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

[AMD] add nightly kv_canary JIT benchmark suite (#28611)

Browse Source
This commit is contained in:
Michael
2026-06-19 19:46:34 -07:00
committed by GitHub
parent 2cbe1e6404
commit 653f6735a0
7 changed files with 82 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/jit/benchmark/kv_canary/bench_plan.py
+3 -1
View File
@@ -23,9 +23,11 @@ from sglang.jit_kernel.benchmark.utils import (
from sglang.jit_kernel.kv_canary.plan import launch_canary_plan_kernels
from sglang.jit_kernel.kv_canary.verify import VerifyPlan
from sglang.jit_kernel.kv_canary.write import WritePlan
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
register_cuda_ci(est_time=900, suite="nightly-kernel-1-gpu", nightly=True)
# AMD mirrors the CUDA nightly registration (nightly-only, no per-PR suite).
register_amd_ci(est_time=900, suite="nightly-amd-kernel-1-gpu", nightly=True)
_TOTAL_TOKENS_AXIS: list[int] = [256, 4096, 65536, 262144]
test/registered/jit/benchmark/kv_canary/bench_scatter_req_token_ids.py
+5 -1
View File
@@ -14,9 +14,13 @@ from sglang.jit_kernel.benchmark.utils import (
from sglang.jit_kernel.kv_canary.scatter_req_token_ids import (
launch_scatter_req_token_ids_kernel,
)
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
register_cuda_ci(est_time=180, suite="nightly-kernel-1-gpu", nightly=True)
# AMD mirrors the CUDA nightly registration (nightly-only, no per-PR suite).
# Note: amd_ci_exec.sh sets SGLANG_IS_IN_CI, so this runs the CI-reduced range
# (_BS_AXIS_CI/_SEQ_LEN_AXIS_CI via get_benchmark_range), same as CUDA nightly.
register_amd_ci(est_time=180, suite="nightly-amd-kernel-1-gpu", nightly=True)
_BS_AXIS_FULL: list[int] = [1, 8, 64, 256]
test/registered/jit/benchmark/kv_canary/bench_verify.py
+5 -1
View File
@@ -30,9 +30,13 @@ from sglang.jit_kernel.kv_canary.verify import (
VerifyPlan,
launch_canary_verify_kernel,
)
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
register_cuda_ci(est_time=900, suite="nightly-kernel-1-gpu", nightly=True)
# AMD mirrors the CUDA nightly registration (nightly-only, no per-PR suite).
# Note: amd_ci_exec.sh sets SGLANG_IS_IN_CI, so this runs the CI-reduced range
# (build_fast_matrix_cases via get_benchmark_range), same as CUDA nightly.
register_amd_ci(est_time=900, suite="nightly-amd-kernel-1-gpu", nightly=True)
_X_NAMES = [
test/registered/jit/benchmark/kv_canary/bench_write.py
+5 -1
View File
@@ -28,9 +28,13 @@ from sglang.jit_kernel.kv_canary.verify import (
VerifyOrWriteContext,
)
from sglang.jit_kernel.kv_canary.write import WritePlan, launch_canary_write_kernel
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
register_cuda_ci(est_time=900, suite="nightly-kernel-1-gpu", nightly=True)
# AMD mirrors the CUDA nightly registration (nightly-only, no per-PR suite).
# Note: amd_ci_exec.sh sets SGLANG_IS_IN_CI, so this runs the CI-reduced range
# (build_fast_matrix_cases via get_benchmark_range), same as CUDA nightly.
register_amd_ci(est_time=900, suite="nightly-amd-kernel-1-gpu", nightly=True)
_X_NAMES = [
test/run_suite.py
+1
View File
@@ -129,6 +129,7 @@ NIGHTLY_SUITES = {
HWBackend.AMD: [
"nightly-amd",
"nightly-amd-1-gpu",
"nightly-amd-kernel-1-gpu",
"nightly-amd-1-gpu-mi35x",
"nightly-amd-1-gpu-zimage-turbo",
"nightly-amd-2-gpu-mi35x-deepseek-r1-mxfp4-tp2",