Skip to content

[AMD] Add DeepSeek-V4.1-Flash MI35x nightly accuracy test #41476

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
HaiShaw merged 4 commits into from
Oct 4, 2026
Merged

[AMD] Add DeepSeek-V4.1-Flash MI35x nightly accuracy test #41476

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
da344d6
[AMD] Add DeepSeek-V4.1-Flash MI35x nightly accuracy test
michaelzhang-ai Sep 27, 2026
fe0afdc
[AMD] Set DeepSeek-V4.1-Flash MI35x GSM8K threshold from measured run
michaelzhang-ai Oct 1, 2026
d2a7ead
[AMD] Trim DeepSeek-V4.1-Flash MI35x est_time to the measured runtime
michaelzhang-ai Oct 1, 2026
208ef1c
[AMD] Remove DeepSeek-V3.2 MI35x nightly jobs
michaelzhang-ai Oct 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
172 changes: 52 additions & 120 deletions .github/workflows/nightly-test-amd.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,9 +65,6 @@ on:
# 8-GPU DeepSeek-V3.2 (MI30x)
- nightly-8-gpu-deepseek-v32
- nightly-8-gpu-deepseek-v32-mtp
# 8-GPU DeepSeek-V3.2 (MI35x)
- nightly-8-gpu-mi35x-deepseek-v32
- nightly-8-gpu-mi35x-deepseek-v32-mtp
# 8-GPU DeepSeek-R1 (MI35x only)
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8
Expand All @@ -81,6 +78,8 @@ on:
- nightly-8-gpu-mi35x-deepseek-v4-pro
- nightly-8-gpu-mi35x-deepseek-v4-pro-mtp
- nightly-8-gpu-mi35x-deepseek-v4-pro-dspark
# 4-GPU DeepSeek-V4.1-Flash (MI35x only - FP4 experts need gfx95x)
- nightly-4-gpu-mi35x-deepseek-v41-flash
# 8-GPU Kimi-K2.6 (MI30x)
- nightly-8-gpu-kimi-k26
# 8-GPU Kimi-K3 (MI35x only - native MXFP4 needs gfx95x)
Expand Down Expand Up @@ -1068,120 +1067,6 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}

# ==============================================================================
# 8-GPU DeepSeek-V3.2 (MI35x)
#
# Accuracy and performance share one job per config, as the MI30x V3.2 jobs
# above already do: both steps serve the same weights, so splitting them cost
# a second container setup and a second cold weight load for nothing.
# ==============================================================================

nightly-8-gpu-mi35x-deepseek-v32:
name: ${{ format('nightly-8-gpu-mi35x-deepseek-v32 ({0}, linux-mi35x-gpu-8)', matrix.rocm_version) }}
strategy:
fail-fast: false
matrix:
rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm10", "rocm724", "rocm720"]') }}
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-deepseek-v32,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}

- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm

- name: Setup docker (${{ matrix.rocm_version }})
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
ENABLE_CACHE_HOST: "1"

- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
# Install tabulate for run_suite.py (missing in MI35x container)
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate

- name: Accuracy Test MI35x ROCm (8-GPU DeepSeek-V3.2)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_MOE_COPY_WEIGHT_VIEWS_BEFORE_H2D=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-deepseek-v32 --nightly --timeout-per-file 3600 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}

- name: Performance Test MI35x (8-GPU DeepSeek-V3.2 Basic)
timeout-minutes: 150
continue-on-error: true
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_MOE_COPY_WEIGHT_VIEWS_BEFORE_H2D=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-deepseek-v32-basic --nightly --timeout-per-file 5400 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}

nightly-8-gpu-mi35x-deepseek-v32-mtp:
name: ${{ format('nightly-8-gpu-mi35x-deepseek-v32-mtp ({0}, linux-mi35x-gpu-8)', matrix.rocm_version) }}
strategy:
fail-fast: false
matrix:
rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm10", "rocm724", "rocm720"]') }}
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-deepseek-v32-mtp,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}

- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm

- name: Setup docker (${{ matrix.rocm_version }})
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
ENABLE_CACHE_HOST: "1"

- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
# Install tabulate for run_suite.py (missing in MI35x container)
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate

- name: Accuracy Test MI35x ROCm (8-GPU DeepSeek-V3.2 TP+MTP)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32-mtp --nightly --timeout-per-file 3600 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}

- name: Performance Test MI35x ROCm (8-GPU DeepSeek-V3.2 MTP)
timeout-minutes: 180
continue-on-error: true
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-deepseek-v32-mtp --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}

# ==============================================================================
# 8-GPU DeepSeek-R1 (MI35x only)
# ==============================================================================
Expand Down Expand Up @@ -1702,6 +1587,54 @@ jobs:
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-deepseek-v4-pro-dspark --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }}
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true

# ==============================================================================
# 4-GPU DeepSeek-V4.1-Flash (MI35x only - FP4 experts need gfx95x)
#
# The cookbook recipe is TP4 + EP4, so this uses four GPUs of an 8-GPU runner.
# ==============================================================================

nightly-4-gpu-mi35x-deepseek-v41-flash:
name: ${{ format('nightly-4-gpu-mi35x-deepseek-v41-flash ({0}, linux-mi35x-gpu-8)', matrix.rocm_version) }}
strategy:
fail-fast: false
matrix:
rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm10", "rocm724", "rocm720"]') }}
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-4-gpu-mi35x-deepseek-v41-flash,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}

- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm

- name: Setup docker (${{ matrix.rocm_version }})
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
ENABLE_CACHE_HOST: "1"

- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
# Install tabulate for run_suite.py (missing in MI35x container)
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate

- name: Accuracy Test MI35x ROCm (4-GPU DeepSeek-V4.1-Flash)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_MOE_COPY_WEIGHT_VIEWS_BEFORE_H2D=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-4-gpu-mi35x-deepseek-v41-flash --nightly --timeout-per-file 5400 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}

# ==============================================================================
# 8-GPU Kimi-K2.6 (MI30x)
#
Expand Down Expand Up @@ -2395,9 +2328,6 @@ jobs:
# 8-GPU DeepSeek-V3.2 (MI30x)
- nightly-8-gpu-deepseek-v32
- nightly-8-gpu-deepseek-v32-mtp
# 8-GPU DeepSeek-V3.2 (MI35x)
- nightly-8-gpu-mi35x-deepseek-v32
- nightly-8-gpu-mi35x-deepseek-v32-mtp
# 8-GPU DeepSeek-R1 (MI35x only)
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8
Expand All @@ -2411,6 +2341,8 @@ jobs:
- nightly-8-gpu-mi35x-deepseek-v4-pro
- nightly-8-gpu-mi35x-deepseek-v4-pro-mtp
- nightly-8-gpu-mi35x-deepseek-v4-pro-dspark
# 4-GPU DeepSeek-V4.1-Flash (MI35x only - FP4 experts need gfx95x)
- nightly-4-gpu-mi35x-deepseek-v41-flash
# 8-GPU Kimi-K2.6 (MI30x)
- nightly-8-gpu-kimi-k26
# 8-GPU Kimi-K3 (MI35x only - native MXFP4 needs gfx95x)
Expand Down
154 changes: 154 additions & 0 deletions test/registered/amd/accuracy/mi35x/test_deepseek_v41_flash_eval_mi35x.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,154 @@
"""MI35x DeepSeek-V4.1-Flash GSM8K Completion Evaluation Test (4-GPU)

Tests deepseek-ai/DeepSeek-V4.1-Flash with the GSM8K few-shot benchmark on
MI35x.

Server arguments follow the MI350X High-Throughput cell of the cookbook recipe
(docs/src/snippets/configs/deepseek-ai/deepseek-v4_1.jsx): TP4 + EP4 with
DSpark speculative decoding, decode CUDA graphs up to the 256-request cap, and
decoder SWA bounded replay. The attention and MoE backends are left to resolve
on HIP, as the recipe does. SGLANG_USE_AITER and SGLANG_MOE_PADDING come from
the ROCm image (docker/rocm.Dockerfile), so neither the recipe nor this test
sets them.

DeepSeek-V4.1-Flash ships FP4 experts, which need gfx95x, so this runs on MI35x
only. The recipe itself is TP4, so the job takes an 8-GPU MI35x runner and uses
four of its GPUs, like the MiniMax TP4 jobs.

The eval uses the few-shot *completion* harness, as the DeepSeek-V4-Flash MI35x
test does. It bypasses the chat template, so V4.1's thinking mode cannot route
the answer into `reasoning_content` and leave `message.content` empty. Scoring
raw completions keeps this test measuring whether the ROCm kernels produce
correct tokens.

Registry: nightly-amd-4-gpu-mi35x-deepseek-v41-flash suite
"""

import os
import unittest
from types import SimpleNamespace

import requests

from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)

# Register for AMD CI - DeepSeek-V4.1-Flash accuracy test on MI35x (~18 min in
# run 36797490320: ~16 min of server startup, then 1319 GSM8K questions in ~1 min)
register_amd_ci(
est_time=1200, suite="nightly-amd-4-gpu-mi35x-deepseek-v41-flash", nightly=True
)

DEEPSEEK_V41_FLASH_MODEL_PATH = os.environ.get(
"DEEPSEEK_V41_FLASH_MODEL_PATH", "deepseek-ai/DeepSeek-V4.1-Flash"
)
SERVER_LAUNCH_TIMEOUT = 5400
# Measured 0.911 on all 1319 questions with the High-Throughput cell at TP4 on
# MI35x (rocm10, sgl-project/sglang Actions run 36797490320). The threshold
# sits 0.02 under that.
ACCURACY_THRESHOLD = 0.89
TP_SIZE = 4


class TestDeepseekV41FlashEvalMI35x(CustomTestCase):
"""DeepSeek-V4.1-Flash GSM8K Completion Evaluation Test for AMD MI35x."""

@classmethod
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "1319"))
cls.max_new_tokens = int(os.environ.get("GSM8K_MAX_NEW_TOKENS", "512"))

def test_deepseek_v41_flash_gsm8k_accuracy(self):
"""Test DeepSeek-V4.1-Flash with GSM8K few-shot completion benchmark."""
other_args = [
"--tp",
str(TP_SIZE),
"--ep-size",
str(TP_SIZE),
"--mem-fraction-static",
"0.8",
"--speculative-algorithm",
"DSPARK",
"--speculative-dspark-block-size",
"5",
"--cuda-graph-max-bs-decode",
"256",
"--max-running-requests",
"256",
"--enable-decoder-swa-bounded-replay",
"--reasoning-parser",
"auto",
"--tool-call-parser",
"auto",
"--trust-remote-code",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
"--watchdog-timeout",
"1200",
]
env = os.environ.copy()
# Both come from the recipe. AITER_FLYDSL_FORCE_REDUCE forces the FlyDSL
# MoE down-projection onto a per-slot reduce instead of atomics, which
# makes the output repeatable run to run.
env["AITER_FLYDSL_FORCE_REDUCE"] = "1"
# At the default bound, small batches take AITER's BF16-activation MoE
# route, which has no kernel for these experts at the current pin: the
# server fails during decode-graph warmup. Keep this until
# ROCm/aiter#5802 is in the AITER pin (see sgl-project/sglang#41308).
# AITER reads it directly, so it will not grep to anything in-tree.
env["AITER_BF16_FP8_MOE_BOUND"] = "0"

process = popen_launch_server(
DEEPSEEK_V41_FLASH_MODEL_PATH,
self.base_url,
timeout=SERVER_LAUNCH_TIMEOUT,
other_args=other_args,
env=env,
)

try:
requests.get(self.base_url + "/flush_cache")

args = SimpleNamespace(
num_shots=8,
data_path=None,
num_questions=self.num_questions,
parallel=self.num_questions,
max_new_tokens=self.max_new_tokens,
host="127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
acc = metrics["accuracy"]

passed = acc >= ACCURACY_THRESHOLD
status = "✅ PASS" if passed else "❌ FAIL"
print(f" accuracy={acc:.3f} threshold={ACCURACY_THRESHOLD} {status}")

if is_in_ci():
summary = "### DeepSeek-V4.1-Flash Model (MI35x)\n\n"
summary += "| Model | TP | Accuracy | Threshold | Status |\n"
summary += "| ----- | -- | -------- | --------- | ------ |\n"
summary += f"| {DEEPSEEK_V41_FLASH_MODEL_PATH} | {TP_SIZE} | {acc:.3f} | {ACCURACY_THRESHOLD} | {status} |\n"
write_github_step_summary(summary)

self.assertGreaterEqual(
acc,
ACCURACY_THRESHOLD,
f"DeepSeek-V4.1-Flash accuracy {acc:.3f} below threshold {ACCURACY_THRESHOLD}",
)
finally:
kill_process_tree(process.pid)


if __name__ == "__main__":
unittest.main()
1 change: 1 addition & 0 deletions test/run_suite.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -159,6 +159,7 @@
"nightly-amd-vlm",
"nightly-amd-accuracy-8-gpu-deepseek-v4-flash",
"nightly-amd-8-gpu-mi35x-deepseek-v4-flash",
"nightly-amd-4-gpu-mi35x-deepseek-v41-flash",
# MI35x 8-GPU suite (different model configs)
"nightly-amd-8-gpu-mi35x",
],
Expand Down
Loading