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

[AMD] Nightly Test Coverage - Minimax-M3-MXFP8 Accuracy Test (#30613)

Browse Source
This commit is contained in:
YC Yen-Ching Tseng
2026-07-25 23:52:22 -07:00
committed by GitHub
parent 2cbddb842d
commit 1d0cd2e473
3 changed files with 351 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
@@ -87,6 +87,8 @@ on:
- nightly-8-gpu-mi35x-glm5-mxfp4-rocm720
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25-rocm720
# 4-GPU MiniMax-M3 MXFP8 (MI35x)
- nightly-4-gpu-mi35x-minimax-m3-rocm720
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27-rocm720
# Diffusion (MI30x)
@@ -1890,6 +1892,44 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 4-GPU MiniMax-M3 MXFP8 (MI35x)
# ==============================================================================
nightly-4-gpu-mi35x-minimax-m3-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-m3-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 }}
ENABLE_CACHE_HOST: "1"
- 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-M3 MXFP8)
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 GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-4-gpu-mi35x-minimax-m3-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 ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 8-GPU MiniMax-M2.7 (MI30x only)
# ==============================================================================
@@ -2045,6 +2085,8 @@ jobs:
- nightly-8-gpu-mi35x-glm5-mxfp4-rocm720
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25-rocm720
# 4-GPU MiniMax-M3 MXFP8 (MI35x)
- nightly-4-gpu-mi35x-minimax-m3-rocm720
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27-rocm720
# Diffusion (MI30x)
.github/workflows/nightly-test-amd.yml
+42
View File
@@ -83,6 +83,8 @@ on:
- nightly-8-gpu-mi35x-glm5-mxfp4
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25
# 4-GPU MiniMax-M3 MXFP8 (MI35x)
- nightly-4-gpu-mi35x-minimax-m3
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27
# Diffusion (MI30x)
@@ -1744,6 +1746,44 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 4-GPU MiniMax-M3 MXFP8 (MI35x)
# ==============================================================================
nightly-4-gpu-mi35x-minimax-m3:
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-m3,'))
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 }}
ENABLE_CACHE_HOST: "1"
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
- name: Accuracy Test MI35x (4-GPU MiniMax-M3 MXFP8)
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 GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-4-gpu-mi35x-minimax-m3-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)
# ==============================================================================
@@ -1895,6 +1935,8 @@ jobs:
- nightly-8-gpu-mi35x-glm5-mxfp4
# 4-GPU MiniMax-M2.5 (MI35x)
- nightly-4-gpu-mi35x-minimax-m25
# 4-GPU MiniMax-M3 MXFP8 (MI35x)
- nightly-4-gpu-mi35x-minimax-m3
# 8-GPU MiniMax-M2.7 (MI30x only)
- nightly-8-gpu-minimax-m27
# Diffusion (MI30x)
test/registered/amd/accuracy/mi35x/test_minimax_m3_tp4_eval_mi35x.py
+267
View File
@@ -0,0 +1,267 @@
"""MI35x MiniMax-M3 MXFP8 GSM8K Chat+Thinking Evaluation Test (4-GPU, TP=4)
Tests MiniMax-M3 (MXFP8 checkpoint) with TP=4 on MI35x. MI35x (gfx950 / CDNA4)
has hardware MX-scaled matmul, so the MXFP8 MoE weights are served natively;
you still pass `--quantization mxfp8`. Serves with the aiter attention backend,
fp8 (e4m3) KV cache, and radix cache disabled — validated accuracy-neutral vs
the bf16-KV / triton-attn baseline (0.972 vs 0.970 on GSM8K chat+thinking).
MiniMax-M3 is a reasoning model: it must be evaluated through the chat template
with thinking enabled (its `<mm:think>` reasoning path). Raw few-shot completion
(no chat template) does NOT engage its reasoning and severely underscores it
(~0.87 vs ~0.96 on GSM8K), so this test uses chat + thinking to match how the
model is meant to be served and the published reference accuracy.
Registry: nightly-amd-4-gpu-mi35x-minimax-m3-tp4 suite
"""
import json
import os
import re
import time
import unittest
import urllib.request
from concurrent.futures import ThreadPoolExecutor
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-m3-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_M3_TP4_MODELS = [
ModelConfig(
model_path="MiniMaxAI/MiniMax-M3-MXFP8",
tp_size=4,
accuracy_threshold=0.95,
timeout=5400,
variant="TP4+MXFP8+aiterAttn+fp8KV",
other_args=[
"--quantization",
"mxfp8",
"--dtype",
"bfloat16",
"--trust-remote-code",
"--attention-backend",
"aiter",
"--kv-cache-dtype",
"fp8_e4m3",
"--disable-radix-cache",
"--chunked-prefill-size",
"8192",
"--mem-fraction-static",
"0.80",
"--watchdog-timeout",
"1200",
],
env_vars={
"SGLANG_USE_AITER": "1",
# ROCm 7.0's rocBLAS/hipBLASLt rejects the bf16-input/fp32-output
# router GEMM (torch.mm(bf16, bf16, out_dtype=float32)); force the
# fp32 router path. Also gives more precise expert routing.
"SGLANG_OPT_USE_BF16_ROUTER_GEMM": "0",
},
),
]
def get_answer_value(answer_str):
"""Extract numerical answer from response (last integer)."""
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,
model_path: str,
num_questions: int = 1319,
parallel: int = 64,
max_tokens: int = 4096,
) -> Tuple[float, float, float]:
"""Run GSM8K in chat + thinking mode (M3's intended reasoning path).
Uses the OpenAI-compatible /v1/chat/completions endpoint so the chat
template is applied, and forces thinking via chat_template_kwargs
(M3's template reads ``thinking_mode``). The final answer is the last
integer in the response; the reasoning trace lives in ``<mm:think>`` tags
inside ``content`` (or in ``reasoning_content`` if a reasoning parser is on),
so both channels are scanned.
"""
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))
instruction = (
"\n\nPlease reason step by step, and give the final answer as a single "
"integer on the last line."
)
n = len(lines[:num_questions])
questions = [lines[i]["question"] for i in range(n)]
labels = [get_answer_value(lines[i]["answer"]) for i in range(n)]
assert all(l != INVALID for l in labels)
def query(question: str) -> str:
body = {
"model": model_path,
"messages": [{"role": "user", "content": question + instruction}],
"temperature": 0,
"max_tokens": max_tokens,
"chat_template_kwargs": {"thinking_mode": "enabled"},
}
req = urllib.request.Request(
base_url + "/v1/chat/completions",
data=json.dumps(body).encode(),
headers={"Content-Type": "application/json"},
)
with urllib.request.urlopen(req, timeout=1800) as resp:
msg = json.loads(resp.read())["choices"][0]["message"]
return (msg.get("content") or "") + " " + (msg.get("reasoning_content") or "")
tic = time.perf_counter()
with ThreadPoolExecutor(max_workers=parallel) as ex:
outputs = list(ex.map(query, questions))
latency = time.perf_counter() - tic
preds = [get_answer_value(o) for o in outputs]
acc = float(np.mean(np.array(preds) == np.array(labels)))
invalid = float(np.mean(np.array(preds) == INVALID))
return acc, invalid, latency
class TestMiniMaxM3TP4EvalMI35x(unittest.TestCase):
"""MiniMax-M3 MXFP8 TP=4 GSM8K Chat+Thinking Evaluation Test for AMD MI35x."""
@classmethod
def setUpClass(cls):
cls.models = MI35X_MINIMAX_M3_TP4_MODELS
cls.base_url = DEFAULT_URL_FOR_TEST
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "1319"))
def test_minimax_m3_tp4_accuracy(self):
"""Test MiniMax-M3 MXFP8 TP=4 with GSM8K chat+thinking benchmark."""
all_results = []
summary = "### MiniMax-M3 MXFP8 TP=4 chat+thinking (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,
config.model_path,
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()