[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
+1
View File
@@ -446,6 +446,7 @@ jobs:
fail-fast: false
matrix:
include:
- { runner_config: 2-gpu-gb300, run_timeout: '30', timeout_per_file: '1800' }
- { runner_config: 4-gpu-gb300, run_timeout: '30', timeout_per_file: '1800' }
uses: ./.github/workflows/_pr-test-stage.yml
with:
+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",
+1
View File
@@ -22,6 +22,7 @@ runner_configs:
1-gpu-large: { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 1-gpu-h100 }
2-gpu-large: { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 2-gpu-h100 }
4-gpu-b200: { install: *default, artifact_version: v6, install_timeout: "20", runs_on: $b200_runner }
2-gpu-gb300: { install: *default, artifact_version: v6, install_timeout: "20", runs_on: 2-gpu-gb300 }
4-gpu-gb300: { install: *default, artifact_version: v6, install_timeout: "20", runs_on: 4-gpu-gb300 }
4-gpu-h100: { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 4-gpu-h100 }
8-gpu-h200: { install: *default, artifact_version: v4, install_timeout: "20", runs_on: 8-gpu-h200 }
@@ -1,4 +1,4 @@
"""DeepSeek-V4.1-Flash + DSPARK accuracy on four GB300 GPUs: DP-attention4 + EP4.
"""DeepSeek-V4.1-Flash + DSPARK accuracy on two GB300 GPUs: DP-attention2 + EP2.
The DP+EP recipe is intentionally exercised rather than skipped: startup or
accuracy failures must be visible in CI while this path is being validated.
@@ -10,10 +10,10 @@ from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.dsv41_flash_common import DSV41FlashAccuracyMixin
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=600, stage="base-c", runner_config="4-gpu-gb300")
register_cuda_ci(est_time=600, stage="base-c", runner_config="2-gpu-gb300")
class TestDSV41FlashDP4EP4DSpark(DSV41FlashAccuracyMixin, CustomTestCase):
class TestDSV41FlashDP2EP2DSpark(DSV41FlashAccuracyMixin, CustomTestCase):
"""DP attention and EP MoE using the built-in gather/reduce path.
DSPARK with DP attention does not support the DeepEP backend.
@@ -21,9 +21,9 @@ class TestDSV41FlashDP4EP4DSpark(DSV41FlashAccuracyMixin, CustomTestCase):
parallel_args = [
"--attn-dp-size",
"4",
"2",
"--ep-size",
"4",
"2",
"--enable-dp-lm-head",
"--moe-a2a-backend",
"none",
@@ -1,20 +1,20 @@
"""DeepSeek-V4.1-Flash + DSPARK accuracy on GB300 with encoder bounded replay.
Run GSM8K and MMMU-Pro with TP4 + EP4, encoder SWA bounded replay, and
Run GSM8K and MMMU-Pro with TP2 + EP2, encoder SWA bounded replay, and
breakable prefill graphs using the shared accuracy recipe.
"""
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.dsv41_flash_common import TP4_EP4_ARGS, DSV41FlashAccuracyMixin
from sglang.test.dsv41_flash_common import TP2_EP2_ARGS, DSV41FlashAccuracyMixin
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=600, stage="base-c", runner_config="4-gpu-gb300")
register_cuda_ci(est_time=600, stage="base-c", runner_config="2-gpu-gb300")
class TestDSV41FlashEncoderBoundedBCG(DSV41FlashAccuracyMixin, CustomTestCase):
parallel_args = TP4_EP4_ARGS
parallel_args = TP2_EP2_ARGS
graph_args = [
*DSV41FlashAccuracyMixin.graph_args,
"--enable-encoder-swa-bounded-replay",
@@ -1,4 +1,4 @@
"""DeepSeek-V4.1-Flash + DSPARK accuracy on four GB300 GPUs: TP4 + EP4.
"""DeepSeek-V4.1-Flash + DSPARK accuracy on two GB300 GPUs: TP2 + EP2.
GSM8K and MMMU-Pro with tensor-parallel attention and expert-parallel MoE,
without DP attention.
@@ -7,16 +7,16 @@ without DP attention.
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.dsv41_flash_common import TP4_EP4_ARGS, DSV41FlashAccuracyMixin
from sglang.test.dsv41_flash_common import TP2_EP2_ARGS, DSV41FlashAccuracyMixin
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=600, stage="base-c", runner_config="4-gpu-gb300")
register_cuda_ci(est_time=600, stage="base-c", runner_config="2-gpu-gb300")
class TestDSV41FlashTP4EP4DSpark(DSV41FlashAccuracyMixin, CustomTestCase):
class TestDSV41FlashTP2EP2DSpark(DSV41FlashAccuracyMixin, CustomTestCase):
"""Tensor-parallel attention and expert-parallel MoE, without DP attention."""
parallel_args = TP4_EP4_ARGS
parallel_args = TP2_EP2_ARGS
if __name__ == "__main__":
@@ -1,6 +1,6 @@
"""DeepSeek-V4.1-Flash + DSPARK accuracy on four GB300 GPUs: TRT-LLM bounded BCG.
"""DeepSeek-V4.1-Flash + DSPARK accuracy on two GB300 GPUs: TRT-LLM bounded BCG.
TP4 + EP4 with TRT-LLM varlen attention and decoder SWA bounded replay inside
TP2 + EP2 with TRT-LLM varlen attention and decoder SWA bounded replay inside
breakable prefill graphs.
"""
@@ -10,7 +10,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.dsv41_flash_common import DSV41FlashAccuracyMixin
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=600, stage="base-c", runner_config="4-gpu-gb300")
register_cuda_ci(est_time=600, stage="base-c", runner_config="2-gpu-gb300")
class TestDSV41FlashTrtllmBoundedBCG(DSV41FlashAccuracyMixin, CustomTestCase):
@@ -20,7 +20,7 @@ class TestDSV41FlashTrtllmBoundedBCG(DSV41FlashAccuracyMixin, CustomTestCase):
"--attn-dp-size",
"1",
"--ep-size",
"4",
"2",
"--dsv4-attn-backend",
"trtllm",
]
+1
View File
@@ -92,6 +92,7 @@ PER_COMMIT_SUITES = {
"base-b-test-diffusion-unit-1-gpu-h100",
"base-c-test-4-gpu-h100",
"base-c-test-4-gpu-b200",
"base-c-test-2-gpu-gb300",
"base-c-test-4-gpu-gb300",
"base-c-test-8-gpu-h20",
"base-c-test-8-gpu-h200",