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

[AMD] [MiniMax-M2.5 Day 0] Add MiniMax-M2.5 nightly accuracy test (#19443)

Browse Source
This commit is contained in:
Michael
2026-02-27 02:39:33 -08:00
committed by GitHub
parent f69ca93d49
commit 403195d59d
6 changed files with 653 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
.github/workflows/nightly-test-amd-rocm720.yml vendored
View File

@@ -34,6 +34,7 @@ on:
- 'nightly-8-gpu-kimi-k25-rocm720'
- 'nightly-8-gpu-qwen3-235b-rocm720'
- 'nightly-8-gpu-glm5-rocm720'
- 'nightly-8-gpu-minimax-m25-rocm720'
# MI35x ROCm 7.2 jobs
- 'nightly-test-1-gpu-mi35x-rocm720'
- 'nightly-8-gpu-mi35x-qwen3-235b-mxfp4-rocm720'
@@ -47,6 +48,7 @@ on:
- 'nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720'
- 'nightly-8-gpu-mi35x-kimi-k25-rocm720'
- 'nightly-8-gpu-mi35x-glm5-rocm720'
- 'nightly-8-gpu-mi35x-minimax-m25-rocm720'
workflow_call:
inputs:
ref:
@@ -561,6 +563,37 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU MiniMax-M2.5 (Accuracy) ROCm 7.2
nightly-8-gpu-minimax-m25-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-minimax-m25-rocm720')
runs-on: linux-mi325-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: ${{ github.workspace }}
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-aiter-build --skip-test-time-deps
- name: Accuracy Test ROCm 7.2 (8-GPU MiniMax-M2.5)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-minimax-m25 --nightly --timeout-per-file 3600 --continue-on-error || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ============================================== MI35x ROCm 7.2 Tests ==============================================
# MI35x 1-GPU ROCm 7.2 tests
nightly-test-1-gpu-mi35x-rocm720:
@@ -959,6 +992,39 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU MiniMax-M2.5 (Accuracy) ROCm 7.2
nightly-8-gpu-mi35x-minimax-m25-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-mi35x-minimax-m25-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: ${{ github.workspace }}
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-aiter-build --skip-test-time-deps
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
- name: Accuracy Test MI35x ROCm 7.2 (8-GPU MiniMax-M2.5)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-minimax-m25 --nightly --timeout-per-file 3600 --continue-on-error || TEST_EXIT_CODE=$?
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_filter == 'all' || inputs.job_filter == 'nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720')
@@ -1013,6 +1079,7 @@ jobs:
- nightly-8-gpu-kimi-k25-rocm720
- nightly-8-gpu-qwen3-235b-rocm720
- nightly-8-gpu-glm5-rocm720
- nightly-8-gpu-minimax-m25-rocm720
# MI35x ROCm 7.2 jobs
- nightly-test-1-gpu-mi35x-rocm720
- nightly-accuracy-8-gpu-mi35x-rocm720
@@ -1026,6 +1093,7 @@ jobs:
- nightly-8-gpu-mi35x-kimi-k25-rocm720
- nightly-8-gpu-mi35x-qwen3-235b-mxfp4-rocm720
- nightly-8-gpu-mi35x-glm5-rocm720
- nightly-8-gpu-mi35x-minimax-m25-rocm720
runs-on: ubuntu-latest
steps:
- name: Check if any job failed

68
.github/workflows/nightly-test-amd.yml vendored
View File

@@ -34,11 +34,13 @@ on:
- 'nightly-8-gpu-kimi-k25'
- 'nightly-8-gpu-qwen3-235b'
- 'nightly-8-gpu-glm5'
- 'nightly-8-gpu-minimax-m25'
# MI35x jobs
- 'nightly-test-1-gpu-mi35x'
- 'nightly-8-gpu-mi35x-qwen3-235b-mxfp4'
- 'nightly-8-gpu-mi35x-kimi-k25'
- 'nightly-8-gpu-mi35x-glm5'
- 'nightly-8-gpu-mi35x-minimax-m25'
- 'nightly-accuracy-8-gpu-mi35x'
- 'nightly-8-gpu-mi35x-grok1-int4'
- 'nightly-8-gpu-mi35x-grok2'
@@ -559,6 +561,37 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU MiniMax-M2.5 (Accuracy)
nightly-8-gpu-minimax-m25:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-minimax-m25')
runs-on: linux-mi325-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: ${{ github.workspace }}
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Accuracy Test (8-GPU MiniMax-M2.5)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-minimax-m25 --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ============================================== MI35x Tests ==============================================
# MI35x 1-GPU tests - platform-agnostic tests that may work on CDNA4 (gfx950)
nightly-test-1-gpu-mi35x:
@@ -959,6 +992,39 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU MiniMax-M2.5 (Accuracy)
nightly-8-gpu-mi35x-minimax-m25:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-mi35x-minimax-m25')
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: ${{ github.workspace }}
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
- name: Accuracy Test MI35x (8-GPU MiniMax-M2.5)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-minimax-m25 --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
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_filter == 'all' || inputs.job_filter == 'nightly-perf-8-gpu-mi35x-deepseek-v32-mtp')
@@ -1013,6 +1079,7 @@ jobs:
- nightly-8-gpu-kimi-k25
- nightly-8-gpu-qwen3-235b
- nightly-8-gpu-glm5
- nightly-8-gpu-minimax-m25
# MI35x jobs
- nightly-test-1-gpu-mi35x
- nightly-accuracy-8-gpu-mi35x
@@ -1024,6 +1091,7 @@ jobs:
- nightly-8-gpu-mi35x-kimi-k25
- nightly-8-gpu-mi35x-qwen3-235b-mxfp4
- nightly-8-gpu-mi35x-glm5
- nightly-8-gpu-mi35x-minimax-m25
# 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

25
docs/basic_usage/minimax_m2.md
View File

@@ -1,13 +1,14 @@
# MiniMax M2.1/M2 Usage
# MiniMax M2.5/M2.1/M2 Usage
[MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1) and [MiniMax-M2](https://huggingface.co/MiniMaxAI/MiniMax-M2) are advanced large language models created by [MiniMax](https://www.minimax.io/).
[MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5), [MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1), and [MiniMax-M2](https://huggingface.co/MiniMaxAI/MiniMax-M2) are advanced large language models created by [MiniMax](https://www.minimax.io/).
MiniMax-M2 series redefines efficiency for agents. It's a compact, fast, and cost-effective MoE model (230 billion total parameters with 10 billion active parameters) built for elite performance in coding and agentic tasks, all while maintaining powerful general intelligence. With just 10 billion activated parameters, MiniMax-M2 provides the sophisticated, end-to-end tool use performance expected from today's leading models, but in a streamlined form factor that makes deployment and scaling easier than ever.
The MiniMax-M2 series redefines efficiency for agents. These compact, fast, and cost-effective MoE models (230 billion total parameters with 10 billion active parameters) are built for elite performance in coding and agentic tasks, all while maintaining powerful general intelligence. With just 10 billion activated parameters, the MiniMax-M2 series provides sophisticated, end-to-end tool use performance expected from today's leading models, but in a streamlined form factor that makes deployment and scaling easier than ever.
## Supported Models
This guide applies to the following models. You only need to update the model name during deployment. The following examples use **MiniMax-M2**:
- [MiniMaxAI/MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)
- [MiniMaxAI/MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1)
- [MiniMaxAI/MiniMax-M2](https://huggingface.co/MiniMaxAI/MiniMax-M2)
@@ -49,6 +50,24 @@ python -m sglang.launch_server \
--mem-fraction-static 0.85
```
### AMD GPUs (MI300X/MI325X/MI355X)
8-GPU deployment command:
```bash
SGLANG_USE_AITER=1 python -m sglang.launch_server \
--model-path MiniMaxAI/MiniMax-M2.5 \
--tp-size 8 \
--ep-size 8 \
--attention-backend aiter \
--tool-call-parser minimax-m2 \
--reasoning-parser minimax-append-think \
--host 0.0.0.0 \
--trust-remote-code \
--port 8000 \
--mem-fraction-static 0.85
```
## Testing Deployment
After startup, you can test the SGLang OpenAI-compatible API with the following command:

2
docs/supported_models/text_generation/generative_models.md
View File

@@ -37,7 +37,7 @@ in the GitHub search bar.
| **MiniCPM** (v3, 4B) | `openbmb/MiniCPM3-4B` | OpenBMB’s series of compact LLMs for edge devices; MiniCPM 3 (4B) achieves GPT-3.5-level results in text tasks. |
| **OLMo** (2, 3) | `allenai/OLMo-3-1125-32B`, `allenai/OLMo-3-32B-Think`, `allenai/OLMo-2-1124-7B-Instruct` | Allen AI’s series of Open Language Models designed to enable the science of language models. |
| **OLMoE** (Open MoE) | `allenai/OLMoE-1B-7B-0924` | Allen AI’s open Mixture-of-Experts model (7B total, 1B active parameters) delivering state-of-the-art results with sparse expert activation. |
| **MiniMax-M2** (M2, M2.1) | `minimax/MiniMax-M2`, `minimax/MiniMax-M2.1` | MiniMax’s SOTA LLM for coding & agentic workflows. |
| **MiniMax-M2** (M2, M2.1, M2.5) | `MiniMaxAI/MiniMax-M2.5`, `MiniMaxAI/MiniMax-M2.1`, `MiniMaxAI/MiniMax-M2` | MiniMax's SOTA LLM for coding & agentic workflows. |
| **StableLM** (3B, 7B) | `stabilityai/stablelm-tuned-alpha-7b` | StabilityAI’s early open-source LLM (3B & 7B) for general text generation; a demonstration model with basic instruction-following ability. |
| **Command-(R,A)** (Cohere) | `CohereLabs/c4ai-command-r-v01`, `CohereLabs/c4ai-command-r7b-12-2024`, `CohereLabs/c4ai-command-a-03-2025` | Cohere’s open conversational LLM (Command series) optimized for long context, retrieval-augmented generation, and tool use. |
| **DBRX** (Databricks) | `databricks/dbrx-instruct` | Databricks’ 132B-parameter MoE model (36B active) trained on 12T tokens; competes with GPT-3.5 quality as a fully open foundation model. |

245
test/registered/amd/accuracy/mi30x/test_minimax_m25_eval_amd.py Normal file
View File

@@ -0,0 +1,245 @@
"""AMD MiniMax-M2.5 GSM8K Completion Evaluation Test (8-GPU)
Tests MiniMax-M2.5 with TP=8 + EP=8 configuration using few-shot completion
benchmark on MI325/MI300X.
Registry: nightly-amd-accuracy-8-gpu-minimax-m25 suite
"""
import ast
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=3600,
suite="nightly-amd-accuracy-8-gpu-minimax-m25",
nightly=True,
)
INVALID = -9999999
@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
MINIMAX_M25_MODELS = [
ModelConfig(
model_path="MiniMaxAI/MiniMax-M2.5",
tp_size=8,
accuracy_threshold=0.93,
timeout=3600,
variant="TP8+EP8",
other_args=[
"--ep-size",
"8",
"--trust-remote-code",
"--attention-backend",
"aiter",
"--mem-fraction-static",
"0.85",
"--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 TestMiniMaxM25EvalAMD(unittest.TestCase):
"""MiniMax-M2.5 GSM8K Completion Evaluation Test for AMD MI325/MI300X."""
@classmethod
def setUpClass(cls):
cls.models = MINIMAX_M25_MODELS
cls.base_url = DEFAULT_URL_FOR_TEST
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
def test_minimax_m25_accuracy(self):
"""Test MiniMax-M2.5 with GSM8K completion benchmark."""
all_results = []
summary = "### MiniMax-M2.5 Models (MI325)\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()

249
test/registered/amd/accuracy/mi35x/test_minimax_m25_eval_mi35x.py Normal file
View File

@@ -0,0 +1,249 @@
"""MI35x MiniMax-M2.5 GSM8K Completion Evaluation Test (8-GPU)
Tests MiniMax-M2.5 with TP=8 + EP=8 configuration using few-shot completion
benchmark on MI35x.
Registry: nightly-amd-8-gpu-mi35x-minimax-m25 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-minimax-m25",
nightly=True,
)
INVALID = -9999999
@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
MI35X_MINIMAX_M25_MODELS = [
ModelConfig(
model_path="MiniMaxAI/MiniMax-M2.5",
tp_size=8,
accuracy_threshold=0.93,
timeout=5400,
variant="TP8+EP8",
other_args=[
"--ep-size",
"8",
"--trust-remote-code",
"--attention-backend",
"aiter",
"--mem-fraction-static",
"0.85",
"--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 TestMiniMaxM25EvalMI35x(unittest.TestCase):
"""MiniMax-M2.5 GSM8K Completion Evaluation Test for AMD MI35x."""
@classmethod
def setUpClass(cls):
cls.models = MI35X_MINIMAX_M25_MODELS
cls.base_url = DEFAULT_URL_FOR_TEST
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
def test_minimax_m25_accuracy(self):
"""Test MiniMax-M2.5 with GSM8K completion benchmark."""
all_results = []
summary = "### MiniMax-M2.5 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()