291 lines
8.8 KiB
Python
291 lines
8.8 KiB
Python
"""AMD GROK GSM8K Completion Evaluation Test (8-GPU)
|
|
|
|
Tests GROK models (Grok-1 FP8, Grok-1 INT4, Grok-2) using
|
|
few-shot completion benchmark on MI300X.
|
|
|
|
Registry: nightly-amd-8-gpu-grok 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
|
|
|
|
# DISABLED: Split into individual files for each model variant
|
|
# See: test_grok1_fp8_eval_amd.py, test_grok1_int4_eval_amd.py, test_grok2_eval_amd.py
|
|
register_amd_ci(
|
|
est_time=2700,
|
|
suite="nightly-amd-8-gpu-grok",
|
|
nightly=True,
|
|
disabled="Split into test_grok1_fp8_eval_amd.py, test_grok1_int4_eval_amd.py, test_grok2_eval_amd.py",
|
|
)
|
|
|
|
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
|
|
tokenizer_path: Optional[str] = None
|
|
timeout: Optional[int] = None
|
|
|
|
def __post_init__(self):
|
|
if self.other_args is None:
|
|
self.other_args = []
|
|
if self.env_vars is None:
|
|
self.env_vars = {}
|
|
|
|
|
|
# GROK models for MI300X
|
|
GROK_MODELS = [
|
|
# GROK1-FP8
|
|
ModelConfig(
|
|
model_path="lmzheng/grok-1",
|
|
tp_size=8,
|
|
accuracy_threshold=0.80,
|
|
timeout=3600,
|
|
tokenizer_path="Xenova/grok-1-tokenizer",
|
|
other_args=[
|
|
"--quantization",
|
|
"fp8",
|
|
"--attention-backend",
|
|
"aiter",
|
|
"--mem-fraction-static",
|
|
"0.85",
|
|
"--trust-remote-code",
|
|
],
|
|
env_vars={
|
|
"RCCL_MSCCL_ENABLE": "0",
|
|
"SGLANG_USE_AITER": "1",
|
|
"SGLANG_INT4_WEIGHT": "0",
|
|
},
|
|
),
|
|
# GROK1-INT4
|
|
ModelConfig(
|
|
model_path="amd/grok-1-W4A8KV8",
|
|
tp_size=8,
|
|
accuracy_threshold=0.80,
|
|
timeout=3600,
|
|
tokenizer_path="Xenova/grok-1-tokenizer",
|
|
other_args=[
|
|
"--quantization",
|
|
"fp8",
|
|
"--attention-backend",
|
|
"aiter",
|
|
"--mem-fraction-static",
|
|
"0.85",
|
|
"--trust-remote-code",
|
|
],
|
|
env_vars={
|
|
"RCCL_MSCCL_ENABLE": "0",
|
|
"SGLANG_USE_AITER": "1",
|
|
"SGLANG_INT4_WEIGHT": "1",
|
|
},
|
|
),
|
|
# GROK2
|
|
ModelConfig(
|
|
model_path="xai-org/grok-2",
|
|
tp_size=8,
|
|
accuracy_threshold=0.915,
|
|
timeout=3600,
|
|
tokenizer_path="alvarobartt/grok-2-tokenizer",
|
|
other_args=[
|
|
"--quantization",
|
|
"fp8",
|
|
"--attention-backend",
|
|
"aiter",
|
|
"--mem-fraction-static",
|
|
"0.85",
|
|
"--trust-remote-code",
|
|
],
|
|
env_vars={
|
|
"RCCL_MSCCL_ENABLE": "0",
|
|
"SGLANG_USE_AITER": "1",
|
|
"SGLANG_INT4_WEIGHT": "0",
|
|
},
|
|
),
|
|
]
|
|
|
|
|
|
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 TestGrokEvalAMD(unittest.TestCase):
|
|
"""GROK GSM8K Completion Evaluation Test for AMD MI300X."""
|
|
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
cls.models = GROK_MODELS
|
|
cls.base_url = DEFAULT_URL_FOR_TEST
|
|
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
|
|
|
|
def test_grok_accuracy(self):
|
|
"""Test GROK models with GSM8K completion benchmark."""
|
|
all_results = []
|
|
summary = "### GROK Models (MI300X)\n\n"
|
|
summary += "| Model | TP | Accuracy | Threshold | Status |\n"
|
|
summary += "| ----- | -- | -------- | --------- | ------ |\n"
|
|
|
|
for config in self.models:
|
|
with self.subTest(model=config.model_path):
|
|
print(f"\n{'='*60}")
|
|
print(f"Testing: {config.model_path}")
|
|
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)])
|
|
if config.tokenizer_path:
|
|
other_args.extend(["--tokenizer-path", config.tokenizer_path])
|
|
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": config.model_path,
|
|
"accuracy": acc,
|
|
"passed": passed,
|
|
}
|
|
)
|
|
summary += f"| {config.model_path} | {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.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
|
|
all_results.append(
|
|
{
|
|
"model": config.model_path,
|
|
"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()
|