Add nightly test for GLM5.2 LayerSplit (#31512)

This commit is contained in:
Baizhou Zhang
2026-07-17 00:29:14 -07:00
committed by GitHub
parent 619609aa5a
commit e835512303
@@ -9,8 +9,7 @@ PD-disaggregated GLM-5.2 deployment: a layer-split prefill worker running
interleave prefill-CP + layer split, and an ordinary decode worker that receives
full cache shards via PD transfer.
Sized for the 4-GPU B200 runner (prefill TP=2 + decode TP=2) rather than an
8-GPU deployment, since the 8-gpu-b200 runner is nightly-only.
Runs nightly on an 8-GPU B200 runner (prefill TP=4 + decode TP=4).
"""
import unittest
@@ -22,10 +21,9 @@ from sglang.test.server_fixtures.disaggregation_fixture import (
)
register_cuda_ci(
est_time=1200,
stage="extra-b",
runner_config="4-gpu-b200",
disabled="Temporarily disabled",
est_time=450,
suite="nightly-8-gpu-b200",
nightly=True,
)
@@ -36,13 +34,15 @@ class TestGLM52DSACacheLayerSplit(PDDisaggregationServerBase, GSM8KMixin):
gsm8k_accuracy_thres = 0.935
gsm8k_num_questions = 1319
gsm8k_num_threads = 200
gsm8k_num_shots = 0
gsm8k_num_shots = 20
# Prefill worker: interleave prefill-CP + DSA cache layer split on 2 GPUs
# (TP=2 -> attn_cp_size=2, so KV/indexer layers shard 2-way across CP ranks).
# Prefill worker: interleave prefill-CP + DSA cache layer split on 4 GPUs
# (TP=4 -> attn_cp_size=4, so KV/indexer layers shard 4-way across CP ranks).
extra_prefill_args = [
"--tp",
"2",
"4",
"--attn-cp-size",
"4",
"--dsa-prefill-backend",
"trtllm",
"--kv-cache-dtype",
@@ -57,20 +57,36 @@ class TestGLM52DSACacheLayerSplit(PDDisaggregationServerBase, GSM8KMixin):
"4096",
"--max-prefill-tokens",
"4096",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"6",
]
# Decode worker: ordinary local decode cache on the other 2 GPUs, receives
# full shards via PD transfer.
# Decode worker: ordinary local decode cache, receives full shards via PD
# transfer.
extra_decode_args = [
"--tp",
"2",
"4",
"--dsa-decode-backend",
"trtllm",
"--kv-cache-dtype",
"fp8_e4m3",
"--mem-fraction-static",
"0.85",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"6",
"--base-gpu-id",
"2",
"4",
]
@classmethod