[AMD] Fix TP2 DeepSeek-R1 nhead=64 MLA decode crash and add nightly coverage (#27188)
Co-authored-by: clintg6 <7388379+clintg6@users.noreply.github.com>
This commit is contained in:
Clint
@@ -0,0 +1,185 @@
|
||||
"""MI35x DeepSeek-R1-MXFP4 TP=2 GSM8K AITER MLA regression.
|
||||
|
||||
DeepSeek-R1 has 128 attention heads, so TP=2 gives 64 heads per rank. This
|
||||
covers the AITER persistent MLA decode metadata path that previously crashed
|
||||
with GPU memory access faults.
|
||||
|
||||
Registry: nightly-amd-2-gpu-mi35x-deepseek-r1-mxfp4-tp2 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 typing import 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_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=1800,
|
||||
suite="nightly-amd-2-gpu-mi35x-deepseek-r1-mxfp4-tp2",
|
||||
nightly=True,
|
||||
)
|
||||
|
||||
INVALID = -9999999
|
||||
|
||||
DEEPSEEK_R1_MXFP4_LOCAL_PATH = "/data2/models/amd-DeepSeek-R1-MXFP4-Preview"
|
||||
DEEPSEEK_R1_MXFP4_HF_MODEL_ID = "amd/DeepSeek-R1-MXFP4-Preview"
|
||||
SERVER_LAUNCH_TIMEOUT = 3600
|
||||
GSM8K_ACCURACY_THRESHOLD = 0.93
|
||||
|
||||
|
||||
def get_model_path() -> str:
|
||||
env_path = os.environ.get("DEEPSEEK_R1_MXFP4_MODEL_PATH")
|
||||
if env_path:
|
||||
return env_path
|
||||
if os.path.exists(DEEPSEEK_R1_MXFP4_LOCAL_PATH):
|
||||
return DEEPSEEK_R1_MXFP4_LOCAL_PATH
|
||||
return DEEPSEEK_R1_MXFP4_HF_MODEL_ID
|
||||
|
||||
|
||||
def get_one_example(lines, i, include_answer):
|
||||
ret = "Question: " + lines[i]["question"] + "\nAnswer:"
|
||||
if include_answer:
|
||||
ret += " " + lines[i]["answer"]
|
||||
return ret
|
||||
|
||||
|
||||
def get_few_shot_examples(lines, k):
|
||||
ret = ""
|
||||
for i in range(k):
|
||||
ret += get_one_example(lines, i, True) + "\n\n"
|
||||
return ret
|
||||
|
||||
|
||||
def get_answer_value(answer_str):
|
||||
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]:
|
||||
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 TestDeepSeekR1MXFP4TP2MI35x(unittest.TestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = get_model_path()
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
|
||||
|
||||
env = os.environ.copy()
|
||||
env["SGLANG_USE_AITER"] = "1"
|
||||
env["SGLANG_AITER_MLA_PERSIST"] = "1"
|
||||
|
||||
cls.process = popen_launch_server(
|
||||
model=cls.model,
|
||||
base_url=cls.base_url,
|
||||
timeout=SERVER_LAUNCH_TIMEOUT,
|
||||
other_args=[
|
||||
"--attention-backend",
|
||||
"aiter",
|
||||
"--tp",
|
||||
"2",
|
||||
"--chunked-prefill-size",
|
||||
"131072",
|
||||
"--disable-radix-cache",
|
||||
"--mem-fraction-static",
|
||||
"0.85",
|
||||
"--trust-remote-code",
|
||||
"--kv-cache-dtype",
|
||||
"fp8_e4m3",
|
||||
"--model-loader-extra-config",
|
||||
'{"enable_multithread_load": true}',
|
||||
],
|
||||
env=env,
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
kill_process_tree(cls.process.pid)
|
||||
|
||||
def test_gsm8k(self):
|
||||
acc, invalid, latency = run_gsm8k_benchmark(
|
||||
self.base_url, num_questions=self.num_questions
|
||||
)
|
||||
print(f"accuracy={acc:.3f} invalid={invalid:.3f} latency={latency:.1f}s")
|
||||
|
||||
if is_in_ci():
|
||||
write_github_step_summary(
|
||||
"### DeepSeek-R1-MXFP4 TP=2 GSM8K (MI35x)\n\n"
|
||||
"| Model | TP | Examples | Accuracy | Invalid | Threshold | Latency |\n"
|
||||
"| ----- | -- | -------- | -------- | ------- | --------- | ------- |\n"
|
||||
f"| {self.model} | 2 | {self.num_questions} | {acc:.3f} | "
|
||||
f"{invalid:.3f} | {GSM8K_ACCURACY_THRESHOLD:.2f} | {latency:.1f}s |\n"
|
||||
)
|
||||
|
||||
self.assertGreaterEqual(acc, GSM8K_ACCURACY_THRESHOLD)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,184 @@
|
||||
"""MI35x DeepSeek-R1-MXFP4 TP=4 GSM8K AITER MLA regression.
|
||||
|
||||
DeepSeek-R1 has 128 attention heads, so TP=4 gives 32 heads per rank. This
|
||||
covers the AITER persistent MLA decode metadata path for the nhead=32 case.
|
||||
|
||||
Registry: nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-tp4 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 typing import 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_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=1800,
|
||||
suite="nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-tp4",
|
||||
nightly=True,
|
||||
)
|
||||
|
||||
INVALID = -9999999
|
||||
|
||||
DEEPSEEK_R1_MXFP4_LOCAL_PATH = "/data2/models/amd-DeepSeek-R1-MXFP4-Preview"
|
||||
DEEPSEEK_R1_MXFP4_HF_MODEL_ID = "amd/DeepSeek-R1-MXFP4-Preview"
|
||||
SERVER_LAUNCH_TIMEOUT = 3600
|
||||
GSM8K_ACCURACY_THRESHOLD = 0.93
|
||||
|
||||
|
||||
def get_model_path() -> str:
|
||||
env_path = os.environ.get("DEEPSEEK_R1_MXFP4_MODEL_PATH")
|
||||
if env_path:
|
||||
return env_path
|
||||
if os.path.exists(DEEPSEEK_R1_MXFP4_LOCAL_PATH):
|
||||
return DEEPSEEK_R1_MXFP4_LOCAL_PATH
|
||||
return DEEPSEEK_R1_MXFP4_HF_MODEL_ID
|
||||
|
||||
|
||||
def get_one_example(lines, i, include_answer):
|
||||
ret = "Question: " + lines[i]["question"] + "\nAnswer:"
|
||||
if include_answer:
|
||||
ret += " " + lines[i]["answer"]
|
||||
return ret
|
||||
|
||||
|
||||
def get_few_shot_examples(lines, k):
|
||||
ret = ""
|
||||
for i in range(k):
|
||||
ret += get_one_example(lines, i, True) + "\n\n"
|
||||
return ret
|
||||
|
||||
|
||||
def get_answer_value(answer_str):
|
||||
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]:
|
||||
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 TestDeepSeekR1MXFP4TP4MI35x(unittest.TestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = get_model_path()
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
|
||||
|
||||
env = os.environ.copy()
|
||||
env["SGLANG_USE_AITER"] = "1"
|
||||
env["SGLANG_AITER_MLA_PERSIST"] = "1"
|
||||
|
||||
cls.process = popen_launch_server(
|
||||
model=cls.model,
|
||||
base_url=cls.base_url,
|
||||
timeout=SERVER_LAUNCH_TIMEOUT,
|
||||
other_args=[
|
||||
"--attention-backend",
|
||||
"aiter",
|
||||
"--tp",
|
||||
"4",
|
||||
"--chunked-prefill-size",
|
||||
"131072",
|
||||
"--disable-radix-cache",
|
||||
"--mem-fraction-static",
|
||||
"0.85",
|
||||
"--trust-remote-code",
|
||||
"--kv-cache-dtype",
|
||||
"fp8_e4m3",
|
||||
"--model-loader-extra-config",
|
||||
'{"enable_multithread_load": true}',
|
||||
],
|
||||
env=env,
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
kill_process_tree(cls.process.pid)
|
||||
|
||||
def test_gsm8k(self):
|
||||
acc, invalid, latency = run_gsm8k_benchmark(
|
||||
self.base_url, num_questions=self.num_questions
|
||||
)
|
||||
print(f"accuracy={acc:.3f} invalid={invalid:.3f} latency={latency:.1f}s")
|
||||
|
||||
if is_in_ci():
|
||||
write_github_step_summary(
|
||||
"### DeepSeek-R1-MXFP4 TP=4 GSM8K (MI35x)\n\n"
|
||||
"| Model | TP | Examples | Accuracy | Invalid | Threshold | Latency |\n"
|
||||
"| ----- | -- | -------- | -------- | ------- | --------- | ------- |\n"
|
||||
f"| {self.model} | 4 | {self.num_questions} | {acc:.3f} | "
|
||||
f"{invalid:.3f} | {GSM8K_ACCURACY_THRESHOLD:.2f} | {latency:.1f}s |\n"
|
||||
)
|
||||
|
||||
self.assertGreaterEqual(acc, GSM8K_ACCURACY_THRESHOLD)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -114,6 +114,8 @@ NIGHTLY_SUITES = {
|
||||
"nightly-amd-1-gpu",
|
||||
"nightly-amd-1-gpu-mi35x",
|
||||
"nightly-amd-1-gpu-zimage-turbo",
|
||||
"nightly-amd-2-gpu-mi35x-deepseek-r1-mxfp4-tp2",
|
||||
"nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-tp4",
|
||||
"nightly-amd-4-gpu",
|
||||
"nightly-amd-8-gpu",
|
||||
"nightly-amd-vlm",
|
||||
|
||||
Reference in New Issue
Block a user