[Core] Clean up array-like msgspec structs (#32688)

This commit is contained in:
Lianmin Zheng
2026-07-28 16:25:22 -07:00
committed by GitHub
parent 70ea37e7e0
commit 9ca4023b13
4 changed files with 10 additions and 20 deletions
@@ -5,7 +5,7 @@
//! `msgspec.msgpack`. Two struct families are involved:
//!
//! * `EventBatch` (the outer payload) — declared with
//! `array_like=True, omit_defaults=True, gc=False` (no tag).
//! `array_like=True, gc=False` (no tag).
//! * `KVCacheEvent` (each inner event variant) — additionally declared
//! with `tag=True`.
//!
@@ -17,9 +17,6 @@
//! of each inner event array, so an event is
//! `[class_name_str, field1, field2, ...]`. The outer `EventBatch`
//! array does **not** carry a tag prefix.
//! * `omit_defaults=True` allows trailing fields whose values equal their
//! declared defaults to be dropped from the array. The decoder therefore
//! accepts variable-length sequences for each struct shape.
//!
//! This module deserializes those bytes into Rust types and exposes a single
//! [`decode_event_batch`] entry point.
@@ -32,7 +29,7 @@ use serde::Deserialize;
/// Top-level batch payload published by SGLang.
///
/// Wire shape (`EventBatch`, `array_like`):
/// `[ts: f64, events: [...], attn_dp_rank: int_or_nil_or_omitted]`.
/// `[ts: f64, events: [...], attn_dp_rank: int_or_nil]`.
/// SGLang declares `attn_dp_rank` as a Python `Optional[int]`; we decode
/// it as `u32` since DP ranks are non-negative and bounded by the
/// publisher's `dp_size`.
@@ -43,7 +40,8 @@ pub struct KvEventBatch {
/// Ordered list of cache events in this batch.
pub events: Vec<KvCacheEvent>,
/// Optional DP-attention rank that produced this batch. `None` if the
/// publisher emitted nil or omitted the field via `omit_defaults`.
/// publisher emitted nil. The decoder also accepts an omitted field for
/// compatibility.
pub attn_dp_rank: Option<u32>,
}
@@ -332,8 +330,8 @@ impl<'de> Deserialize<'de> for BoundedU32Vec {
// ---------------------------------------------------------------------------
// Custom Deserialize impls — msgspec encodes these structs as msgpack arrays
// (not maps), and `omit_defaults=True` means trailing optional fields may be
// absent. We therefore implement `Deserialize` by hand against `SeqAccess`.
// (not maps). The visitors also accept absent trailing optional fields for
// compatibility.
// ---------------------------------------------------------------------------
impl<'de> Deserialize<'de> for KvEventBatch {
@@ -731,8 +729,7 @@ mod tests {
#[test]
fn attn_dp_rank_omitted_decodes_as_none() {
// msgspec's `omit_defaults=True` may drop attn_dp_rank entirely from
// the wire array when it equals its default of None.
// Accept a payload that omits the optional trailing rank field.
let event = build_all_blocks_cleared_bytes();
let bytes = build_batch_bytes(5.0, &[event], None, /* include_dp_field */ false);
@@ -61,7 +61,6 @@ def select_kv_publisher_dp_rank(
class EventBatch(
msgspec.Struct,
array_like=True, # type: ignore[call-arg]
omit_defaults=True, # type: ignore[call-arg]
gc=False, # type: ignore[call-arg]
):
ts: float
@@ -72,7 +71,6 @@ class EventBatch(
class KVCacheEvent(
msgspec.Struct,
array_like=True, # type: ignore[call-arg]
omit_defaults=True, # type: ignore[call-arg]
gc=False, # type: ignore[call-arg]
tag=True,
):
@@ -72,7 +72,7 @@ def raise_if_tokenizer_required(
)
class SamplingParams(msgspec.Struct, kw_only=True, omit_defaults=True):
class SamplingParams(msgspec.Struct, kw_only=True, array_like=True):
"""
The sampling parameters.
@@ -106,10 +106,10 @@ class SamplingParams(msgspec.Struct, kw_only=True, omit_defaults=True):
skip_special_tokens: bool = True
spaces_between_special_tokens: bool = True
no_stop_trim: bool = False
custom_params: Optional[Dict[str, CustomParamValue]] = None
stream_interval: Optional[int] = None
logit_bias: Optional[Dict[str, float]] = None
sampling_seed: Optional[int] = None
custom_params: Optional[Dict[str, CustomParamValue]] = None
# --- Internal fields (populated by __post_init__ or normalize(), not API-facing) ---
stop_strs: Optional[Union[str, List[str]]] = None # from stop
@@ -269,7 +269,7 @@ class SamplingParams(msgspec.Struct, kw_only=True, omit_defaults=True):
tokenizer, self.stop_strs, self.stop_regex_strs, self.min_new_tokens
)
# Clear API input aliases so omit_defaults=True drops them from the wire.
# Clear API input aliases after normalizing them into internal fields.
self.stop = None
self.stop_regex = None
self.is_normalized = True
@@ -411,11 +411,6 @@ class TestSamplingParamsMsgspecStruct(CustomTestCase):
self.assertTrue(sp.spaces_between_special_tokens)
self.assertFalse(sp.no_stop_trim)
def test_msgpack_omits_default_fields(self):
encoded = msgspec.msgpack.encode(SamplingParams())
self.assertEqual(msgspec.msgpack.decode(encoded), {})
def test_msgpack_round_trip_preserves_normalized_state(self):
tokenizer = MagicMock()
tokenizer.encode.side_effect = lambda s, add_special_tokens=False: {