Shared CP KV materialization remained on the critical path for cached NSA/MLA prefill batches. This change introduces a one-layer-ahead prefetcher that materializes the cached prefix for the next layer on a separate CUDA stream and consumes it when that layer reaches attention. The prefetch path keeps the existing dense page-table semantics, defers waiting until the prefetched buffer is actually consumed, and uses the TAI optimized materialize/remap helpers when enabled before falling back to the torch implementation. The implementation is intentionally gated by environment variables and keeps layer-2-only probe logging for functional confirmation without making normal profiling noisy. Constraint: Prefill CP shared KV must preserve existing page-table and dense KV semantics for NSA paged topk attention Constraint: The production performance path requires SGLANG_CP_SHARED_KV_USE_TAI_MATERIALIZE=1 and logging disabled Rejected: Wait immediately after the producer layer attention | this truncated the overlap window and hid less work Rejected: Torch-only prefetch materialize | it bypassed the optimized TAI materialize/remap path and could erase the expected win Confidence: medium Scope-risk: moderate Directive: Do not evaluate Phase8 throughput with SGLANG_CP_SHARED_KV_LOG_MLA_PREFETCH=1; use it only to confirm create/start/consume_hit behavior Tested: Local AST parse for modified Python files Tested: Local git diff --check Tested: Remote g0034 container AST parse for modified files under /sgl-workspace/sglang-tai Tested: Remote g0034 container pytest target covering Phase8 log env, TAI range materialize, optimized slot inverse/remap, and existing token TAI path Not-tested: Full prefill/decode/router throughput after the TAI prefetch-path fix
Unit Tests
Component-level tests that do not launch a server or load model weights. Tests can use CPU or GPU — the key criterion is no server process.
Quick Start
- Find the source file under
python/sglang/srt/. - Create the corresponding test here, mirroring the source tree:
srt/mem_cache/radix_cache.py → unit/mem_cache/test_radix_cache.py srt/sampling/sampling_params.py → unit/sampling/test_sampling_params.py - Register for CI at the top of the file (after imports, before test classes):
from sglang.test.ci.ci_register import register_cpu_ci register_cpu_ci(est_time=5, suite="stage-a-test-cpu") # or: register_cuda_ci(est_time=10, suite="stage-b-test-1-gpu-small") - Run locally:
pytest test/registered/unit/ -v # all unit tests pytest test/registered/unit/mem_cache/ -v # one module - Run with coverage:
# summary pytest test/registered/unit/ --cov --cov-config=.coveragerc -v # PR incremental check (require ≥60% on changed lines) pytest test/registered/unit/ --cov --cov-config=.coveragerc --cov-report=xml diff-cover coverage.xml --compare-branch=origin/main --fail-under=60
Example
"""Unit tests for <module> — no server, no model loading."""
from sglang.test.ci.ci_register import register_cpu_ci
register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
import unittest
from sglang.srt.<module> import TargetClass
from sglang.test.test_utils import CustomTestCase
class TestTargetClass(CustomTestCase):
def test_basic_behavior(self):
obj = TargetClass(...)
self.assertEqual(obj.method(), expected)
if __name__ == "__main__":
unittest.main()
Rules
- No
popen_launch_server()orEngine(...). - No model weight loading.
- Use
CustomTestCase(fromsglang.test.test_utils, adds CI retry). - Use
unittest.mockfor dependencies that are expensive to construct.