97dbcaf919
### What this PR does / why we need it? This PR introduces a caching mechanism for CPU-based `torch.Generator` objects in the `_random_sample_310p` function to optimize sampling performance. It includes unit tests for cache persistence and state recovery. Feedback highlights a critical bug where keying the cache by batch index instead of generator ID can break RNG reproducibility during request re-scheduling, and notes a potential memory leak in the global cache. ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? Tested via new unit tests in `tests/ut/_310p/sample/test_sampler_310.py` verifying cache logic and error handling. --------- Signed-off-by: csoulnd <daidaicurry@foxmail.com>
80 lines
3.2 KiB
Python
80 lines
3.2 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. 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.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
|
|
import torch
|
|
from vllm.model_executor.layers.batch_invariant import vllm_is_batch_invariant
|
|
|
|
from vllm_ascend.sample.sampler import (
|
|
DEFAULT_LOGPROBS_MODE,
|
|
AscendSampler,
|
|
AscendTopKTopPSampler,
|
|
)
|
|
from vllm_ascend.utils import global_stream, npu_stream_switch
|
|
|
|
_CPU_GENERATOR_CACHE_310P: dict[int, tuple[torch.Generator, int]] = {}
|
|
|
|
|
|
def _random_sample_310p(
|
|
probs: torch.Tensor,
|
|
generators: dict[int, torch.Generator],
|
|
) -> torch.Tensor:
|
|
"""310P-specific random sampling with CPU exponential generation for q."""
|
|
with npu_stream_switch(global_stream()):
|
|
q = torch.empty_like(probs)
|
|
q = q.cpu()
|
|
if len(generators) != q.shape[0]:
|
|
q.exponential_()
|
|
if generators:
|
|
for i, generator in generators.items():
|
|
cache_entry = _CPU_GENERATOR_CACHE_310P.get(i)
|
|
if cache_entry is None or cache_entry[1] != id(generator):
|
|
cpu_generator = torch.Generator(device="cpu")
|
|
try:
|
|
# Keep RNG stream consistent with the original generator.
|
|
cpu_generator.set_state(generator.get_state())
|
|
except Exception:
|
|
cpu_generator.manual_seed(generator.initial_seed())
|
|
cache_entry = (cpu_generator, id(generator))
|
|
_CPU_GENERATOR_CACHE_310P[i] = cache_entry
|
|
cpu_generator, _ = cache_entry
|
|
q[i].exponential_(generator=cpu_generator)
|
|
q = q.npu()
|
|
torch.npu.current_stream().wait_stream(global_stream())
|
|
return probs.div_(q).argmax(dim=-1).view(-1)
|
|
|
|
|
|
class AscendTopKTopPSampler310(AscendTopKTopPSampler):
|
|
def forward_native(self, logits, generators, k, p):
|
|
if vllm_is_batch_invariant():
|
|
return super().forward_native(logits, generators, k, p)
|
|
|
|
logits = self.apply_top_k_top_p(logits, k, p)
|
|
logits_to_return = None
|
|
if self.logprobs_mode == "processed_logits":
|
|
logits_to_return = logits
|
|
elif self.logprobs_mode == "processed_logprobs":
|
|
logits_to_return = logits.log_softmax(dim=-1, dtype=torch.float32)
|
|
|
|
probs = logits.softmax(dim=-1, dtype=torch.float32)
|
|
return _random_sample_310p(probs, generators), logits_to_return
|
|
|
|
|
|
class AscendSampler310(AscendSampler):
|
|
def __init__(self, logprobs_mode=DEFAULT_LOGPROBS_MODE):
|
|
super().__init__(logprobs_mode=logprobs_mode)
|
|
self.topk_topp_sampler = AscendTopKTopPSampler310(logprobs_mode=logprobs_mode)
|