minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[MLX] Fix single-token chunked-prefill continuation misrouted as decode (#30181)

Browse Source
This commit is contained in:
NOOB
2026-07-06 20:31:33 -07:00
committed by GitHub
parent 3a679459e5
commit c3da0a2582
2 changed files with 293 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/mlx/tp_worker.py
+49 -26
View File
@@ -132,6 +132,29 @@ class MlxTpModelWorker(TpModelWorker):
# insert. Any older tracked slot is released during component cleanup.
req.mamba_last_track_seqlen = None
def _route_extend_request(self, rid: str, decoding_rids: set[str]) -> str:
"""Classify a request within an extend / mixed batch.
Shared by the sync (:meth:`_forward_batch_generation_mlx`) and async
(:meth:`_async_extend_batch`) paths so both route identically.
Returns one of:
* ``"prefill"`` -- not seen before; start a fresh prefill.
* ``"decode"`` -- a genuine single-token decode step mixed into
this batch (present in ``batch.decoding_reqs``).
* ``"continuation"`` -- a chunked-prefill continuation. Routing keys on
request state, **not** ``seq_len``: a final continuation chunk can be
exactly one token, which must still extend. Routing it as a decode
would drop the real token and feed the model its own previous-chunk
prediction, silently corrupting the output.
"""
if not self._mlx_runner.has_request(rid):
return "prefill"
if rid in decoding_rids:
return "decode"
return "continuation"
def _forward_batch_generation_mlx(
self, batch: ScheduleBatch
) -> GenerationBatchResult:
@@ -164,6 +187,9 @@ class MlxTpModelWorker(TpModelWorker):
prefill_rids: list[tuple[str, int]] = []
extend_rids: list[tuple[str, int]] = []
decode_rids: list[str] = []
# Genuine decode steps mixed into this extend batch; see
# _route_extend_request.
decoding_rids = {r.rid for r in (batch.decoding_reqs or [])}
for i, req in enumerate(reqs):
seq_len = extend_seq_lens[i]
@@ -172,18 +198,15 @@ class MlxTpModelWorker(TpModelWorker):
offset += seq_len
slot_offset += seq_len
if self._mlx_runner.has_request(req.rid):
if seq_len > 1:
# Chunked prefill continuation
next_token = self._mlx_runner.extend(
req.rid, req_token_ids, req_new_slots
)
extend_rids.append((req.rid, next_token))
else:
# MIXED mode: single-token decode
decode_rids.append(req.rid)
else:
# New prefill
route = self._route_extend_request(req.rid, decoding_rids)
if route == "continuation":
next_token = self._mlx_runner.extend(
req.rid, req_token_ids, req_new_slots
)
extend_rids.append((req.rid, next_token))
elif route == "decode":
decode_rids.append(req.rid)
else: # "prefill"
prefix_slot_ids = req.prefix_indices.tolist()
full_token_ids = list(req.get_fill_ids())
next_token = self._mlx_runner.prefill(
@@ -306,6 +329,9 @@ class MlxTpModelWorker(TpModelWorker):
pending_prefills: list[MlxPendingPrefill] = []
pending_extends: list[MlxPendingExtend] = []
mixed_decode_rids: list[str] = []
# Genuine decode steps mixed into this extend batch; see
# _route_extend_request.
decoding_rids = {r.rid for r in (batch.decoding_reqs or [])}
for i, req in enumerate(reqs):
seq_len = extend_seq_lens[i]
@@ -314,21 +340,18 @@ class MlxTpModelWorker(TpModelWorker):
offset += seq_len
slot_offset += seq_len
if self._mlx_runner.has_request(req.rid):
if seq_len > 1:
# Chunked prefill continuation
pending_extends.append(
self._mlx_runner.extend_start(
req_id=req.rid,
new_token_ids=req_token_ids,
new_slot_ids=req_new_slots,
)
route = self._route_extend_request(req.rid, decoding_rids)
if route == "continuation":
pending_extends.append(
self._mlx_runner.extend_start(
req_id=req.rid,
new_token_ids=req_token_ids,
new_slot_ids=req_new_slots,
)
else:
# MIXED mode: single-token decode
mixed_decode_rids.append(req.rid)
else:
# New prefill
)
elif route == "decode":
mixed_decode_rids.append(req.rid)
else: # "prefill"
prefix_slot_ids = req.prefix_indices.tolist()
full_token_ids = list(req.get_fill_ids())
pending_prefills.append(