Dispatch EAGLE radix bigram builder to tai-kernel

The pure-Python convert_to_bigram_key list comprehension runs on every
cache_finished_req / cache_unfinished_req of every EAGLE radix variant
(radix_cache, hiradix_cache, swa_radix_cache), with token-list lengths
that scale with prompt + output.  Scheduler profiles consistently flag it
as the largest mem_cache-side CPU hotspot.

This commit wires sglang.srt.mem_cache.utils.convert_to_bigram_key to
tai_kernel.radix.convert_to_bigram_key when the extension is importable,
falling back to the pure-Python implementation otherwise.  The tai-kernel
path uses a pybind11 module that calls the CPython C API directly
(PyTuple_New + PyTuple_SET_ITEM with ref-stealing) rather than rebuilding
the list comprehension's bytecode-level tuple allocations.  Measured 1.4x
at n=131k and up to 2.5x for n=1k on g0034 Python 3.11; allocator-bound
at large n because CPython's 2-tuple freelist already amortises the
construction.  The int64-packing follow-up that bypasses tuple allocation
entirely is parked as a separate work item.

Runtime safety:
- The dispatcher catches any first-call JIT compile / runtime failure,
  logs once, and falls through to the pure-Python path for the rest of
  the process — JIT failures must degrade rather than crash a serving
  loop.
- SGLANG_DISABLE_TAI_BIGRAM forces the Python path for bisecting.

Constraint: Output must be a real Python List[Tuple[int, int]] because
downstream radix dicts use the tuples as hashable keys.

Rejected: Pre-allocated tuple slab pool | CPython's per-interpreter 2-tuple
freelist already serves this case, and we cannot recycle tuples that
become radix-tree keys without changing the consumer.

Rejected: int64-packed keys this round | requires changes to RadixKey,
get_child_key_fn, key_match_fn, and EAGLE bigram detection; deserves its
own plan.

Confidence: high

Scope-risk: low

Directive: Keep _python_convert_to_bigram_key reachable; if the tai-kernel
path is ever removed, the EAGLE radix cache must continue to work
unchanged.

Tested: tai-kernel side validated on the cluster
(``python benchmark/radix/benchmark_convert_to_bigram_key.py --check``
prints byte-exact correctness then 1.4-2.5x speedup across sizes
128..131072 on g0034).  The new
``test/registered/unit/mem_cache/test_convert_to_bigram_key.py``
exercises both dispatch paths via SGLANG_DISABLE_TAI_BIGRAM patching.

Not-tested: End-to-end EAGLE serving accuracy + scheduler-time delta on
a real workload.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-24 03:21:52 +08:00
parent ec7e9fbc57
commit d655fad040
3 changed files with 117 additions and 1 deletions

View File

@@ -215,6 +215,7 @@ class Envs:
SGLANG_CP_SHARED_KV_MLA_PREFETCH_MIN_PREFIX_PAGES = EnvInt(-1)
SGLANG_CP_DRAFT_SHARED_KV = EnvBool(False)
SGLANG_CP_DRAFT_SHARED_KV_DEBUG = EnvBool(False)
SGLANG_DISABLE_TAI_BIGRAM = EnvBool(False)
SGLANG_TEST_REQUEST_TIME_STATS = EnvBool(False)
SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK = EnvBool(False)
SGLANG_SIMULATE_ACC_LEN = EnvFloat(-1)

View File

@@ -13,6 +13,7 @@
# ==============================================================================
"""Common utilities."""
import logging
from typing import Any, List, Optional, Tuple
import torch
@@ -21,6 +22,8 @@ import triton.language as tl
from sglang.srt.environ import envs
logger = logging.getLogger(__name__)
@triton.jit
def set_mla_kv_buffer_kernel(
@@ -264,7 +267,7 @@ def maybe_init_custom_mem_pool(
return False, None, None
def convert_to_bigram_key(tokens: List[int]) -> List[Tuple[int, int]]:
def _python_convert_to_bigram_key(tokens: List[int]) -> List[Tuple[int, int]]:
# EAGLE uses bigram keys in the radix tree since draft sequence is the one-token-shifted version of target
# [1, 2, 3, 4] -> [(1,2), (2,3), (3,4)]
if len(tokens) and isinstance(tokens[0], tuple):
@@ -272,3 +275,28 @@ def convert_to_bigram_key(tokens: List[int]) -> List[Tuple[int, int]]:
if len(tokens) < 2:
return []
return [(tokens[i], tokens[i + 1]) for i in range(len(tokens) - 1)]
try:
from tai_kernel.radix import convert_to_bigram_key as _tai_bigram_impl
except Exception:
_tai_bigram_impl = None
def convert_to_bigram_key(tokens: List[int]) -> List[Tuple[int, int]]:
global _tai_bigram_impl
if _tai_bigram_impl is not None and not envs.SGLANG_DISABLE_TAI_BIGRAM.get():
try:
return _tai_bigram_impl(tokens)
except Exception as e:
# JIT compile or runtime failures (missing compiler, stale cache,
# CPython ABI mismatch) must degrade to the Python path rather
# than crash a serving loop. Disable after one failure so we
# don't keep paying the broken-extension cost on every call.
logger.warning(
"tai-kernel convert_to_bigram_key failed, falling back to "
"Python implementation for the rest of this process: %s",
e,
)
_tai_bigram_impl = None
return _python_convert_to_bigram_key(tokens)

View File

@@ -0,0 +1,87 @@
"""Unit tests for sglang.srt.mem_cache.utils.convert_to_bigram_key.
Covers both the pure-Python fallback and the tai-kernel pybind11 path
(`tai_kernel.radix.convert_to_bigram_key`) when the extension is importable.
Run:
python -m pytest test/registered/unit/mem_cache/test_convert_to_bigram_key.py -v
"""
import sys
import unittest
import unittest.mock
for _mod in ("sgl_kernel", "sgl_kernel.kvcacheio"):
if _mod not in sys.modules:
sys.modules[_mod] = unittest.mock.MagicMock()
from sglang.srt.environ import envs
from sglang.srt.mem_cache import utils as mem_cache_utils
from sglang.srt.mem_cache.utils import (
_python_convert_to_bigram_key,
convert_to_bigram_key,
)
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
register_cuda_ci(est_time=2, suite="stage-b-test-1-gpu-small")
register_amd_ci(est_time=2, suite="stage-b-test-1-gpu-small-amd")
class TestConvertToBigramKey(unittest.TestCase):
"""The dispatcher must agree with the pure-Python reference on every input,
regardless of whether the tai-kernel extension is loaded."""
cases = [
[],
[42],
[1, 2],
[1, 2, 3, 4, 5],
list(range(1024)),
# Already-bigram input should be returned as-is (no re-pairing).
[(1, 2), (2, 3), (3, 4)],
]
def _assert_equivalent(self, tokens):
expected = _python_convert_to_bigram_key(list(tokens))
got = convert_to_bigram_key(list(tokens))
self.assertEqual(got, expected)
self.assertIsInstance(got, list)
for item in got:
self.assertIsInstance(item, tuple)
self.assertEqual(len(item), 2)
def test_python_fallback(self):
# Force the fallback regardless of whether tai_kernel is available.
with unittest.mock.patch.object(
envs.SGLANG_DISABLE_TAI_BIGRAM, "get", return_value=True
):
for tokens in self.cases:
with self.subTest(tokens=tokens):
self._assert_equivalent(tokens)
def test_tai_kernel_path(self):
if mem_cache_utils._tai_bigram_impl is None:
self.skipTest("tai_kernel.radix not importable in this environment")
with unittest.mock.patch.object(
envs.SGLANG_DISABLE_TAI_BIGRAM, "get", return_value=False
):
for tokens in self.cases:
with self.subTest(tokens=tokens):
self._assert_equivalent(tokens)
def test_distinct_tuple_identity(self):
# Bigrams that share token values must still be distinct objects so
# downstream radix dicts don't alias keys.
for force_fallback in (True, False):
if not force_fallback and mem_cache_utils._tai_bigram_impl is None:
continue
with unittest.mock.patch.object(
envs.SGLANG_DISABLE_TAI_BIGRAM, "get", return_value=force_fallback
):
got = convert_to_bigram_key([7, 7, 7])
self.assertEqual(got, [(7, 7), (7, 7)])
self.assertIsNot(got[0], got[1])
if __name__ == "__main__":
unittest.main()