[AMD] Implement QuarkW4A8MXFp4MoE to support amd/gpt-oss-120b-w-mxfp4-a-fp8 (#27204)
Signed-off-by: Stanley Winata <stanley.winata@amd.com>
This commit is contained in:
Stanley Winata
@@ -0,0 +1,251 @@
|
||||
"""MI35x GPT-OSS W4A8 MXFP4-FP8 GSM8K Completion Evaluation Test (8-GPU)
|
||||
|
||||
Tests the AMD Quark `gpt-oss-120b-w-mxfp4-a-fp8` checkpoint (MXFP4
|
||||
weights + static per-tensor FP8 activations) using few-shot completion
|
||||
benchmark on MI35x.
|
||||
|
||||
Registry: nightly-amd-8-gpu-mi35x suite
|
||||
"""
|
||||
|
||||
import ast
|
||||
import os
|
||||
|
||||
# Set HF cache for MI35x
|
||||
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 for AMD CI - MI35x GPT-OSS W4A8 MXFP4-FP8 accuracy test (~15 min)
|
||||
register_amd_ci(est_time=900, suite="nightly-amd-8-gpu-mi35x", 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
|
||||
|
||||
def __post_init__(self):
|
||||
if self.other_args is None:
|
||||
self.other_args = []
|
||||
if self.env_vars is None:
|
||||
self.env_vars = {}
|
||||
|
||||
|
||||
# GPT-OSS Quark W4A8 MXFP4-FP8 models for MI35x
|
||||
MI35X_GPT_OSS_W4A8_MXFP4_MODELS = [
|
||||
ModelConfig(
|
||||
model_path="amd/gpt-oss-120b-w-mxfp4-a-fp8",
|
||||
tp_size=8,
|
||||
# Match the peer `openai/gpt-oss-120b` GSM8K threshold in
|
||||
# `test_gpt_oss_eval_mi35x.py` (0.79). The AMD Quark W4A8 MXFP4-FP8
|
||||
# measures ~0.855 on this host, comfortably above 0.79.
|
||||
accuracy_threshold=0.79,
|
||||
timeout=900,
|
||||
other_args=[
|
||||
"--chunked-prefill-size",
|
||||
"130172",
|
||||
"--max-running-requests",
|
||||
"128",
|
||||
"--mem-fraction-static",
|
||||
"0.85",
|
||||
"--attention-backend",
|
||||
"triton",
|
||||
"--trust-remote-code",
|
||||
],
|
||||
# AITER MXFP4 fused-MoE for gpt-oss uses the SEPARATED gate/up tile
|
||||
# layout (matches `gptoss_fp4_tuned_fmoe.csv` flydsl entries, the
|
||||
# `QuarkW4A8MXFp4MoE.process_weights_after_loading` shuffle and the
|
||||
# Mxfp4MoEMethod weight shuffle, both uses `shuffle_weight_a16w4`
|
||||
# and + `shuffle_scale_a16w4`) Other AITER MXFP4 callers default
|
||||
# to INTERLEAVE, so opt out explicitly here.
|
||||
env_vars={
|
||||
"SGLANG_USE_AITER": "1",
|
||||
"SGLANG_USE_AITER_MOE_GU_ITLV": "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 TestGptOssW4A8Mxfp4EvalMI35x(unittest.TestCase):
|
||||
"""GPT-OSS W4A8 MXFP4-FP8 GSM8K Completion Evaluation Test for AMD MI35x."""
|
||||
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.models = MI35X_GPT_OSS_W4A8_MXFP4_MODELS
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
|
||||
|
||||
def test_gpt_oss_accuracy(self):
|
||||
"""Test GPT-OSS W4A8 MXFP4-FP8 models with GSM8K completion benchmark."""
|
||||
all_results = []
|
||||
summary = "### GPT-OSS W4A8 MXFP4-FP8 Models (MI35x)\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)])
|
||||
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()
|
||||
Reference in New Issue
Block a user