[Fix] weight_checker: Make the checksum device-agnostic (#45918)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
bb9e9d5632
commit
77f90ade97
@@ -9,6 +9,7 @@ from pydantic import BaseModel, ConfigDict
|
||||
|
||||
from sglang.srt.managers.mm_utils import tensor_hash
|
||||
from sglang.srt.mem_cache.storage.mmap.mmap_allocator import alloc_mmap
|
||||
from sglang.srt.platforms import current_platform
|
||||
from sglang.srt.runtime_context import get_parallel
|
||||
from sglang.srt.utils.weight_checker_comparator import (
|
||||
CHUNK_NUMEL,
|
||||
@@ -205,7 +206,7 @@ class WeightChecker:
|
||||
def _compute_checksum(
|
||||
self, skip_tensor_list: Optional[List[str]] = None, *, role: str
|
||||
) -> Dict:
|
||||
torch.cuda.synchronize()
|
||||
current_platform.synchronize()
|
||||
start = time.perf_counter()
|
||||
|
||||
quantized_set = _build_quantized_set(self._get_model())
|
||||
@@ -222,7 +223,7 @@ class WeightChecker:
|
||||
|
||||
overall = overall_checksum(checksums)
|
||||
|
||||
torch.cuda.synchronize()
|
||||
current_platform.synchronize()
|
||||
elapsed = time.perf_counter() - start
|
||||
logger.info(
|
||||
f"[WeightChecker] checksum computed for {len(checksums)} tensors in {elapsed:.3f}s"
|
||||
|
||||
@@ -0,0 +1,32 @@
|
||||
"""WeightChecker checksum on a non-CUDA device (CPU stands in for TPU)."""
|
||||
|
||||
import unittest
|
||||
from unittest.mock import patch
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
|
||||
from sglang.srt.utils.weight_checker import ChecksumInfo, WeightChecker
|
||||
from sglang.test.ci.ci_register import register_cpu_ci
|
||||
from sglang.test.test_utils import CustomTestCase, enter_scope, published_topology
|
||||
|
||||
register_cpu_ci(est_time=5, suite="base-a-test-cpu")
|
||||
|
||||
|
||||
class TestChecksumWithoutCuda(CustomTestCase):
|
||||
def test_checksum_does_not_require_cuda(self):
|
||||
"""The checksum action must work on hosts without CUDA, e.g. TPU engines
|
||||
checked by an RL trainer after every weight update."""
|
||||
model = nn.Linear(4, 4)
|
||||
enter_scope(self, published_topology())
|
||||
checker = WeightChecker(get_model=lambda: model)
|
||||
with patch.object(
|
||||
torch.cuda, "synchronize", side_effect=RuntimeError("no CUDA device")
|
||||
):
|
||||
out = checker.handle("checksum", role="target")
|
||||
info = ChecksumInfo.model_validate(out)
|
||||
self.assertEqual(set(info.checksums), {"weight", "bias"})
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user