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

[DeepSeekV3.2] Add pure TP+MTP test (#15088)

Browse Source 
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
This commit is contained in:
Ashton Chew
2025-12-16 21:48:12 -08:00
committed by GitHub
parent 31d48d7f6f
commit 2bdbaef18e
2 changed files with 107 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
docs/basic_usage/deepseek_v32.md
View File

@@ -64,10 +64,16 @@ python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8
## Multi-token Prediction
SGLang implements Multi-Token Prediction (MTP) for DeepSeek V3.2 based on [EAGLE speculative decoding](https://docs.sglang.io/advanced_features/speculative_decoding.html#EAGLE-Decoding). With this optimization, the decoding speed can be improved significantly on small batch sizes. Please look at [this PR](https://github.com/sgl-project/sglang/pull/11652) for more information.
Example usage:
Example usage with DP Attention:
```bash
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
```
Example usage with Pure TP:
```bash
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
```
- The best configuration for `--speculative-num-steps`, `--speculative-eagle-topk` and `--speculative-num-draft-tokens` can be searched with [bench_speculative.py](https://github.com/sgl-project/sglang/blob/main/scripts/playground/bench_speculative.py) script for given batch size. The minimum configuration is `--speculative-num-steps 1 --speculative-eagle-topk 1 --speculative-num-draft-tokens 2`, which can achieve speedup for larger batch sizes.
- The default value of `--max-running-requests` is set to `48` for MTP. For larger batch sizes, this value should be increased beyond the default value.

106
test/nightly/test_deepseek_v32_tp.py
View File

@@ -2,9 +2,12 @@ import os
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
@@ -14,7 +17,7 @@ from sglang.test.test_utils import (
write_github_step_summary,
)
register_cuda_ci(est_time=600, suite="nightly-8-gpu-h200", nightly=True)
register_cuda_ci(est_time=900, suite="nightly-8-gpu-h200", nightly=True)
DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2-Exp"
@@ -134,10 +137,99 @@ class TestDeepseekV32_Partial_TP(CustomTestCase):
"accuracy": metrics["accuracy"],
}
)
self.assertGreater(metrics["accuracy"], 0.935)
class TestDeepseekV32_TP_MTP(CustomTestCase):
"""Test DeepSeek V3.2 with pure TP + MTP (EAGLE speculative decoding)."""
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V32_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--tp",
"8",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-frac",
"0.7",
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
num_shots=20,
data_path=None,
num_questions=1400,
parallel=1400,
max_new_tokens=512,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"{metrics=}")
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
if is_in_ci():
TEST_RESULTS.append(
{
"variant": "tp_mtp",
"prefill_backend": "flashmla_sparse",
"decode_backend": "flashmla_kv",
"kv_cache": "fp16",
"accuracy": metrics["accuracy"],
"avg_spec_accept_length": avg_spec_accept_length,
}
)
self.assertGreater(metrics["accuracy"], 0.935)
self.assertGreater(avg_spec_accept_length, 2.5)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
acc_length, speed = send_one_prompt(args)
print(f"{acc_length=:.2f} {speed=:.2f}")
if is_in_ci():
# Update last result with speed data
if TEST_RESULTS and TEST_RESULTS[-1]["variant"] == "tp_mtp":
TEST_RESULTS[-1]["speed"] = speed
# Write the summary table after all tests complete
_write_summary_table()
self.assertGreater(metrics["accuracy"], 0.935)
self.assertGreater(acc_length, 2.5)
self.assertGreater(speed, 110)
def _format_optional_metric(value, fmt=".2f", suffix=""):
"""Format an optional metric value, returning '-' if not available."""
if value is None:
return "-"
return f"{value:{fmt}}{suffix}"
def _write_summary_table():
@@ -147,19 +239,21 @@ def _write_summary_table():
gpu_config = os.getenv("GPU_CONFIG", "8-gpu-h200")
# Build table header
# Build table header - keep original columns + add MTP-specific ones
summary = (
f"### {DEEPSEEK_V32_MODEL_PATH} GSM8K Accuracy (TP Tests) [{gpu_config}]\n\n"
)
summary += "| Variant | Prefill Backend | Decode Backend | KV Cache | Accuracy |\n"
summary += "|---------|-----------------|----------------|----------|----------|\n"
summary += "| Variant | Prefill Backend | Decode Backend | KV Cache | Accuracy | Spec Acc Len | Speed |\n"
summary += "|---------|-----------------|----------------|----------|----------|--------------|-------|\n"
# Add each result as a row
for result in TEST_RESULTS:
summary += (
f"| {result['variant']} | {result['prefill_backend']} | "
f"{result['decode_backend']} | {result['kv_cache']} | "
f"{result['accuracy']:.3f} |\n"
f"{result['accuracy']:.3f} | "
f"{_format_optional_metric(result.get('avg_spec_accept_length'))} | "
f"{_format_optional_metric(result.get('speed'), '.1f', ' tok/s')} |\n"
)
write_github_step_summary(summary)