diff --git a/python/sglang/benchmark/datasets/common.py b/python/sglang/benchmark/datasets/common.py
index 29bafff6b30c..e606e9ed8827 100644
--- a/python/sglang/benchmark/datasets/common.py
+++ b/python/sglang/benchmark/datasets/common.py
@@ -67,11 +67,11 @@ def compute_random_lens(full_len: int, range_ratio: float, num: int) -> List[int
 @lru_cache(maxsize=1)
 def get_available_tokens(tokenizer):
     """Get valid token ids from the tokenizer vocabulary."""
-    return [
+    return sorted(
         token_id
         for token_id in tokenizer.get_vocab().values()
         if isinstance(token_id, int)
-    ]
+    )
 
 
 def gen_prompt(tokenizer, token_num):
diff --git a/python/sglang/benchmark/datasets/image.py b/python/sglang/benchmark/datasets/image.py
index 5deffe0a4ea5..f44a0a786384 100644
--- a/python/sglang/benchmark/datasets/image.py
+++ b/python/sglang/benchmark/datasets/image.py
@@ -1,4 +1,5 @@
 import io
+import random
 import warnings
 from argparse import Namespace
 from dataclasses import dataclass
@@ -30,6 +31,7 @@ class ImageDataset(BaseDataset):
     image_resolution: str
     backend: str
     random_image_count: bool
+    seed: int
 
     @classmethod
     def from_args(cls, args: Namespace) -> "ImageDataset":
@@ -44,10 +46,15 @@ def from_args(cls, args: Namespace) -> "ImageDataset":
             image_resolution=args.image_resolution,
             backend=args.backend,
             random_image_count=args.random_image_count,
+            seed=args.seed,
         )
 
     def load(self, tokenizer=None, model_id=None) -> List[DatasetRow]:
         processor = get_processor(model_id)
+        # Processor initialization may consume global RNG state. Reset it here so
+        # --seed fixes the generated prompts, image sizes, and image contents.
+        random.seed(self.seed)
+        np.random.seed(self.seed)
         return sample_image_requests(
             num_requests=self.num_requests,
             image_count=self.image_count,
diff --git a/test/registered/bench_fn/test_benchmark_datasets_api.py b/test/registered/bench_fn/test_benchmark_datasets_api.py
index 4c7c3d7f4f8f..b25a337d786d 100644
--- a/test/registered/bench_fn/test_benchmark_datasets_api.py
+++ b/test/registered/bench_fn/test_benchmark_datasets_api.py
@@ -34,6 +34,7 @@
     sample_generated_shared_prefix_requests,
 )
 from sglang.benchmark.datasets.image import (
+    ImageDataset,
     parse_random_image_resolution,
     sample_image_requests,
 )
@@ -479,6 +480,43 @@ def test_image_sampler_random_resolution(self):
             self.assertGreaterEqual(height, 8)
             self.assertLessEqual(height, 16)
 
+    def test_image_dataset_seed_is_independent_of_processor_initialization(self):
+        dataset = ImageDataset.from_args(
+            make_args(
+                num_prompts=3,
+                image_resolution="random:8x16-16x32",
+                seed=20260717,
+            )
+        )
+        processor_init_count = 0
+
+        def get_processor_with_rng_side_effects(_model_id):
+            nonlocal processor_init_count
+            processor_init_count += 1
+            random.random()
+            np.random.random(processor_init_count)
+            return self.processor
+
+        with patch(
+            "sglang.benchmark.datasets.image.get_processor",
+            side_effect=get_processor_with_rng_side_effects,
+        ):
+            first = dataset.load(model_id="test-model")
+            random.seed(999)
+            np.random.seed(999)
+            second = dataset.load(model_id="test-model")
+
+        self.assertEqual(
+            [
+                (row.prompt, row.prompt_len, row.output_len, row.image_data)
+                for row in first
+            ],
+            [
+                (row.prompt, row.prompt_len, row.output_len, row.image_data)
+                for row in second
+            ],
+        )
+
     def test_parse_random_image_resolution(self):
         self.assertEqual(
             parse_random_image_resolution("random:256x384-1024x1536"),
@@ -520,6 +558,30 @@ def fake_choices(population, k):
         self.assertFalse(special_token_ids & sampled_pool)
         self.assertTrue(sampled_pool)
 
+    def test_gen_mm_prompt_is_independent_of_vocab_order(self):
+        class OrderedVocabTokenizer:
+            all_special_ids = []
+
+            def __init__(self, items):
+                self.vocab = dict(items)
+
+            def get_vocab(self):
+                return self.vocab
+
+            def decode(self, token_ids):
+                return " ".join(map(str, token_ids))
+
+        items = [(f"token_{token_id}", token_id) for token_id in range(32)]
+        first = OrderedVocabTokenizer(items)
+        second = OrderedVocabTokenizer(reversed(items))
+
+        random.seed(20260717)
+        first_prompt = gen_mm_prompt(first, image_pad_id=None, token_num=16)
+        random.seed(20260717)
+        second_prompt = gen_mm_prompt(second, image_pad_id=None, token_num=16)
+
+        self.assertEqual(first_prompt, second_prompt)
+
     def test_mmmu_sampler(self):
         fake_records = [
             {"image_1": Image.new("RGB", (4, 4), color="white"), "question": "q1"},