[CI] [NVIDIA] gb300: Run DSV4.1-Flash tests on 2 GPUs with engram host tables (#45294)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
em-dreyer
2026-10-21 18:53:24 -07:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 67bbfcee3c
commit 3d508d614f
8 changed files with 30 additions and 21 deletions
+9 -3
View File
@@ -4,6 +4,11 @@ Each server configuration lives in its own file under
``test/registered/e2e/models/test_deepseek_v41_flash_gb300_*.py`` so the
configurations partition independently across GB300 runners. All use default
decoder bounded replay and explicitly enabled breakable prefill graphs.
They run on two GB300 GPUs (TP2) with the engram tables in host memory
(SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE). The two engram tables are ~189 GiB in FP8
(2 x ~384M rows x 264 B), so at TP2 offloading them frees ~94 GiB per GPU and the
remaining weights (~143 GiB/GPU) leave room for KV cache, graphs and DSPARK.
"""
import requests
@@ -19,7 +24,7 @@ from sglang.test.test_utils import (
MODEL = "deepseek-ai/DeepSeek-V4.1-Flash"
# Tensor-parallel attention and expert-parallel MoE, without DP attention.
TP4_EP4_ARGS = ["--attn-dp-size", "1", "--ep-size", "4"]
TP2_EP2_ARGS = ["--attn-dp-size", "1", "--ep-size", "2"]
class DSV41FlashAccuracyMixin(GSM8KMixin):
@@ -49,7 +54,8 @@ class DSV41FlashAccuracyMixin(GSM8KMixin):
"8192",
]
parallel_args = []
server_env = {}
# Keep the engram tables in host memory so DSV4.1-Flash fits on two GB300s.
server_env = {"SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE": "1"}
@classmethod
def setUpClass(cls):
@@ -62,7 +68,7 @@ class DSV41FlashAccuracyMixin(GSM8KMixin):
other_args=[
"--trust-remote-code",
"--tp",
"4",
"2",
"--speculative-algorithm",
"DSPARK",
"--speculative-dspark-block-size",