[mm] rust-server: native multimodal processing for Qwen VL (integrate sglang-mm, e2e) (#32365)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

rust/sglang-server/Cargo.toml

@@ -43,14 +43,25 @@ dynamo-parsers = "7.0.1"
 dynamo-protocols = "5.1.0"
 dynamo-renderer = "5.0.0"
 flume = "0.12.0"
+# Safe POD slice casts (feature buffers viewed as bytes for the shm copy).
+bytemuck = "1"
 itertools = "0.14"
 hf-hub = { version = "0.4", default-features = false }
+# POSIX shm for the MM feature fan-out (`mm::ShmSegment`).
+libc = "0.2"
+# Same major as the workspace pyo3: the zero-copy MM drain (`take_mm`) moves
+# Rust vectors into numpy arrays.
+numpy = "0.29.0"
 rmp-serde = "1"
 rmpv = { version = "1", features = ["with-serde"] }
 # Pinned EXACTLY: this crate's accepted grammar defines the
 # "anything Rust admits, Python can compile" invariant in `message::sampling`.
 # A minor bump can widen it and silently reopen a scheduler-killing hole.
 regex-syntax = "=0.8.11"
+# The pure-Rust core of the MM pipeline. `default-features = false` drops the
+# pyo3 bindings so it links as a plain rlib; renamed so `use sglang_mm::…` reads
+# naturally while the crate keeps its own artifact name in the shared target/.
+sglang_mm = { package = "sglang-mm", path = "../sglang-mm", default-features = false }
 
 [dev-dependencies]
 # `Router::oneshot` for handler-level router tests.