sglang: bump to 0.5.11 and prune 0.5.9 fallbacks in _compat

ishandhanani · ishandhanani · commit 2224cb161ebc · 2026-05-06T21:50:18.000Z
All 9 testable launch scripts in examples/backends/sglang/launch/ pass on
0.5.11 with no handler/init code changes. (disagg_router needs 4 GPUs and
multimodal_disagg needs 3 GPUs; both skipped on the test box's 2x L40S.)

The new support window is N=0.5.11, N-1=0.5.10, so the 0.5.9-targeted
branches in _compat.py come out as part of this bump per the
N/N-1 policy in components/src/dynamo/sglang/CLAUDE.md.

Removed:
* `NetworkAddress` polyfill class + `try/except` around its import
  (`sglang.srt.utils.network` is canonical from 0.5.10 onward).
* `mm_encode()` wrapper. Both 0.5.10 and 0.5.11 take
  `_encode(mm_items, modality)` and return the same 3-tuple, so the
  call sites in `encode_worker_handler.py` now invoke
  `await self.encoder._encode(...)` directly.
* `enable_disjoint_streaming_output` `stream_output` fallback. The
  field has been `incremental_streaming_output` since at-or-before
  0.5.10 (verified: `ServerArgs.__dataclass_fields__` in 0.5.11 has
  `incremental_streaming_output` and not `stream_output`). The
  wrapper itself stays — diffusion `SimpleNamespace` stubs need the
  attribute-absent no-op path.

Kept (not version-bound):
* `ensure_sglang_top_level_exports()`, `filter_supported_async_generate_kwargs`,
  `get_scheduler_info` (the latter still probes fork/experimental
  attribute paths).

CLAUDE.md example version bumped to reflect 0.5.10 as the new N-1.
diff --git a/components/src/dynamo/sglang/CLAUDE.md b/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 <= 0.5.9. Remove when min supported version is 0.6.0+`
+   `# Fallback for sglang <= 0.5.10. Remove when min supported version is 0.5.12+`
 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.
@@ -341,7 +341,7 @@ Checklist for adding a new worker (e.g., a new modality or serving mode):
 
 ```
 sglang/
-  _compat.py               # SGLang version compat shim (network imports + NetworkAddress polyfill)
+  _compat.py               # SGLang version compat shim (signature probing for async_generate kwargs)
   __main__.py              # Entry point
   main.py                  # Worker dispatch
   args.py                  # Config parsing (ServerArgs vs SimpleNamespace)
diff --git a/components/src/dynamo/sglang/_compat.py b/components/src/dynamo/sglang/_compat.py
@@ -16,12 +16,16 @@
 """
 
 import inspect
-import ipaddress
 import logging
-import socket
 from functools import lru_cache
 from typing import Any
 
+from sglang.srt.utils.network import (  # noqa: F401 — re-exported for callers
+    NetworkAddress,
+    get_local_ip_auto,
+    get_zmq_socket,
+)
+
 logger = logging.getLogger(__name__)
 
 
@@ -105,117 +109,6 @@ def filter_supported_async_generate_kwargs(
     return {key: value for key, value in kwargs.items() if key in supported_kwarg_names}
 
 
-# ---------------------------------------------------------------------------
-# Network utilities: NetworkAddress, get_local_ip_auto, get_zmq_socket
-#
-# 0.5.10+: sglang.srt.utils.network (canonical)
-# 0.5.9:   sglang.srt.utils (get_local_ip_auto, get_zmq_socket only;
-#           NetworkAddress did not exist)
-# ---------------------------------------------------------------------------
-try:
-    from sglang.srt.utils.network import (  # noqa: F401
-        NetworkAddress,
-        get_local_ip_auto,
-        get_zmq_socket,
-    )
-except ImportError:
-    # Fallback for sglang 0.5.9. Remove when min supported version is 0.5.10+
-    from sglang.srt.utils import (  # type: ignore[no-redef]  # noqa: F401
-        get_local_ip_auto,
-        get_zmq_socket,
-    )
-
-    logger.info(
-        "sglang.srt.utils.network not found (sglang 0.5.9); "
-        "using compatibility shim for NetworkAddress"
-    )
-
-    class NetworkAddress:  # type: ignore[no-redef]
-        """Minimal polyfill for sglang.srt.utils.network.NetworkAddress."""
-
-        def __init__(self, host: str, port: int) -> None:
-            self.host = host
-            self.port = port
-
-        @property
-        def is_ipv6(self) -> bool:
-            try:
-                ipaddress.IPv6Address(self.host)
-                return True
-            except ValueError:
-                return False
-
-        @classmethod
-        def parse(cls, addr: str) -> "NetworkAddress":
-            """Parse 'host:port', '[IPv6]:port', or bare host."""
-            addr = addr.strip()
-            if addr.startswith("["):
-                end = addr.find("]")
-                host = addr[1:end] if end != -1 else addr.strip("[]")
-                rest = addr[end + 1 :] if end != -1 else ""
-                if rest.startswith(":") and rest[1:].isdigit():
-                    return cls(host, int(rest[1:]))
-                return cls(host, 0)
-            if addr.count(":") == 1:
-                host_part, port_part = addr.rsplit(":", 1)
-                if port_part.isdigit():
-                    return cls(host_part, int(port_part))
-            return cls(addr, 0)
-
-        def resolved(self) -> "NetworkAddress":
-            """DNS-resolve the host, preserving port."""
-            try:
-                infos = socket.getaddrinfo(
-                    self.host, None, family=socket.AF_UNSPEC, type=socket.SOCK_STREAM
-                )
-                resolved_ip = infos[0][4][0]
-                return NetworkAddress(resolved_ip, self.port)
-            except socket.gaierror:
-                return self
-
-        def to_host_port_str(self) -> str:
-            """Return '[IPv6]:port' or 'host:port'."""
-            if self.is_ipv6:
-                return f"[{self.host}]:{self.port}"
-            return f"{self.host}:{self.port}"
-
-        def to_tcp(self) -> str:
-            """Return 'tcp://[IPv6]:port' or 'tcp://host:port'."""
-            if self.is_ipv6:
-                return f"tcp://[{self.host}]:{self.port}"
-            return f"tcp://{self.host}:{self.port}"
-
-
-# ---------------------------------------------------------------------------
-# MMEncoder._encode() adapter
-#
-# 0.5.10+: _encode(mm_items, modality) -> (grid_dim, embedding, aux_data)
-# 0.5.9:   _encode(mm_items)           -> (grid_dim, embedding)
-#
-# Imports are deferred to avoid pulling sgl_kernel (CUDA-only) at module
-# level, which breaks test collection on arm64 CPU-only CI nodes.
-# ---------------------------------------------------------------------------
-
-
-async def mm_encode(encoder: Any, mm_items: Any, modality: Any) -> tuple:
-    """Version-safe wrapper around MMEncoder._encode().
-
-    Always returns (grid_dim, embedding, aux_data). On sglang 0.5.9
-    _encode takes no modality arg and returns a 2-tuple; on 0.5.10+ it
-    takes modality and returns a 3-tuple. We try the new signature first
-    and fall back to the old one.
-    """
-    try:
-        result = await encoder._encode(mm_items, modality)
-    except TypeError:
-        # sglang 0.5.9: _encode(mm_items) -> (grid_dim, embedding)
-        result = await encoder._encode(mm_items)
-
-    if len(result) == 2:
-        return (*result, None)
-    return result
-
-
 def get_scheduler_info(engine: Any) -> dict:
     """Return the scheduler-info dict for rank-0 of an ``sgl.Engine``.
 
@@ -249,36 +142,13 @@ def get_scheduler_info(engine: Any) -> dict:
 
 
 def enable_disjoint_streaming_output(server_args: Any) -> None:
-    """
-    Enable SGLang's disjoint streaming output across ServerArgs field renames.
+    """Enable SGLang's disjoint streaming output.
 
-    Covers sglang <= 0.5.x (`stream_output`) and newer releases
-    (`incremental_streaming_output`).
+    Diffusion workers pass a ``SimpleNamespace`` stub that does not carry the
+    field, so this is a no-op when the attribute is absent.
     """
-    fields = getattr(type(server_args), "__dataclass_fields__", None)
-    if isinstance(fields, dict):
-        if "incremental_streaming_output" in fields:
-            server_args.incremental_streaming_output = True
-            return
-        if "stream_output" in fields:
-            server_args.stream_output = True
-            return
-        raise AttributeError(
-            "SGLang ServerArgs has neither 'incremental_streaming_output' nor "
-            "'stream_output'"
-        )
-
     if hasattr(server_args, "incremental_streaming_output"):
         server_args.incremental_streaming_output = True
-        return
-    if hasattr(server_args, "stream_output"):
-        server_args.stream_output = True
-        return
-
-    logger.debug(
-        "Skipping streaming output compatibility for non-ServerArgs object: %s",
-        type(server_args).__name__,
-    )
 
 
 __all__ = [
@@ -289,5 +159,4 @@ def enable_disjoint_streaming_output(server_args: Any) -> None:
     "get_local_ip_auto",
     "get_scheduler_info",
     "get_zmq_socket",
-    "mm_encode",
 ]
diff --git a/components/src/dynamo/sglang/request_handlers/multimodal/encode_worker_handler.py b/components/src/dynamo/sglang/request_handlers/multimodal/encode_worker_handler.py
@@ -26,7 +26,6 @@
 )
 from dynamo.common.multimodal import EMBEDDING_SENDER_FACTORIES
 from dynamo.common.utils import nvtx_utils as _nvtx
-from dynamo.sglang._compat import mm_encode
 from dynamo.sglang.args import Config
 from dynamo.sglang.protocol import (
     MultiModalGroup,
@@ -218,8 +217,8 @@ async def _encode_with_cache(
         # SGLang's _encode outputs are already on CPU; use CPU as target for consistency
         target_device = torch.device("cpu")
         if uncached_urls:
-            grid_dim, new_embeddings, _aux = await mm_encode(
-                self.encoder, uncached_urls, Modality.IMAGE
+            grid_dim, new_embeddings, _aux = await self.encoder._encode(
+                uncached_urls, Modality.IMAGE
             )
             # Verify SGLang output is on CPU as expected
             if new_embeddings.device != target_device:
@@ -344,7 +343,7 @@ async def generate(
                         image_grid_dim,
                         precomputed_embeddings,
                         _aux,
-                    ) = await mm_encode(self.encoder, image_urls, Modality.IMAGE)
+                    ) = await self.encoder._encode(image_urls, Modality.IMAGE)
 
             image_grid_thw_list = (
                 image_grid_dim.tolist()
diff --git a/pyproject.toml b/pyproject.toml
@@ -64,7 +64,7 @@ vllm = [
 
 sglang = [
     "uvloop",
-    "sglang[diffusion]==0.5.10.post1",
+    "sglang[diffusion]==0.5.11",
     "nixl[cu12]>=1.0.0,<1.1.0",
     "cupy-cuda12x>=13.0.0",
 ]

Original file line number	Diff line number	Diff line change
`@@ -64,7 +64,7 @@ vllm = [`
`64`	`64`
`65`	`65`	`sglang = [`
`66`	`66`	`"uvloop",`
`67`		`- "sglang[diffusion]==0.5.10.post1",`
	`67`	`+ "sglang[diffusion]==0.5.11",`
`68`	`68`	`"nixl[cu12]>=1.0.0,<1.1.0",`
`69`	`69`	`"cupy-cuda12x>=13.0.0",`
`70`	`70`	`]`