Fix the hash weight of HashsEncoder

Author: AooooooA-C

docs/source/getting-started/quickstart_vllm.md

@@ -149,7 +149,7 @@ You may directly edit the example file at `unified-cache-management/examples/ucm
 
 ### Feature 2:  Sparsity
 
-The sparse module was not compiled by default. To enable it, set the environment variable `export ENABLE_SPARSE=TRUE` and re-compile the code you built. And uncomment `ucm_sparse_config` code block in `unified-cache-management/examples/ucm_config_example.yaml`. Additionally, if you want to run GSAOnDevice, you also need to set the environment variable `export VLLM_HASH_ATTENTION=1`.
+The sparse module was not compiled by default. To enable it, set the environment variable `export ENABLE_SPARSE=TRUE` and re-compile the code you built. And uncomment `ucm_sparse_config` code block in `unified-cache-management/examples/ucm_config_example.yaml`. Additionally, if you want to run GSAOnDevice, you also need to set the environment variable `export VLLM_HASH_ATTENTION=1`, and you can fix the hash weight of the HashEncoder by setting the environment variable `export HASHENCODERSEED=<the seed number>`.
 
 ## Step 3: Launching Inference
 

ucm/sparse/gsa_on_device/hash_encoder.py

@@ -22,6 +22,7 @@
 THE SOFTWARE.
 """
 
+import os
 import torch
 
 if hasattr(torch, "npu") and torch.npu.is_available():
@@ -317,6 +318,13 @@ def __init__(
 
     def _init_hash_weights(self):
         # Step 1: 随机高斯矩阵
+        seed_str =  os.getenv("HASHENCODERSEED")
+        if seed_str is not None and seed_str.strip() != "":
+            seed = int(seed_str)
+            torch.manual_seed(seed)
+            if self.device.type == "cuda":
+                torch.cuda.manual_seed(seed)
+                torch.cuda.manual_seed_all(seed)
         random_weights = torch.normal(
             mean=0,
             std=2,