Build Rust extensions on demand in source checkouts (#34994)

rust/sglang-mm/Cargo.toml

@@ -11,17 +11,17 @@ license.workspace = true
 # features: the wheel build needs the PyO3 bindings, which are NOT default (see
 # [features] below).
 [package.metadata.sglang]
-python-module = "sglang.srt.multimodal._core"
+python-module = "sglang.srt.rust_extensions._multimodal"
 debug = false
 features = ["python", "parallel"]
 
 [lib]
 # Unique per-crate artifact name (the shared workspace target/ dir cannot hold
-# two `lib_core` cdylibs). The importable Python module is still `_core`: the
-# name comes from the `#[pymodule]` entry point and the setuptools-rust
-# `target` in python/pyproject.toml, which renames the built artifact.
+# multiple `lib_core` cdylibs). The importable Python module is `_multimodal`: the
+# name comes from the `#[pymodule]` entry point and the setuptools-rust target,
+# which renames the built artifact.
 name = "sglang_mm_core"
-# cdylib: the PyO3 module (`sglang.srt.multimodal._core`).
+# cdylib: the PyO3 module (`sglang.srt.rust_extensions._multimodal`).
 # rlib:   pure-Rust core linked by sglang-server's native MM path.
 crate-type = ["cdylib", "rlib"]
 

rust/sglang-mm/README.md

@@ -6,9 +6,9 @@ GIL-released.
 
 Built two ways:
 
-- **PyO3 extension** `sglang.srt.multimodal._core` (features `python,parallel`,
-  requested by the wheel build) via setuptools-rust when installing sglang —
-  used by Python processors and parity tests.
+- **PyO3 extension** `sglang.srt.rust_extensions._multimodal` (features
+  `python,parallel`, requested by the wheel build) via setuptools-rust when
+  installing sglang — used by Python processors and parity tests.
 - **Pure-Rust `rlib`** (default features, i.e. neither) linked by
   `sglang-server`'s MM worker path — that copy needs no pyo3, no libpython, and
   no rayon: it spawns no threads and runs inline on the calling thread, because
@@ -19,7 +19,7 @@ Built two ways:
 
 ```
 src/
-├── lib.rs                    # module root; PyO3 module (_core) feature-gated
+├── lib.rs                    # module root; PyO3 module (_multimodal) feature-gated
 ├── pipeline.rs               # the server-pipeline contract: MmFamilyProcessor
 │                             #   trait + the carriers (Tensor, TokenLayout, ...)
 ├── driver.rs                 # model-independent request driver (fetch →
@@ -115,7 +115,7 @@ un-stripped URL to `open()`).
 ## Python API
 
 ```python
-from sglang.srt.multimodal._core import common, inkling
+from sglang.srt.rust_extensions._multimodal import common, inkling
 
 # Common (model-agnostic)
 common.resize_rgb(arr, out_w, out_h)
@@ -174,7 +174,8 @@ impl ImageProcessorSpec for MyModelProcessor {
 
 4. Wire up in `src/lib.rs`: `mod my_model;` and `my_model::register(m)?;`.
 
-5. Add Python processor class that calls `from sglang.srt.multimodal._core import my_model`.
+5. Add Python processor class that calls
+   `from sglang.srt.rust_extensions._multimodal import my_model`.
 
 ## Available transform primitives (`common::transforms`)
 

rust/sglang-mm/bench/bench_parity.py

@@ -5,7 +5,6 @@ import numpy as np
 import torch
 from PIL import Image
 
-from sglang.srt.multimodal._core import inkling as _rs_inkling
 from sglang.srt.multimodal.inkling.image_processing import (
     IMAGE_MEAN,
     IMAGE_STD,
@@ -13,6 +12,7 @@ from sglang.srt.multimodal.inkling.image_processing import (
     _encode_image_bytes,
     _fill_patches_numba,
 )
+from sglang.srt.rust_extensions._multimodal import inkling as _rs_inkling
 
 PS = 40
 

rust/sglang-mm/pyproject.toml

@@ -9,4 +9,4 @@ description = "Rust-accelerated multimodal preprocessing for SGLang"
 requires-python = ">=3.10"
 
 [tool.maturin]
-module-name = "_core"
+module-name = "_multimodal"

rust/sglang-mm/src/lib.rs

@@ -1,7 +1,7 @@
 //! sglang-mm: Rust-accelerated multimodal preprocessing for SGLang.
 //!
 //! Built two ways:
-//! * PyO3 extension `sglang.srt.multimodal._core` (feature `python`, default),
+//! * PyO3 extension `sglang.srt.rust_extensions._multimodal` (feature `python`),
 //!   used by Python processors (e.g. Inkling) and by parity tests.
 //! * Pure-Rust `rlib` (`default-features = false`), linked by `sglang-server`'s
 //!   MM worker path — no pyo3 in that dependency graph.
@@ -18,7 +18,7 @@ use pyo3::prelude::*;
 
 #[cfg(feature = "python")]
 #[pymodule]
-fn _core(m: &Bound<'_, PyModule>) -> PyResult<()> {
+fn _multimodal(m: &Bound<'_, PyModule>) -> PyResult<()> {
     common::register(m)?;
     inkling::register(m)?;
     qwen_vl::register(m)?;

rust/sglang-mm/tests/test_golden.py

@@ -4,7 +4,7 @@ import os
 import numpy as np
 import pytest
 
-from sglang.srt.multimodal._core import inkling as _rs_inkling
+from sglang.srt.rust_extensions._multimodal import inkling as _rs_inkling
 
 GOLDEN_DIR = os.environ.get(
     "INKLING_MM_GOLDEN_DIR",

rust/sglang-mm/tests/test_hash_fetch.py

@@ -15,12 +15,12 @@ sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "bench"))
 from bench_parity import make_photo_like
 
 from sglang.srt.managers.mm_utils import data_hash, hash_feature
-from sglang.srt.multimodal._core import common as _rs_common
 from sglang.srt.multimodal.inkling import InklingProcessor
 from sglang.srt.multimodal.inkling.image_processing_rust import (
     InklingRustImageProcessor,
 )
 from sglang.srt.multimodal.processors import inkling as prc
+from sglang.srt.rust_extensions._multimodal import common as _rs_common
 
 
 def png_bytes(arr):

rust/sglang-mm/tests/test_resize_parity.py

@@ -6,8 +6,8 @@ import numpy as np
 import pytest
 from PIL import Image
 
-from sglang.srt.multimodal._core import common as _rs_common
-from sglang.srt.multimodal._core import inkling as _rs_inkling
+from sglang.srt.rust_extensions._multimodal import common as _rs_common
+from sglang.srt.rust_extensions._multimodal import inkling as _rs_inkling
 
 
 def py_scaled_dims(