mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2026-04-23 08:21:53 +08:00
sunxin
560 lines
21 KiB
Python
560 lines
21 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 draft_model_update kernel.
|
|
|
|
Kernel semantics (from draft_model_update.cu):
|
|
- Launched as <<<1, 512>>>, one thread per batch slot (bsz = seq_lens_this_time.shape[0]).
|
|
- inter_next_tokens is a packed 1-D array; slot tid's tokens start at
|
|
cu_seqlens_q_output[tid].
|
|
- Branch A — encoder step (seq_len_encoder > 0):
|
|
seq_lens_decoder[tid] += seq_len_encoder; seq_lens_encoder[tid] = 0
|
|
token = inter_next_tokens[cu_seqlens_q_output[tid]] (first token)
|
|
pre_ids[tid][1] = token; step_idx[tid] += 1; draft_tokens[tid][0] = token
|
|
if step_idx[tid] < max_dec_len[tid]:
|
|
base_model_draft_tokens[tid][substep+1] = token
|
|
- Branch B — decoder step (seq_len_encoder == 0, seq_len_decoder > 0):
|
|
if step_idx[tid] >= max_dec_len[tid] - 1:
|
|
base_model_draft_tokens[tid][substep+1] = -1 # near limit, no update
|
|
else:
|
|
seq_lens_decoder[tid] += seq_len_this_time
|
|
token = inter_next_tokens[cu_seqlens_q_output[tid] + seq_len_this_time - 1]
|
|
draft_tokens[tid][0] = token
|
|
base_model_draft_tokens[tid][substep+1] = token
|
|
step_idx[tid] += seq_len_this_time
|
|
pre_ids[tid][step_idx[tid]] = token
|
|
- Stopped slot (!stop_flags[tid] is False):
|
|
draft_tokens[tid][0] = -1; base_model_draft_tokens[tid][substep+1] = -1
|
|
- EOS check is removed from current kernel (commented out).
|
|
- Post-branch seq_lens_this_time update:
|
|
running → seq_lens_this_time[tid] = 1
|
|
stopped → seq_lens_this_time[tid] = 0; seq_lens_encoder[tid] = 0
|
|
- not_need_stop[0] = (stop_sum < bsz) via BlockReduce.
|
|
"""
|
|
|
|
import unittest
|
|
from typing import Any, Dict
|
|
|
|
import numpy as np
|
|
import paddle
|
|
import pytest
|
|
|
|
from fastdeploy.model_executor.ops.gpu import draft_model_update
|
|
|
|
CUDA_PLACE = paddle.CUDAPlace(0) if paddle.is_compiled_with_cuda() else paddle.CPUPlace()
|
|
CPU_PLACE = paddle.CPUPlace()
|
|
|
|
|
|
# ============================================================
|
|
# Layer 1: Helpers
|
|
# ============================================================
|
|
|
|
|
|
def to_paddle_inputs(inputs: Dict[str, Any]) -> Dict[str, Any]:
|
|
"""Convert numpy dict → 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 draft_model_update kernel."""
|
|
draft_model_update(
|
|
paddle_inputs["inter_next_tokens"],
|
|
paddle_inputs["draft_tokens"],
|
|
paddle_inputs["pre_ids"],
|
|
paddle_inputs["seq_lens_this_time"],
|
|
paddle_inputs["seq_lens_encoder"],
|
|
paddle_inputs["seq_lens_decoder"],
|
|
paddle_inputs["step_idx"],
|
|
paddle_inputs["cu_seqlens_q_output"],
|
|
paddle_inputs["stop_flags"],
|
|
paddle_inputs["not_need_stop"],
|
|
paddle_inputs["max_dec_len"],
|
|
paddle_inputs["end_ids"],
|
|
paddle_inputs["base_model_draft_tokens"],
|
|
paddle_inputs["max_seq_len"],
|
|
paddle_inputs["substep"],
|
|
)
|
|
|
|
|
|
# In-place modified outputs (matches SetInplaceMap in .cu)
|
|
OUTPUT_KEYS = [
|
|
"draft_tokens",
|
|
"pre_ids",
|
|
"seq_lens_this_time",
|
|
"seq_lens_encoder",
|
|
"seq_lens_decoder",
|
|
"step_idx",
|
|
"stop_flags",
|
|
"not_need_stop",
|
|
"base_model_draft_tokens",
|
|
]
|
|
|
|
|
|
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(
|
|
bsz: int = 8,
|
|
max_draft_token: int = 4,
|
|
pre_id_length: int = 64,
|
|
max_base_model_draft_token: int = 5,
|
|
substep: int = 1,
|
|
seed: int = 42,
|
|
stop_flags: np.ndarray = None,
|
|
seq_lens_encoder: np.ndarray = None,
|
|
seq_lens_decoder: np.ndarray = None,
|
|
) -> Dict[str, Any]:
|
|
"""Generate randomized test inputs.
|
|
|
|
inter_next_tokens is a packed 1-D int64 array. Each slot tid contributes
|
|
seq_lens_this_time[tid] tokens starting at cu_seqlens_q_output[tid].
|
|
"""
|
|
rng = np.random.default_rng(seed)
|
|
|
|
if stop_flags is None:
|
|
stop_flags = np.zeros(bsz, dtype=bool)
|
|
# ~25% stopped
|
|
n_stop = max(0, bsz // 4)
|
|
if n_stop > 0:
|
|
stop_idxs = rng.choice(bsz, size=n_stop, replace=False)
|
|
stop_flags[stop_idxs] = True
|
|
|
|
if seq_lens_encoder is None:
|
|
seq_lens_encoder = rng.integers(0, 4, size=bsz, dtype=np.int32)
|
|
|
|
if seq_lens_decoder is None:
|
|
seq_lens_decoder = rng.integers(1, 20, size=bsz, dtype=np.int32)
|
|
|
|
# seq_lens_this_time: 1 token per running slot (decode), encoder slots also 1
|
|
seq_lens_this_time = np.ones(bsz, dtype=np.int32)
|
|
seq_lens_this_time[stop_flags] = 0
|
|
|
|
# cu_seqlens_q_output: packed offsets for inter_next_tokens
|
|
# running slots contribute seq_lens_this_time tokens; stopped slots 0 tokens
|
|
tokens_per_slot = seq_lens_this_time.astype(np.int32)
|
|
cu_seqlens_q_output = np.zeros(bsz, dtype=np.int32)
|
|
cu_seqlens_q_output[1:] = np.cumsum(tokens_per_slot[:-1])
|
|
total_tokens = int(np.sum(tokens_per_slot))
|
|
|
|
inter_next_tokens = rng.integers(100, 50000, size=max(total_tokens, 1), dtype=np.int64)
|
|
draft_tokens = rng.integers(0, 1000, size=(bsz, max_draft_token), dtype=np.int64)
|
|
pre_ids = rng.integers(0, 1000, size=(bsz, pre_id_length), dtype=np.int64)
|
|
step_idx = rng.integers(1, 20, size=bsz, dtype=np.int64)
|
|
max_dec_len = rng.integers(50, 100, size=bsz, dtype=np.int64)
|
|
end_ids = rng.integers(1, 100, size=3, dtype=np.int64)
|
|
base_model_draft_tokens = rng.integers(0, 1000, size=(bsz, max_base_model_draft_token), dtype=np.int64)
|
|
not_need_stop = np.array([True], dtype=bool)
|
|
|
|
return {
|
|
"inter_next_tokens": inter_next_tokens,
|
|
"draft_tokens": draft_tokens,
|
|
"pre_ids": pre_ids,
|
|
"seq_lens_this_time": seq_lens_this_time,
|
|
"seq_lens_encoder": seq_lens_encoder,
|
|
"seq_lens_decoder": seq_lens_decoder,
|
|
"step_idx": step_idx,
|
|
"cu_seqlens_q_output": cu_seqlens_q_output,
|
|
"stop_flags": stop_flags,
|
|
"not_need_stop": not_need_stop,
|
|
"max_dec_len": max_dec_len,
|
|
"end_ids": end_ids,
|
|
"base_model_draft_tokens": base_model_draft_tokens,
|
|
"max_seq_len": 0, # unused in kernel (legacy attr, kept for API compat)
|
|
"substep": substep,
|
|
# meta
|
|
"bsz": bsz,
|
|
"max_draft_token": max_draft_token,
|
|
"pre_id_length": pre_id_length,
|
|
"max_base_model_draft_token": max_base_model_draft_token,
|
|
}
|
|
|
|
|
|
# ============================================================
|
|
# Layer 3: Reference implementation (1:1 with CUDA kernel)
|
|
# ============================================================
|
|
|
|
|
|
def reference_impl(inputs: Dict[str, Any]) -> Dict[str, Any]:
|
|
"""Python reference of draft_model_update_kernel.
|
|
|
|
Mirrors the CUDA kernel logic exactly:
|
|
- Kernel only processes slots where seq_lens_this_time[tid] > 0
|
|
- Branch A: seq_len_encoder > 0
|
|
- Branch B: seq_len_encoder == 0 and seq_len_decoder > 0
|
|
- Stopped: stop_flags[tid] is True (only processed if seq_lens_this_time > 0)
|
|
|
|
Note: CUDA kernel calculates next_tokens_start_id by counting
|
|
running slots (seq_lens_this_time[i] > 0) before tid, not using
|
|
cu_seqlens_q_output directly.
|
|
"""
|
|
# Deep-copy mutable in-place arrays
|
|
draft_tokens = inputs["draft_tokens"].copy()
|
|
pre_ids = inputs["pre_ids"].copy()
|
|
seq_lens_this_time = inputs["seq_lens_this_time"].copy()
|
|
seq_lens_encoder = inputs["seq_lens_encoder"].copy()
|
|
seq_lens_decoder = inputs["seq_lens_decoder"].copy()
|
|
step_idx = inputs["step_idx"].copy()
|
|
stop_flags = inputs["stop_flags"].copy()
|
|
base_model_draft_tokens = inputs["base_model_draft_tokens"].copy()
|
|
|
|
# Read-only
|
|
inter_next_tokens = inputs["inter_next_tokens"]
|
|
max_dec_len = inputs["max_dec_len"]
|
|
substep = inputs["substep"]
|
|
bsz = inputs["bsz"]
|
|
|
|
stop_sum = 0
|
|
|
|
for tid in range(bsz):
|
|
stop_flag_now_int = 0
|
|
|
|
seq_len_this_time = int(seq_lens_this_time[tid])
|
|
seq_len_encoder = int(seq_lens_encoder[tid])
|
|
seq_len_decoder = int(seq_lens_decoder[tid])
|
|
|
|
# Calculate next_tokens_start_id: count running slots before tid
|
|
# This matches the CUDA kernel logic
|
|
next_tokens_start_id = 0
|
|
for i in range(tid):
|
|
if seq_lens_this_time[i] > 0:
|
|
next_tokens_start_id += 1
|
|
|
|
# CUDA kernel: if (tid < bsz && seq_lens_this_time[tid] > 0)
|
|
# Only process slots where seq_lens_this_time[tid] > 0
|
|
if seq_len_this_time > 0:
|
|
if not stop_flags[tid]:
|
|
token_this_time = -1
|
|
if seq_len_encoder > 0:
|
|
# Branch A: encoder step
|
|
token_this_time = inter_next_tokens[next_tokens_start_id]
|
|
seq_lens_decoder[tid] = seq_len_encoder + seq_len_decoder
|
|
seq_lens_encoder[tid] = 0
|
|
pre_ids[tid, 1] = token_this_time
|
|
step_idx[tid] += 1
|
|
draft_tokens[tid, 0] = token_this_time
|
|
if step_idx[tid] < max_dec_len[tid]:
|
|
base_model_draft_tokens[tid, substep + 1] = token_this_time
|
|
elif seq_len_decoder > 0:
|
|
# Branch B: decoder step
|
|
if step_idx[tid] >= max_dec_len[tid] - 1:
|
|
# near max_dec_len: only mark -1, no state update
|
|
base_model_draft_tokens[tid, substep + 1] = -1
|
|
else:
|
|
seq_lens_decoder[tid] += seq_len_this_time
|
|
# CUDA kernel uses next_tokens_start[0] (first token)
|
|
token_this_time = inter_next_tokens[next_tokens_start_id]
|
|
draft_tokens[tid, 0] = token_this_time
|
|
base_model_draft_tokens[tid, substep + 1] = token_this_time
|
|
step_idx[tid] += seq_len_this_time
|
|
pre_ids[tid, step_idx[tid]] = token_this_time
|
|
else:
|
|
# Stopped slot (but seq_lens_this_time > 0)
|
|
draft_tokens[tid, 0] = -1
|
|
base_model_draft_tokens[tid, substep + 1] = -1
|
|
stop_flag_now_int = 1
|
|
|
|
# 2. set seq_lens_this_time (only for processed slots)
|
|
if not stop_flags[tid]:
|
|
seq_lens_this_time[tid] = 1
|
|
else:
|
|
seq_lens_this_time[tid] = 0
|
|
seq_lens_encoder[tid] = 0
|
|
|
|
stop_sum += stop_flag_now_int
|
|
|
|
not_need_stop = np.array([stop_sum < bsz], dtype=bool)
|
|
|
|
return {
|
|
"draft_tokens": draft_tokens,
|
|
"pre_ids": pre_ids,
|
|
"seq_lens_this_time": seq_lens_this_time,
|
|
"seq_lens_encoder": seq_lens_encoder,
|
|
"seq_lens_decoder": seq_lens_decoder,
|
|
"step_idx": step_idx,
|
|
"stop_flags": stop_flags,
|
|
"not_need_stop": not_need_stop,
|
|
"base_model_draft_tokens": base_model_draft_tokens,
|
|
}
|
|
|
|
|
|
# ============================================================
|
|
# Layer 4a: TEST_CONFIGS
|
|
# ============================================================
|
|
|
|
TEST_CONFIGS = [
|
|
{
|
|
"name": "all_encoder",
|
|
"bsz": 4,
|
|
"seed": 42,
|
|
"seq_lens_encoder": np.array([2, 3, 1, 4], dtype=np.int32),
|
|
"seq_lens_decoder": np.array([5, 5, 5, 5], dtype=np.int32),
|
|
"stop_flags": np.zeros(4, dtype=bool),
|
|
},
|
|
{
|
|
"name": "all_decoder",
|
|
"bsz": 4,
|
|
"seed": 42,
|
|
"seq_lens_encoder": np.zeros(4, dtype=np.int32),
|
|
"seq_lens_decoder": np.array([5, 10, 3, 7], dtype=np.int32),
|
|
"stop_flags": np.zeros(4, dtype=bool),
|
|
},
|
|
{
|
|
"name": "mixed_enc_dec",
|
|
"bsz": 8,
|
|
"seed": 100,
|
|
},
|
|
{
|
|
"name": "all_stopped",
|
|
"bsz": 4,
|
|
"seed": 42,
|
|
"stop_flags": np.ones(4, dtype=bool),
|
|
"seq_lens_encoder": np.zeros(4, dtype=np.int32),
|
|
"seq_lens_decoder": np.array([5, 5, 5, 5], dtype=np.int32),
|
|
},
|
|
{
|
|
"name": "near_max_dec_len",
|
|
"bsz": 4,
|
|
"seed": 42,
|
|
"seq_lens_encoder": np.zeros(4, dtype=np.int32),
|
|
"seq_lens_decoder": np.array([5, 5, 5, 5], dtype=np.int32),
|
|
"stop_flags": np.zeros(4, dtype=bool),
|
|
},
|
|
{
|
|
"name": "large_batch",
|
|
"bsz": 64,
|
|
"seed": 200,
|
|
},
|
|
]
|
|
|
|
|
|
# ============================================================
|
|
# Layer 4b: Test suite
|
|
# ============================================================
|
|
|
|
|
|
class TestDraftModelUpdate(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)
|
|
# near_max_dec_len: set step_idx close to max_dec_len - 1
|
|
if cfg["name"] == "near_max_dec_len":
|
|
inputs["step_idx"][:] = inputs["max_dec_len"] - 1
|
|
self._run_and_compare(inputs)
|
|
|
|
def test_encoder_step(self):
|
|
"""Branch A: encoder step updates decoder, clears encoder, writes pre_ids[1]."""
|
|
enc = np.array([3, 0, 5, 0], dtype=np.int32)
|
|
dec = np.array([10, 20, 10, 20], dtype=np.int32)
|
|
stop = np.zeros(4, dtype=bool)
|
|
inputs = gen_inputs(bsz=4, seed=42, seq_lens_encoder=enc, seq_lens_decoder=dec, stop_flags=stop)
|
|
ref_before_enc = enc.copy()
|
|
ref_before_dec = dec.copy()
|
|
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
# encoder slots: encoder cleared, decoder incremented
|
|
for tid in [0, 2]:
|
|
self.assertEqual(ref["seq_lens_encoder"][tid], 0)
|
|
self.assertEqual(ref["seq_lens_decoder"][tid], ref_before_enc[tid] + ref_before_dec[tid])
|
|
self.assertEqual(ref["step_idx"][tid], inputs["step_idx"][tid] + 1)
|
|
|
|
def test_decoder_step(self):
|
|
"""Branch B: decoder step increments decoder and step_idx."""
|
|
enc = np.zeros(4, dtype=np.int32)
|
|
dec = np.array([5, 10, 3, 7], dtype=np.int32)
|
|
stop = np.zeros(4, dtype=bool)
|
|
inputs = gen_inputs(bsz=4, seed=42, seq_lens_encoder=enc, seq_lens_decoder=dec, stop_flags=stop)
|
|
# ensure step_idx well below max_dec_len
|
|
inputs["step_idx"][:] = 5
|
|
inputs["max_dec_len"][:] = 50
|
|
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
for tid in range(4):
|
|
self.assertEqual(ref["seq_lens_this_time"][tid], 1)
|
|
self.assertGreater(ref["seq_lens_decoder"][tid], dec[tid])
|
|
|
|
def test_stopped_slots(self):
|
|
"""Stopped slots: draft_tokens[0]=-1, base_model_draft_tokens[substep+1]=-1.
|
|
|
|
Note: Kernel only processes slots where seq_lens_this_time[tid] > 0.
|
|
Stopped slots with seq_lens_this_time == 0 are skipped by the kernel
|
|
(draft_tokens remain unchanged from initial values).
|
|
"""
|
|
# Test with default inputs where ~25% of slots are stopped
|
|
# gen_inputs sets seq_lens_this_time=0 for stopped slots
|
|
inputs = gen_inputs(bsz=8, seed=42)
|
|
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
|
|
# Verify: slots with stop_flags=True should have draft_tokens[:, 0] unchanged
|
|
# (because seq_lens_this_time=0, kernel skips them)
|
|
# Slots with stop_flags=False should have updated draft_tokens
|
|
for tid in range(inputs["bsz"]):
|
|
if inputs["stop_flags"][tid]:
|
|
# Stopped slot: draft_tokens unchanged (seq_lens_this_time was 0)
|
|
pass # No assertion - draft_tokens can be anything
|
|
else:
|
|
# Running slot: draft_tokens should be updated
|
|
self.assertNotEqual(ref["draft_tokens"][tid, 0], -1)
|
|
self.assertEqual(ref["seq_lens_this_time"][tid], 1)
|
|
|
|
def test_not_need_stop_all_running(self):
|
|
"""All running → not_need_stop[0] = True."""
|
|
enc = np.zeros(4, dtype=np.int32)
|
|
inputs = gen_inputs(
|
|
bsz=4,
|
|
seed=42,
|
|
stop_flags=np.zeros(4, dtype=bool),
|
|
seq_lens_encoder=enc,
|
|
seq_lens_decoder=np.array([5, 5, 5, 5], dtype=np.int32),
|
|
)
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
self.assertTrue(ref["not_need_stop"][0])
|
|
|
|
def test_not_need_stop_all_stopped(self):
|
|
"""All stopped (but processed) → not_need_stop[0] = False.
|
|
|
|
Note: Kernel only processes slots where seq_lens_this_time[tid] > 0.
|
|
To test not_need_stop, we need slots that have seq_lens_this_time > 0
|
|
but are marked as stopped (i.e., they were running but got stopped).
|
|
"""
|
|
# Create inputs where slots are stopped but have seq_lens_this_time > 0
|
|
# This simulates the case where slots were running but got stopped
|
|
stop = np.array([True, True, True, True], dtype=bool)
|
|
enc = np.zeros(4, dtype=np.int32)
|
|
inputs = gen_inputs(
|
|
bsz=4,
|
|
seed=42,
|
|
stop_flags=stop,
|
|
seq_lens_encoder=enc,
|
|
seq_lens_decoder=np.array([5, 5, 5, 5], dtype=np.int32),
|
|
)
|
|
# Manually set seq_lens_this_time to ensure kernel processes these slots
|
|
inputs["seq_lens_this_time"] = np.array([1, 1, 1, 1], dtype=np.int32)
|
|
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
self.assertFalse(ref["not_need_stop"][0])
|
|
|
|
def test_near_max_dec_len_decoder(self):
|
|
"""Branch B: step_idx >= max_dec_len-1 → base_model_draft_tokens=-1, no state update."""
|
|
enc = np.zeros(4, dtype=np.int32)
|
|
dec = np.array([5, 5, 5, 5], dtype=np.int32)
|
|
stop = np.zeros(4, dtype=bool)
|
|
inputs = gen_inputs(bsz=4, seed=42, seq_lens_encoder=enc, seq_lens_decoder=dec, stop_flags=stop)
|
|
inputs["max_dec_len"][:] = 10
|
|
inputs["step_idx"][:] = 9 # >= max_dec_len - 1
|
|
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
substep = inputs["substep"]
|
|
for tid in range(4):
|
|
self.assertEqual(ref["base_model_draft_tokens"][tid, substep + 1], -1)
|
|
# seq_lens_decoder should NOT be incremented in this branch
|
|
self.assertEqual(ref["seq_lens_decoder"][tid], dec[tid])
|
|
|
|
def test_encoder_step_at_max_dec_len(self):
|
|
"""Branch A: step_idx >= max_dec_len after increment → no base_model_draft write."""
|
|
enc = np.array([2, 2, 2, 2], dtype=np.int32)
|
|
dec = np.array([5, 5, 5, 5], dtype=np.int32)
|
|
stop = np.zeros(4, dtype=bool)
|
|
inputs = gen_inputs(bsz=4, seed=42, seq_lens_encoder=enc, seq_lens_decoder=dec, stop_flags=stop)
|
|
inputs["max_dec_len"][:] = 5
|
|
# After step_idx += 1, step_idx[tid] = max_dec_len → no write
|
|
inputs["step_idx"][:] = 4
|
|
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
substep = inputs["substep"]
|
|
for tid in range(4):
|
|
# step_idx becomes 5 == max_dec_len[tid] → guard fails → no write
|
|
orig_bm = inputs["base_model_draft_tokens"][tid, substep + 1]
|
|
self.assertEqual(ref["base_model_draft_tokens"][tid, substep + 1], orig_bm)
|
|
|
|
def test_seq_lens_this_time_reset(self):
|
|
"""Running slots get seq_lens_this_time=1; stopped slots get 0."""
|
|
stop = np.array([False, True, False, True], dtype=bool)
|
|
enc = np.zeros(4, dtype=np.int32)
|
|
inputs = gen_inputs(
|
|
bsz=4,
|
|
seed=42,
|
|
stop_flags=stop,
|
|
seq_lens_encoder=enc,
|
|
seq_lens_decoder=np.array([5, 5, 5, 5], dtype=np.int32),
|
|
)
|
|
self._run_and_compare(inputs)
|
|
|
|
ref = reference_impl(inputs)
|
|
for tid in [0, 2]:
|
|
self.assertEqual(ref["seq_lens_this_time"][tid], 1)
|
|
for tid in [1, 3]:
|
|
self.assertEqual(ref["seq_lens_this_time"][tid], 0)
|
|
|
|
@pytest.mark.gpu
|
|
def test_bsz_within_block_size(self):
|
|
"""bsz = 512 (max block size) should work without error."""
|
|
inputs = gen_inputs(bsz=512, seed=42)
|
|
paddle_inputs = to_paddle_inputs(inputs)
|
|
run_kernel(paddle_inputs) # should not raise
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|