Reserve multimodal runtime allocations and keep padded inputs aligned (#34141)

Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com>
Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com>
Co-authored-by: wangwenchen0407 <wangwenchen@meta.com>
Co-authored-by: Hanming Lu <hanminglu@meta.com>

python/sglang/srt/mem_cache/kv_cache_configurator.py

@@ -114,6 +114,26 @@ def _should_enable_lazy_compaction() -> bool:
     return not envs.SGLANG_DISABLE_LAZY_COMPACTION.get()
 
 
+def mm_runtime_reservation_gb(
+    *, is_multimodal: bool, mm_feature_transport: Optional[str]
+) -> float:
+    """Multimodal GPU memory allocated only after the KV pool is sized
+    (mm embedding cache + GPU feature-transport pools); reserve it out of
+    the KV budget so it doesn't have to fit in the runtime slack."""
+    if not is_multimodal:
+        return 0.0
+    reserved_mb = envs.SGLANG_VLM_CACHE_SIZE_MB.get()
+    if mm_feature_transport in ("cuda_ipc", "cuda_vmm"):
+        reserved_mb += envs.SGLANG_MM_FEATURE_CACHE_MB.get()
+    if reserved_mb > 0:
+        logger.info(
+            "Reserving %.2f GB of the KV budget for post-sizing multimodal "
+            "allocations (feature-transport pools + embedding cache).",
+            reserved_mb / 1024,
+        )
+    return reserved_mb / 1024
+
+
 # base ratio of mamba pool size to max_running_requests. Under
 # SGLANG_OPT_MAMBA_SKIP_DECODE_LOCK the decode-time skip frees one resident slot
 # per running request, so the base drops by 1 (overlap 5->4, lazy 4->3). no_buffer
@@ -1732,7 +1752,11 @@ class KVCacheConfigurator:
                 )
                 / 1024,
             )
-        rest_memory = available_gpu_memory - slack_gb
+        mm_reservation_gb = mm_runtime_reservation_gb(
+            is_multimodal=self.model_config.is_multimodal,
+            mm_feature_transport=self.server_args.mm_feature_transport,
+        )
+        rest_memory = available_gpu_memory - slack_gb - mm_reservation_gb
         if self.mambaish_config is not None:
             rest_memory = self._handle_max_mamba_cache(rest_memory)
 

python/sglang/srt/model_executor/forward_batch_info.py

@@ -1439,6 +1439,9 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
         # padding
         self._original_num_tokens = self.positions.shape[0]
         self.input_ids = self._pad_tensor_to_size(self.input_ids, num_tokens)
+        if self.input_embeds is not None:
+            # Keep token-aligned inputs consistent after padding.
+            self.input_embeds = self._pad_tensor_to_size(self.input_embeds, num_tokens)
         self.req_pool_indices = self._pad_tensor_to_size(self.req_pool_indices, bs)
         if self.lora_ids is not None:
             self.lora_ids.extend((bs - len(self.lora_ids)) * [None])

python/sglang/srt/model_executor/model_runner_components/kv_pool_runtime.py

@@ -8,6 +8,7 @@ import torch
 
 from sglang.srt.configs.hybrid_arch import mambaish_config
 from sglang.srt.distributed import get_world_group
+from sglang.srt.mem_cache.kv_cache_configurator import mm_runtime_reservation_gb
 from sglang.srt.model_executor.cuda_graph_config import Backend
 from sglang.srt.platforms import current_platform
 from sglang.srt.utils.common import get_available_gpu_memory, get_device_memory_capacity
@@ -76,8 +77,12 @@ def compute_post_capture_kv_resize(
             )
             / 1024,
         )
+    mm_reservation_gb = mm_runtime_reservation_gb(
+        is_multimodal=model_runner.model_config.is_multimodal,
+        mm_feature_transport=model_runner.server_args.mm_feature_transport,
+    )
     budget_bytes = (
-        int(max(0.0, free_gb - headroom_gb) * (1 << 30))
+        int(max(0.0, free_gb - headroom_gb - mm_reservation_gb) * (1 << 30))
         + pool.post_capture_backed_bytes
     )
     config = model_runner.kv_cache_configurator.config_from_budget(

test/registered/vlm/test_vision_openai_server_a.py

@@ -8,6 +8,7 @@ import unittest
 
 import openai
 
+from sglang.srt.environ import envs
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -46,6 +47,11 @@ class TestQwen3VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
     model = "Qwen/Qwen3-VL-30B-A3B-Instruct"
     extra_args = ["--cuda-graph-max-bs-decode=4"]
 
+    @classmethod
+    def setUpClass(cls):
+        with envs.SGLANG_MM_FEATURE_CACHE_MB.override(512):
+            super().setUpClass()
+
 
 class TestQwen2VLContextLengthServer(CustomTestCase):
     # --context-length 300 is calibrated to this model's mm-token expansion: