chore(sglang): bump to 0.5.18

components/src/dynamo/sglang/CLAUDE.md

@@ -20,7 +20,7 @@ support the current version plus 1 version back (N and N-1). The pattern:
    enough surface area to cover what Dynamo actually calls.
 4. Each fallback branch in `_compat.py` MUST have a comment noting which SGLang
    version it supports and when it can be removed, e.g.:
-   `# Fallback for sglang <= 0.5.16. Remove when min supported version is 0.5.18+`
+   `# Fallback for sglang <= 0.5.17. Remove when min supported version is 0.5.19+`
 5. When a new SGLang version is released and the old N-1 falls outside the support
    window, delete the corresponding fallback branches and polyfills from `_compat.py`.
    If `_compat.py` becomes trivial re-exports, inline the imports and delete the file.
@@ -70,10 +70,11 @@ Worker dispatch (main.py:60-132):
    have `max_running_requests`, `dllm_algorithm_config`, or other LLM-specific fields.
    Use `getattr()` when accessing fields that may not exist on the stub.
 
-SGLang 0.5.17 makes a resolved `ServerArgs` unconditionally read-only. Apply Dynamo's
-post-resolution startup overrides through `_compat.override_server_args()`; control-plane
-updates after engine creation should use the tokenizer manager's update API instead of
-assigning fields on `server_args`.
+SGLang 0.5.17 and 0.5.18 make a resolved `ServerArgs` unconditionally read-only.
+Apply Dynamo's post-resolution, pre-publish startup updates through
+`_compat.override_server_args()`; it selects 0.5.17's `override()` or 0.5.18's
+`_late_resolution()` API. Control-plane updates after engine creation should use the
+tokenizer manager's update API instead of assigning fields on `server_args`.
 
 **DynamoConfig** combines `DynamoRuntimeConfig` (common flags like `--namespace`,
 `--output-modalities`, `--media-output-fs-url`) with `DynamoSGLangConfig` (sglang-specific

components/src/dynamo/sglang/_compat.py

@@ -41,7 +41,7 @@ def _warn_require_reasoning_unsupported() -> None:
 def ensure_sglang_tensor_image_size() -> None:
     """Allow SGLang's image-token resolver to handle decoded image tensors.
 
-    SGLang 0.5.13 through 0.5.17 assume every decoded image exposes the PIL
+    SGLang 0.5.13 through 0.5.18 assume every decoded image exposes the PIL
     ``height``/``width`` attributes. Its CUDA JPEG decoder instead returns a
     CHW tensor, causing multimodal requests to fall back to retokenization.
 
@@ -82,15 +82,20 @@ def resolve_image_token_counts(self: Any, images: list[Any]) -> list[int]:
 
 
 def override_server_args(server_args: Any, source: str, **fields: Any) -> None:
-    """Apply a post-resolution SGLang configuration update.
-
-    SGLang 0.5.17 makes ``ServerArgs`` unconditionally read-only after
-    resolution. Both supported CUDA releases expose ``ServerArgs.override`` as
-    the audited mutation API, so Dynamo must use it instead of assigning fields.
-    The separately pinned XPU image still uses SGLang 0.5.11, which predates
-    that API; preserve its legacy assignment behavior until its engine pin is
-    upgraded.
+    """Apply a post-resolution, pre-publish SGLang configuration update.
+
+    SGLang 0.5.18 replaced ``ServerArgs.override`` with
+    ``ServerArgs._late_resolution`` for launcher-stage updates that every holder
+    of the instance must observe. SGLang 0.5.17 exposes the former API. The
+    separately pinned XPU image still uses SGLang 0.5.11, which predates both;
+    preserve its legacy assignment behavior until its engine pin is upgraded.
     """
+    late_resolution = getattr(server_args, "_late_resolution", None)
+    if callable(late_resolution):
+        late_resolution(source, **fields)
+        return
+
+    # Fallback for SGLang 0.5.17. Remove when minimum supported SGLang is 0.5.18+.
     override = getattr(server_args, "override", None)
     if callable(override):
         override(source, **fields)

components/src/dynamo/sglang/init_llm.py

@@ -14,7 +14,6 @@
 from dynamo.common.utils.endpoint_types import parse_endpoint_types
 from dynamo.llm import ModelInput, ModelType, WorkerType
 from dynamo.runtime import DistributedRuntime
-from dynamo.sglang._compat import override_server_args
 from dynamo.sglang.args import Config
 from dynamo.sglang.health_check import (
     SglangDisaggHealthCheckPayload,
@@ -67,15 +66,9 @@ async def init_decode(
         engine = snapshot_engine
         load_time = 0.0
         if getattr(server_args, "enable_forward_pass_metrics", False):
-            logging.warning(
-                "Forward pass metrics disabled in snapshot mode: the engine was "
-                "created before the endpoint existed, so its FPM publisher bound "
-                "a different IPC path than the relay would subscribe to."
-            )
-            override_server_args(
-                server_args,
-                "dynamo.snapshot",
-                enable_forward_pass_metrics=False,
+            raise RuntimeError(
+                "Snapshot ServerArgs must disable forward-pass metrics before "
+                "engine creation"
             )
     else:
         set_forward_pass_metrics_worker_id(server_args, generate_endpoint)
@@ -227,15 +220,9 @@ async def init_prefill(
         engine = snapshot_engine
         load_time = 0.0
         if getattr(server_args, "enable_forward_pass_metrics", False):
-            logging.warning(
-                "Forward pass metrics disabled in snapshot mode: the engine was "
-                "created before the endpoint existed, so its FPM publisher bound "
-                "a different IPC path than the relay would subscribe to."
-            )
-            override_server_args(
-                server_args,
-                "dynamo.snapshot",
-                enable_forward_pass_metrics=False,
+            raise RuntimeError(
+                "Snapshot ServerArgs must disable forward-pass metrics before "
+                "engine creation"
             )
     else:
         set_forward_pass_metrics_worker_id(server_args, generate_endpoint)

components/src/dynamo/sglang/snapshot.py

@@ -139,7 +139,13 @@ async def prepare_snapshot_engine(
     # Enable memory_saver so GPU memory can be released for CRIU.
     # When using GMS, weights use VA-stable unmap/remap (no CPU backup); GMS
     # forbids enable_weights_cpu_backup. Otherwise use CPU backup for weights.
-    snapshot_overrides = {"enable_memory_saver": True}
+    snapshot_overrides = {
+        "enable_memory_saver": True,
+        # Snapshot engines are created before the Dynamo endpoint exists, so
+        # their FPM publisher cannot be wired to the relay. Disable it before
+        # SGLang publishes ServerArgs; 0.5.18 forbids changing it afterwards.
+        "enable_forward_pass_metrics": False,
+    }
     try:
         from gpu_memory_service.integrations.sglang import is_gms_active
 

components/src/dynamo/sglang/tests/test_sglang_unit.py

@@ -99,6 +99,27 @@ def test_diffusion_generator_kwargs_omits_unset_master_port():
     assert "master_port" not in kwargs
 
 
+def test_override_server_args_supports_sglang_0_5_17():
+    calls = []
+
+    class ServerArgs:
+        def override(self, source, **fields):
+            calls.append((source, fields))
+            for name, value in fields.items():
+                object.__setattr__(self, name, value)
+
+    server_args = ServerArgs()
+
+    override_server_args(
+        server_args,
+        "dynamo.test",
+        enable_memory_saver=True,
+    )
+
+    assert calls == [("dynamo.test", {"enable_memory_saver": True})]
+    assert server_args.enable_memory_saver is True
+
+
 def test_override_server_args_supports_legacy_xpu_pin():
     server_args = SimpleNamespace(enable_memory_saver=False)
 

docs/fern/components/releases.data.ts

@@ -71,7 +71,7 @@ export const CURRENT_TAG = "1.4.0";
 export const CURRENT_WHEEL = "1.4.0";
 
 export const MAIN_TOT: BackendPins = {
-  sglang: "0.5.17",
+  sglang: "0.5.18",
   trtllm: "1.3.0rc24",
   vllm: "0.27.1",
   nixlSglang: "1.3.2",
@@ -1177,7 +1177,7 @@ export const FEATURE_INTERACTIONS: BackendInteractions[] = [
       // KV Block Manager
       [{ status: "wip" }, { status: "wip" }, { status: "wip" }, { status: "na" }],
       // Multimodal
-      [{ status: "yes", label: "Supported serving patterns", note: "Supports aggregated EPD, E/PD, and E/P/D patterns. Traditional disaggregated EP/D is not supported.", source: "/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/sglang-multimodal" }, { status: "yes", label: "Image-aware routing on Dynamo's SGLang image", note: "Hash forwarding is upstream in SGLang 0.5.13+ and Dynamo pins 0.5.17, so the shipped image routes on image overlap. A custom build without that patch still serves the request but degrades to text-prefix routing.", source: "/dynamo/dev/multimodal/multimodal-kv-routing" }, { status: "na" }, { status: "wip" }, { status: "na" }],
+      [{ status: "yes", label: "Supported serving patterns", note: "Supports aggregated EPD, E/PD, and E/P/D patterns. Traditional disaggregated EP/D is not supported.", source: "/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/sglang-multimodal" }, { status: "yes", label: "Image-aware routing on Dynamo's SGLang image", note: "Hash forwarding is upstream in SGLang 0.5.13+ and Dynamo pins 0.5.18, so the shipped image routes on image overlap. A custom build without that patch still serves the request but degrades to text-prefix routing.", source: "/dynamo/dev/multimodal/multimodal-kv-routing" }, { status: "na" }, { status: "wip" }, { status: "na" }],
       // Request Migration
       [{ status: "yes" }, { status: "yes" }, { status: "yes" }, { status: "wip" }, { status: "yes" }, { status: "na" }],
       // Request Cancellation

docs/fern/pages/developer-guide/knowledge-base/modular-components/backends/sglang/multimodal.md

@@ -122,7 +122,7 @@ The launcher configures KV events on each worker and sets `--router-mode kv` wit
 The Dynamo SGLang image includes both routing prerequisites:
 
 - Dynamo is built with the `mm-routing` Rust feature.
-- SGLang 0.5.13 or later includes `GenerateReqInput.mm_hashes` support. Dynamo currently pins 0.5.17.
+- SGLang 0.5.13 or later includes `GenerateReqInput.mm_hashes` support. Dynamo currently pins 0.5.18.
 
 Custom installations on SGLang 0.5.12 or earlier need the `mm_hashes` change
 from [sgl-project/sglang#25300](https://github.com/sgl-project/sglang/pull/25300).

docs/fern/pages/reference/general/compatibility.mdx

@@ -156,7 +156,7 @@ Current stable release: v1.4.0 (container tag `1.4.0`, wheel version `1.4.0`).
 
 | Dynamo | Type | SGLang | TensorRT-LLM | vLLM | NIXL (SGL / TRT / vLLM) | UCX |
 | --- | --- | --- | --- | --- | --- | --- |
-| main (ToT) | development head | 0.5.17 | 1.3.0rc24 | 0.27.1 | 1.3.2 / 1.3.1 / 1.3.2 | - |
+| main (ToT) | development head | 0.5.18 | 1.3.0rc24 | 0.27.1 | 1.3.2 / 1.3.1 / 1.3.2 | - |
 | v1.4.0 | stable | 0.5.16 | 1.3.0rc22 | 0.26.0 | 1.3.0 / 1.3.1 / 1.3.2 | 1.21.x |
 | v1.3.1 | patch | 0.5.14 | 1.3.0rc19 | 0.23.0 | 1.3.2 / 1.0.1 / 1.1.0 | 1.20.x |
 | v1.3.0 | stable | 0.5.14 | 1.3.0rc19 | 0.23.0 | 1.3.0 / 1.0.1 / 1.1.0 | 1.20.x |
@@ -306,10 +306,10 @@ Each cell states whether the row feature works together with the column feature.
 | Feature | Disaggregated Serving | KV-Aware Routing | SLA-Based Planner | KV Block Manager | Multimodal | Request Migration | Request Cancellation | LoRA | Tool Calling | Speculative Decoding |
 | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
 | Disaggregated Serving | n/a | Yes | Yes | Experimental | Yes — Supports aggregated EPD, E/PD, and E/P/D patterns. Traditional disaggregated EP/D is not supported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/sglang-multimodal) | Yes | Experimental — Cancellation during remote prefill is not supported in disaggregated mode. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | Experimental — Code hooks exist, but examples and documentation are not yet available. |
-| KV-Aware Routing | Yes | n/a | Yes | Experimental | Yes — Hash forwarding is upstream in SGLang 0.5.13+ and Dynamo pins 0.5.17, so the shipped image routes on image overlap. A custom build without that patch still serves the request but degrades to text-prefix routing. (https://docs.nvidia.com/dynamo/dev/multimodal/multimodal-kv-routing) | Yes | Yes | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | Experimental |
+| KV-Aware Routing | Yes | n/a | Yes | Experimental | Yes — Hash forwarding is upstream in SGLang 0.5.13+ and Dynamo pins 0.5.18, so the shipped image routes on image overlap. A custom build without that patch still serves the request but degrades to text-prefix routing. (https://docs.nvidia.com/dynamo/dev/multimodal/multimodal-kv-routing) | Yes | Yes | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | Experimental |
 | SLA-Based Planner | Yes | Yes | n/a | Experimental | n/a | Yes | Yes | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | n/a |
 | KV Block Manager | Experimental | Experimental | Experimental | n/a | Experimental | Experimental | Experimental | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Experimental | Experimental |
-| Multimodal | Yes — Supports aggregated EPD, E/PD, and E/P/D patterns. Traditional disaggregated EP/D is not supported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/sglang-multimodal) | Yes — Hash forwarding is upstream in SGLang 0.5.13+ and Dynamo pins 0.5.17, so the shipped image routes on image overlap. A custom build without that patch still serves the request but degrades to text-prefix routing. (https://docs.nvidia.com/dynamo/dev/multimodal/multimodal-kv-routing) | n/a | Experimental | n/a | Yes | Experimental | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | n/a |
+| Multimodal | Yes — Supports aggregated EPD, E/PD, and E/P/D patterns. Traditional disaggregated EP/D is not supported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/sglang-multimodal) | Yes — Hash forwarding is upstream in SGLang 0.5.13+ and Dynamo pins 0.5.18, so the shipped image routes on image overlap. A custom build without that patch still serves the request but degrades to text-prefix routing. (https://docs.nvidia.com/dynamo/dev/multimodal/multimodal-kv-routing) | n/a | Experimental | n/a | Yes | Experimental | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | n/a |
 | Request Migration | Yes | Yes | Yes | Experimental | Yes | n/a | Yes | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | Experimental |
 | Request Cancellation | Experimental — Cancellation during remote prefill is not supported in disaggregated mode. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | Yes | Experimental | Experimental | Yes | n/a | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | Yes | n/a |
 | LoRA | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | n/a | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) | No — SGLang does not support LoRA in Dynamo, so every LoRA pairing is unsupported. (https://docs.nvidia.com/dynamo/dev/knowledge-base/modular-components/backends/sg-lang/overview) |

docs/fern/pages/reference/general/releases-machine-readable.mdx

@@ -15,7 +15,7 @@ Current stable release: v1.4.0 (Aug 14, 2026; container tag `1.4.0`, wheel versi
 
 | Version | Kind | Date | SGLang | TensorRT-LLM | vLLM | NIXL (SGL / TRT / vLLM) | UCX | Notes | Delta |
 | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
-| main (ToT) | development head | - | 0.5.17 | 1.3.0rc24 | 0.27.1 | 1.3.2 / 1.3.1 / 1.3.2 | - | - | - |
+| main (ToT) | development head | - | 0.5.18 | 1.3.0rc24 | 0.27.1 | 1.3.2 / 1.3.1 / 1.3.2 | - | - | - |
 | v1.4.0 | stable | Aug 14, 2026 | 0.5.16 | 1.3.0rc22 | 0.26.0 | 1.3.0 / 1.3.1 / 1.3.2 | 1.21.x | [release notes](https://docs.nvidia.com/dynamo/dev/reference/releases/v1-4-0) | Audit subsystem migrated into request trace (DYN_AUDIT_* honored as legacy aliases); HTTP header capture in trace records is an explicit fail-closed allowlist; deprecated multimodal worker flags and vLLM worker-role flags removed; runtime images no longer bundle software video decoders (H.264/H.265 decodes via NVDEC); UCX 1.21.x. |
 | v1.3.1 | patch | Aug 5, 2026 | 0.5.14 | 1.3.0rc19 | 0.23.0 | 1.3.2 / 1.0.1 / 1.1.0 | 1.20.x | [release notes](https://docs.nvidia.com/dynamo/dev/reference/releases/v1-3-0#v131) | Patch release. Fixes disaggregated SGLang serving over AWS EFA on GB200: the SGLang EFA runtime moves to NIXL 1.3.2 and all three EFA images to EFA Installer 1.49.0. Backend pins are otherwise unchanged from v1.3.0. |
 | v1.3.0 | stable | Jul 20, 2026 | 0.5.14 | 1.3.0rc19 | 0.23.0 | 1.3.0 / 1.0.1 / 1.1.0 | 1.20.x | [release notes](https://docs.nvidia.com/dynamo/dev/reference/releases/v1-3-0) | CUDA 12 container images discontinued; EFA variants retagged from -efa-amd64 to -efa (the images were already multi-arch — the old suffix was misleading); GA wheels published as 1.3.0.post1 (containers stay :1.3.0); UCX 1.20.x. |