],
diff --git a/fastdeploy/worker/gcu_worker.py b/fastdeploy/worker/gcu_worker.py
@@ -22,7 +22,7 @@
 
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
-from fastdeploy.utils import get_logger
+from fastdeploy.utils import get_logger, set_random_seed
 from fastdeploy.worker.gcu_model_runner import GCUModelRunner
 from fastdeploy.worker.output import ModelRunnerOutput
 from fastdeploy.worker.worker_base import WorkerBase
@@ -60,6 +60,7 @@ 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: GCUModelRunner = GCUModelRunner(
             fd_config=self.fd_config,
@@ -128,6 +129,7 @@ def graph_optimize_and_warm_up_model(self) -> None:
             self.model_runner.sot_warmup()
         # 2. Triger cuda grpah capture
         self.model_runner.capture_model()
+        set_random_seed(self.fd_config.model_config.seed)
 
     def check_health(self) -> bool:
         """ """
diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
@@ -131,6 +131,7 @@ def __init__(
             fill_value=4,
             dtype="int64",
         )
+
         self.restore_chunked_prefill_request = dict()
 
         # Initialize attention Backend
@@ -813,6 +814,7 @@ def _prepare_inputs(self) -> None:
             top_p=self.share_inputs["top_p"],
             top_k=self.share_inputs["top_k"],
             min_p=self.share_inputs["min_p"],
+            seed=self.share_inputs["infer_seed"],
             step_idx=self.share_inputs["step_idx"],
             pre_token_ids=self.share_inputs["pre_ids"],
             prompt_ids=self.share_inputs["prompt_ids"],
diff --git a/fastdeploy/worker/gpu_worker.py b/fastdeploy/worker/gpu_worker.py
@@ -27,7 +27,7 @@
 from fastdeploy.engine.request import Request
 from fastdeploy.platforms import current_platform
 from fastdeploy.plugins.model_runner import load_model_runner_plugins
-from fastdeploy.utils import get_logger
+from fastdeploy.utils import get_logger, set_random_seed
 from fastdeploy.worker.model_runner_base import ModelRunnerBase
 from fastdeploy.worker.output import ModelRunnerOutput
 from fastdeploy.worker.worker_base import WorkerBase
@@ -75,6 +75,7 @@ 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: ModelRunnerBase = ModelRunner(
             fd_config=self.fd_config,
@@ -129,6 +130,7 @@ 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)
diff --git a/fastdeploy/worker/iluvatar_model_runner.py b/fastdeploy/worker/iluvatar_model_runner.py
@@ -509,6 +509,7 @@ def _prepare_inputs(self) -> None:
             temperature=self.share_inputs["temperature"],
             top_p=self.share_inputs["top_p"],
             top_k=self.share_inputs["top_k"],
+            seed=self.share_inputs["seed"],
             step_idx=self.share_inputs["step_idx"],
             pre_token_ids=self.share_inputs["pre_ids"],
             prompt_ids=self.share_inputs["prompt_ids"],
diff --git a/fastdeploy/worker/iluvatar_worker.py b/fastdeploy/worker/iluvatar_worker.py
@@ -23,7 +23,7 @@
 
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
-from fastdeploy.utils import get_logger
+from fastdeploy.utils import get_logger, set_random_seed
 from fastdeploy.worker.iluvatar_model_runner import IluvatarModelRunner
 from fastdeploy.worker.output import ModelRunnerOutput
 from fastdeploy.worker.worker_base import WorkerBase
@@ -60,6 +60,7 @@ 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: IluvatarModelRunner = IluvatarModelRunner(
             fd_config=self.fd_config,
@@ -130,6 +131,7 @@ def graph_optimize_and_warm_up_model(self) -> None:
 
         # 2. Triger cuda grpah capture
         self.model_runner.capture_model()
+        set_random_seed(self.fd_config.model_config.seed)
 
     def check_health(self) -> bool:
         """ """
diff --git a/fastdeploy/worker/xpu_model_runner.py b/fastdeploy/worker/xpu_model_runner.py
@@ -677,6 +677,7 @@ def _prepare_inputs(self, is_dummy_run=False) -> None:
             top_p=self.share_inputs["top_p"],
             top_k=self.share_inputs["top_k"],
             min_p=self.share_inputs["min_p"],
+            seed=self.share_inputs["infer_seed"],
             step_idx=self.share_inputs["step_idx"],
             pre_token_ids=self.share_inputs["pre_ids"],
             frequency_penalties=self.share_inputs["frequency_score"],
diff --git a/fastdeploy/worker/xpu_worker.py b/fastdeploy/worker/xpu_worker.py
@@ -23,7 +23,7 @@
 from fastdeploy import envs
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request
-from fastdeploy.utils import get_logger
+from fastdeploy.utils import get_logger, set_random_seed
 from fastdeploy.worker.output import ModelRunnerOutput
 from fastdeploy.worker.worker_base import WorkerBase
 from fastdeploy.worker.xpu_model_runner import XPUModelRunner
@@ -60,6 +60,7 @@ 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: XPUModelRunner = XPUModelRunner(
             fd_config=self.fd_config,
@@ -110,6 +111,7 @@ def determine_available_memory(self) -> int:
 
         self.model_runner.prepare_profile()
         self.model_runner.profile_run()
+        set_random_seed(self.fd_config.model_config.seed)
 
         total_available_memory = int(total_memory * self.cache_config.gpu_memory_utilization)
         used_memory = xpu_get_used_global_memory(int(self.device_ids[self.local_rank]))
diff --git a/test/ci_use/Qwen2-7B-Instruct_offline/test_Qwen2-7B-Instruct_offline.py b/test/ci_use/Qwen2-7B-Instruct_offline/test_Qwen2-7B-Instruct_offline.py
@@ -191,6 +191,29 @@ def test_chat_completion(llm):
             pytest.fail(f"Chat case {i + 1} failed")
 
 
+def test_seed(llm):
+    """
+    Test chat completion with same seed
+    """
+    prompt = "请介绍下中国的四大发明，用一句话概述每个发明。"
+    sampling_params = SamplingParams(temperature=0.1, seed=1, max_tokens=100)
+    num_runs = 5
+
+    results = []
+    try:
+        for i in range(num_runs):
+            outputs = llm.generate(prompt, sampling_params)
+            results.append(outputs[0].outputs.text)
+
+        assert all([result == results[0] for result in results]), "Results are not identical."
+        print("All results are identical.")
+
+    except Exception:
+        print("Failed during prompt generation.")
+        traceback.print_exc()
+        pytest.fail("Prompt generation test failed")
+
+
 if __name__ == "__main__":
     """
     Main entry point for the test script.
diff --git a/test/ci_use/Qwen3-MoE/test_sampling_consistency.py b/test/ci_use/Qwen3-MoE/test_sampling_consistency.py
