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

[AMD] Add MiniMax-M2.5 TP=4 nightly accuracy test for MI355X (#27126)

Browse Source
This commit is contained in:
YC Yen-Ching Tseng
2026-06-03 22:17:40 -07:00
committed by GitHub
parent ff5c4d7b57
commit 6dcd78a37f
3 changed files with 339 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+42
View File
@@ -82,6 +82,8 @@ on:
- nightly-8-gpu-mi35x-glm51-rocm720
# 8-GPU GLM-5-MXFP4 (MI35x only)
- nightly-8-gpu-mi35x-glm5-mxfp4-rocm720
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25-rocm720
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27-rocm720
# Diffusion (MI30x)
@@ -1681,6 +1683,44 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 4-GPU MiniMax-M2.5 (MI35x)
# ==============================================================================
nightly-4-gpu-mi35x-minimax-m25-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-minimax-m25-rocm720,'))
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 (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
- name: Accuracy Test MI35x ROCm 7.2 (4-GPU MiniMax-M2.5)
timeout-minutes: 120
run: |
> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e SGLANG_USE_AITER_UNIFIED_ATTN=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-4-gpu-mi35x-minimax-m25-tp4 --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 ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 8-GPU MiniMax-M2.7 (MI30x only)
# ==============================================================================
@@ -1831,6 +1871,8 @@ jobs:
- nightly-8-gpu-mi35x-glm51-rocm720
# 8-GPU GLM-5-MXFP4 (MI35x only)
- nightly-8-gpu-mi35x-glm5-mxfp4-rocm720
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25-rocm720
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27-rocm720
# Diffusion (MI30x)
.github/workflows/nightly-test-amd.yml
+42
View File
@@ -79,6 +79,8 @@ on:
- nightly-8-gpu-mi35x-glm51
# 8-GPU GLM-5-MXFP4 (MI35x only)
- nightly-8-gpu-mi35x-glm5-mxfp4
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27
# Diffusion (MI30x)
@@ -1612,6 +1614,44 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 4-GPU MiniMax-M2.5 (MI35x)
# ==============================================================================
nightly-4-gpu-mi35x-minimax-m25:
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-minimax-m25,'))
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 (ROCm 7.0)
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 --skip-test-time-deps
- name: Accuracy Test MI35x (4-GPU MiniMax-M2.5)
timeout-minutes: 120
run: |
> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e SGLANG_USE_AITER_UNIFIED_ATTN=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-4-gpu-mi35x-minimax-m25-tp4 --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 MiniMax-M2.7 (MI30x only)
# ==============================================================================
@@ -1759,6 +1799,8 @@ jobs:
- nightly-8-gpu-mi35x-glm51
# 8-GPU GLM-5-MXFP4 (MI35x only)
- nightly-8-gpu-mi35x-glm5-mxfp4
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27
# Diffusion (MI30x)
test/registered/amd/accuracy/mi35x/test_minimax_m25_tp4_eval_mi35x.py
+255
View File
@@ -0,0 +1,255 @@
"""MI35x MiniMax-M2.5 GSM8K Completion Evaluation Test (4-GPU, TP=4)
Tests MiniMax-M2.5 with TP=4 + EP=1 configuration using few-shot completion
benchmark on MI35x. This configuration uses unified attention with FP8 KV cache.
Registry: nightly-amd-4-gpu-mi35x-minimax-m25-tp4 suite
"""
import os
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-4-gpu-mi35x-minimax-m25-tp4",
nightly=True,
)
INVALID = -9999999
@dataclass
class ModelConfig:
"""Configuration for a model to test."""
model_path: str
tp_size: int = 4
accuracy_threshold: float = 0.93
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
MI35X_MINIMAX_M25_TP4_MODELS = [
ModelConfig(
model_path="MiniMaxAI/MiniMax-M2.5",
tp_size=4,
accuracy_threshold=0.93,
timeout=5400,
variant="TP4+UnifiedAttn+FP8KV",
other_args=[
"--ep-size",
"1",
"--dtype",
"bfloat16",
"--trust-remote-code",
"--attention-backend",
"aiter",
"--mem-fraction-static",
"0.85",
"--disable-radix-cache",
"--kv-cache-dtype",
"fp8_e4m3",
"--page-size",
"16",
"--watchdog-timeout",
"1200",
],
env_vars={
"SGLANG_USE_AITER": "1",
"SGLANG_USE_AITER_UNIFIED_ATTN": "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."""
if not isinstance(answer_str, str):
return INVALID
answer_str = answer_str.replace(",", "")
numbers = re.findall(r"\d+", answer_str)
if not numbers:
return INVALID
try:
return int(numbers[-1])
except ValueError:
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 TestMiniMaxM25TP4EvalMI35x(unittest.TestCase):
"""MiniMax-M2.5 TP=4 GSM8K Completion Evaluation Test for AMD MI35x."""
@classmethod
def setUpClass(cls):
cls.models = MI35X_MINIMAX_M25_TP4_MODELS
cls.base_url = DEFAULT_URL_FOR_TEST
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "1319"))
def test_minimax_m25_tp4_accuracy(self):
"""Test MiniMax-M2.5 TP=4 with GSM8K completion benchmark."""
all_results = []
summary = "### MiniMax-M2.5 TP=4 (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}"
)
print(f" invalid={invalid:.3f} latency={latency:.1f}s")
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()