"git@developer.sourcefind.cn:renzhc/diffusers_dcu.git" did not exist on "1d1e1a2888bd65b51f13272de2f709fd91e0beb1"
Unverified Commit c3eac1b0 authored by Lianmin Zheng's avatar Lianmin Zheng Committed by GitHub
Browse files

Fix torch.compile for MoE (#2033)

parent b275ce00
from typing import Optional from typing import Callable, Optional
import torch import torch
from torch.nn import functional as F from torch.nn import functional as F
...@@ -98,7 +98,9 @@ def fused_moe_forward_native( ...@@ -98,7 +98,9 @@ def fused_moe_forward_native(
renormalize: bool, renormalize: bool,
topk_group: Optional[int] = None, topk_group: Optional[int] = None,
num_expert_group: Optional[int] = None, num_expert_group: Optional[int] = None,
custom_routing_function: Optional[Callable] = None,
) -> torch.Tensor: ) -> torch.Tensor:
assert custom_routing_function is None
topk_weights, topk_ids = select_experts_native( topk_weights, topk_ids = select_experts_native(
hidden_states=x, hidden_states=x,
router_logits=router_logits, router_logits=router_logits,
...@@ -114,4 +116,4 @@ def fused_moe_forward_native( ...@@ -114,4 +116,4 @@ def fused_moe_forward_native(
x1 = F.silu(torch.einsum("ti,taoi -> tao", x, w1_weights)) x1 = F.silu(torch.einsum("ti,taoi -> tao", x, w1_weights))
x3 = torch.einsum("ti, taoi -> tao", x, w3_weights) x3 = torch.einsum("ti, taoi -> tao", x, w3_weights)
expert_outs = torch.einsum("tao, taio -> tai", (x1 * x3), w2_weights) expert_outs = torch.einsum("tao, taio -> tai", (x1 * x3), w2_weights)
return torch.einsum("tai,ta -> ti", expert_outs, topk_weights) return torch.einsum("tai,ta -> ti", expert_outs, topk_weights.to(expert_outs.dtype))
...@@ -28,8 +28,9 @@ from sglang.utils import get_exception_traceback ...@@ -28,8 +28,9 @@ from sglang.utils import get_exception_traceback
DEFAULT_FP8_MODEL_NAME_FOR_TEST = "neuralmagic/Meta-Llama-3.1-8B-FP8" DEFAULT_FP8_MODEL_NAME_FOR_TEST = "neuralmagic/Meta-Llama-3.1-8B-FP8"
DEFAULT_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.1-8B-Instruct" DEFAULT_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.1-8B-Instruct"
DEFAULT_SMALL_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.2-1B-Instruct" DEFAULT_SMALL_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.2-1B-Instruct"
DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST = "Alibaba-NLP/gte-Qwen2-1.5B-instruct"
DEFAULT_MOE_MODEL_NAME_FOR_TEST = "mistralai/Mixtral-8x7B-Instruct-v0.1" DEFAULT_MOE_MODEL_NAME_FOR_TEST = "mistralai/Mixtral-8x7B-Instruct-v0.1"
DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST = "Qwen/Qwen1.5-MoE-A2.7B"
DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST = "Alibaba-NLP/gte-Qwen2-1.5B-instruct"
DEFAULT_MLA_MODEL_NAME_FOR_TEST = "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct" DEFAULT_MLA_MODEL_NAME_FOR_TEST = "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct"
DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST = "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8" DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST = "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8"
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH = 600 DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH = 600
...@@ -740,7 +741,7 @@ def run_mmlu_test( ...@@ -740,7 +741,7 @@ def run_mmlu_test(
try: try:
metrics = run_eval(args) metrics = run_eval(args)
print(f"{metrics=}") print(f"{metrics=}")
assert metrics["score"] >= 0.65 self.assertGreaterEqual(metrics["score"], 0.65)
finally: finally:
pass pass
......
...@@ -27,6 +27,7 @@ suites = { ...@@ -27,6 +27,7 @@ suites = {
"test_srt_engine.py", "test_srt_engine.py",
"test_srt_endpoint.py", "test_srt_endpoint.py",
"test_torch_compile.py", "test_torch_compile.py",
"test_torch_compile_moe.py",
"test_torchao.py", "test_torchao.py",
"test_triton_attention_kernels.py", "test_triton_attention_kernels.py",
"test_triton_attention_backend.py", "test_triton_attention_backend.py",
......
...@@ -40,7 +40,7 @@ class TestDataParallelism(unittest.TestCase): ...@@ -40,7 +40,7 @@ class TestDataParallelism(unittest.TestCase):
) )
metrics = run_eval(args) metrics = run_eval(args)
assert metrics["score"] >= 0.65 self.assertGreaterEqual(metrics["score"], 0.65)
def test_update_weight(self): def test_update_weight(self):
response = requests.post( response = requests.post(
......
...@@ -55,7 +55,7 @@ class TestDoubleSparsity(unittest.TestCase): ...@@ -55,7 +55,7 @@ class TestDoubleSparsity(unittest.TestCase):
) )
metrics = run_eval(args) metrics = run_eval(args)
assert metrics["score"] >= 0.65 self.assertGreaterEqual(metrics["score"], 0.65)
if __name__ == "__main__": if __name__ == "__main__":
......
...@@ -35,7 +35,7 @@ class TestEvalAccuracyMini(unittest.TestCase): ...@@ -35,7 +35,7 @@ class TestEvalAccuracyMini(unittest.TestCase):
) )
metrics = run_eval(args) metrics = run_eval(args)
assert metrics["score"] >= 0.65 self.assertGreaterEqual(metrics["score"], 0.65)
if __name__ == "__main__": if __name__ == "__main__":
......
...@@ -34,7 +34,7 @@ class TestRetractDecode(unittest.TestCase): ...@@ -34,7 +34,7 @@ class TestRetractDecode(unittest.TestCase):
) )
metrics = run_eval(args) metrics = run_eval(args)
assert metrics["score"] >= 0.65 self.assertGreaterEqual(metrics["score"], 0.65)
if __name__ == "__main__": if __name__ == "__main__":
......
...@@ -39,7 +39,7 @@ class TestTorchCompile(unittest.TestCase): ...@@ -39,7 +39,7 @@ class TestTorchCompile(unittest.TestCase):
) )
metrics = run_eval(args) metrics = run_eval(args)
assert metrics["score"] >= 0.65 self.assertGreaterEqual(metrics["score"], 0.65)
def run_decode(self, max_new_tokens): def run_decode(self, max_new_tokens):
response = requests.post( response = requests.post(
...@@ -49,8 +49,8 @@ class TestTorchCompile(unittest.TestCase): ...@@ -49,8 +49,8 @@ class TestTorchCompile(unittest.TestCase):
"sampling_params": { "sampling_params": {
"temperature": 0, "temperature": 0,
"max_new_tokens": max_new_tokens, "max_new_tokens": max_new_tokens,
"ignore_eos": True,
}, },
"ignore_eos": True,
}, },
) )
return response.json() return response.json()
...@@ -66,7 +66,7 @@ class TestTorchCompile(unittest.TestCase): ...@@ -66,7 +66,7 @@ class TestTorchCompile(unittest.TestCase):
print(res["text"]) print(res["text"])
throughput = max_tokens / (tok - tic) throughput = max_tokens / (tok - tic)
print(f"Throughput: {throughput} tokens/s") print(f"Throughput: {throughput} tokens/s")
assert throughput >= 152 self.assertGreaterEqual(throughput, 152)
if __name__ == "__main__": if __name__ == "__main__":
......
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_child_process
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
popen_launch_server,
)
class TestTorchCompile(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_SMALL_MOE_MODEL_NAME_FOR_TEST
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=["--enable-torch-compile", "--torch-compile-max-bs", "1"],
)
@classmethod
def tearDownClass(cls):
kill_child_process(cls.process.pid, include_self=True)
def test_mmlu(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mmlu",
num_examples=64,
num_threads=32,
)
metrics = run_eval(args)
self.assertGreaterEqual(metrics["score"], 0.50)
def run_decode(self, max_new_tokens):
response = requests.post(
self.base_url + "/generate",
json={
"text": "The capital of France is",
"sampling_params": {
"temperature": 0,
"max_new_tokens": max_new_tokens,
"ignore_eos": True,
},
},
)
return response.json()
def test_throughput(self):
import time
max_tokens = 256
tic = time.time()
res = self.run_decode(max_tokens)
tok = time.time()
print(f"{res=}")
throughput = max_tokens / (tok - tic)
print(f"Throughput: {throughput} tokens/s")
self.assertGreaterEqual(throughput, 290)
if __name__ == "__main__":
unittest.main()
...@@ -48,7 +48,7 @@ class TestTritonAttnBackend(unittest.TestCase): ...@@ -48,7 +48,7 @@ class TestTritonAttnBackend(unittest.TestCase):
) )
metrics = run_eval(args) metrics = run_eval(args)
assert metrics["score"] >= 0.65 self.assertGreaterEqual(metrics["score"], 0.65)
finally: finally:
kill_child_process(process.pid, include_self=True) kill_child_process(process.pid, include_self=True)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment