move the PD bootstrap registry under api_server::disaggregation (#33895)

This commit is contained in:
Rain Jiang
2026-08-13 02:13:43 -07:00
committed by GitHub
parent 34206c0017
commit ba23846ccf
8 changed files with 328 additions and 184 deletions
+6 -5
View File
@@ -32,7 +32,10 @@ tracing-subscriber = { workspace = true }
tracing-appender = { workspace = true }
uuid = { workspace = true }
arc-swap = "1"
axum = { version = "0.8.9", features = ["json", "tokio"] }
# Safe POD slice casts (feature buffers viewed as bytes for the shm copy).
bytemuck = "1"
core_affinity = "0.8"
# the dynamo-tokenizers is deps on hf-hub, should bump version together.
dynamo-tokenizers = "1.7.0"
@@ -43,21 +46,19 @@ 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 }
itertools = "0.14"
# 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"
rmp-serde = "1"
rmpv = { version = "1", features = ["with-serde"] }
# 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/.