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

[AMD][CI] Add GLM-5-MXFP4 accuracy and perf nightly tests for MI35x (#21773)

Browse Source
This commit is contained in:
Michael
2026-04-14 18:55:36 -07:00
committed by GitHub
parent adb310b976
commit 39c6bf730c
4 changed files with 564 additions and 166 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+48 -82
View File
@@ -43,7 +43,6 @@ on:
- nightly-8-gpu-kimi-k25-rocm720
- nightly-8-gpu-qwen3-235b-rocm720
- nightly-8-gpu-qwen35-rocm720
- nightly-8-gpu-glm5-rocm720
- nightly-8-gpu-glm51-rocm720
- nightly-8-gpu-minimax-m27-rocm720
- nightly-1-gpu-zimage-turbo-rocm720
@@ -61,8 +60,8 @@ on:
- nightly-8-gpu-mi35x-kimi-k25-rocm720
- nightly-8-gpu-mi35x-qwen3-235b-mxfp4-rocm720
- nightly-8-gpu-mi35x-qwen35-rocm720
- nightly-8-gpu-mi35x-glm5-rocm720
- nightly-8-gpu-mi35x-glm51-rocm720
- nightly-8-gpu-mi35x-glm5-mxfp4-rocm720
job_filter:
description: 'Or type comma-separated job names (overrides dropdown if non-empty)'
required: false
@@ -665,50 +664,6 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU GLM-5 (Accuracy + Performance combined) ROCm 7.2
nightly-8-gpu-glm5-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-glm5-rocm720,'))
runs-on: linux-mi325-8gpu-sglang
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.ref }}
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git@96f807a33b75
- name: Accuracy Test ROCm 7.2 (8-GPU GLM-5 NSA)
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-glm5 --nightly --timeout-per-file 3600 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test ROCm 7.2 (8-GPU GLM-5)
timeout-minutes: 120
continue-on-error: true # Perf test failure doesn't fail the job if accuracy passed
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-glm5 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU GLM-5.1 (Accuracy + Performance combined) ROCm 7.2
nightly-8-gpu-glm51-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-glm51-rocm720,'))
@@ -1332,40 +1287,6 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU GLM-5 (Accuracy only) ROCm 7.2
nightly-8-gpu-mi35x-glm5-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-glm5-rocm720,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.ref }}
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- 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
bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git@96f807a33b75
- name: Accuracy Test MI35x ROCm 7.2 (8-GPU GLM-5 NSA)
timeout-minutes: 180
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-8-gpu-mi35x-glm5 --nightly --timeout-per-file 7200 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU GLM-5.1 (Accuracy + Performance combined) ROCm 7.2
nightly-8-gpu-mi35x-glm51-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-glm51-rocm720,'))
@@ -1410,6 +1331,52 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU GLM-5-MXFP4 (Accuracy + Performance combined) ROCm 7.2
nightly-8-gpu-mi35x-glm5-mxfp4-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-glm5-mxfp4-rocm720,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.ref }}
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git@96f807a33b75
- name: Accuracy Test MI35x ROCm 7.2 (8-GPU GLM-5-MXFP4)
timeout-minutes: 180
run: |
> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-glm5-mxfp4 --nightly --timeout-per-file 7200 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test MI35x ROCm 7.2 (8-GPU GLM-5-MXFP4)
timeout-minutes: 300
continue-on-error: true
run: |
> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 registered/amd/perf/mi35x/test_glm5_mxfp4_perf_mi35x.py || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU DeepSeek-V3.2 Performance Test (MTP) ROCm 7.2
nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720,'))
@@ -1467,7 +1434,6 @@ jobs:
- nightly-8-gpu-kimi-k25-rocm720
- nightly-8-gpu-qwen3-235b-rocm720
- nightly-8-gpu-qwen35-rocm720
- nightly-8-gpu-glm5-rocm720
- nightly-8-gpu-glm51-rocm720
- nightly-8-gpu-minimax-m27-rocm720
# MI30x ROCm 7.2 Diffusion Tests
@@ -1487,8 +1453,8 @@ jobs:
- nightly-8-gpu-mi35x-kimi-k25-rocm720
- nightly-8-gpu-mi35x-qwen3-235b-mxfp4-rocm720
- nightly-8-gpu-mi35x-qwen35-rocm720
- nightly-8-gpu-mi35x-glm5-rocm720
- nightly-8-gpu-mi35x-glm51-rocm720
- nightly-8-gpu-mi35x-glm5-mxfp4-rocm720
runs-on: ubuntu-latest
steps:
- name: Check if any job failed
.github/workflows/nightly-test-amd.yml
+48 -84
View File
@@ -43,9 +43,7 @@ on:
- nightly-8-gpu-kimi-k25
- nightly-8-gpu-qwen3-235b
- nightly-8-gpu-qwen35
- nightly-8-gpu-glm5
- nightly-8-gpu-glm51
- nightly-8-gpu-glm51-mxfp4
- nightly-8-gpu-minimax-m27
- nightly-1-gpu-zimage-turbo
- nightly-test-1-gpu-mi35x
@@ -62,9 +60,8 @@ on:
- nightly-8-gpu-mi35x-kimi-k25
- nightly-8-gpu-mi35x-qwen3-235b-mxfp4
- nightly-8-gpu-mi35x-qwen35
- nightly-8-gpu-mi35x-glm5
- nightly-8-gpu-mi35x-glm51
- nightly-8-gpu-mi35x-glm51-mxfp4
- nightly-8-gpu-mi35x-glm5-mxfp4
job_filter:
description: 'Or type comma-separated job names (overrides dropdown if non-empty)'
required: false
@@ -671,50 +668,6 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU GLM-5 (Accuracy + Performance combined)
nightly-8-gpu-glm5:
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-glm5,'))
runs-on: linux-mi325-8gpu-sglang
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.ref }}
- name: Setup docker
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh
bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git@96f807a33b75
- name: Accuracy Test (8-GPU GLM-5 NSA)
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-glm5 --nightly --timeout-per-file 3600 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test (8-GPU GLM-5)
timeout-minutes: 120
continue-on-error: true # Perf test failure doesn't fail the job if accuracy passed
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-glm5 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU GLM-5.1 (Accuracy + Performance combined)
nightly-8-gpu-glm51:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-glm51,'))
@@ -1341,40 +1294,6 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU GLM-5 (Accuracy only)
nightly-8-gpu-mi35x-glm5:
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-glm5,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.ref }}
- name: Setup docker
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh
# Install tabulate for run_suite.py (missing in MI35x container)
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git@96f807a33b75
- name: Accuracy Test MI35x (8-GPU GLM-5 NSA)
timeout-minutes: 180
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-8-gpu-mi35x-glm5 --nightly --timeout-per-file 7200 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU GLM-5.1 (Accuracy + Performance combined)
nightly-8-gpu-mi35x-glm51:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-glm51,'))
@@ -1419,6 +1338,52 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU GLM-5-MXFP4 (Accuracy + Performance combined)
nightly-8-gpu-mi35x-glm5-mxfp4:
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-glm5-mxfp4,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.ref }}
- name: Setup docker
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git@96f807a33b75
- name: Accuracy Test MI35x (8-GPU GLM-5-MXFP4)
timeout-minutes: 180
run: |
> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-glm5-mxfp4 --nightly --timeout-per-file 7200 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test MI35x (8-GPU GLM-5-MXFP4)
timeout-minutes: 300
continue-on-error: true
run: |
> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 registered/amd/perf/mi35x/test_glm5_mxfp4_perf_mi35x.py || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU DeepSeek-V3.2 Performance Test (MTP)
nightly-perf-8-gpu-mi35x-deepseek-v32-mtp:
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-perf-8-gpu-mi35x-deepseek-v32-mtp,'))
@@ -1476,7 +1441,6 @@ jobs:
- nightly-8-gpu-kimi-k25
- nightly-8-gpu-qwen3-235b
- nightly-8-gpu-qwen35
- nightly-8-gpu-glm5
- nightly-8-gpu-glm51
- nightly-8-gpu-minimax-m27
# MI30x Diffusion Tests
@@ -1494,8 +1458,8 @@ jobs:
- nightly-8-gpu-mi35x-kimi-k25
- nightly-8-gpu-mi35x-qwen3-235b-mxfp4
- nightly-8-gpu-mi35x-qwen35
- nightly-8-gpu-mi35x-glm5
- nightly-8-gpu-mi35x-glm51
- nightly-8-gpu-mi35x-glm5-mxfp4
# MI35x perf jobs excluded from check - perf failures don't block CI
# - nightly-perf-8-gpu-mi35x-deepseek-v32-basic
# - nightly-perf-8-gpu-mi35x-deepseek-v32-mtp
test/registered/amd/accuracy/mi35x/test_glm5_mxfp4_eval_mi35x.py
+281
View File
@@ -0,0 +1,281 @@
"""MI35x GLM-5-MXFP4 GSM8K Completion Evaluation Test (8-GPU)
Tests the AMD Quark MXFP4-quantized GLM-5 model using few-shot
completion benchmark on MI35x.
Model: amd/GLM-5-MXFP4 (MOE-only MXFP4 quantization of zai-org/GLM-5)
Reference: https://huggingface.co/amd/GLM-5-MXFP4
Registry: nightly-amd-8-gpu-mi35x-glm5-mxfp4 suite
"""
import ast
import os
os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
import re
import time
import unittest
from dataclasses import dataclass
from typing import List, Optional, Tuple
import numpy as np
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
from sglang.utils import download_and_cache_file, read_jsonl
register_amd_ci(
est_time=5400,
suite="nightly-amd-8-gpu-mi35x-glm5-mxfp4",
nightly=True,
)
INVALID = -9999999
GLM5_MXFP4_LOCAL_PATH = "/data2/models/amd-GLM-5-MXFP4"
GLM5_MXFP4_HF_MODEL_ID = "amd/GLM-5-MXFP4"
def get_model_path() -> str:
"""Get effective model path: env var > local path > HF model ID."""
env_path = os.environ.get("GLM5_MXFP4_MODEL_PATH")
if env_path:
return env_path
if os.path.exists(GLM5_MXFP4_LOCAL_PATH):
return GLM5_MXFP4_LOCAL_PATH
return GLM5_MXFP4_HF_MODEL_ID
@dataclass
class ModelConfig:
"""Configuration for a model to test."""
model_path: str
tp_size: int = 8
accuracy_threshold: float = 0.50
other_args: Optional[List[str]] = None
env_vars: Optional[dict] = None
timeout: Optional[int] = None
variant: Optional[str] = None
def __post_init__(self):
if self.other_args is None:
self.other_args = []
if self.env_vars is None:
self.env_vars = {}
def get_display_name(self) -> str:
if self.variant:
return f"{self.model_path} ({self.variant})"
return self.model_path
def get_glm5_mxfp4_models() -> List[ModelConfig]:
"""Get GLM-5-MXFP4 model configurations for MI35x."""
model_path = get_model_path()
return [
ModelConfig(
model_path=model_path,
tp_size=8,
accuracy_threshold=0.90,
timeout=5400,
variant="mxfp4",
other_args=[
"--trust-remote-code",
"--chunked-prefill-size",
"131072",
"--disable-radix-cache",
"--mem-fraction-static",
"0.85",
"--context-length",
"4096",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
"--watchdog-timeout",
"1200",
],
env_vars={"SGLANG_USE_AITER": "1"},
),
]
def get_one_example(lines, i, include_answer):
"""Format a single GSM8K example."""
ret = "Question: " + lines[i]["question"] + "\nAnswer:"
if include_answer:
ret += " " + lines[i]["answer"]
return ret
def get_few_shot_examples(lines, k):
"""Get k few-shot examples for prompting."""
ret = ""
for i in range(k):
ret += get_one_example(lines, i, True) + "\n\n"
return ret
def get_answer_value(answer_str):
"""Extract numerical answer from response."""
answer_str = answer_str.replace(",", "")
numbers = re.findall(r"\d+", answer_str)
if len(numbers) < 1:
return INVALID
try:
return ast.literal_eval(numbers[-1])
except SyntaxError:
return INVALID
def run_gsm8k_benchmark(
base_url: str,
num_questions: int = 200,
num_shots: int = 5,
parallel: int = 64,
) -> Tuple[float, float, float]:
"""Run GSM8K few-shot completion benchmark."""
import sglang as sgl
from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
data_path = download_and_cache_file(url)
lines = list(read_jsonl(data_path))
few_shot_examples = get_few_shot_examples(lines, num_shots)
questions = []
labels = []
for i in range(len(lines[:num_questions])):
questions.append(get_one_example(lines, i, False))
labels.append(get_answer_value(lines[i]["answer"]))
assert all(l != INVALID for l in labels)
arguments = [{"question": q} for q in questions]
@sgl.function
def few_shot_gsm8k(s, question):
s += few_shot_examples + question
s += sgl.gen(
"answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
)
backend = RuntimeEndpoint(base_url)
sgl.set_default_backend(backend)
tic = time.perf_counter()
states = few_shot_gsm8k.run_batch(
arguments, temperature=0, num_threads=parallel, progress_bar=True
)
latency = time.perf_counter() - tic
preds = [get_answer_value(states[i]["answer"]) for i in range(len(states))]
acc = np.mean(np.array(preds) == np.array(labels))
invalid = np.mean(np.array(preds) == INVALID)
return float(acc), float(invalid), float(latency)
class TestGLM5MXFP4EvalMI35x(unittest.TestCase):
"""GLM-5-MXFP4 GSM8K Completion Evaluation Test for AMD MI35x."""
@classmethod
def setUpClass(cls):
cls.models = get_glm5_mxfp4_models()
cls.base_url = DEFAULT_URL_FOR_TEST
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
def test_glm5_mxfp4_accuracy(self):
"""Test GLM-5-MXFP4 with GSM8K completion benchmark."""
model_path = get_model_path()
is_local_path = model_path.startswith("/")
if is_local_path and not os.path.exists(model_path):
print(f"\nSKIPPING: Local model not found at {model_path}")
self.skipTest(f"Local model not found at {model_path}")
return
if is_local_path:
print(f"Using local model: {model_path}")
else:
print(f"Using HuggingFace model: {model_path}")
all_results = []
summary = "### GLM-5-MXFP4 Models (MI35x)\n\n"
summary += "| Model | Variant | TP | Accuracy | Threshold | Status |\n"
summary += "| ----- | ------- | -- | -------- | --------- | ------ |\n"
for config in self.models:
display_name = config.get_display_name()
with self.subTest(model=display_name):
print(f"\n{'='*60}")
print(f"Testing: {display_name}")
print(f"{'='*60}")
env = os.environ.copy()
for key, value in config.env_vars.items():
env[key] = value
other_args = list(config.other_args)
other_args.extend(["--tp", str(config.tp_size)])
timeout = config.timeout or DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
try:
process = popen_launch_server(
model=config.model_path,
base_url=self.base_url,
timeout=timeout,
other_args=other_args,
env=env,
)
try:
acc, invalid, latency = run_gsm8k_benchmark(
self.base_url, num_questions=self.num_questions
)
passed = acc >= config.accuracy_threshold
status = "PASS" if passed else "FAIL"
print(
f" accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
)
all_results.append(
{
"model": display_name,
"accuracy": acc,
"passed": passed,
}
)
summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
finally:
kill_process_tree(process.pid)
except Exception as e:
summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ERROR |\n"
all_results.append(
{
"model": display_name,
"accuracy": None,
"passed": False,
"error": str(e),
}
)
if is_in_ci():
write_github_step_summary(summary)
failed = [r for r in all_results if not r["passed"]]
if failed:
raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
if __name__ == "__main__":
unittest.main()
test/registered/amd/perf/mi35x/test_glm5_mxfp4_perf_mi35x.py
+187
View File
@@ -0,0 +1,187 @@
"""MI35x Nightly performance benchmark for GLM-5-MXFP4 model.
Benchmarks the AMD Quark MXFP4-quantized GLM-5 model on MI35x with 8 GPUs.
Model: amd/GLM-5-MXFP4 (MOE-only MXFP4 quantization of zai-org/GLM-5)
Reference: https://huggingface.co/amd/GLM-5-MXFP4
Registry: nightly-perf-8-gpu-mi35x-glm5-mxfp4 suite
"""
import os
os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
import unittest
from typing import List
from sglang.test.ci.ci_register import register_amd_ci
from sglang.test.nightly_bench_utils import BenchmarkResult
from sglang.test.nightly_utils import NightlyBenchmarkRunner
from sglang.test.test_utils import DEFAULT_URL_FOR_TEST, _parse_int_list_env
register_amd_ci(
est_time=18000,
suite="nightly-perf-8-gpu-mi35x-glm5-mxfp4",
nightly=True,
)
def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
"""Generate a simplified markdown report without traces and cost columns.
Skips the first result if it's a warmup run (duplicate batch_size).
"""
model_header = results[0].model_path
if results[0].run_name and results[0].run_name != "default":
model_header += f" ({results[0].run_name})"
gpu_config = os.getenv("GPU_CONFIG", "MI35x")
if gpu_config:
model_header += f" [{gpu_config}]"
summary = f"### {model_header}\n"
summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
report_results = (
results[1:]
if len(results) > 1 and results[0].batch_size == results[1].batch_size
else results
)
for result in report_results:
itl = (
1 / (result.output_throughput / result.batch_size) * 1000
if result.output_throughput > 0
else 0
)
summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
return summary
GLM5_MXFP4_LOCAL_PATH = "/data2/models/amd-GLM-5-MXFP4"
GLM5_MXFP4_HF_MODEL_ID = "amd/GLM-5-MXFP4"
PROFILE_DIR = "performance_profiles_glm5_mxfp4_mi35x"
def get_model_path() -> str:
"""Get effective model path: env var > local path > HF model ID."""
env_path = os.environ.get("GLM5_MXFP4_MODEL_PATH")
if env_path:
return env_path
if os.path.exists(GLM5_MXFP4_LOCAL_PATH):
return GLM5_MXFP4_LOCAL_PATH
return GLM5_MXFP4_HF_MODEL_ID
class TestGLM5MXFP4PerfMI35x(unittest.TestCase):
"""MI35x Nightly performance benchmark for GLM-5-MXFP4 model."""
@classmethod
def setUpClass(cls):
cls.model = get_model_path()
print(f"Using model path: {cls.model}")
cls.base_url = DEFAULT_URL_FOR_TEST
cls.batch_sizes = [1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "1024"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "1024"))
cls.variants = [
{
"name": "basic",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--chunked-prefill-size",
"131072",
"--disable-radix-cache",
"--mem-fraction-static",
"0.85",
"--context-length",
"4096",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
"--watchdog-timeout",
"1200",
"--reasoning-parser",
"glm45",
"--tool-call-parser",
"glm47",
],
},
]
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
cls.runner.setup_profile_directory()
cls.runner.full_report = f"## {cls.__name__}\n"
def test_bench_one_batch(self):
"""Run benchmark across all configured variants."""
failed_variants = []
is_local_path = self.model.startswith("/")
if is_local_path and not os.path.exists(self.model):
print(f"\nSKIPPING: Local model not found at {self.model}")
self.runner.full_report += (
f"\nTest skipped: Local model not found at {self.model}\n"
)
self.runner.write_final_report()
return
if is_local_path:
print(f"Using local model: {self.model}")
else:
print(
f"Using HuggingFace model: {self.model} (will download if not cached)"
)
old_env = {}
env_vars = {"SGLANG_USE_AITER": "1"}
for key, value in env_vars.items():
old_env[key] = os.environ.get(key)
os.environ[key] = value
try:
for variant_config in self.variants:
with self.subTest(variant=variant_config["name"]):
result_tuple = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=variant_config["other_args"],
variant=variant_config["name"],
extra_bench_args=["--trust-remote-code"],
enable_profile=False,
)
results = result_tuple[0]
success = result_tuple[1]
if not success:
failed_variants.append(variant_config["name"])
if results:
self.runner.full_report += (
generate_simple_markdown_report(results) + "\n"
)
finally:
for key, value in old_env.items():
if value is None:
os.environ.pop(key, None)
else:
os.environ[key] = value
self.runner.write_final_report()
if failed_variants:
raise AssertionError(
f"Benchmark failed for {self.model} with the following variants: "
f"{', '.join(failed_variants)}"
)
if __name__ == "__main__":
unittest.main()