mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2026-04-23 00:17:25 +08:00
freeliuzc
e87ce4b8cd
* support new mtp * refactor(speculate_decoding and mtp): optimize mtp sturcture logic. Update spec-branch status-process * fix cuda-graph for spec-decoding * fix xpu mtp and fix some note * fix unittest and optmize note * fix model status update in eos-branch
333 lines
12 KiB
Python
333 lines
12 KiB
Python
# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
|
|
"""
|
|
Unit tests for naive_update_model_status kernel.
|
|
|
|
Kernel semantics (from naive_update_model_status.cu):
|
|
- Launched as <<<1, 1024>>>, one thread per real batch slot.
|
|
- Guard: seq_lens_this_time[bid] > 0 (already zeroed for stopped/paused
|
|
slots by pre_process before this kernel runs).
|
|
- Scatters sampled token from packed next_tokens (indexed by cu_seqlens_q_output)
|
|
into accept_tokens[bid, 0].
|
|
- Sets accept_num[bid] = 1 for running slots, 0 otherwise.
|
|
- Sets seq_lens_this_time[bid] = 1 for running, 0 otherwise.
|
|
|
|
cu_seqlens_q_output layout:
|
|
next_tokens[cu_seqlens_q_output[i] .. cu_seqlens_q_output[i+1]-1]
|
|
are the output tokens for request i (exactly 1 for decode, 0 for stopped).
|
|
"""
|
|
|
|
import unittest
|
|
from typing import Any, Dict
|
|
|
|
import numpy as np
|
|
import paddle
|
|
import pytest
|
|
|
|
from fastdeploy.model_executor.ops.gpu import naive_update_model_status
|
|
|
|
CUDA_PLACE = paddle.CUDAPlace(0) if paddle.is_compiled_with_cuda() else paddle.CPUPlace()
|
|
|
|
|
|
# ============================================================
|
|
# Layer 1: Helpers
|
|
# ============================================================
|
|
|
|
|
|
def to_paddle_inputs(inputs: Dict[str, Any]) -> Dict[str, Any]:
|
|
"""Convert numpy dict → GPU paddle tensors."""
|
|
paddle_inputs = {}
|
|
for k, v in inputs.items():
|
|
if isinstance(v, (int, bool, float)):
|
|
paddle_inputs[k] = v
|
|
elif v is not None:
|
|
paddle_inputs[k] = paddle.to_tensor(v, place=CUDA_PLACE)
|
|
return paddle_inputs
|
|
|
|
|
|
def run_kernel(paddle_inputs: Dict[str, Any]):
|
|
"""Call naive_update_model_status kernel (5 tensor args)."""
|
|
naive_update_model_status(
|
|
paddle_inputs["accept_tokens"],
|
|
paddle_inputs["accept_num"],
|
|
paddle_inputs["seq_lens_this_time"],
|
|
paddle_inputs["next_tokens"],
|
|
paddle_inputs["cu_seqlens_q_output"],
|
|
)
|
|
|
|
|
|
OUTPUT_KEYS = [
|
|
"accept_tokens",
|
|
"accept_num",
|
|
"seq_lens_this_time",
|
|
]
|
|
|
|
|
|
def get_outputs(paddle_inputs: Dict[str, Any]) -> Dict[str, np.ndarray]:
|
|
return {k: paddle_inputs[k].numpy() for k in OUTPUT_KEYS}
|
|
|
|
|
|
# ============================================================
|
|
# Layer 2: Input generation
|
|
# ============================================================
|
|
|
|
|
|
def gen_inputs(
|
|
real_bsz: int = 8,
|
|
max_step_tokens: int = 4,
|
|
seed: int = 42,
|
|
seq_lens_this_time: np.ndarray = None,
|
|
) -> Dict[str, Any]:
|
|
"""Generate randomized test inputs.
|
|
|
|
seq_lens_this_time: per-slot values pre-set by caller (mirrors what
|
|
pre_process produces). Slots with value > 0 are treated as running;
|
|
slots with value == 0 are stopped/paused. If None, a random mix is
|
|
generated (~75% running).
|
|
|
|
cu_seqlens_q_output: cumulative token offsets. Running slots (seq_lens > 0)
|
|
get 1 token; stopped/paused slots get 0 tokens.
|
|
"""
|
|
rng = np.random.default_rng(seed)
|
|
|
|
if seq_lens_this_time is None:
|
|
# ~75% running: values in [1, 9]; ~25% stopped: value = 0
|
|
seq_lens_this_time = rng.integers(1, 10, size=real_bsz, dtype=np.int32)
|
|
n_stop = max(0, real_bsz // 4)
|
|
if n_stop > 0:
|
|
stop_idxs = rng.choice(real_bsz, size=n_stop, replace=False)
|
|
seq_lens_this_time[stop_idxs] = 0
|
|
|
|
is_running = seq_lens_this_time > 0
|
|
|
|
# Build cu_seqlens_q_output: running slots contribute 1 token each
|
|
tokens_per_slot = is_running.astype(np.int32)
|
|
cu_seqlens_q_output = np.zeros(real_bsz + 1, dtype=np.int32)
|
|
cu_seqlens_q_output[1:] = np.cumsum(tokens_per_slot)
|
|
total_tokens = int(cu_seqlens_q_output[-1])
|
|
|
|
# Sample tokens for each running slot (packed)
|
|
next_tokens = rng.integers(0, 50000, size=max(total_tokens, 1), dtype=np.int64)
|
|
|
|
# Pre-allocate accept_tokens/accept_num with noise
|
|
accept_tokens = rng.integers(0, 100, size=(real_bsz, max_step_tokens), dtype=np.int64)
|
|
accept_num = rng.integers(0, 5, size=real_bsz, dtype=np.int32)
|
|
|
|
return {
|
|
"accept_tokens": accept_tokens,
|
|
"accept_num": accept_num,
|
|
"seq_lens_this_time": seq_lens_this_time.copy(),
|
|
"next_tokens": next_tokens,
|
|
"cu_seqlens_q_output": cu_seqlens_q_output,
|
|
# meta (not passed to kernel)
|
|
"real_bsz": real_bsz,
|
|
"max_step_tokens": max_step_tokens,
|
|
"is_running": is_running,
|
|
}
|
|
|
|
|
|
# ============================================================
|
|
# Layer 3: Reference implementation (1:1 with CUDA kernel)
|
|
# ============================================================
|
|
|
|
|
|
def reference_impl(inputs: Dict[str, Any]) -> Dict[str, Any]:
|
|
"""Python reference of naive_update_model_status_kernel.
|
|
|
|
Guard: seq_lens_this_time[bid] > 0 (matches simplified kernel).
|
|
"""
|
|
accept_tokens = inputs["accept_tokens"].copy()
|
|
accept_num = inputs["accept_num"].copy()
|
|
seq_lens_this_time = inputs["seq_lens_this_time"].copy()
|
|
|
|
next_tokens = inputs["next_tokens"]
|
|
cu_seqlens_q_output = inputs["cu_seqlens_q_output"]
|
|
real_bsz = inputs["real_bsz"]
|
|
|
|
for bid in range(real_bsz):
|
|
if seq_lens_this_time[bid] > 0:
|
|
# Write last (only) token for this slot
|
|
accept_tokens[bid, 0] = next_tokens[cu_seqlens_q_output[bid + 1] - 1]
|
|
accept_num[bid] = 1
|
|
seq_lens_this_time[bid] = 1
|
|
else:
|
|
accept_num[bid] = 0
|
|
seq_lens_this_time[bid] = 0
|
|
|
|
return {
|
|
"accept_tokens": accept_tokens,
|
|
"accept_num": accept_num,
|
|
"seq_lens_this_time": seq_lens_this_time,
|
|
}
|
|
|
|
|
|
# ============================================================
|
|
# Layer 4a: TEST_CONFIGS
|
|
# ============================================================
|
|
|
|
TEST_CONFIGS = [
|
|
{
|
|
"name": "all_running",
|
|
"real_bsz": 8,
|
|
"max_step_tokens": 4,
|
|
"seed": 42,
|
|
"seq_lens_this_time": np.array([3, 5, 1, 7, 2, 4, 6, 8], dtype=np.int32),
|
|
},
|
|
{
|
|
"name": "mixed_stop",
|
|
"real_bsz": 8,
|
|
"max_step_tokens": 4,
|
|
"seed": 100,
|
|
"seq_lens_this_time": np.array([1, 0, 3, 0, 5, 0, 2, 4], dtype=np.int32),
|
|
},
|
|
{
|
|
"name": "all_stopped",
|
|
"real_bsz": 4,
|
|
"max_step_tokens": 4,
|
|
"seed": 42,
|
|
"seq_lens_this_time": np.zeros(4, dtype=np.int32),
|
|
},
|
|
{
|
|
"name": "single_slot",
|
|
"real_bsz": 1,
|
|
"max_step_tokens": 4,
|
|
"seed": 42,
|
|
"seq_lens_this_time": np.array([1], dtype=np.int32),
|
|
},
|
|
{
|
|
"name": "large_batch",
|
|
"real_bsz": 64,
|
|
"max_step_tokens": 8,
|
|
"seed": 200,
|
|
"seq_lens_this_time": None, # randomly generated
|
|
},
|
|
{
|
|
"name": "bsz_1024",
|
|
"real_bsz": 1024,
|
|
"max_step_tokens": 4,
|
|
"seed": 42,
|
|
"seq_lens_this_time": None, # randomly generated
|
|
},
|
|
]
|
|
|
|
|
|
# ============================================================
|
|
# Layer 4b: Test suite
|
|
# ============================================================
|
|
|
|
|
|
class TestNaiveUpdateModelStatus(unittest.TestCase):
|
|
|
|
def setUp(self):
|
|
if not paddle.is_compiled_with_cuda():
|
|
self.skipTest("Requires CUDA")
|
|
|
|
def _run_and_compare(self, inputs: Dict[str, Any]):
|
|
"""Run reference + kernel, compare all outputs."""
|
|
ref = reference_impl(inputs)
|
|
paddle_inputs = to_paddle_inputs(inputs)
|
|
run_kernel(paddle_inputs)
|
|
outputs = get_outputs(paddle_inputs)
|
|
|
|
for key in OUTPUT_KEYS:
|
|
np.testing.assert_array_equal(
|
|
outputs[key],
|
|
ref[key],
|
|
err_msg=f"{key} mismatch",
|
|
)
|
|
|
|
def test_configs(self):
|
|
"""Run all TEST_CONFIGS via subTest."""
|
|
for cfg in TEST_CONFIGS:
|
|
with self.subTest(name=cfg["name"]):
|
|
test_cfg = {k: v for k, v in cfg.items() if k != "name"}
|
|
inputs = gen_inputs(**test_cfg)
|
|
self._run_and_compare(inputs)
|
|
|
|
def test_running_slots_get_token(self):
|
|
"""Running slots should have accept_tokens[bid, 0] = next_tokens[cu_q[bid+1]-1]."""
|
|
seq_lens = np.array([1, 0, 1, 1], dtype=np.int32)
|
|
inputs = gen_inputs(real_bsz=4, seed=42, seq_lens_this_time=seq_lens)
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
cu_q = inputs["cu_seqlens_q_output"]
|
|
next_tokens = inputs["next_tokens"]
|
|
for bid in range(4):
|
|
if seq_lens[bid] > 0:
|
|
expected_token = next_tokens[cu_q[bid + 1] - 1]
|
|
self.assertEqual(ref["accept_tokens"][bid, 0], expected_token)
|
|
self.assertEqual(ref["accept_num"][bid], 1)
|
|
self.assertEqual(ref["seq_lens_this_time"][bid], 1)
|
|
|
|
def test_stopped_slots_cleared(self):
|
|
"""Stopped slots (seq_lens_this_time=0): accept_num=0, seq_lens_this_time=0."""
|
|
seq_lens = np.array([0, 1, 0, 1], dtype=np.int32)
|
|
inputs = gen_inputs(real_bsz=4, seed=42, seq_lens_this_time=seq_lens)
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
for bid in [0, 2]:
|
|
self.assertEqual(ref["accept_num"][bid], 0)
|
|
self.assertEqual(ref["seq_lens_this_time"][bid], 0)
|
|
|
|
def test_all_stopped(self):
|
|
"""All stopped: all accept_num=0, seq_lens_this_time=0."""
|
|
seq_lens = np.zeros(8, dtype=np.int32)
|
|
inputs = gen_inputs(real_bsz=8, seed=42, seq_lens_this_time=seq_lens)
|
|
self._run_and_compare(inputs)
|
|
ref = reference_impl(inputs)
|
|
np.testing.assert_array_equal(ref["accept_num"], 0)
|
|
np.testing.assert_array_equal(ref["seq_lens_this_time"], 0)
|
|
|
|
def test_mixed_prefill_decode(self):
|
|
"""Mixed prefill+decode: stopped slots (seq_lens=0) get 0 tokens in packed next_tokens."""
|
|
# slots 0,2 are decode (seq_lens > 0); slots 1,3 are stopped (seq_lens = 0)
|
|
seq_lens = np.array([1, 0, 1, 0], dtype=np.int32)
|
|
inputs = gen_inputs(real_bsz=4, seed=77, seq_lens_this_time=seq_lens)
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
for bid in [0, 2]:
|
|
self.assertEqual(ref["accept_num"][bid], 1)
|
|
for bid in [1, 3]:
|
|
self.assertEqual(ref["accept_num"][bid], 0)
|
|
|
|
def test_seq_lens_normalized_to_one(self):
|
|
"""Running slots with seq_lens_this_time > 1 are normalized to 1 after kernel."""
|
|
# pre_process may set values > 1 for some slots; kernel normalizes to 1
|
|
seq_lens = np.array([3, 7, 0, 5], dtype=np.int32)
|
|
inputs = gen_inputs(real_bsz=4, seed=99, seq_lens_this_time=seq_lens)
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
for bid in [0, 1, 3]:
|
|
self.assertEqual(ref["seq_lens_this_time"][bid], 1)
|
|
self.assertEqual(ref["seq_lens_this_time"][2], 0)
|
|
|
|
@pytest.mark.gpu
|
|
def test_bsz_exceeds_block_size(self):
|
|
"""real_bsz > 1024 should raise."""
|
|
with self.assertRaises(Exception):
|
|
seq_lens = np.ones(1025, dtype=np.int32)
|
|
inputs = gen_inputs(real_bsz=1025, seed=42, seq_lens_this_time=seq_lens)
|
|
paddle_inputs = to_paddle_inputs(inputs)
|
|
run_kernel(paddle_inputs)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|